vol2 no6

(IJCNS) International Journal of Computer and Network Security, Vol. 2, No. 6, June 2010

1

Universal Tool for University Course Schedule Using Genetic Algorithm

Mohamed Tahar Ben Othman, Senior Member IEEE

College of Computer, Qassim University, Saudi Arabia.

[email protected] [email protected]

Abstract: Time tabling is one of the most important administrative activities that take place in academic institutions. Timetable Scheduling takes a huge effort at the beginning of each year/semester for every university program. It may result also – if it is not well done – to a problem in the registration process as well as an abnormal class start. For this reason, several researches were conducted for Timetable Scheduling automation. The number of constraints that have to be satisfied for such problem make Timetable scheduling difficult task. This problem is known as NP Hard. A large number of variants of the time tabling problem have been proposed in the literature, which differ from each other based on the type of institution involved (university or school) and the type of constraints. In this paper, we have resolved the problem of Timetable scheduling using Genetic Algorithm. The originality of our work comes from its universality. The different variables and constraints are defined by the end-user and then the genetic algorithm is processed to provide the optimal or the near-optimal solution.

Keywords: Scheduling, Time tabling, Genetic Algorithm, hard and soft constraints.

1. Introduction Scheduling is the process of allocating tasks to a set of resources. Generally, these tasks require the use of varied resources while the resources, on the other hand, are limited in nature with respect to quantity as well as the time that they are available for. This allocation has to respect a number of constraints. The scheduling of time tabling is one of the most areas that were studied over the last decade. Among the wide variety of time tabling problems, educational time tabling is one of the mostly studied from a practical viewpoint. It is one of the most important and time-consuming tasks which occur periodically (i.e. each year/semester) in all academic institutions. The quality of the time tabling has a great impact on a broad range of different parties including lecturers, students and administrators [20]. The university's timetable scheduling is the assignment to each course the timeslot, the teacher, the classroom… etc so that certain conditions for these resources and for the students' registrations are respected. Because of the number of constraints that have to be dealt with and the limitation in resources this problem is considered as NP-hard. The constraints that have to be respected are called hard constraints. On the other hand, a good schedule should also be efficient such minimize wasted time between classes for students as well as teachers. These

constraints generally are considered as soft constraints that may not be all respected in an acceptable solution.

Timetable scheduling can be seen as a finite Constraint Satisfaction Problem (CSP). CSP is a problem with a finite set of variables, where each variable is associated with a finite domain. Relationships between variables constraint the possible instantiations they can take at the same time. A solution to a CSP is an instantiation of all the variables with values from their respective domains, satisfying all the constraints. In many CSPs, such as resource planning and Timetable scheduling, some solutions are better than others [21]. To solve a CSP, one must find the solution tuple that instantiate variables with values of their respective domains, and that these instantiations do not violate any of the constraints. Most of researches that were conducted are either studying a solution to a particular type of institution or providing a system or language to describe the different constraints and variables and generate the time table. In this paper we provide a tool that in one side, gives the possibility to define any number of variables and constraints and run genetic algorithm to provide a number of optimal or near-optimal solutions and in the other side interacts with the user to modify any solution by providing an assistance to check if any constraint is broken due to such modification.

The rest of this paper is organized as follow: the next section we give a brief description of genetic algorithm. Section 3 presents some related works. The problem description and the proposed solution are given in section 4. finally, the discussion and conclusion are presented in section 5.

2. Genetic Algorithm The genetic algorithm is a method for solving both constrained and unconstrained optimization problems that is based on natural selection, the process that drives biological evolution. The genetic algorithm repeatedly modifies a population of individual solutions. At each step, the genetic algorithm selects individuals at a given selected method (eg. random) from the current population to be parents and uses them produce the children for the next generation. Over successive generations, the population "evolves" toward an optimal solution. Genetic algorithm can be used to solve a variety of optimization problems that are not well suited for standard optimization algorithms, including problems in


2

which the objective function is discontinuous, stochastic, or highly nonlinear [22]. The genetic algorithm uses three main types of rules at each step to create the next generation from the current population. Selection rules: select the individuals called parents, that contribute to the population at the next generation. Crossover rules combine two parents to form children for the next generation. Mutation rules apply random changes to individual parents to form children. Figure 1 shows the structure of the tool we built. It consists of two main layers: the first layer gives the ability to define variables and constraints and after interfacing with the genetic algorithm layer the different solutions (Fig 2) can be manipulated via a user-friendly interface. The second layer is responsible of the genetic algorithm processing. The fitness function and then the initial population are auto-generated upon variables and constraints definition (Fig 4). Figs 2-3 show the plot of the best fitness got after 200 generations over which the GA algorithm ran.

3. Related works Several research papers have been published in the last decade, which address different aspects of “the time tabling problem”. Indeed, there is even an international conference on the “Practice and Theory of Automated Time tabling” (PATAT) which has taken place biannually since 1995. A Large variety of solving techniques has been tried out for solution of time tabling problems [1-23]. Most research has concentrated on automatically solving difficult static time tabling problems. Here we present some that are related to genetic algorithm. J. Matthews concentrated on solving dynamic time tabling problems [24], in which a change of circumstance would require the modification of an existing time tabling solution. The paper [25] describes a genetic algorithm-based approach with two main stages for solving the course time tabling problem. A local search is applied to the algorithm at each stage. The first stage eliminates the violations of hard constraints, and the second one attempts to minimize the violations of soft constraints while keeping

the number of hard constraint violations zero. J. J. Moreira in [26] presents a solution method to the problem of automatic construction timetables for the exams. Among several mathematical models of representation, the final option was for a model matrix, which is justified by the benefits that this model presents when used in the algorithm solution. The method of solution is a meta-heuristics that includes a genetic algorithm.

4. Problem description

Making course schedule is one of those NP hard problems. The problem can be solved using heuristic search algorithm to find optimal solution, but it works for simple cases. For more complex inputs and requirements finding considerably good solution can take a while or it may be impossible. This is where genetic algorithms come in the game. When making course schedule many constraints (number of professors, students, classes and classrooms, size of classroom, laboratory equipment in classroom and many other) should be taken into consideration. These constraints can be divided into two major types: hard and soft constraints.

• Hard constraints: if one is broken the provided schedule is not considered as acceptable solution.

• Soft constraints: if one is broken the provided schedule is considered as acceptable solution:

These constraints are represented by a set of predicates P. These predicates are generally set as parameters and should be satisfied to get the best solution. Generally a predicate returns true if it is satisfied and false if it is not. Each predicate Pi is assigned a penalty αi if Pi is not satisfied. This penalty depends on the type of the constraint (hard or soft) and on the importance of the requirement within the same type (ex. generally having a professor doing different courses in the same time is more critical than having different courses with different professors in the same classroom at the same time).

4.1. Chromosome

A chromosome is represented by the set of classes. A class is a set of course (c), day (d), time (t), professor (p), classroom (r), level (l) and a list of students (s). As the students' registration is done generally after the schedule is built, we will omit in the first phase the list of the students from the Class.

Chrom = {(ci, di, ti, pi, ri, li) / i=1..N} (1)

where N is the number of courses.

A course can have more than one time weekly. The chromosome is represented using a matrix NxV where N is the number of courses and V is the number of variables (V=6 in this example).

4.2. Fitness

The objective function is calculated for each constraint C over the chromosome:

Initial population

Generation

Gen

erat

ed

fitne

ss

Selection

Crossover and mutation

new population

finish

Variables and constraints definition

Solutions manipulation

Interactive schedule interface

Figure 1: Tool Structure.

Layer 1

Layer 2


3

1 if the constraint is violatedf(C)= (2)0 otherwize

The general fitness is calculated over all courses as the summation of violated constraints times their weights as follows:

= ( ) (3)Fitness f Ci ic i

α∑ ∑

where αi represents the weight attached to the constraint Ci. Ci is a set of predicates that have to be all satisfied for Ci to be satisfied. This fitness function has to be minimized to obtain a better solution. We chose a constant weight α for all hard constraints and a constant weight β for all soft constraints. This yields to the following fitness equation in this particular case:

= ( ) ( )h sFitness f C f C (4)i ic i c i

α β∑ ∑ ∑ ∑∗ + ∗

In our measurements the value of α is taken 2 and the value of β is taken 1. For example if we have the following conditions:

• A professor can has only one class at a time (C1, Hard)

• A Classroom can handle only one course at a time (C2, Hard).

• Two time-slots for he same course cannot be the same day (C3, Soft)

• Courses in the same level cannot be in the same time (C4, Hard)

• No course on Monday between 11:00-12:30 (reserved slot) (C5, Hard)

Then these constraints will be written in the form where they are violated:

(c , d , t , p , r , l ) & (c , d , t , p , r , l )i i i i i i j j j j j j∀

=( ) & ( ) & ( )1 i j i j i j=( ) & ( ) & ( )2 i j i j i j=(c =c ) & ( )3 i j i j

(l =l ) & ( = t )4 i j i j(d =Mon) & ( 11:00 12:30)5 i i

C d d t t p pC d d t t r rC d d (5)C tC t

= = == = =

=== = −

4.3. Crossover

Crossover operation combines two parent chromosomes, and then creates an offspring that will be inserted in the new generation. The two parent chromosomes are selected either randomly or according to their fitness (the two having the best fitness from the remaining pool). A set of rows are selected randomly for both parents. The rows indexes should be the same for both parents so always the variables domains will be respected. The crossover is made by switching the values of some variables taken randomly from

the variables that can be modified between two rows with the same index.

4.4. Mutation

Mutation operation is very simple. It just takes randomly a number of rows in the chromosome matrix and moves it to another randomly chosen slot (change the day and/or the time) or location. In every operation, the domain of different variables is respected. Note that the tool works with variable values as indexes to the real values. These real values for examples courses' names, professors' names, etc can be bound to these variables. The relationship between the different variables should be set at the beginning of the process (eg. which professor or set of professors are willing to teach a given course).

5. Results Figure 2 shows the average number of clashes for each

constraint during the 200 generations. Figure 4 presents the best fitness plot for different population during 200 generations. In order to test the effectiveness of our objective function for solving the timetabling problem which is built by given different weights for hard and soft constraints, we provide in Figure 4 the history of constraints violation per type (Hard and Soft). Figure 4 demonstrates, in one side, that the convergence faster for hard than for soft constraints and, in the other side, the only clashes that may remain (Figure 3) are due to soft constraints.

Figure 2: Average number of Clashes per Constraint


4

Figure 3: Best Fitness

Figure 4: Constraint violation per type

6. Discussion and Conclusion

In this paper we introduced a tool to resolve the timetable scheduling problem with different constraints. This tool uses genetic algorithm to provide the optimal or near-optimal solution. Genetic Algorithm was widely used for time tabling but as many studies indicated the solution to time tabling is different from one institution to another because of the different types of institutions and different constraints within the same type. The originality in our work comes from the hybrid solution the tool supports. In fact, most researches were looking for the optimal solution to time tabling problem in respect of a set of constraints. This NP-hard problem required Heuristic methods among which GA solution. The problem is that GA may, in one side, runs into minima and, in the other side; the solution may not be the wanted solution – even if it has the best fitness – because some soft constraints were not satisfied. For this reason, our tool shows a number of solutions got from GA and then gives the possibility for the user to manipulate the timetable in a way to get the needed solution by recalculating the fitness to see if a constraint is broken each time a

modification is done. By this hybrid solution, it is not required that a GA solution has to be an optimal one. On the other hand, the other originality given by this tool is that it can be adapted to any institution as it gives the possibility to define the different variables and different constraints. The next step for this tool is to be able to support more types of constraints and then it can be used not only for course schedule but for most types of timetable scheduling generation.

Acknowledgements: I would like to acknowledge the financial support of this work from the Deanship of Scientific Research, Qassim University. Also, I would like to thank Professor Abdullah Ibrahiem Al-Shoshan and Dr. Abdulnaser Rachid for their remarks.

References [1] D. Abramson, “Constructing school timetables using

simulated annealing: sequential and parallel algorithms”, Management Science,v37(1),January 1991, pp. 98-113

[2] P. Adamidis and P. Arapakis, “Evolutionary Algorithms in Lecture Time tabling”, Proceedings of the 1999 IEEE Congress on Evolutionary Computation (CEC ’99), IEEE, 1999, pp. 1145-1151.

[3] S.C. Brailsford, C.N. Potts and B.M. Smith, “Constraint Satisfaction Problems: Algorithms and Applications”, European Journal of Operational Research, vol 119, 1999, pp. 557-581.

[4] E. K. Burke and J. P. Newall, "A New Adaptive Heuristic Framework for Examination Time tabling Problems", University of Nottingham, Working Group on Automated Time tabling, TR-2002-1 http://www.cs.nott.ac.uk/TR-cgi/TR.cgi?tr=2002-1

[5] M.W. Carter, “A Survey of Practical Applications of Examination Time tabling Algorithms”, Operations Research vol. 34, 1986, pp. 193-202.

[6] M.W. Carter and G. Laporte, “Recent Developments in Practical Course Time tabling”, In: Burke, E., Carter, M. (Eds.), The Practice and Theory of Automated Time tabling II: Selected Papers from the 2nd Int'l Conf. on the Practice and Theory of Automated Time tabling, Springer Lecture Notes in Computer Science Series, Vol. 1408, 1998, pp. 3-19.

[7] A. Colorni, M. Dorigo, and V. Maniezzo. “Genetic algorithms – A new approach to the timetable problem”, In Lecture Notes in Computer Science - NATO ASI Series, Vol. F 82, Combinatorial Optimization, (Akgul et al eds), Springer-Verlag, 1990, pp. 235-239.

[8] A. Hertz, “Tabu search for large scale time tabling problems,” European journal of operations research, vol. 54, 1991, pp. 39-47.

[9] B. Paechter, A. Cumming, M.G.Norman, and H. Luchian, “Extensions to a memetic time tabling system”, In E.K. Burke and P.M. Ross, eds., Proceedings of the 1st International Conference on the Practice and Theory of Automated Time tabling, 1995.

[10] A. Schaerf, “A Survey of Automated Time tabling”, Artificial Intelligence Review, vol 13 (2), 1999, 87-127.

[11] Arabinda Tripathy, “A lagrangian relaxation approach to course time tabling”, Journal of the Operational Research Society, vol. 31, 1980, pp. 599-603

[12] G.M. White and P.W. Chan, “Towards the Construction of Optimal Examination Timetables”, INFOR 17, 1979, p.p. 219-229.

[13] M. Ben Othman, G. A. Azim, A. Hamdi-Cherif, “Pairwise Sequence Alignment Revisited – Genetic Algorithms and


5

Cosine Functions”, NAUN conference, Attica, Greece, June 1-3, 2008.

[14] M. Ben Othman, A. Hamdi-Cherif, “Genetic algorithm and scalar product for pairwise sequence alignment”, IJC International Journal of Computer, NAUN North Atlantic University Union, November 2008.

[15] A. Bhaduri "University Timetable Scheduling using Genetic Artificial Immune Network", 2009 International Conference on Advances in Recent Technologies in Communication and Computing 978-0-7695-3845-7/09

[16] Wren, “Scheduling, Time tabling and Rostering – A Special Relationship”, in The Practice and Theory of Automated Time tabling: Selected Papers from the 1st Int'l Conf. on the Practice and Theory of Automated Time tabling, Burke, E., Ross, P. (Eds.) Springer Lecture Notes in Computer Science Series, Vol. 1153, 1996, pp. 46-75.

[17] S.C. Brailsford, C.N. Potts and B.M. Smith, “Constraint Satisfaction Problems: Algorithms and Applications,” European Journal of Operational Research, vol 119, 1999, pp. 557-581.

[18] Kazarlis S, Petridis V and Fragkou P, “Solving University Time tabling Problems Using Advanced Genetic Algorithms” Science Series, Vol. 1408, 1998, pp. 3-19.

[19] Bhaduri A, “Genetic Artificial Immune Network for Training Artificial Neural Networks”, In Proc. of the IEEE International Advance Computing Conference (2009) pp. 1998-2001

[20] R. Qu, E. Burke, B. McCollum, L. T.G. Merlot and S. Y. Lee, "A Survey of Search Methodologies and Automated Approaches for Examination Time tabling", Computer Science Technical Report No. NOTTCS-TR-2006-4, University of Nottingham.

[21] L.T. Leng, "Guided Genetic Algorithm", A thesis submitted for the degree of Ph.D in Computer Science, Department of computer science, University of Essex, Colchester, United Kingdom, 1999.

[22] www.systemtechnik.tu-ilmenau.de/~pohlheim/GA_Toolbox/ [23] S. D. ÖNCÜL, D. S. Ö. AYKAÇ, D. BAYRAKTAR and D.

ÇELEBÝ, "A Review Of Time tabling And Ressource

Allocation Models For Light-Rail Transaction Systems", International Conference for Prospects for Research in Transport and Logistics on a Regional Global Perspective, February 12-14, 2009, Istanbul, Turkey

[24] J. Matthews, "A Constraint Modeling Language for Timetables", Project Report, June 2006.

[25] S. Massoodian, A. Esteki, "A Hybrid Genetic Algorithm for Curriculum Based Course Time tabling", February 2008.

[26] J. J. Moreira, "A System for Automatic Construction of Exam Timetable Using Genetic Algorithms", ISLA - Rua de Cabo Borges, 55, 4430-032 V. N. de Gaia, Portugal. Revista de Estudos Politécnicos, Polytechnical Studies Review, 2008, Vol VI, nº 9, ISSN: 1645-9911

Mohamed Tahar Ben Othman received his Ph.D. in Computer Science from The National Institute Polytechnic of Grenoble INPG France in 1993, His Master degree in Computer Science from ENSIMAG "École Nationale Supérieure d'Informatique et de Mathématiques Appliquées de Grenoble" in 1989. He received a degree of Senior Engineer Diploma in Computer Science

from Faculty of Science of Tunis. This author became a Member (M) of IEEE in 1997, and a Senior Member (SM) in 2007. He worked as post-doc researcher in LGI (Laboratoire de Genie Logiciel) in Grenoble, France between 1993 and 1995, Dean of the Faculty of Science and Engineering between 1995 and 1997 at the University of Science and Technology in Sanaa, Yemen, as Senor Software Engineer in Nortel Networks, Canada, between 1998 and 2001 and Assistant Professor in Computer College at Qassim University in Saudi Arabia from 2002 until present. His research interest areas are wireless networks, Adhoc Networks, Communication Protocols, Artificial Intelligence, and Bioinformatics.

Appendix

Figure 2: Different solutions.


6

Figure 3: Available classrooms.

Figure 4: Constraint definition.


7

Formal Specification for Implementing Atomic Read/Write Shared Memory in Mobile Ad Hoc

Networks Using the Mobile Unity Fatma.A.Omara 1, Reham.A.Shihata 2

1Computer Science Department, Information Systems and Computers Faculty

Cairo University, Cairo -Egypt. Fatma_omara@ yahoo.com

2 Math's Department. Science Faculty, EL Minufiya University, Shebin-El Kom -Egypt.

[email protected]

Abstract: The Geoquorum approach for implementing atomic read/write shaved memory in mobile ad hoc networks. This problem in distributed computing is revisited in the new setting provided by the emerging mobile computing technology. A simple solution tailored for use in ad hoc networks is employed as a vehicle for demonstrating the applicability of formal requirements and design strategies to the new field of mobile computing. The approach of this paper is based on well understood techniques in specification refinement, but the methodology is tailored to mobile applications and help designers address novel concerns such as logical mobility, the invocations, specific conditions constructs. The proof logic and programming notation of mobile UNITY provide the intellectual tools required to carryout this task. Also, the quorum systems are investigated in highly mobile networks in order to reduce the communication cost associated with each distributed operation.

Keywords: Formal Specification, Mobility, Mobile Ad Hoc Networks, the Quorum Systems.

1. Introduction

Formal notations led to the development of specification languages; formal verification contributed to the application of mechanical theorem proffers to program checking; and formal derivation is a class of techniques that ensure correctness by construction, has the potential to reshape the way software will be developed in the future program derivation is less costly than post factum verification, is incremental in nature, and can be applied with varying degrees of rigor in conjunction with or completely apart from program verification. More significantly, while verification is tied to analysis and support tools, program derivation deals with the very essence of the design process, the way one thinks about problems and constructs solutions [1]. An initial highly- abstract specification is gradually refined up to the point when it contains so much detail that writing a correct program becomes trivial. Program refinement uses a correct program as starting point and alters it until a new program satisfying some additional desired properties is produced. Mobile systems, in general, consist of components that may move in a physical or logical space if the components that move are hosts, the system exhibits physical mobility. If the components are code fragments, the system is said to display logical

mobility, also referred to as code mobility. Code on demand, remote evaluation, and mobile agents are typical forms of code mobility. Of course, many systems entail a combination of both logical and physical mobility (as explained in our related work). The potentially very large number of independent computing units, a decoupled computing style, frequent disconnections, continuous position changes, and the location – dependent nature of the behavior and communication patterns present designers with unprecedented challenges[1][2]. While formal methods may not be ready yet to deliver complete practical systems, the complexity of the undertaking clearly can benefit enormously from the rigor associated with a precise design process, even if employed only in the design of the most critical aspects of the system. The attempt to answer the question raised earlier consists of a formal specification and derivation for our communication protocol for ad hoc mobile systems, carrying out this exercise by employing the mobile unity proof logic and programming notation. Mobile unity provides a notation for mobile system components, coordination language for expressing interactions among the components and an associated proof Logic. This highly modular extension of the UNITY model extends both the physical and logical notations to accommodate specification of and reasoning about mobile programs that exhibit dynamic reconfiguration. Ensuring the availability and the consistency of shared data is a fundamental task for several mobile network applications. For instance, nodes can share data containing configuration information, which is crucial for carrying out cooperative tasks. The shared data can be used, for example, to coordinate the duty cycle of mobile nodes to conserve energy while maintaining network connectivity. The consistency and the availability of the data plays a crucial role in that case since the loss of information regarding the sleep/awake cycle of the nodes might com-promise network connectivity. The consistency and availability of the shared data is also relevant when tracking mobile objects, or in disaster relief applications where mobile nodes have to coordinate distributed tasks without the aid of a fixed communication infrastructure. This can be attained via read/write shared memory provided each node maintains a copy of data regarding the damage assessment and dynamically updates it by issuing write operations. Also in this case it is important that the data produced by the


8

mobile nodes does not get lost, and that each node is able to Retrieve the most up-to-date information. Strong data con-sistency guarantees have applications also to road safety, detection and avoidance of traffic accidents, or safe driving assistance [3].The atomic consistency guarantee is widely used in distributed systems because it ensures that the distributed operations (e.g., read and write operations) performed on the shared memory are ordered consistently with the natural order of their invocation and response time, and that each local copy is conforming to such an order. Intuitively, this implies that each node is able to retrieve a copy showing the last completed update, which is crucial in cooperative tasks. However, the implementation of a fault-tolerant atomic read/write shared memory represents a challenging task in highly mobile networks because of the lack of a fixed infrastructure or nodes that can serve as a backbone. In fact, it is hard to ensure that each update reaches a subset of nodes that is sufficiently large to be retrieved by any node and at any time, if nodes move along unknown paths and at high speed. The focal point model provides a first answer to this challenge since it masks the dynamic nature of mobile ad hoc networks by a static model. More precisely, it associates mobile nodes to fixed geographical locations called focal points. According to this model, a focal point is active at some point in time if its geographical location contains at least one active mobile node. As a result, a focal point becomes faulty when each mobile node populating that sub-region leaves it or crashes. The merit of this model is to study node mobility in terms of failures of stationary abstract points, and to design coordination protocols for mobile networks in terms of static abstract nodes [1] [3].

2. Related Work

In this section, the quorum systems in highly mobile networks are investigated in order to reduce the com-munication cost associated with each distributed operation. Our analysis is driven by two main reasons: (1) guarantee data availability, and (2) reduce the amount of message transmissions, thus conserving energy. The availability of the data is strictly related to the liveness and response time of the recovery protocol since the focal point failures occur continuously, as they are triggered by the motion of nodes. Quorum systems are well-known techniques designed to enhance the performance of distributed systems, such as to reduce the access cost per operation and the load. A quorum system of a universe U is a set of subsets of U, called quorums, such that any pair of quorums does intersect. In this paper, the analyzing of quorum systems is in condition of high node mobility. Note that the universe U of our quorum systems is (Focal Point) FP, a set of n stationary focal points. This choice allows us to study node mobility in terms of continuous failures of stationary nodes. In the next Section, two examples of quorum systems are analyzed and show that they are not always able to guarantee data consistency and availability under the mobility constraints, and provide in Lemma 1 a condition on the size of the minimum quorum intersection that is necessary to guarantee these properties [10].

2.1 Quorum Systems under Mobility Model

This section shows here that quorums proposed for static networks are not able to guarantee data consistency and availability if assumptions A1 and A2 hold, because the minimum quorum intersection is not sufficiently large to cope with the mobility of the nodes. In fact, since read/write operations are performed over a quorum set, in order to guarantee data consistency each read quorum must intersect a quorum containing the last update. We show that there are scenarios that invalidate this condition in case of quorum systems Qg with non-empty quorum intersection, and in case of dissemination quorum systems Qd with minimum quorum intersection equal to f + 1, where f is the maximum number of failures [10].

2.1.1 Generic Quorum System

It is a set of subsets of a finite universe U such that, any two subsets (quorums) intersect (consistency property) and, there exists at least one subset of correct nodes (availability property). The second condition ensures data availability and poses the constraint f < (n/2). In our system model where nodes continuously fail and recover, this condition is not sufficient to guarantee data availability. For instance, in an implementation of a read/write atomic memory based on Qg, the liveness of the read protocol can be violated since it terminates only after receiving a reply from a full quorum of active focal point. Therefore, since the recovery operation involves a read operation, data can become unavailable [10][11].

2.1.2 Dissemination Quorum Systems

They satisfy a stronger consistency property, but insufficient if failures occur continuously. An f-fail-prone system ß C 2U of U is defined as a set of subsets of faulty nodes of U none of which is contained in another, and such

that some B ∈ ß contains all the faulty nodes (whose number does not exceed f).

Definition 1. A dissemination quorum system Q d of U for a f -fail-prone system ß, is set of subsets of U with the following properties:

(i) | Q1 ∩ Q2 | ¢ B √ Q1 , Q2 ∈ Q d, √ B ∈ ß

(ii) √ B ∈ ß э Q ∈ Q d : Q ∩ B = Ø.

Dissemination quorum systems tolerate less than n/3 failures. Unfortunately, since in our system model an additional focal point might fail between the invocation and the response time of a distributed operation, more than f focal points in a quorum set can be non-active at the time they receive the request. As a result, data availability can be violated. The following lemma provides a condition on the minimum quorum intersection size (lower bound) that is necessary to guarantee data consistency and availability under our system model, provided nodes fail and recover [12].


9

Lemma 1. An implementation φ of a read/write shared memory built on top of a quorum system δ. of a universe FP of stationary nodes that fail and recover according to assumptions A1, A2 guarantees data availability only if |Q1

∩ Q2| > f+ 1 for any Q1 , Q2 ∈ Q. It ensures atomic

consistency only if |Q1 ∩ Q2 > f+ 2 for any Q1 , Q2 ∈ Q.

2.2The MDQ Quorum Systems

This section introduces here a new class of quorum systems, called Mobile Dissemination Quorum system (MDQ) that satisfies the condition in Lemma 1.

Definition 2. A MDQ system Q m of FP is a set of subsets of

FP such that |Q1 ∩ Q2|> f + 2 for any Q1 , Q2 ∈ Q. Note that in contrast with Q g the liveness of the distributed operations performed over a quorum set is guaranteed by the minimum number of alive nodes contained in any quorum. As a result, in case of failures the sender does not need to access another quorum in order to complete the operation. This improves the response time in case of faulty nodes and reduces the message transmissions. Let us consider now the following MDQ system:

Qopt = Q : (Q C FP ) ^ ( | Q | =

Lemma 2. Qopt is a MDQ system and f ≤ n - 3.

Proof: Since |Q1 U Q2| = |Q1| + |Q2| - |Q1 ∩ Q2| for any

And |Q1 U Q2| ≤ n, then |Q1 ∩ Q2| ≥ n+f+3-n.

In addition, Q opt tolerates up to n — 3 failures since the size of a quorum cannot exceed n, that is

This implies (f+3-n) / 2 ≤ 0 note that Q opt is highly resilient (in the trivial case f = n -3, Q opt = {U}). Clearly, there is a trade-off between resiliency and access cost since the access cost per operation increases with the maximum number of failures. Moreover, our assumption of connectivity among active focal points becomes harder to guarantee as f becomes larger. It is important to note that the minimum intersection size between two quorums of Q opt is equal to f + 3. We prove in the following section that there exists an implementation of atomic memory built on top of Q opt This shows that f + 3 is the minimum quorum intersection size necessary to guarantee data consistency and data availability under our mobility model. Therefore, Q opt is optimal in the size of the minimum quorum intersection, that is in terms of message transmissions since the sender can compute a quorum consisting of its closest nodes. This is particularly advantageous in sensor networks because it can lead to energy savings [12].

2.3 An Implementation of Read/Write Atomic Memory

In this section, the Q opt is the quorum system with minimum intersection size f+3 that is able to guarantee data consistency and availability under our system model and mobility constraints. We prove that by showing that there exists an implementation φ of atomic read/write memory built on top of Q opt. Our implementation consists of a suite of read, write and recovery protocols and built on top of the focal points and the Qbcast abstraction [13].

2.3.1 The Qbcast Service

We say that a focal point Fi is faulty at time t if focal point region Gi does not contain any active node at time t or Fi is not connected to a quorum of focal points. In our implementation each read, write and recovery request is forwarded to a quorum of focal points. This task is performed by the Qbcast service. It is tailored for the MDQ system and designed for hiding lower level details. Similarly to Qbcast guarantees reliable delivery. It is invoked using interface qbcast (m), where m is the message to transmit containing one of these request tags write, read, confirm. The notation {si}i ∈ Q qbcast (m, Q denotes the Qbcast invocation over quorum Q, {si} i ∈ Q the set of replies, where Q C Q. We call the subset Q the reply set associated with request m. This set plays a crucial role to prove data availability and atomic consistency. Upon receiving a request m, Qbcast computes a quorum Q ∈ Qopt and transmits message m to each focal point in Q. It is important to note that qbcast (m) returns only if the node receives within T time units at least |Q| — (f + 1) replies from Q. If this does not occur, it waits for a random delay and retries later since if this happens the focal point is faulty by our definition. Note that if read (or write) operations occur more frequently than write (or read) operations, we can reduce message transmissions by distinguishing between read and write and making read (or write) quorums smaller. However, for simplicity of presentation we do not distinguish between read and write quorums [12] [13].

2.3.2 Protocols

The high level description of the read/write/ recovery protocols is illustrated in Figure 1. Each mobile node maintains a copy of the state s associated with the shared variable x, which is a compound object containing the value s.val of x, a timestamp s.t representing the time at which a node issued update s.val, and a confirmed tag that indicates if s.val was propagated to a quorum of focal points. Each node can issue write, read and recovery operations. A new state is generated each time a node issues. a write operation.

QQQ opt,2,1 ∈

nfn≤

++

23

++

23fn

n+ f+3

2


10

Figure 1. Write / Read/Recovery Protocols. 2.3.3 Write Protocol

A node C requesting a write v computes a new state s consisting of value v, the current timestamp, tag unconfirmed, and a random identification rand. It transmits its update to a quorum of focal points via the Qbcast service by invoking qbcast (<write, s>) and successively qbcast (< confirm, s>) to make sure that a quorum of focal points received such an update. Upon receiving a write request, each non-faulty focal point (including recovering) replaces its state with the new state s only if the associate timestamp s.t is higher than the timestamp of its local state, and sets its write tag to unconfirmed. This tag is set to confirmed upon receiving the confirm request sent in the second phase of the write protocol, or sent in the second phase of the read protocol in case the node that issued the write operation could not complete the write operation due to failure[13].

2.3.4 Read Protocol

In the read protocol, a node C invokes qbcast (<read>), which forwards the read request to a quorum Q of focal points. Each non-faulty focal point in Q replies by sending a copy of its local state s. Upon receiving a set of replies from the Qbcast service, node C computes the state with highest timestamp and returns the corresponding value. If the tag of s is equal to unconfirmed, it sends a confirm request. This is to guarantee the linearizabitity of the operations performed on the shared data in case a write operation did not complete due to client failure [11].

2.3.5 Recovery Protocol

It is invoked by a node C upon entering an empty region Gi. More precisely, C broadcasts a join request as soon as it enters a new focal point region and waits for replies. If it does not receive any reply within 2d time units, where d is the maximum transmission delay, it invokes the recovery protocol which works in the same way as the read protocol [12].

2.4 Analysis In this section, the key steps to prove the atomic consistency and data availability of the implementation presented in this paper is shown.

A. Data Availability

The availability of the data is a consequence of our failure model and of the Qbcast service. The following lemmas are useful to prove it and will be also used in showing atomic consistency [13].

Lemma 3. The Qbcast service invoked by an active focal point terminates within Τ time units since the invocation time. Proof: This is true since an active focal point or is able to communicate with a quorum of focal points because of Definition 2, and because at most f + 1 focal points in Q can be faulty when the request reaches their focal point regions. In fact, because of assumptions A1 and A2 at most f + 1 focal points can appear to be faulty during T time units. Therefore, at least |Q| — (f+ 1) focal points in a quorum reply. This proves our thesis since the QBcast service guarantees reliable delivery, and the maximum round-trip transmission delay is equal to T. The following lemma and Theorem 1 is a straightforward derivation of the liveness of the Qbcast service. Lemma 4. An active focal point recovers within T time units.

Theorem 1. This implementation of atomic read/write shared memory guarantees data availability.

Lemma 5. At any time in the execution there are at most f+ I faulty and recovering focal points. Proof. Because of Assumptions A1 and A2, and Lemma 4, there are at most f + 1 faulty and recovering focal points during any time interval [t,t + τ ] for any time t in the execution. This can occur if there are f faulty focal points before t and during [t, t + τ] one of these faulty focal points recovers and another one fails.

B. Atomic Consistency

There exists a total ordering of the operations with certain properties. We need to show that the total order is consistent with the natural order of invocations and response. That is, if o1 completes before o2 begins, then o1 < a o2. Lemma 6. The reply set Q associated with a request satisfies the following properties:

Proof. The first property holds because the QBcast service completes only upon receiving at least |Q| - (f+ 1) replies from a quorum of servers. Therefore,

−

≥+−

≥−

221 fnfn

Q

Q for QQ ii

f nQi

optQ any 2)(

;2

)(

∈≥

−≥

−

∩

Write (v):

s ← {v, t, unconfirmed, rand}

{acki} i ∈ Q ← qbcast (<write s>)

{acki} i ∈ Q ← qbcast (<confirm s>)

Read/recovery ( ):

{si} i ∈ Q ← qbcast (<read>)

s ← state ({si}) i ∈ Q

if (s not confirmed)

{acki} i ∈ Q ← qbcast (<confirm s>)

return s.val


11

Since |QUQ| = |Q| + |Q| - |Q∩Q| and |QUQ|

≤ n,

Then,

Therefore, since for any a, b ∈ R,

then

Lemma 7. Let o1 be a write operation with associated state s1. Then, at any time t in the execution with t > res (o1) there exists a subset Mt of active focal points such that,

(i) (Equality holds only if

/focal points are faulty and one is recovering);

(ii) The state s of its active focal points at time t is such

that s1 ≤s s. Proof. Let us denote t1 — res (o1), and I = [t1, t] . We prove the lemma by induction on the number k of subintervals W1, . . ., Wi, . . ., Wk of I of size ≤τ, such that Wi = [t1 + (i - 1) τ , t1 + iτ] for i = 1, . . .,k, and [t1, t2] C Uk

i=1Wi .We want to show that at any time t there exists a subset Mt satisfying definitions 1. And 2.

If k = 1, there exists a subset Mt of active focal points whose state is ≥ s1. It consists of the reply set Q associated with o1, less an eventual additional failure occurred in [t1, t] .Therefore, because of Lemma 6 and Assumption 1 and 2 of our failure model,

The equality holds only if f + 1 focal points in Q did not receive o1request and one of the focal points in Q fails during [t1 , t] . This can occur only if one focal point recovers, because of Assumption 1. In addition, the state of any recovering focal point in W1 is ≥ s1 because M ∩ Q ≠ Ø for each Q∈ Qm In fact

Therefore each focal point that recovered during W1 can be accounted in set Mt after its recovery. Therefore, Only if f focal points are faulty and one is recovering.

3. Methodology and Notation of Mobile Unity

This section provides a gentle introduction to Mobile UNITY. A significant body of published work is available for the reader interested in a more detailed understanding of the model and its applications to the specification and ver-ification of Mobile IP [4], and to the modeling and verifi-cation of mobile code, among others. Each UNITY .program comprises a declare, always, initially, and assign section. The declare section contains a set of variables that will be used by the program. Each is given a name and a type. The always section contains definitions that may be used for convenience in the remainder of the program or in proofs. The initially section contains a set of state predicates which must be true of the program before execution begins. Finally, the assign section contains a set of assignment statements. In each section, the symbol is" is used to separate the individual elements (declarations, definitions, predicates, or statements). Each assignment statement is of the form x = e if p, where x is a list of program variables, e is a list of expressions, and p is a state predicate called the guard [5]. When a statement is selected, if the guard is satisfied, the right-hand side expressions are evaluated in the current state and the resulting values are stored in the variables on the left-hand side. The standard UNITY execution model involves a non-deterministic, weakly-fair execution of the statements in the assign section. The execution of a program starts in a state satisfying the constraints imposed by the initially section. At each step, one of the assignment statements is selected and executed. The selection of the statements is arbitrary but weakly fair, i.e., each statement is selected infinitely often in an infinite execution [5] [6].All executions are infinite. The Mobile UNITY execution model is slightly different, due to the presence of several new kinds of statements, e.g., the reactive statement and the inhibit statement described later. A toy example of a Mobile UNITY program is shown below.

Program host (i) at λ Declare Token: integer Initially Token = 0 Assign Count token: = token + 1 Move:: λ: =Move (i, λ) End host The name of the program is host, and instances are indexed by i. The first assignment statement in host increases the token count by one. The second statement models movement of the host from one location to another. In Mobile UNITY, movement is reduced to value assignment of a special variable λ that denotes the location of the host. We use Move (i, λ) to denote some expression that captures the motion patterns of host (i) [6]. The overall behavior of this toy example host is to count tokens while moving. The program host (i) actually defines a class of programs parameterized by the identifier i. To create a complete system, we must create instances of this program. As shown below, the Components section of

+

≥

+

222baba

12

−

−

≥fn

M t

12

−

−

≥fn

M t

12

−

−

=fn

M t

nf nf n

Q Q −

+++

−≥

−

2

3

2∩

.22

3=−

+≥

−nnQ Q ∩

112

3

2≥−−

+++

−≥ n

f nf nM tQ ∩


12

the Mobile UNITY program accomplishes this. In our example we create two hosts and place them at initial locations λ0 and λ1. System ad-hoc network Program host (i) at λ …………… End host Components host (0) at λ0 host (1) at λ1 Interactions host (0).token, host (1).token:=host (0).token, host (1).token, 0 When (host (0) λ = host (1).token. λ) ^ (host (1).token = 0) Inhibit host (l).move and host (0).move When (host (0). λ = host (1). λ) ^ (host (l).token > 10) End ad-hoc network

Unlike UNITY, in Mobile UNITY all variables are local to each component. A separate section specifies coordination among components by defining when and how they share data. In mobile systems, coordination is typically location dependent. Furthermore, in order to define the coordination rules, statements in the Interactions section can refer to variables belonging to the components themselves using a dot notation. The section may be viewed as a model of physical reality (e.g., communication takes place only when hosts are within a certain range) or as a specification for desired system services. The operational semantics of the inhibit construct is to strengthen the guard of the affected statements whenever the when clause is true. The statements in the Interactions section are selected for execution in the same way as those in the component programs. Thus, without the inhibit statement, host(0) and host(l) may move away from each other before the token collection takes place, i.e., before the first interaction statement is selected for execution. With the addition of the inhibit statement, when two hosts are co-located, and host(l) holds more than ten tokens, both hosts are prohibited from moving, until host(l) has fewer than eleven tokens[7]. The inhibit construct adds both flexibility and control over the program execution.In addition to its programming notation, Mobile UNITY also provides a proof logic, a specialization of temporal logic. As in UNITY, safety properties specify that certain state transitions are not possible, while progress properties specify that certain actions will eventually take place. The safety properties include unless, invariant, and stable [7]:

• p unless q asserts that if the program reaches a state in which the predicate (p ^ ¬q) holds, p will continue to hold at least as long as q does not, which may be forever. • Stable p is defined as p unless false, which states that once p holds; it will continue to hold forever. • Inv p means ((INIT) p) ^ stable p), i.e., p holds

initially and throughout the execution of the program. INIT characterizes the program's initial state.

The basic progress properties include ensures, leads-to, until, and detects: • p ensures q simply states that if the program reaches a state where p is true, p remains true as long as q is false, and there is one statement that, if selected, is guaranteed to make the predicate q true- This is used to define the most basic progress property of programs. • p leads-to q states that if program reaches a state where p is true, it will eventually reach a state in which q is true. Notice that in the leads-to, p is not required to hold until q is established. • p until q defined as ((p leads-to q) ^ (p unless q)), is used to describe a progress condition which requires p to hold up to the point when q is established. • p detects q is defined as (p q) ^ (q leads-to p) All of the predicate relations defined above represent a short-hand notation for expressions involving Hoare triples quantified over the set of statements in the system. Mobile UNITY and UNITY logic share the same predicate rela-tions. Differences become apparent only when one examines the definitions of unless and ensures and the manner in which they handle the new programming constructs of Mobile UNITY. Here are some properties the toy-example satisfies: (1) (host (0).token + host (l).token = k) Unless (host (0).token + host (l).token > k) — The total count will not decrease (2) host (0).token = k leads-to host (0).token > k — The number of tokens on host (0) will eventually increase

In the next section we employ the Mobile UNITY proof logic to give a formal requirements definition to the geoquorum approach (the application of the paper).

4. The Geoquorum-Approach (The Application)

In this paper the Geoquorum algorithm is presented for implementing the atomic read/write in shared memory of mobile ad hoc networks. This approach is based on associating abstract atomic objects with certain geographic locations. It is assumed that the existence of Focal Points, geographic areas that are normally "populated" by mobile nodes. For example: a focal point may be a road Junction, a scenic observation point. Mobile nodes that happen to populate a focal point participate in implementing a shared atomic object, using a replicated state machine approach. These objects, which are called focal point objects, are prone to Occasional failures when the corresponding geographic areas are depopulated. The Geoquorums algorithm uses the fault-prone focal point objects to implement atomic read/write operations on a fault-tolerant virtual shared object. The Geoquorums algorithm uses a quorum- based strategy in which each quorum consists of a set of focal point objects. The quorums are used to maintain the consistency of the shared memory and to tolerate limited failures of the focal point objects, which may be caused by depopulation of the corresponding geographic areas. The mechanism for changing the set of quorums has presented, thus improving efficiency [8]. Overall, the new Geoquorums


13

algorithm efficiently implements read/write operations in a highly dynamic, mobile network. In this chapter, a new approach to designing algorithms for mobile ad hoc networks is presented. An ad hoc network uses no pre-existing infrastructure, unlike cellular networks that depend on fixed, wired base stations. Instead, the network is formed by the mobile nodes themselves, which co-operate to route communication from sources to destinations. Ad hoc communication networks are by nature, highly dynamic. Mobile nodes are often small devices with limited energy that spontaneously join and leave the network. As a mobile node moves, the set of neighbors with which at can directly communicate may change completely. The nature of ad hoc networks makes it challenging to solve the standard problems encountered in mobile computing, such as location management using classical tools. The difficulties arise from the lack of a fixed infrastructure to serve as the backbone of the network. In this section developing a new approach that allows existing distributed algorithm to be adapted for highly dynamic ad hoc environments one such fundamental problem in distributed computing is implementing atomic read/ write shared memory [8]. Atomic memory is a basic service that facilitates the implementation of many higher level algorithms. For example: one might construct a location service by requiring each mobile node to periodically write its current location to the memory. Alternatively, a shared memory could be used to collect real – time statistics, for example: recording the number of people in a building here, a new algorithm for atomic multi writes/multi- reads memory in mobile ad hoc networks. The problem of implementing atomic read/write memory is divided into two parts; first, we define a static system model, the focal point object model that associates abstract objects with certain fixed geographic locales. The mobile nodes implement this model using a replicated state machine approach. In this way, the dynamic nature of the ad hoc network is masked by a static model. Moreover it should be noted that this approach can be applied to any dynamic network that has a geographic basis. Second, an algorithm is presented to implement read/write atomic memory using the focal point object model. The implementation of the focal point object model depends on a set of physical regions, known as focal points .The mobile nodes within a focal point cooperate to simulate a single virtual object, known as a focal point object. Each focal point supports a local broadcast service, LBcast which provides reliable, totally ordered broadcast. This service allows each node in the focal point to communicate reliably with every other node in the focal point. The focal broadcast service is used to implement a type of replicated state machine, one that tolerates joins and leaves of mobile nodes. If a focal point becomes depopulated, then the associated focal point object fails. (Note that it doesn't matter how a focal point becomes depopulated, be it as a result of mobile nodes failing, leaving the area, going to sleep. etc. Any depopulation results in the focal point failing). The Geoquorums algorithm implements an atomic read/write memory algorithm on top of the geographic abstraction, that is, on top of the focal point object model. Nodes implementing the atomic memory use a Geocast service to communicate with the focal point objects. In order to achieve fault tolerance and availability, the

algorithm replicates the read/write shared memory at a number of focal point objects. In order to maintain consistency, accessing the shared memory requires updating certain sets of focal points known as quorums. An important aspect of our approach is that the members of our quorums are focal point objects, not mobile nodes. The algorithm uses two sets of quorums (I) get-quorums (II) put- quorums with property that every get-quorum intersects every put-quorum. There is no requirement that put-quorums intersect other put-quorums, or get-quorums intersect other get-quorums. The use of quorums allows the algorithm to tolerate the failure of a limited number of focal point objects. Our algorithm uses a Global Position System (GPS) time service, allowing it to process write operations using a single phase, prior single-phase write algorithm made other strong assumptions, for example: relying either on synchrony or single writers. This algorithm guarantees that all read operations complete within two phases, but allows for some reads to be completed using a single phase: the atomic memory algorithm flags the completion of a previous read or write operation to avoid using additional phases, and propagates this information to various focal paint objects[9]. As far as we know, this is an improvement on previous quorum based algorithms. For performance reasons, at different times it may be desirable to use different times it may be desirable to use different sets of get quorums and put-quorums. For example: during intervals when there are many more read operations than write operations, it may be preferable to use smaller get- quorums that are well distributed, and larger put-quorums that are sparsely distributed. In this case a client can rapidly communicate with a get-quorum while communicating with a put – quorum may be slow. If the operational statistics change, it may be useful to reverse the situation. The algorithm presented here includes a limited "reconfiguration" Capability: it can switch between a finite number of predetermined quorum systems, thus changing the available put-quorums and get –quorums. As a result of the static underlying focal point object model, in which focal point objects neither join nor leave, it isn't a severe limitation to require the number of predetermined quorum systems to be finite (and small). The resulting reconfiguration algorithm, however, is quite efficient compared to prior reconfigurable atomic memory algorithms. Reconfiguration doesn't significantly delay read or write operations, and as no consensus service is required, reconfiguration terminates rapidly.

The mathematical notation for the geoquorum approach - I the totally- ordered set of node identifiers.

- I0 є I, a distinguished node identifier in I that is smaller than all order identifiers in I.

- S, the set of port identifiers, defined as N>0× OP×I, Where OP= {get, put, confirm, recon- done}. - O, the totally- ordered, finite set of focal point

identifiers. - T, the set of tags defined as R ≥0 × I. - U, the set of operation identifiers, defined as R ≥0 ×

S. - X, the set of memory locations for each x є X: - Vx the set of values for x - v0,x є Vx , the initial value of


14

X - M, a totally-ordered set of configuration names - c0 є M, a distinguished configuration in M that is smaller than all other names in M. - C, totally- ordered set of configuration identifies, as defined as: R ≥0 ×I ×M - L, set of locations in the plane, defined as R× R

Fig .2 Notations Used in The Geoquorums Algorithm.

Variable Types for Atomic Read/Write object in Geoquorum Approach for Mobile Ad Hoc Network The specification of a variable type for a read/write object in geoquorum approach for mobile ad hoc network is presented. A read/write object has the following variable type (see fig .3) [8].

Put/get variable type τ State

Tag ∈ T, initially< 0.i0> Value ∈ V, initially v0 Config-id ∈ C, initially< 0, i0, c0> Confirmed-set C T, initially Ø Recon-ip, a Boolean, initially false Operations Put (new-tag, new-value, new-Config-id) If (new-tag> tag) then Value ←new-value Tag ← new-tag If (new-Config-id > Config-id) then Config-id ← new-config-id Recon-ip ← true Return put-Ack (Config-id, recon-ip) Get (new-config-id) If (new-config-id >Config-id) then Config-id ← new-Config-id Recon-ip ←true Confirmed ← (tag ∈ confirmed-set) Return get-ack (tag, value, confirmed, Config-id, recon-ip) Confirm (new-tag) Confirmed-set ←confirmed –set U {new-tag} Return confirm-Ack Recon –done (new-Config-id) If (new-Config-id=Config-id) then Recon-ip ←false Return recon-done-Ack ( )

Fig .3 Definition of the Put/Get Variable Type τ

4.1 Operation Manager

In this section the Operation Manger (OM) is presented, an algorithm built on the focal/point object Model. As the focal point Object Model contains two entities, focal point objects and Mobile nodes, two specifications is presented , on for the objects and one for the application running on the mobile nodes [8] [9].

4.1.1 Operation Manager Client

This automaton receives read, write, and recon requests from clients and manages quorum accesses to implement

these operations (see fig .4). The Operation Manager (OM) is the collection of all the operation manager clients (OMi, for all i in I).it is composed of the focal point objects, each of which is an atomic object with the put/get variable type:

Operation Manager Client Transitions Input write (Val) i Effect: Current-port-number Current-port-number +1 Op < write, put, <clock, i>, Val, recon-ip, <0, i0, c0>, Ø> Output write-Ack ( ) i Precondition: Conf-id=<time-stamp, Pid, c> If op .recon-ip then √ C/ ∈ M, э P ∈put-quorums(C/): P C op. acc Else Э P ∈put-quorums(C): P C Op. acc Op .phase=put Op. type=write Effect: Op. phase idle Confirmed confirmed U {op. tag} Input read ( ) i Effect: Current-port-number Current-port-number +1 Op < read, get, ┴, ┬, recon-ip, <0, i0, c0>, Ø> Output read-ack (v) i Precondition: Conf-id=<time-stamp, Pid, c> If op. recon-ip then √ C/ ∈ M, э G ∈get-quorums(C/): G C op. acc Else Э G ∈get-quorums(C): G C op. acc Op. phase=get Op. type=read Op. tag ∈confirmed v= op. value Effect: Op .phase idle Internal read-2( )i Precondition: Conf-id=<time-stamp, Pid, c> √ C/ ∈ M, э G ∈get-quorums(C/): G C op. acc Else Э G ∈get-quorums(C): G C op. acc Op. phase=get Op. type=read

Op. tag ∈ confirmed Effect: Current-port-number Current-port-number +1 Op. phase put Op. Recon. ip recon-ip Op. acc Ø Output read-Ack (v)i Precondition: Conf-id=<time-stamp, Pid, c>


15

If op. recon-ip then √ C/ ∈ M, э P ∈put-quorums(C/): P C op. acc Else Э P ∈put-quorums(C): P C op. acc Op. phase=put Op. type=read v=op. value Effect: Op. phase idle Confirmed confirmed U {op. tag} Input recon (conf-name)i Effect: Conf-id <clock, i, conf-name> Recon-ip true Current-port-number Current-port-number +1 Op < recon, get, ┴, ┴, true, conf-id, Ø> Internal recon-2(cid) i Precondition √ C/ ∈ M, э G ∈get-quorums(C/): G C op. acc √ C/ ∈ M, э P ∈put-quorums(C/): P C op. acc Op. type=recon Op. phase=get Cid=op. recon-conf-id Effect Current-port-number Current-port-number +1 Op. phase put Op. acc Ø Output recon-Ack(c) i Precondition Cid=op. recon-conf-id Cid= <time-stamp, Pid, c> Э P ∈put-quorums(C): P C op. acc Op. type=recon Op. phase=put Effect: If (conf-id=op. recon-conf-id) then Recon-ip false Op. phase idle Input geo-update (t, L) i Effect: Clock 1

Fig .4 Operation Manager Client Read/Write/Recon and Geo-update Transitions for Node

4.2 Focal Point Emulator Overview

The focal point emulator implements the focal point object Model in an ad hoc mobile network. The nodes in a focal point (i.e. in the specified physical region) collaborate to implement a focal point object. They take advantage of the powerful LBcast service to implement a replicated state machine that tolerates nodes continually joining and leaving .This replicated state machine consistently maintains the state of the atomic object, ensuring that the invocations are performed in a consistent order at every mobile node [8].In this section an algorithm is presented to implement the focal point object model. the algorithm allows mobile nodes moving in and out of focal points, communicating with

distributed clients through the geocast service, to implement an atomic object (with port set q=s)corresponding to a particular focal point. We refer to this algorithm as the Focal Point Emulator (FPE). The FPE client has three basic purposes. First, it ensures that each invocation receives at most one response (eliminating duplicates).Second, it abstracts away the geocast communication, providing a simple invoke/respond interface to the mobile node [9]. Third, it provides each mobile node with multiple ports to the focal point object; the number of ports depends on the atomic object being implemented. The remaining code for the FPE server is in fig .5.When a node enters the focal point, it broadcasts a join-request message using the LBcast service and waits for a response. The other nodes in the focal point respond to a join-request by sending the current state of the simulated object using the LBcast service. As an optimization, to avoid unnecessary message traffic and collisions, if a node observes that someone else has already responded to a join-request, and then it does not respond. Once a node has received the response to its join-request, then it starts participating in the simulation, by becoming active. When a node receives a Geocast message containing an operation invocation, it resends it with the Lbcast service to the focal point, thus causing the invocation to become ordered with respect to the other LBcast messages (which are join-request messages, responses to join requests, and operation invocations ).since it is possible that a Geocast is received by more than one node in the focal point ,there is some bookkeeping to make sure that only one copy of the same invocation is actually processed by the nodes. There exists an optimization that if a node observes that an invocation has already been sent with LBcast service, then it does not do so. Active nodes keep track of operation invocations in the order in which they receive them over the LBcast service. Duplicates are discarded using the unique operation ids. The operations are performed on the simulated state in order. After each one, a Geocast is sent back to the invoking node with the response. Operations complete when the invoking node with the response. Operations complete when the invoking node remains in the same region as when it sent the invocation, allowing the geocast to find it. When a node leaves the focal point, it re-initializes its variables .A subtle point is to decide when a node should start collecting invocations to be applied to its replica of the object state. A node receives a snapshot of the state when it joins. However by the time the snapshot is received, it might be out of date, since there may have been some intervening messages from the LBcast service that have been received since the snapshot was sent. Therefore the joining node must record all the operation invocations that are broadcast after its join request was broadcast but before it received the snapshot .this is accomplished by having the joining node enter a "listening" state once it receives its own join request message; all invocations received when a node is in either the listening or the active state are recorded, and actual processing of the invocations can start once the node has received the snapshot and has the active status. A precondition for performing most of these actions that the node is in the relevant focal point. This property is covered in most cases by the integrity requirements of the LBcast and Geocast services, which


16

imply that these actions only happen when the node is in the appropriate focal point [8].

Focal Point Emulator Server Transitions Internal join ( ) Obj , i Precondition: Location ∈ FP-location Status=idle Effect: Join-id ←<clock, i> Status← joining Enqueue (Lbcast-queue,<join-req, join-id>) Input Lbcast- rcv (< join-req, jid>) obj, i Effect: If ((status=joining)) ^ (jid=Join-id)) then Status ←listening If ((status=active)) ^) jid ∉answered-join-reqs)) then Enqueue (LBcast-queue, < join-ack, jid, val>) Input Lbcast- rcv (<join-ack, jid, v>) obj, i Effect: Answered-join-reqs ← answered-join-reqs U {jid} If ((status=listening) ^ (jid =join-id)) then Status ← active val V Input Geocast –rcv (< invoke, inv, oid, loc, FP-loc>) obj,i Effect: If (FP-loc=FP-location) then If (<inv, oid, loc>∉ pending-ops U completed ops) then Enqueue (Lbcast-queue, <invoke, inv, oid, loc>) Input LBcast –rcv (< invoke, inv, oid, loc>) obj,i Effect: If ((status=listening V active) ^ (<inv, oid, loc>∉pending-ops U completed-ops)) Then Enqueue (pending-ops, <inv, oid, loc>) Internal simulate-op (inv) obj, i Precondition: Status=active Peek (pending-ops) =<inv, oid, loc> Effect: (Val, resp)← δ (inv, val) Enqueue (geocost- queue, < response, resp, oid, loc>) Enqueue (completed-ops, Dequeue (pending-ops)) Internal leave ( ) obj, i Precondition: Location ∉ fp-location Status ≠ idle Effect: Status← idle Join-id← <0, i0> Val ← v0 Answered -join- reqs← Ø Pending –ops ← Ø Completed-ops ← Ø Lbcast-queue ← Ø Geocast-queue ← Ø Output Lbcast (m) obj, i Precondition: Peek (Lbcast-queue) =m Effect: Duqueue (Lbcast- queue) Output geocast (m) obj, i

Precondition: Peek (geocast-queue) =m Effect: Dequeue (geocost- queue) Input get-update (l, t) obj,i Effect: Location ← l Clock← t Fig. 5 FPE server transitions for client i and object Obj of variable

type τ = <V, v0, invocations, responses, δ > 5. Problem Specification

The methodology for formal specification of the geoquorum approach is illustrated by considering a set of mobile nodes with identifiers with values from 0 through (N-1) moving through space. Initially some of the nodes are idle while others are active. Nodes communicate with each other while in range. A node can becomes idle at any time but can be reactivated if it encounters an active node. The basic requirement is that of determining that all nodes are idle and storing that information in a Boolean flag (claim) located on some specific node say node (i0), formally the problem reduces.

Stable W (S1)

Claim detects W (P1)

Where W is the condition

W=< ^ i: 0 < i ≤N:: idle [i] > (D1)

(S1) is a safety property stating that once all nodes are idle, no node ever becomes active again. (P1) is a progress property requiring the flag claim to eventually record the system's quiescence. Using idle [i] to express the quiescence of a node and define active [i] to be its negation. It is important to note that the problem definition in this case does not depend on the nature of the underlying computation.

5.1 Formal Derivation

In this section, specification refinement techniques are employed toward the goal of generating a programming solution that accounts for the architectural features of ad hoc networks which form opportunistically as nodes move in and out of range.

The refinement process starts by capturing high level behavioral features of the underlying application [14] [15]. In each case, we provide the informal motivation behind that particular step and show the resulting changes to the specification. As an aid to the reader, each refinement concludes with a list of specification statement labels that captures the current state of the refinement process, as in:


17

Refinement 0: P1, S1

5.1.1 Refinement 1: Activation Principle

A node invocation may become put, get, confirm, reconfig – done. The safety property of these computations can be expressed as:

Get [i] (S2)

Unless

Put [i] (S3)

Unless

Confirm [i] (S4)

Unless

Unless

< Э id, ip: config-id ≠ new-config- id:: recon-ip=false>

In previous, it is determined that all invocations and its conditions of the application. Refinement 1: P1, S2, S3, S4, S5.

5.1.2 Refinement 2: Parameters Based Accounting

Frequent node movement makes direct counting inconvenient, but we can accomplish the same thing by associating id, port-number with each invocation node. Independent of whether it is currently idle or active, each node in the system holds zero or more ids. The advantage of this approach is that ids can be collected and then counted at the collection point. If we define D to be the number of ids in the system and I to be the number of confirm nodes, i.e.

)3( 1 :: ][ :

)2( ][::

DD

>+≡<

>+≡<

iconfirmiI

iidiD

The relationship between the two is established by the invariant:

Inv.D = I (S6)

By adding this constraint to the specification, the quiescence property (W) may be replaced by the predicate (D = N),

where N is the number of nodes in the system. Property (P1) is then replaced by:

Claim detects D = N (P2)

With the collection mechanism left undefined for the time being.

Refinement 2: P2, S2, S3, S4, S5, S6.

5.1.3 Refinement 3: Config-Ids Increasing

To maintain the invariant that the number of ids in the system reflects the number of idle nodes, activation of an idle node requires that the number of ids increase by one. Therefore, when an active node put invoke an idle node, they must increase config-id between them. To express this, we add a history variable config-id / [i], which maintains the value config-id [i], held before it changed last. And the put invocation is a history state of the put-ack state node; this for all states of nodes of the related work and the safety property of the put invocation is as follow:

Put [i] (S7)

Put-ack [j]

Unless

Captures the requirement that, when node i activates and becomes node j, the config-id of node j must be increase in the same step. Clearly, this new property strengthens property (S2), (S3), (S6).

Refinement 3: P2, S4, S5, S7.

5.1.4 Refinement: Operations –Id Collection

According to FPE server, FPE client algorithms, the completed operations is arranged in rank of operations that have been simulated initially ø and there exist Val ∈ V, holds current value of the simulated atomic node, initially v0. oid is a history id of the operation; new-oid is the current id of the operation. To simplify our narration, a host with a higher id is said to rank higher than a node with a lower id. Oid (v0) should eventually collect all N oids. We will accomplish this by forcing nodes to pass their oids to lower ranked nodes for this, we introduce two definitions:

L=<+i: obj [i]:: oid [i] > (D4)

Count the number of oids idle agents hold. Obviously, L = N, when all nodes are idle. We also add

tag)-new Tag id,-config-newid-config ::tag- new Tag id,-config-new id-config :Tid,

>>≠≠∃<

>>≠∃< id-config - new id-config ::id-config-new id-config :id

)(S [i] done-Recontag- new Tag :: tag-new Tag :T

5

>>≠∃<


18

w= < max i: L=N ^ oid [i] > 0:: i > (D5)

To keep track of the highest ranked node holding oids. After all nodes are idle, we will force w to decrease until it reaches 0.when w=0 and L=N, obj (v0) will have collected all the oids. At this stage we add a new progress property,

)(k wuntil 0 P3<>= kw

That begins to shape the progress of oid passing. As mentioned, the until property is a combination of a safety property (the unless portion) and a progress property (the leads – to portion). As long as the highest ranked oid holder passes its oids to a lower ranked node, we can show that all the oids will reach obj (V0= 0) without having to restrict the behavior of any node except node (w) = obj (w). Some can replace (P2) with

Claim detects (w=0) (P4)

Refinement 4: P2, P3, S4, S5, S7

5.1.5 Refinement 5: Pairwise Communication

According to the code for the FPE client and FPE server which discussed in section 3.2 clearly, a node can only activate another node or pass (join-ids=jid) to another node if the two parties can communicate. To accomplish this, we introduce the predicate, com (i, j) that holds if and only if nodes i and j can communicate.

We begin this refinement with a new safety property:

Idle [i] (S8)

Unless

< Э j: j = i : : active/[ j] ^ active [j] ^ active [i ] ^ (join-id[i] + join-id [j]=(join-id)/[i] +(join-id)/[j]-1)> 0 ^ com (i,j)>

This requires that nodes i and J be in communication when J activates i. Also, adding the property:

Join-id [j] > 0 ^ L=N ^ j = 0 (P5)

Until

Join-id [j] = 0 ^ L= N ^ < Э i< j:: com (i, j)>

This requires a node to pass its jids and when it does, to have been able to communicate with a lower ranked node. As we leave this refinement, we separate property (P5) into its two parts; a progress property,

Join –id [J] > 0 ^ L = N ^ J ≠ 0 (P6)

Leads-to

Join-id[j] > 0 ^ L = N ^ < Э I < j:: com (i, j)> , And a safety property,

Join-id [j] > 0 ^ L= N ^ j = 0 (S9)

Unless

Join-id[j] = 0 ^ L=N ^ < Э i< j:: com (i, j)> Refinement 5: P2, P3, P6, S4, S5, S7, S8, S9

5.1.6 Refinement 6: Contact Guarantee

Property (P6) conveys two different things. First; it ensures that a node with join-ids will meet a lower ranked node, if one exists. Second, it requires the join-ids to be passed to the lower ranked node. The former requires us to either place restrictions on the movement of the mobile nodes or make assumptions about the movement. For this reason, we refine property (P6) into two obligations. The first

Join-id [J] > 0 ^ L = N (P7)

Leads – to

Join-id [j] > 0 ^ L = N ^ < Э i< j:: com (i,j)>

Guarantees that a node with join-ids will meet a lower ranked node. The second,

Join-id [j] > 0 ^ L = N ^ < Э i< j:: com (i, j)>

Lead-to

Join-id [j] = 0 ^ L = N ^ < Э i< j:: com (i,j)> Forces a node that has met a lower ranked node, to pass its join-ids. At the point of passing, communication is still available. There two new properties replace property (P6).

Refinement 5: P2, P3, P7, P8, S4, S5, S6, S7, S8, S9

6. Conclusions and Future Work In this paper we presented the formal derivation of geoquorum approach for mobile ad hoc networks. This formal specification is built from mobile unity primitives and proof logic .also this paper provides strong evidence that a formal treatment of mobility and its applications isn't only feasible but, given the complexities of mobile computing. There is an open area for using this specification to mechanistically construct the program text as: first, defining the program components, and then deriving the program statements directly from the final specification, the resulting program (called a system in mobile UNITY). Also, we have viewed a small set of mobility constraints that are necessary to ensure strong data guarantees in highly mobile networks. We have also discussed quorum systems in highly mobile networks and devised a condition that is necessary for a quorum system to guarantee data consistency and availability under our mobility constraints as a survey. This work leaves several open questions such as the problem of


19

dealing with network partitions and periods of network instability in which our set of assumptions are invalid.

Acknowledgment

The authors would like to thank the Conference INFOS 2010 (Cairo University) reviewers for their constructive comments and suggestions.

References

[1]A. Smith, H. Balakrishnan, M. Goraczko, N. Priyantha, "Support for Location: Tracking Moving Devices with the Cricket Location System", in: Proceedings of the 2nd International Conference on Mobile Systems, Applications, and Services, Jun 2004.

[2]S. Gilbert, N. Lynch, A. Shvartsman, "RAMBO II: Rapidly Reconfigurable Atomic Memory for Dynamic Networks," in: Proceedings of the International Conference on Dependable Systems and Networks, June 2003, PP. 259-269.

[3]B. Liu, P. Brass, O. Dousse, P. Nain, D. Towsley, "Mobility Improves Coverage of Sensor Networks", in: Proceedings of Mobile Ad Hoc, May 2005, PP. 300-308.

[4]R. Friedman, M. Gradinariu, G. Simon, "Locating Cache Proxies in MANETs", in: Proceedings of the 5th International Symposium of Mobile Ad Hoc Networks, 2004, PP. 175-186.

[5]J. Luo, J-P. Hubaux, P. Eugster," Resource Management: PAN: Providing Reliable Storage in Mobile Ad Hoc Networks with Probabilistic Quorum Systems", in: Proceedings of the 4th International Symposium on Mobile Ad Hoc Networking and Computing, 2003, PP. 1-12.

[6]D. Tulone, Mechanisms for Energy Conservation in Wireless Sensor Networks. Ph.D. Thesis, Department of Computer Science, University of Pisa, Dec 2005.

[7]W. Zhao, M. Ammar, E. Zegura,"A Message Ferrying Approach for Data Delivery in Sparse Mobile Ad Hoc Networks", in: Proceedings of the 5th International Symposium on Mobile Ad hoc Networking and Computing, May 2004, PP.187-198.

[8]S. Dolev, S. Gilbert, N. Lynch, A. Shvartsman, J. Welch, "GeoQuorums: Implementing Atomic Memory in Mobile Ad Hoc Networks", in: Proceedings of the 17th International Conference on Distributed Computing, October 2003, PP. 306-320.

[9]H. Wu, R. Fujimoto, R. Guensler, M. Hunter, "MDDV: A Mobility-Centric Data Dissemination Algorithm for Vehicular Networks", in: Proceedings of the1st International Workshop on Vehicular Ad hoc Networks, Oct 2004, PP. 47-56.

[10]J. Polastre, J. Hill, D. Culler, "Versatile Low Power Media Access for Wireless Sensor Networks", in: Proceedings of SenSys, 2004.

[11]S. PalChanduri, J.-Y. Le Boudec, M. Vojnovic, "Perfect Simulations for Random Mobility Models", in: Annual Simulation Symposium 2005, PP. 72-79. Available from: http://www.cs.rice.edu/santa/ research/mobility

[12]J.Y. Le Boudec, M. Vojnovic, "Perfect Simulation and Stationarity of A Class Of Mobility Models", Infocom (2005).

[13]T. Hara, "Location Management of Replication Considering Data Update in Ad Hoc Networks, in: 20th International Conference AINA, 2006, PP. 753-758.

[14]T. Hara," Replication Management for Data Sharing In Mobile Ad Hoc Networks", Journal of Interconnection Networks 7(1) (2006), PP.75-90.

[15]Y Sawai, M. Shinohara, A. Kanzaki, T. Hara, S. Nishio, "Consistency Management Among Replicas Using A Quorum System in Ad Hoc Networks", MDM (2006), PP.128-132.


20

Power Management in Wireless ad-hoc Networks with Directional Antennas

Mahasweta Sarkar1 and Ankur Gupta2

1Department of Electrical and Computer Engineering San Diego State University, San Diego, California, USA

Email: [email protected]

2Department of Electrical and Computer Engineering San Diego State University, San Diego, California, USA

Email: [email protected]

Abstract: Power control in wireless ad-hoc networks has been known to significantly increase network lifetime and also impact system throughput positively. In this paper, we investigate the virtues of power control in an ad-hoc network where nodes are equipped with directional antennas. We use a cross-layer power optimization scheme. The novelty of our work lies in the fact that we exploit the knowledge of the network topology to dynamically change our power control scheme. The information about the network topology is retrieved from the Network Layer. Specifically, we consider three different network topologies – (i) a sparse topology, (ii) a dense topology and a (iii) cluster-based topology. The MAC layer, implements a dynamic power control scheme which regulates the transmission power of a transmitting node based on the node density at the receiver node. In addition a transmitting node can adjust its transmission power to the optimal level if the SINR value at the receiving node is known. Our cross-layer power control algorithm takes the above fact into account. Extensive simulation in NS2 shows that in all the three topologies our cross layer design (use of directional antennas with power control) leads to prolonged network lifetime and significantly increases the system throughput.

Keywords: ad hoc network, power management, directional antenna, cross layer

1. Introduction

An ad-hoc or short-live network is the network of two or more mobile devices connected to each other without the help of intervening infrastructure. In contrast to a fixed wireless network, an ad-hoc network can be deployed in remote geographical locations and requires minimum setup and administration costs. Moreover, the integration of an ad-hoc network with a bigger network-such as the Internet-or a wireless infrastructure network increases the coverage area and application domain of the ad-hoc network. However, communication in an ad-hoc network between different hosts that are not directly linked is an issue not only for search and rescue operations, but also for educational and business purposes.

An ad-hoc network can be classified into two main types: mobile ad-hoc network and mobile ad-hoc sensors network. Unlike typical sensor networks, which communicate directly with the centralized controller, a mobile ad-hoc sensor network follows a broader sequence of operational scenarios, thus demanding a less complex setup procedure. A mobile

ad-hoc sensor or hybrid ad-hoc network consists of a number of sensor spreads in a geographical area. Each sensor is capable of mobile communication and has some level of intelligence to process signals and to transmit data. In order to support routed communications between two mobile nodes, the routing protocol determines the node connectivity and routes packets accordingly. This makes a mobile ad-hoc sensor network highly adaptable so that it can be deployed in almost all environments like military operations, security in shopping malls and hotels or to locate a vacant parking spot.

Each node in the network is equipped with an antenna (omni directional or directional) which enables it to find its neighbors and communicate with them. Since the antenna plays such an important role in the communication, the choice of the antenna becomes critical. With the existing schemes, an omni directional antenna was used both at the transmitter and at the receiver end. Although using this configuration does enable us to achieve the task of communication, there are many drawbacks of using an omni direction antenna as compared to using a directional one. Extensive research in MAC and PHY layer protocols is being conducted to support these antennas.With the advancement in antenna technology it has become possible to use array antennas and purely directional antennas in ad-hoc networks. Some of the advantages of using directional antennas as compared to omni directional antennas are increased range, spatial reuse and reduced collisions. Since the antenna beam can be pointed to a particular direction and all the energy can be focused there, the range is increased. Unlike omni directional antennas, by using directional antennas we can have more nodes communicating in the same space, thereby exploiting spatial reuse and increasing the throughput of the system. Since omni directional antennas transmit and receive in all the directions, they are more prone to collisions, which is not the case when using directional antennas. Now that the choice of antennas has been discussed, we discuss the main focus of this paper which is power control. The first drawback of using the fixed-power approach is that it negatively impacts channel utilization by not allowing concurrent transmissions to take place over the reserved floor. The second drawback of the fixed-power approach is that the received power may be far more than necessary to achieve the required signal-to-interference-and-noise ratio (SINR), thus wasting the node’s energy and shortening its lifetime. Therefore, there is a need for a solution, possibly a


21

cross layer one that allows concurrent transmissions to take place in the same vicinity and simultaneously conserves energy. Ad-hoc networks could be broadly classified into three topologies namely, spare, dense and cluster form. Since the MAC and Network layers play an important role in conserving power and increasing the system throughput, it is essential to have a cross layer design that can support different topologies. The MAC layer in a system controls how and when a node accesses the medium. This also includes the transmission power used for communication. If a transmitting node is aware of the ideal SINR value at the receiver, it can adjust its transmitting power such that only required amount of power is used for communication. In doing so not only does the transmitter save power (compared to transmitting at full power) but also interferes with lesser nodes (as increased power would mean increased range). In case of sparse network, we introduce the “locally aware” power management protocol in which a receiver can inform its transmitter of the minimum required SINR (to be able to decode the signal) so that the transmitter can adjust its transmitting power accordingly. In a dense topology a receiver may face interference from more nodes than in a sparse topology. We propose the “globally aware” protocol here. In such a case it is essential that the receiver transmits its minimum required SINR value in all the directions. This will ensure that all the transmitters in this receiver’s neighborhood will control their power such that they do not interfere with this receiver anymore but can still carry on with their own communications. Both the sparse and the dense topology designs are supported by a power aware routing protocol at the Network layer. Since a sensor network is a network in which nodes gather relevant data about an activity and send it to an actor node which processes the data further; it is important that efficient routing is performed from the sensor nodes to the actor nodes. Cluster topologies are very common in ad-hoc sensor networks to record signals of interest at certain places only. Fig. 1 shows the difference between a homogeneous and non-homogeneous spatial distribution of nodes. We study the changes in power consumed and throughput when all cluster nodes transmit at the same power to when there is power control. Power control is implemented by making sure that each cluster uses its own power level such that it does not interfere with other cluster nodes. This simple cross layer design helps prolong the network lifetime and also shows increase in throughput.

Figure 1. Difference between a Homogeneous and Cluster Form Spatial Distribution of nodes

In this paper we start by looking at the power management schemes used in ad-hoc networks in section II. We then present our cross layer design in section III. Simulation setup in section IV is followed by results and analysis in the Vth section. We conclude this paper in the VIth section followed by the references in the VIIth section.

2. Previous Work

The research community has done a lot of work to suggest potential ways of power control in ad-hoc networks. In [1] Rong Zheng and Robin Kravets proposed an on demand power management scheme for ad-hoc networks. It is an extensible on-demand power management framework for ad-hoc networks that adapts to traffic load. Nodes maintain soft-state timers that determine power management transitions. By monitoring routing control messages and data transmission, these timers are set and refreshed on-demand. Nodes that are not involved in data delivery may go to sleep as supported by the MAC protocol. This soft state is aggregated across multiple flows and its maintenance requires no additional out-of-band messages. The implementation is a prototype of the framework in the NS-2 simulator that uses the IEEE 802.11 MAC protocol. Simulation studies using this scheme with the Dynamic Source Routing protocol show a reduction in energy consumption near 50% when compared to a network without power management under both long-lived CBR traffic and on-off traffic loads, with comparable throughput and latency. Preliminary results also show that it outperforms existing routing backbone election approaches. In [2] Zongpeng Li and Baochun Li present a probabilistic power management scheme for ad-hoc networks. It introduces Odds, an integrated set of energy-efficient and fully distributed algorithms for power management in wireless ad-hoc networks. Odds is built on the observation that explicit and periodic re-computation of the backbone topology is costly with respect to its additional bandwidth overhead, especially when nodes are densely populated or highly mobile. Building on a fully probabilistic approach, Odds seek to make a minimum overhead, perfectly balanced, and fully localized decision on each node with respect to when and how long it needs to enter standby mode to conserve energy. Such a decision does not rely on periodic message broadcasts in the local neighborhood, so that Odds are scalable as node density increases. Detailed mathematical analysis, discussions and simulation results have shown that Odds are indeed able to achieve our objectives while operating in a wide range of density and traffic loads. Power control schemes using purely directional transmission and reception have not been researched in detail. Exploiting the advantages of directional antennas [13, 14, 15, 20] shows considerable improvement in the overall network performance. MAC layer solutions have been proposed from a long time. In [3] a scheme for controlling the transmission power is presented. More specifically to the effects of using different transmit powers on the average power consumption and end-to-end network throughput in a wireless ad-hoc environment. This power management approach reduces the


22

system power consumption and thereby prolongs the battery life of mobile nodes. Furthermore, the invention improves the end-to-end network throughput as compared to other ad-hoc networks in which all mobile nodes use the same transmit power. The improvement is due to the achievement of a tradeoff between minimizing interference ranges, reduction in the average number of hops to reach a destination, reducing the probability of having isolated clusters, and reducing the average number of transmissions including retransmissions due to collisions. The present invention provides a network with better end-to-end throughput performance and lowers the transmit power. [6] and [8] also talk about transmission power control. Topology control [4], [9] schemes have also been proposed. In [4] authors present Span, a power saving technique for multi-hop ad-hoc wireless networks that reduces energy consumption without significantly diminishing the capacity or connectivity of the network. Span builds on the observation that when a region of a shared-channel wireless network has a sufficient density of nodes, only a small number of them need be on at any time to forward traffic for active connections. Span is a distributed, randomized algorithm where nodes make local decisions on whether to sleep, or to join a forwarding backbone as a coordinator. Each node bases its decision on an estimate of how many of its neighbors will benefit from it being awake and the amount of energy available to it. Authors use a randomized algorithm where coordinators rotate with time, demonstrating how localized node decisions lead to a connected, capacity-preserving global topology. Improvement in system lifetime due to Span increases as the ratio of idle-to-sleep energy consumption increases, and increases as the density of the network increases. Simulations show that with a practical energy model, system lifetime of an 802.11 network in power saving mode with Span is a factor of two better than without. Span integrates nicely with 802.11. When run in conjunction with the 802.11 power saving mode, Span improves communication latency, capacity, and system lifetime. A cone-based solution that guarantees network connectivity was proposed in [9]. Each node i gradually increases its transmission power until it finds at least one neighbor in every cone of angle. Node i starts the algorithm by broadcasting a “Hello” message at low transmission power and collecting replies. It gradually increases the transmission power to discover more neighbors and continuously caches the direction in which replies are received. It then checks whether each cone of angle contains a node. The protocol assumes the availability of directional information (angle-of-arrival), which requires extra hardware. This scheme does not seem to work for directional antennas [16] as most of the replies would be lost depending on where the source’s antenna is pointing. Some researchers proposed the use of a synchronized global signaling channel to build a global network topology database, where each node communicates only with its nearest N neighbors (N is a design parameter). This approach, however, requires a signaling channel in which each node is assigned a dedicated slot. A scheme based on clustering [5] in ad-hoc networks was also presented. This paper discusses the problem of power

control when nodes are non-homogeneously dispersed in space. In such situations, one seeks to employ per packet power control depending on the source and destination of the packet. This gives rise to a joint problem which involves not only power control but also clustering. Three solutions for joint clustering and power control are presented. The first protocol, CLUSTERPOW, aims to increase the network capacity by increasing spatial reuse. The second, Tunnelled CLUSTERPOW, allows a finer optimization by using encapsulation. The last, MINPOW, whose basic idea is not new, provides an optimal routing solution with respect to the total power consumed in communication. The contribution includes a clean implementation of MINPOW at the network layer without any physical layer support. All three protocols ensure that packets ultimately reach their intended destinations. Power aware routing protocols such as Power Aware AODV [10] also contribute to conserving the network power. Changes have been suggested to the NS2 routing structure to accommodate for a power aware routing protocol that is aimed at increasing the network lifetime. [17, 18, 19] also propose energy efficient schemes for wireless networks. 3. Proposed Scheme 3.1 Power Control: The MAC Perspective

In this paper we propose a cross layer design in which the MAC layer and the Network layer functions can be modified to achieve lesser power consumption and increase throughput. The system mimics the 802.11 standard and uses directional antennas at the physical layer. Since different situations demand different control mechanisms, the solutions in this paper are aimed at the three most common topologies in wireless ad-hoc networks, namely sparse, dense and cluster forms. A power controlled MAC protocol decides the power level for transmission for a particular node. It is the power used to access the channel and also to carry on subsequent transmissions. Researchers have used different metrics to determine the optimum transmission power; the SINR of the receiving node being the most popular [add ref]. We assume that there is a mechanism like GPS that enables the nodes to be aware of the network topology. This enables the receivers to calculate the SINR level based on the number of transmitters and their distance from this receiver. Not only does this prolong network lifetime but also allows more nodes to communicate at the same time. In the proposed protocols the collision avoidance information does not prevent interfering nodes from accessing the channel. Instead, it bounds the transmission powers of future packets generated by these nodes. This is unlike the RTS/CTS packets used in the 802.11 scheme.

To understand what this collision avoidance information is and how nodes can make use of it, consider the transmission of a packet from some node i to some node j. Let SINR (i,j) be the SINR at node j for the desired signal from node i. Then,


23

(1) where, P (i,j)= received power at node j for a transmission from node i ∑P (k,j)= power received as noise at j from its neighbors hj= thermal noise (default value in NS2 is 101 dBm for the 802.11 std.) at node j. A packet is correctly received if the SINR is above a certain threshold (say, SINRth) that reflects the QoS of the link. By allowing nearby nodes to transmit concurrently, the interference power at receiver j increases, and so SINR (i,j) decreases. Therefore, to be able to correctly receive the intended packet at node j, the transmission power at node i must be computed while taking into account potential future transmissions in the neighborhood of receiver j. This is achieved by incorporating an interference margin in the computation of SINR (i,j). This margin represents the additional interference power that receiver j can tolerate while ensuring coherent reception of the upcoming packet from node i. Nodes at some interfering distance from j can now start new transmissions while the transmission i → j is taking place. The interference margin is incorporated by scaling up the transmission power at node i beyond what is minimally needed to overcome the current interference at node j. For our simulations, a SINR margin of 10 dB ensures correct reception. Also, it is to be kept in mind that the minimum transmission power should not be allowed to drop below a threshold given by

(2) where, Pt (i,j)= power transmitted by node i such that the transmission range does not exceed node j Pr (j,i)= power received by node j when node i transmits Pmax = maximum power for this configuration 1) Spare Networks: The Locally Aware Protocol This protocol is aimed at a solution to networks which are sparse. Since the network is not crowded, one receiver does not have multiple transmitters in its range. The communication between a pair of nodes starts when the transmitter senses the channel to be idle and sends a RTS to the intended receiver at maximum power. This ensures that the signal strength is high enough to maintain the SINR level at the receiver incase of interference. This also helps incase the intended receiver is farther away from the transmitter. In such a case, increased power would mean increased range. When the receiver hears the RTS successfully, it waits for SIFS amount of time and responds back with a CTS frame. The receiver calculates the ideal transmitting power that the transmitter should use to carry on further communication. This calculation is based on the noise level at the receiver and the required SINR level to process a signal correctly. This information is included in the CTS frame. On receiving the CTS frame, the transmitter adjusts its power to the new level as specified by the CTS

frame and continues further communication. Incase the receiver needs to update the transmitter of any changes in the SINR level, it can do so by including the new power level information in any CTS frame which it would send out in response to a data frame. Since this protocol tries to control the power levels of node pairs, we call it local power control. Also, since the nodes are sufficiently apart, the receiver does not need to inform all the nodes in the network of its new SINR level. 2) Dense Networks: A Globally Aware Protocol The protocol which we have simulated states that each node in the network transmits with optimum power (which is the minimum required SINR) in an attempt to maximize the throughput of the network. When a transmitting node sends an RTS to the receiver successfully, the receiver piggy backs the SINR information in the CTS frame which can be further used to decide the ideal transmitting power at the transmitter. It also broadcasts the Collision Avoidance Information in all the directions. When the neighboring nodes receive this information, they adjust their transmission power such that they can still carry on their conversation and not disrupt the other nodes. When a pair of communicating nodes experience interference from some other node (SINR goes down at the reciever), the Collision Avoidance information is updated and can be sent to the transmitter with a CTS (in response to a data frame). This allows the nodes which are at a fair distance from other nodes to transmit at maximum power and also restrict the power of nodes which are closer to each other. Now a node with a packet to transmit is allowed to proceed with its transmission if the transmission power will not disturb the ongoing receptions in the node’s neighborhood beyond the allowed interference margin. Allowing for concurrent transmissions increases the network throughput and decreases contention delay.

Figure 2. A receiving node transmitting its collision avoidance information in all the directions for the “Globally

Aware” protocol 3) Clusters: A Power Management Protocol


24

The cluster approach is defined for networks which tend to operate in small groups. If all the nodes were to transmit at a random level of transmission power, the network throughput would degrade since the neighboring nodes would cause interference. It would be appropriate in such a scenario to restrict the power levels of different groups. Each member in a group uses the same power level. Different groups use different power levels depending on the size of the group. Since new members may be added to the cluster and old members might leave, a special node called the Cluster Head (CH) is elected. This node has knowledge of the topology of the network. It calculates the minimum required power so that all node pairs can communicate. This value is beaconed after short durations so that all the nodes have the updated information and are transmitting with only the required power. Fig. 3 shows the reason for having each group use its own power level. Nodes can still communicate within the cluster and use much less power. Since reducing the power also means a reduction in range for the directional antennas, we see a better throughput result in our case. This is a result of better spatial reuse. Next we study the changes made to the Network layer.

Figure 3. Nodes inside a cluster can use lesser power to communicate

3.2 Power Control: The Network Perspective Network layer solutions are energy-oriented and aim at reducing energy consumption, with throughput being a secondary factor. It includes intelligent routing based on immediate link costs i.e. power aware routing [10]. Every node in the network can compute the power consumed to send a packet to its one hop neighbors. This allows nodes to choose the path which consumes least power, thereby preserving the system power. We implement the power aware AODV protocol and simulation results show that network throughput increased marginally and system lifetime was preserved. The advantage of power aware routing is that it can be integrated with the changed MAC layer, to enhance the system performance. The novelty here lies in the physical layer, with directional antennas.

The Ad-hoc On Demand Distance Vector (AODV) routing algorithm is a routing protocol designed for ad-hoc mobile networks. AODV is capable of both unicast and multicast routing. It is an on demand algorithm, meaning that it builds routes between nodes only as desired by source nodes. It maintains these routes as long as they are needed by the sources. Additionally, AODV forms trees which connect multicast group members. The trees are composed of the group members and the nodes needed to connect the

members. AODV uses sequence numbers to ensure the freshness of routes. It is loop-free, self-starting, and scales to large numbers of mobile nodes.

AODV builds routes using a route request / route reply query cycle. When a source node desires a route to a destination for which it does not already have a route, it broadcasts a route request (RREQ) packet across the network. Nodes receiving this packet update their information for the source node and set up backwards pointers to the source node in the route tables. In addition to the source node's IP address, current sequence number, and broadcast ID, the RREQ also contains the most recent sequence number for the destination of which the source node is aware. A node receiving the RREQ may send a route reply (RREP) if it is either the destination or if it has a route to the destination with corresponding sequence number greater than or equal to that contained in the RREQ. If this is the case, it unicasts a RREP back to the source. Otherwise, it rebroadcasts the RREQ. Nodes keep track of the RREQ's source IP address and broadcast ID. If they receive a RREQ which they have already processed, they discard the RREQ and do not forward it. As the RREP propagates back to the source, nodes set up forward pointers to the destination. Once the source node receives the RREP, it may begin to forward data packets to the destination. If the source later receives a RREP containing a greater sequence number or contains the same sequence number with a smaller hop count, it may update its routing information for that destination and begin using the better route. As long as the route remains active, it will continue to be maintained. A route is considered active as long as there are data packets periodically travelling from the source to the destination along that path. Once the source stops sending data packets, the links will time out and eventually be deleted from the intermediate node routing tables. If a link break occurs while the route is active, the node upstream of the break propagates a route error (RERR) message to the source node to inform it of the now unreachable destination(s). After receiving the RERR, if the source node still desires the route, it can reinitiate route discovery.

With power control, AODV manages to preserve the network life as it routes packets through nodes which have higher power remaining. This means when the AODV protocol calculates the next hop on receiving a RREQ packet, it considers the remaining power of all its neighboring nodes before sending a RREP packet. In such a case every node must be periodically updated of the power levels of the neighboring nodes. We do so by periodically broadcasting the remaining power information to immediate neighbors. If a node receives information that a neighboring node is running out of power, it must immediately calculate a new route to the destination. Although this protocol may face slightly larger delays (since longer routes may be chosen depending on the remaining power of neighbors), it manages to preserve network lifetime with the same throughput efficiency.


25

4. Simulation Setup Table 1 lists the parameters and their respective values that have been used in the simulations. We simulate the three topologies in NS2 on a Unix platform. The codes have been written in Tcl. In the next section we evaluate the performance of the cross layer design. We perform simulations to obtain the network throughput and the average consumption of power in the network using our design. The packet size is 512 bytes unless specified otherwise. Each flow in the network transmits CBR traffic. We do not consider mobility in our simulations. For the radio propagation model, the two ray path loss model is used. NS2 uses the 802.11 model for wireless networks. The power consumption has been calculated by considering the power used for transmission only, taking into consideration the number of directional antennas used. Perl language has been used to calculate values from the trace files. The power consumed for receiving and carrier sensing have not been calculated. For our simulations, we consider a mobile ad-hoc network in a square area of dimensions 500m X 500m.

Table 1: Parameters and their values used in simulations

TOPOLOGY PARAMETER VALUE

Sparse Network Number of nodes 8 Size of network 500 X 500 Simulation time 500 secs Packet size 512 bytes Bandwidth 1Mb Traffic Type CBR Pmax 100mW MinRecvPwr 1mW

Dense Network Number of nodes 30 Size of network 500 X 500 Simulation time 500 secs Packet size 512 bytes Bandwidth 2Mb Traffic Type CBR Pmax 100mW MinRecvPwr 1mW

Cluster form Number of nodes 16 Size of network 500 X 500 Simulation time 500 secs Packet size 512 bytes Bandwidth 2Mb Traffic Type CBR Pmax 100mW MinRecvPwr 1mW

5. Results and Analysis We start this section by testing our protocol and studying the behavior of the network in different situations. We obtained data by varying different parameters like the CBR rate and changing the number of nodes in a network. The CBR rate corresponds to how fast the packets are being generated and transmitted from the transmitter. Controlling this parameter means controlling the network load. Since a given network has only limited bandwidth, by increasing the CBR, the throughput should first increase as more and more packets are sent successfully over the network. As the bandwidth is constrained in a network, on increasing the CBR rate further, collisions occur, packets get dropped and throughput decreases. So the network behaves best for a particular CBR for a given scenario. This claim is validated by Fig. 4, 5 and 6. It is clear that the throughput in all the three types of networks increases as the CBR rate is increased, but starts to decrease and attains a steady state when the rate is increased further. When the available bandwidth of a network is fixed and the number of nodes is varied, the affect is very similar to changing the CBR. With a constant CBR, when nodes are added to a particular network, they increase the throughput of the network as more packets are sent over the medium. It is noticed that when too many nodes start talking in the same space, collisions occur and the throughput decreases. From Fig. 7, 8 and 9 we can see that as we increase the number of nodes in the three cases, the throughput increases to a maximum value. When a high number of nodes are present is any network, the value of throughput is much low. For Fig. 7, 8 and 9 the number of nodes in each network is specified in Table I. The CBR rate for the sparse network is 20 Kbps and for the dense and cluster network is 40 Kbps. These tests show that even after making changes to the MAC, Network and Physical Layer of the network, it has good behavior. It should be noted that in Fig. 7, when there are 18 nodes in the network, the throughput is very low. This is attributed to the fact that the cross-layer design for the sparse network is not effective when a large number of nodes are present. This is the point where we switch from the Locally Aware Scheme to the Globally Aware Scheme and we say our network is no longer sparse. Using the Globally Aware Scheme helps more nodes in the network be aware of the SINR values at the receivers and they can then adjust their transmission power accordingly. Different networks may have a different value for number of nodes in sparse and dense network.


26

Figure 4. Sparse Network: Throughput values for different

CBR rate.

Figure 5. Dense Network: Throughput values for different

CBR rate.

Figure 6. Cluster Network: Throughput values for different

CBR rate.

Figure 7. Sparse Network: Throughput values with different

number of nodes.

Figure 8. Dense Network: Throughput values with different

number of nodes.

Figure 9. Cluster Network: Throughput values with

different number of nodes.


27

Since we are aiming at conserving power and increasing the throughput, it is necessary to compare the performance of these 3 types of topologies when they are not empowered by the cross layer design. In traditional networks, omni directional antennas are used. We are going to make the following comparisons: 1) between a network equipped with omni directional antennas (802.11b) to a network using directional antennas with CLPM implementation, 2) between a network equipped with directional antennas (802.11b-like) to a network using directional antennas with CLPM implementation. The terms SN, DN and CN imply Sparse Network, Dense Network and Cluster Network respectively. In the cases where there is no power control, all the nodes transmit at the maximum allowable transmitting power. Simulation results for five different networks in each topology case (namely sparse, dense and cluster) have also been shown at the end (Fig. 14, 15 and 16) to give the reader an idea of the variation in results. Fig. 10 compares the throughput between the three topologies. It is to be kept in mind that the available bandwidth for the sparse network was 1Mb while it was 2Mb for the dense network and the cluster form. The cluster form network shows maximum improvement in throughput (13.33% more) which indicates that when all nodes in the cluster are allowed to transmit without power control, it leads to many collisions.

Figure 10. Throughput comparison between the three

topologies when using a omni directional antenna and a directional antenna with our power scheme implemented.

Fig. 11 shows that implementing the power control design increases the throughput by nearly 6% (in the cluster form case) when compared to a directional antenna without power control. This is attributed to the fact that reducing the transmission power allowed for more node pairs to communicate and thus more packets were transmitted successfully. The dense network and the cluster form topologies also show better performance once the cross layer design in used.

Figure 11. Throughput comparison between the three

topologies when using a regular directional antenna and a directional antenna with our power scheme implemented.

In Fig. 12 we have compared the power consumed between the three topologies between a network having omni directional antennas and a network having directional antennas with power control. It is clear from the results that in case of all the three topologies, the network with a power control scheme performs better than a network with omni directional antennas. The maximum allowable transmission power for a node in any case was 100mW and the MinRecvPwr was 1mW.

Figure 12. Comparison in power consumed in different

topologies when using a omni directional antenna to using a directional antenna with power control.

Fig. 13 shows the difference in power consumed in case of the three topologies when using a regular directional antenna to when a directional antenna with power control is used. Controlling the transmission power shows a lower value of consumed power as compared to a network when all the nodes are transmitting at the maximum power. The simulations show that reducing the transmission power of the nodes did not affect the throughput of the system negatively and also helped in conserving the overall network power. Throughout the simulation, the receiver keeps the transmitter aware of the ideal transmitting power by sending its SINR value to the transmitter. The network without


28

power control uses nodes that transmit with maximum power all the time.

Figure 13. Comparison in power consumed in different

topologies when using a regular directional antenna to using a directional antenna with power control.

This leads to wastage of energy and also hampers conversations going on in the nearby space. The overall network power is saved by almost 50% in the sparse topology and also by considerable amounts in the dense and the cluster form.

Figure 14. Simulation results for power consumption for five different network scenarios used in the Sparse topology case.

Figure 15. Simulation results for power consumption for

five different network scenarios used in the Dense topology case.

Figure 16. Simulation results for power consumption for five different network scenarios used in the Cluster form

topology case. 6. Conclusion

We proposed a CLPM (cross layer design for power management) and throughput enhancement for wireless ad-hoc networks. Our primary objective was to exploit the advantages of directional antennas and develop a power control scheme for a sparse, a dense and a clustered network. The underlying MAC layer is responsible for controlling the transmission power of a node depending upon the minimum required SINR level at the receiver. This information is sent by the receiver to only the intended transmitter incase of sparse networks. In case of dense networks, this collision avoidance information is transmitted in all the directions in lieu of more transmitters in a receivers range. This helped in not only conserving the power of nodes but also increased the throughput as more communications could take place in the same space. The Network layer performs a power aware routing protocol called Power AODV which aims at reducing the overall power consumed in routing packets. Performance comparison of this cross layer design with a 802.11b network and a 802.11b-like network using directional antennas, showed that it reduced the overall network power consumption and also increased the throughput. Thus, a network with a power management scheme implemented will have better performance than a network without such a scheme.

References [1] Rong Zheng, Robin Kravets, “On-demand Power

Management for Ad-hoc Networks”, INFOCOM 2003.

[2] Zongpeng Li and Baochun Li, “Probabilistic Power Management for Wireless Ad-hoc Networks”, Mobile Networks and Applications, October 2005.

[3] Krishnamurthy, Srikanth ElBatt, Tamer Connors, Dennis, “Power management for throughput enhancement in wireless ad-hoc networks”, In Proceedings of the IEEE ICC 2000.

[4] Benjie Chen, Kyle Jamieson, Hari Balakrishnan, and Robert Morris, “Span: An Energy-Efficient Coordination Algorithm for Topology Maintenance in Ad-hoc Wireless Networks”, Wireless Networks archive Vol. 8, Issue 5, September 2002.


29

[5] Vikas Kawadia and P. R. Kumar, “Power Control and Clustering in Ad-hoc Networks”, INFOCOM 2003.

[6] El-Osery, A.I. Baird, D. Bruder, S, “Transmission power management in ad-hoc networks: issues and advantages”, In Proceedings of the IEEE Networking, Sensing and Control, 2005.

[7] Marwan Krunz and Alaa Muqattash, “Transmission Power Control in Wireless Ad-hoc Networks: Challenges, Solutions, and Open Issues”, Network, IEEE Sept-Oct 2004.

[8] Matthias Rosenschon1, Markus Heurung1, Joachim Habermann, “New Implementations into Simulation Software NS-2 for Routing in Wireless Ad-Hoc Networks”, 2004.

[9] http://www.isi.edu/nsnam/ns/tutorial/ [10] Asis Nasipuri, Kai Li and Uma Reddy Sappidi, “Power

Consumption and Throughput in Mobile Ad-hoc Networks using Directional Antennas”, In Preceedings of the Eleventh International Conference, Computer Communications and Networks, 2002.

[11] Thanasis Korakis, Gentian Jakllari and Leandros Tassiulas, “A MAC protocol for full explotiation of Directional Antennas in Ad-hoc Wireless Networks”, MobiHoc, June 1-3, 2003, Annapolis, Maryland, USA.

[12] ANSI/IEEE Std 802.11, “Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) specifications”, 1999.

[13] Z. Huang, Z. Zhang, and B. Ryu, "Power control for directional antenna-based mobile ad-hoc networks", in Proc., Intl. Conf on Wireless Comm. and Mobile Computing, pp. 917-922, 2006.

[14] Ram Ramanathan, "On the performance of ad-hoc networks with beamforming antennas", Proc. Of MobiHoc '200 1, pages 95-105, October 2001.

[15] B. Alawieh, C. Assi, W. Ajib, "A Power Control Scheme for Directional MAC Protocols in MANET", IEEE Conf, WCNC2007, pp.258-263, 2007.

[16] Nader S. Fahmy, Terence D. Todd, and Vytas Kezys, "Ad-hoc Networks with Smart Antennas Using IEEE 802.11-Based Protocols", In Proceedings of IEEE International Conference on Communications (ICC), April 2002.

[17] P. Karn., “MACA -A New Channel Access Method for Packet Radio”, in Proc. 9th ARRL Computer Networking Conference, 1990.

[18] E.-S. Jung and N. Vaidya, “A power control MAC protocol for ad-hoc networks”, In 8th ACM Int. Conf on Mobile Çomputing and Networking (MobiCom02), pages 36-47, Atlanta, GA, September 2002.

[19] J.-P. Ebert, B. Stremmel, E. Wiederhold, and A. Wolisz, “An Energy-efficient Power Control Approach for WLANs”, Journal of Communications and Networks (JCN), 2(3): 197-206, September 2000.

[20] Mineo Takai, Jay Martin, Aifeng Ren, and Rajive Bagrodia, "Directional virtua1 carrier sensing for directiona1 antennas in mobile ad-hoc networks", MOBIHOC'02, June 2002.

Authors Profile

Mahasweta Sarkar joined the department of Electrical and Computer Engineering at San Diego

State University as an Assistant Professor in August 2006, after receiving her Ph.D. from UCSD in December 2005. She received her Bachelor’s degree in Computer Science (Suma Cum Laude) from San Diego State University, San Diego in June, 2000. Prior to joining SDSU, she worked as a Research Scientist at SPAWAR Systems Center,

Point Loma, San Diego. Her research interests include protocol designs, flow control, scheduling and power management issues in Wireless Local Area Networks. Ankur Gupta is currently a graduate research student under Dr.

Mahasweta Sarkar at San Diego State University, San Diego. He received his Bachelor’s degree in Electronics Engineering in June 2001 from Pune University, India. His research interest includes analysing and studying the effect of power management protocols in wireless devices that emply directional antennas.


30

A Bidding based Resource Allocation Scheme in WiMAX

Mahasweta Sarkar1 and Padmasini Chandramouli2

1Department of Electrical and Computer Engineering, San Diego State University,

San Diego, California, 92182, United States [email protected]

2Department of Electrical and Computer Engineering, San Diego State University,

San Diego, California, 92182, United States [email protected]

Abstract: WiMAX has traditionally focused on delivering Quality of Service (QoS) to users based on a pre-determined traffic classification model. It broadly classifies traffic into four categories In this paper, we argue that since applications today are diverse, clubbing them in merely four different groups does not do justice to their QoS requirements. We propose that instead QoS should be delivered on the basis of an application’s individual requirement. We propose resource allocation scheme for WiMAX networks. Our scheme abides by two basic principles: (A) We distinguish applications based on their resource demands. Unlike existing WiMAX schemes, our scheme allocates network resources to them as per their individual QoS requirement (B) We grant "extra" resources to applications that require it, for a price, by conducting bidding. We show, through extensive analysis and simulation results, that efficiency of our resource management scheme surpasses that of WiMAX's in addition to yielding higher revenues.

Keywords: WiMAX, Resource allocation, bidding strategy, revenue maximization.

1. Introduction Fueled by a technologically savvy new generation of mobile and stationery end users, the wireless marketplace is on the cusp of rapid change and limitless opportunities. WiMAX is a broadband wireless technology which promises to provide ubiquitous service to mobile data users. The current standard on WiMAX (802.16) classifies users into four following QoS service classes [1]: (a)Unsolicited Grant Services (UGS) to support Constant Bit Rate (CBR) traffic like voice applications (b)Real-Time Polling Services (RtPS) designed to support real-time traffic that generate variable size data packets on a periodic basis, such as MPEG video. (c) Non-Real-Time Polling Services (NRtPS) designed to support non-real-time and delay tolerant services that require variable size data grant burst types on a regular basis such as FTP traffic and (d)Best Effort (BE) Service designed to support data streams that do not require stringent QoS support like HTTP traffic (web browsing application).

However, we argue that since applications today are diverse, classifying them in merely four different groups does not do justice to their QoS requirements. Such broad classification eliminates the finer requirements of an application, thereby providing it with a non-customized

service which might lead to over or under serving the application. For example, WiMAX traffic classification (IEEE 802.16), would club both a video conferencing and a streaming video application in the “RtPS” traffic category and thereby extend identical network resource and service to both the applications. In reality, a video conferencing application is much more sensitive to jitter and delay than a streaming video application. Therefore a video conferencing application should require and be provided with higher network resources and prompt network services over a streaming video application. However, WiMAX chooses to treat both these applications in the same fashion and allocates equal resources to both thus under-serving the video-conferencing application and over-serving the streaming-video application. Another set of application currently classified under the UGS scheme suffers from similar disparity – namely a voice chat application and a streaming audio application. Traditional resource allocation schemes do not acknowledge the varying resource and QoS requirements of various applications within the same service class as enumerated by the examples above.

As multimedia traffic in the Internet starts dominating data traffic, the need to extend network resource and service to an application on a “per-need” basis will become not only necessary but also crucial. Without such distinction, Internet gaming applications will benefit the same service quality as an important video conferencing application thereby rendering the pseudo-customized network service model of IEEE 802.16 useless. This paper addresses this particular issue and proposes a resource allocation scheme which mitigates the problem at hand.

However, in doing so, we do not want to deviate from the existing WiMAX standard [1] as that would surely lead to backward compatibility problems. Also, in order to prevent any and every user from making a claim on “extra” resource, we grant extra resources to applications only for a price. This would deter undeserving applications from claiming valuable network resources. To that extent we introduce a resource auction principle [2] to formulate our resource allocation scheme. We assume that QoS-intense traffic will be willing to pay a higher price for their “enhanced” services – hence the concept of bidding for resources which not only prevents anarchy in the system but also leads to maximization of revenue earned by the ISPs.


31

Our resource allocation scheme splits the available resource, namely bandwidth into a “fixed” pool and a “bidding” pool. At the onset, resources from the fixed pool are allocated to users based on a standard WiMAX-like resource allocation algorithm. However if the users still have unsatisfied demands they bid for extra resource which might be allocated to them (from the bidding pool) if they happen to win the bid. An eligible user who bids the highest amount for a bandwidth pool, is the winner. We will present a detailed description of the bidding process and a user’s eligibility to bid in Section 3.

The rest of the paper is organized as follows: Section 2 describes the system model and the premise of our problem. Section 3 describes the resource allocation scheme. In Section 4 we present numerical results and a discussion of those results. We conclude the paper in Section 5.

2. System Model A WiMAX network comprises of a Base Station (BS) that communicates with several Subscriber Stations (SS) wirelessly. Each of the SSs in turn, communicates with several users. A BS is primarily responsible for allocating available resources to its SSs. A SS further allocates these resources, to users that it caters to, as depicted in Figure 1. We have quantified the UGS, RtPS, nRtPS and BE demands with respect to each SSs in Table1. We will be using the system in Figure1 as basis for explaining our resource allocation scheme. This allocation of resources by BS to SS, and further by a SS to its users is achieved via WiMAX’s scheduling scheme. Note that the WiMAX standard (IEEE 802.16 [1]) does not specify any standard scheduling scheme. It is left to the judgment and decision of individual vendors.

Figure 1. WiMAX architecture depicting one BS catering to five SSs each of which caters to a set of users. The resource demands of each of these users are also outlined in the figure

Table 1: Bandwidth requirements of the SSs in Figure 1

A distinctive feature of IEEE 802.16 is its Quality-of-Service (QoS) support [1]. As discussed in Section I, it has four service classes to support real time and non-real time communication, namely, UGS, RtPS, NRtPS and BE. We identify the fact that today’s applications are extremely diverse with varied QoS requirements. Applications that are categorized under the same service class as per the conventional IEEE 80.16 traffic classification rules [1] might in reality require different services as we have illustrated with the example of a video conferencing application and a streaming video application in Section I. This led us to a fine grained traffic classification especially within the UGS and RtPS classes. Note that, since NRtPS and BE traffic classes comprise of applications that does not have stringent QoS requirements, finer grained traffic classification in these classes would probably be unnecessary.

Our traffic classification scheme is based at the user level. We classify the users into three categories namely GOLD, SILVER and BRONZE. The GOLD users are the ones that have a mix of all four traffic classes namely UGS, RtPS, NRtPS and BE. The SILVER users are those who have a mix of three traffic classes – RtPS, NRtPS and BE. The BRONZE users are those who only have NRtPS and BE traffic at a particular point in time. We further classify GOLD and SILVER class users into priority classes 1, 2, 3 and 4. A SILVER user who is running an application with the most stringent QoS requirement will fall under Priority Class 1, a SILVER user with comparatively lesser stringent QoS requirement will opt for priority class 2, 3 or 4.For e.g. a live video conference for an official meeting would probably demand a Class 1 service whereas a friendly Internet chat with a friend might be clubbed as Class 2 or 3 traffic.

Our scheme assumes that users with applications which have stringent QoS requirements will also be willing to pay for the extra resource that they would demand from the network. A real life scenario of this concept could be that of an end user paying a higher fee to its ISP for increased bandwidth allocation. We assume that GOLD and SILVER class users with a high component of delay-sensitive UGS and RtPS traffic within their traffic mix, are the ones who participate in the “bidding” process to acquire extra resources from the network to support their “high maintenance” traffic. We also assume that these users have the willingness and the budget to bid/pay for the extra resource (bandwidth) that they would be demanding from the network. We define a “bid” as an amount of money that a particular user is willing to pay for the extra resource that it would demand from the network. We exploit a user’s willingness to pay for preferential treatment from the network as the basis of maximizing revenue that can be earned by the ISP for provisioning for “additional” network resources to the “high priority class” users.

Demand requirement

SS1 SS2 SS3 SS4 SS5

UGS+RTPS 65 4 12 15 40

NRTPS+BE 15 6 8 5 30


32

3. The Resource Allocation Scheme We first present an overview of our resource allocation scheme before delving into the technical details. Resource is allocated to all users such that no user is ever starved even if it does not have delay-sensitive traffic [3]-[6]. Total available resource (namely bandwidth) is split into two portions: (a) “fixed” pool and (b) “bidding” pool. First, resources from the fixed pool are allocated by the BS to the SSs based on a simple proportionality concept which we will describe shortly. However, after this initial resource allocation, if demand for resources is not met for the GOLD and SILVER class users, then they have the option of “bidding” for extra resources from the network. The eligible users send their bid to their respective SS who passes on the bid to the BS. The BS on the other hand, computes its own selling price for a chunk of resource that it has put up for auction. Once it accumulates the prices that various users are willing to pay for a particular resource, it decides upon which user to allocate the resource to. That particular user is then declared to be the winner of that “bid”. We now present a formal description of our resource allocation scheme. Table 2 enumerates the variables and notations that we use in explaining the scheme. In addition, we also provide a numerical example which we believe will further help the reader in understanding our scheme.

Table 2 : Notations, Variables and their Descriptions

Variable/Notation Description of Notations

m Number of SS in the network

FP(i), i=1,2,..m Fixed Pool chunks numbered from 1 through m for each SS

n Number of “Bidding Pool” chunks

BP(i),for i=1,2,..n

Bidding Pool chunks numbered from 1 through n

[n,m] Initial Cost Matrix contains cost incurred by BS for allocating each bidding pool portion to each SS irrespective of their eligibility to bid.

[n,m] The final cost matrix contains Cost incurred by BS for allocating each bidding pool BP(n) to eligible SSs(m) .

CC* Lowest total cost incurred by BS for allocating all its Bidding chunks for auctioning

CC*BP(i) Lowest total cost incurred by BS for allocating its Bidding chunks excluding BP(i) for auctioning.

Π BP(i) profit margin ($) associated with bidding pool portion ‘i’ {or BP(i)} ∀ i = 1,2,……n

ER BP(i) Expected Revenue estimated by BS by auctioning bandwidth pool ‘i’ {or BP(i}) ∀ i = 1,2,……n

Let us assume that the total available bandwidth (which is the resource of concern) is 100 MB out of which 60MB is ear-marked for the fixed pool and 40 MB for the bidding pool. Note that this division of the total resource pool is not fixed and can have an impact on the performance of our scheme. In general, we have observed that when traffic constitutes of mostly UGS and RtPS categories, it makes sense to have a 40% or higher resource allocation in the “bidding” pool. In scenarios where traffic comprises mainly of nRtPS and BE categories - assigning around 20% resource to the bidding pool shows good results. Quantifying these relationships is a future research project that we are going to undertake.

3.1 Resource Allocation from Fixed Pool The scheme allocates a minimum bandwidth to all class of users who are in need of bandwidth regardless of the traffic type they are carrying. It is realistic to assume that the minimum bandwidth limit would be a variable dictated by the demand of the application and the resources of the ISP. However, priority is always given to UGS and RtPS traffic classes over nRtPS and BE traffic classes.

The following four steps outline the resource allocation method from the fixed pool:

1. After allocating minimum bandwidth to all users, the BS broadcasts a message to all SSs requesting them to transmit their individual resource demands corresponding to each service type. SS accumulates demand information from its users and transmits them to the BS.

2. The BS first considers UGS and RtPS demand of each SS and splits the fixed pool bandwidth as follows:

(1)

Using the network in Figure 1 as an example, SS1 has 65 MB of UGS and RtPS bandwidth requirement. Total sum of UGS and RtPS demand of all SSs is 136 MB. Thus going by the above equation, bandwidth allocated to SS1 from the Fixed Pool is equal to FP(1)=(65*60)/136=31MB. Note that the bandwidth allocated to a SS never exceeds the demand of that SS. Similarly, bandwidth for UGS and RtPS traffic allocated to SS2 through SS5 in Figure1 would be FP(2)=2MB, FP(3)=6MB, FP(4)=3MB and FP(5)=19MB respectively. Figure 2 depicts the proportional splitting of Fixed pool resources ,among five SSs , with regards to their UGS and RtPS traffic.


33

3. However, if there are no users with UGS and RtPS traffic, the fixed pool will be split proportionally according to BE and nRtPS.

(2)

4. After allocating bandwidth to UGS and RtPS traffic for all SSs, if there are resources left then these resources are proportionally split and allocated to the SSs to meet their BE and nRtPS traffic requirements.

5. Once BS allocates bandwidth to the SSs, the SSs repeat the same proportionality principle (as explained in Step 2) to split the resources granted to it by the BS, amongst its users.

Let us again consider SS1 in Figure 1. Total (UGS+RtPS) demand of SS1 is 65MB. (UGS+RtPS) demand of user1 is 20MB. The bandwidth allocated to the SS by the BS for (UGS+RtPS) traffic is 31MB (as per our previous computation). Then bandwidth allocated by SS1 to user1 = 20*31/65=9.5MB. Similarly, the other users get the following chunk of bandwidth: user2=25*31/65=12; user3=15*31/65 =7MB; user4=5*31/65=2.5MB. In order to ensure fairness in terms of resource allocation, each and every user in the network will be allocated a minimum chunk of bandwidth (as long as it has a bandwidth requirement) regardless of the traffic type it is carrying. This is especially important to prevent resource starvation thereby leading to unsatisfied users in the system.

51%

3%

10%5%

31%

BW available from fixed poolBW allocated to SS1 BW allocated to SS2BW allocated to SS3 BW allocated to SS4BW allocated to SS5

Figure 2: A pie-chart depicting the bandwidth allocation to the five SS’s from the “Fixed Pool” based on their UGS and RtPS traffic demands.

3.2 Resource Allocation from Bidding Pool If resources from the fixed pool are not sufficient to satisfy user demands of GOLD and SILVER class users, then they resort to “bidding” for extra resources. This extra resource is allocated to the winner of the bid by the BS from the “bidding” pool of resources. Going by our example, unsatisfied demand of the SSs are: SS1=49MB, SS2=8MB, SS3=14MB, SS4= 17MB, and SS5= 50MB. Note that unsatisfied demand of an SS only comprises of resource demand of GOLD and SILVER user belonging to that particular SS. Unsatisfied traffic demand of BRONZE users are not considered since they do not bid for resources.

In order to implement some amount of fairness in the scheme, a SS is deterred from bidding repeatedly for extra resource (thus decreasing the chance for other SSs to win a bid) by introducing a “penalty” cost. A user/SS who has won a bidding pool portion in previous bidding game will incur a fixed penalty cost if they participate in a new bidding game again. This is mainly to dissuade same user/SS from winning pool portions frequently. There is also a “flat fee” to cover the overhead costs.

Assuming that the resource allocation scheme is run every T=3 seconds, and the SSs are required to pay a “flat fee” of $0.15 and a penalty fee of $0.01 for using 1 MB of resource from the bidding pool, we calculate the Initial Amount payable by each SS as below :

Initial AmountSS(i)= flat fee + penalty fee*(number of previous winnings of SS(i)) where i=1,2..#of SSs (3)

Table 3 lists the unsatisfied bandwidth requirements and the Initial Amount payable by each SS based on an assumed number of prior victories in winning the bid by each SS.

Table 3: Unsatisfied bandwidth and associated cost SS

Unsatisfied Demand

# of times SS won previously (assumption)

Amount to be paid for using 1MB of resource for T=3 sec incl. penalty

SS1 49MB 3 0.15+0.01*3= 0.18

SS2 8.9MB 0 0.15 SS3 14MB 1 0.15+0.01*1=

0.16 SS4 17MB 4 0.15+0.01*4=

0.19 SS5 50MB 2 0.15+0.01*2=

0.17

3.2.1 Resource Partitioning in Bidding Pool The BS first collects the information regarding unsatisfied resource demands from all its SSs. The BS then divides the bidding pool into portions, proportional to unsatisfied demand of each SS. For e.g. the bidding pool with available bandwidth of 40MB will be divided into 5 parts -P1 through P5 (for each of the SSs that reports an unsatisfied demand) as follows:

(4)

Thus with a bidding pool of 40MB and a total unsatisfied BW request of 138.9 MB, SS1 will be allocated a BP(1)=49*40/138.9=14.1MB. Similar calculations will yield BP(2) = 2.56MB, BP(3)=4.03 MB, BP(4) = 4.89 MB and BP(5) = 14.39 MB respectively for SS2 through SS5 in Figure 1 respectively - (Step 2).


34

3.2.2 Determining Initial Cost Matrix

The BS first determines the minimum price for 1MB of bandwidth for a time duration of 3 seconds (as outlined in Step 1). The BS then determines the initial cost matrix . This matrix represents the cost of allocating a chunk (portion) of the bidding pool to a SS. Recall that ‘n’ represents the number of chunks that the bidding pool resource has been split up into while ‘m’ represents the number of SS in the system. is calculated as follows:

where i= 1,2…m and j=1,2..n. (5)

For example; SS1 requires $0.18 to access 1MB of bandwidth for the duration of 3 seconds as calculated in Table 3. Thus the associated cost for SS1 to access BP(1) which comprises of a chunk of 14.1MB of bandwidth, would be equal to $0.18* 14.1 = $2.53. Similarly the minimum costs incurred by SS(1) through SS(5) for bidding for bandwidth chunks denoted as BP(1) through BP(5) are depicted in the Initial Cost matrix (Table 4). We will henceforth use for determining eligibility of an SS to participate in the bidding process.

Table 4: Initial Cost Matrix

3.2.3 Eligibility of SS to Participate in Bidding The BS first determines the eligibility of a SS to participate in the bidding process based on the following

factors:

(a) Budget Factor Budget Factor is the first criterion that determines eligibility of a SS to participate in the bidding process. A SS that cannot afford a specific (or any) pool portion(s) is not eligible to bid. Each SS has a certain budget for bidding. Let us suppose that SS1 through SS5 has a budget of $4, $2, $1, $1.5 and $3 respectively. Since a SS cannot exceed its budget during bidding, a SS can only bid for pool portions that cost less than or equal to its budget. Note that a SS with sufficient monetary resource can bid for multiple chunks of bandwidth.

Thus SS1 with a budget of $4 can bid for all the chunks from BP(1) through BP(5) whereas SS2 with a budget of $2

can bid only for BP(2), BP(3) and BP(4). Table5 enumerates the cost that a SS encounters to bid for a certain chunk of bandwidth (BP(1),…,BP(5)). A ‘*’ denotes the SS’s inability to bid for that particular chunk of bandwidth. Thus due to budget constraints, SS2 will not be able to bid for BP(1) and BP(5) while SS1 and SS5 will be able to bid for all bandwidth chunks from BP(1) through BP(5).

Table 5: SS eligibility to bid based on its budget

(b) BW Demand Factor In addition to the budgetary constraint, SSs are allowed to

bid only for bandwidth chunks that offer bandwidth less than or equal to the demand of the SS. This is enforced in order to prevent resource wastage. Total bandwidth (BW) provided by each chunk or bidding-pool portions (BP) for a duration of 3 seconds is calculated as BP(i)*3 thus leading to 42.3MB, 7.68 MB, 12.09MB, 14.67 MB and 43.17 MB chunks of bandwidth available in the 5 pool portions for 3 seconds each as depicted in Table 6.

Table 6: Total Bandwidth available in each pool for 3 seconds

Thus SS4 with an unsatisfied BW of 17 MB can bid only for bandwidth-pools BP(2), BP(3) and BP(4) while SS1 and SS5 can bid for all pool portions. Similarly SS2 with unsatisfied BW of 8.9MB can bid only for BP(2)while SS3 with an unsatisfied BW of 14MB can bid for BP(2) and BP(3) only. Taking into consideration the budget and the BW constraint, we compute the final matrix (Table 7). Once again, a ‘*’ in the table entry denotes the corresponding SS’s ineligibility to bid for the particular bandwidth pool (BP). Henceforth we will be using Final cost matrix for all future calculations.

SS1 SS2 SS3 SS4 SS5

BP1

2.52 * * * 2.4

BP2

0.45 0.375 0.4 0.475 0.425

BP3

0.72 0.6 0.64 0.76 0.68

BP4

0.9 0.75 0.8 0.95 0.85

P5 2.6 * * * 2.45

SS1 SS2 SS3 SS4 SS5

BP1

0.18*14= 2.52

0.15*14= 2.1

0.16*14= 2.24

0.19*14 =2.66

0.17* 14=2.4

BP2

0.18*2.56= 0.45

0.15*2.56= 0.375

0.16*2.56= 0.4

0.19*2.56= 0.475

0.17*2.56= 0.425

BP3

0.18*4=0.72

0.15*4= 0.6

0.16*4=0.64

0.19*4= 0.76

0.17*4= 0.68

BP4

0.18*5=0.9

0.15*5= 0.75

0.16*5=0.8

0.19*5= 0.95

0.17*5=0.85

BP5

0.18*14.4= 2.6

0.15*14.4= 2.16

0.16*14.4=2.3

0.19*14.4=2.7

0.17*14.4=2.45

BP1*3 BP2*3 BP3*3

BP4*3

BP5*3

42.3 MB

7.68 MB

12.09 MB

14.67 MB

43.2 MB


35

Table 7: Final Cost Matrix

SS1 SS2 SS3 SS4 SS5

BP1

2.52 * * * 2.4

BP2

0.45 0.375 0.4 0.475 0.425

BP3

0.72 * 0.64 0.76 0.68

BP4

0.9 * * 0.95 0.85

BP5

2.6 * * * 2.45

3.2.1 Determining Expected Revenue, Initial Bid Price and Profit Margin

Next step is to estimate minimum revenue i.e. expected revenue (Step 4c) that the BS expects make in the bidding game. For estimating minimum revenue we need to calculate initial bid price(Step 4a) and the profit margin (Step 4b).The BS calculates the initial bidding price for each bandwidth chunk (or BP) and also estimates the profit margin corresponding to each BP that it expects to gain at the end of the bidding. For example, using Table 7 as a reference, the BS will estimate the profit margin that it will acquire by allocating the bandwidth chunk BP(1) to both SS(1) and SS(5) (note that both these SSs are eligible to bid for resource BP(1)). The BP will of course allocate the chunk to the SS that will maximize its profit making that particular SS, the winner of the bid.

(a) Initial Bid Price

The initial bid price for each pool portion is equal to the minimum cost that an eligible SS has to pay for that particular pool portion. Consider the Final cost matrix from Table 7. The initial bid prices can be calculated as follows:

= min{ [row, column] } ∀column {1,…,m} and [row, column] ≠ ‘*’ . (6)

Thus, = min { [1,SS1], [1,SS5]} = $2.4

= $0.375 = $0.64

=min{ [4,SS1], [4,SS4],CM[4,SS5]} = $0.85

= min { [5,SS1], [5, SS5]} = $2.45

(b) Profit Margin Our profit margin calculation is inspired by the standard profit margin calculations in a Nash-Bidding Strategy as outlined in any standard Economics text book as in [7]. The BS calculates its profit margin for each chunk of bandwidth pool (or BP) as follows. (i) First, the minimum cost that a BS can earn by auctioning its entire bandwidth in the reserved pool is calculated. This cost is represented as (CC*). In order to calculate CC*, the BPs are arranged in decreasing order of magnitude. Thus referring to Table 6, we have BP(5) > BP(1) > BP(4) > BP(3) > BP(2). Each

bandwidth-pool (or BP) is now assigned a SS who is eligible to bid for that particular BP AND offers the minimum price for that BP.

Note that, an SS cannot be associated with two different BPs even if it is the lowest bidder on both of them. Again, this is enforced to ensure fairness. Thus referring to the Final Cost Matrix in Table 7, BP(5) is associated with the minimum bidder – SS5- amongst its eligible bidders SS1 and SS5. The second biggest bandwidth-pool is BP(1) which is associated with SS1. [Note that though SS5 had a lower bid on BP(1) than SS1, but since SS5 has already been associated with BP(5), it cannot be associated again with BP(1)]. Similarly, applying the above policy, BP(4), BP(3) and BP(2) are associated with SS4, SS3 and SS2 respectively and consecutively.

Thus, the minimum cost incurred by BS for allocating all its BP for auctioning =

CC*= SS5’s cost for BP(5) + SS1’s cost for BP(1)+ SS4’s cost for BP(4)+ SS3’s cost for BP(3)+ SS2’s cost for BP(2) (7) Plugging the above values from Table 7, we have; CC*= $2.45 + $2.52 + $0.95 +$ 0.64 + $0.375 = $6.885. (ii) The BS now iteratively calculates a CC* for each of the bandwidth-pool portions or BPs. In order to do so, it eliminates that particular bandwidth chunk or BP and computes yet another CC* value without those two elements. We denote these CC* calculations for each particular BP as CC*BP(i) ∀i {1,….,n} CC*BP(1) = SS5’s cost for BP(5) + SS1’s cost for BP(4)+ SS3’s cost for BP(3)+ SS2’s cost for BP(2) (8) Plugging the above values from Table 7, we have; CC*BP(1) = $2.45 + $0.90 +$ 0.64 + $0.375 = $4.315. (iii) Finally, the profit margin associated with each bandwidth chunk is calculated as:

Π BP(i) = CC* - CC*BP(i) ∀i {1,….,n} (9)

Thus, profit margin associated with BP(1) = Π BP(1) = CC* - CC*BP(1) = 6.885 – 4.315 = $2.57

(c) Expected Revenue The BS then calculates a expected revenue (ER) associated with each bandwidth-pool chunk that is expects to earn by auctioning that particular chunk of bandwidth. It calculates the expected revenue as follows:

ER BP(i) = Π BP(i) + ∀i {1,….,n} (10)

Thus, ER BP(1) = Π BP(1) + = $2.4 + $2.57= $4.97. Similarly expected revenue for the other bandwidth-

pools are determined. We use expected revenue with the objective to discourage a SS from excessively overbidding for a particular pool portion. During the bidding game, once a SS reaches the expected revenue associated with the bandwidth-pool it is bidding for, the BS allocates that bandwidth-pool to that SS. However if there are more than one SSs who reach the expected revenue for a pool portion, then the highest bidder amongst them gets the bandwidth. Thus we prevent the scenario of “survival of the wealthiest” .


36

3.2.2 The Auction The BS broadcasts the initial bidding prices (calculated in Step 4a) for each BP to all the eligible SSs with unsatisfied bandwidth demands. Each SS adds a random amount to the initial bid price of each pool portion that it is eligible to bid for. While doing this, a SS can never exceed its budget. The SS then makes a price offer to the BS. If the price meets the BS’s calculated ER value for that specific bandwidth-pool, the SS is granted access to that bandwidth-pool and that particular bandwidth-pool is taken off the auction. Note that though a SS can bid for different pool portions it is allowed to win only one bandwidth-pool per game.

3.2.3 Resource Distribution by SS Once a SS wins a bandwidth-pool that it bid for, it distributes it amongst the users based on the offers made by individual users. Gold class users with delay sensitive and resource intensive applications (RtPS or UGS) have the highest budget hence bid higher than other users and hence has higher probability of winning the biggest chunk of the resource. It is actually at this level that budget based on priority class plays an important role. Also note that users and hence SS can bid for multiple resource pool but can win only one. This is done to prevent a particular user from hogging the entire network resource and ensure fairness amongst users with various priority classes.

3.3 Nash Equilibrium is reached in our Resource Allocation Scheme - Discussion

The above resource allocation scheme has been designed such that it attains Nash Equilibrium. A rigorous proof of this statement can be found in [8]. In this section we provide a qualitative discussion. Nash Equilibrium (NE) is a concept in Game theory that has been widely used in various areas of engineering for attaining optimal solutions to a variety of problems. NE is a profile of strategies in which each player’s strategy is an optimal response with respect to all the other players’ strategies. Therefore, when a system attains NE, a player’s payoff does not increase if this player unilaterally deviates from the NE strategy. In this paper, we ensure that the bidding “game” reaches NE by enforcing the following rules and regulations that have to be followed by users, SSs and the BS participating in this game:

1. Users with stringent delay and jitter requirements are allocated higher budget, increasing their chances to win because while bidding a user cannot exceed its budget. 2. Users are allowed to bid for only those bidding pools that meet their demand. They are not allowed to bid for more resources than they require which avoids unnecessary resource wastage.

3. BS calculates the Expected Revenue (ER) at the beginning of the bidding game to prevent users and SSs from overbidding or underbidding.

4. Incase the BS’s ER price could not be met by any SS, the BS still has to allocate the bandwidth-pool to the highest bidder.

5. The BS penalizes SS and SS further penalizes users for participating in bidding game again when they have won resources from the previous games. This is to ensure fair allocation of resources to all SS and hence users. The penalty can be reset after a certain period of time.

Qualitatively, we attempt to prove to the reader that our resource allocation scheme attains NE by showing that:

1. The bidding format leads to resource allocation such that it generates the maximum revenue that could be earned from a bandwidth-pool.

2. No user can benefit by changing its bid offer unilaterally

3.3.1 Necessary Condition for NE From Game Theory literature, we know that our resource allocation scheme can claim to attain Nash Equilibrium if the scheme can guarantee that it will generate the maximum revenue that can be possibly generated by allocating a certain BP to a certain SS [8],[9]. No other allocation could have maximized the revenue earned by the BS. Let us assume that our resource allocation scheme allocates BP(i) to SS(a) ∀i {1,….,n} and ∀a {1,….,m}. For our scheme to have attained NE;

(11) for any combination of ‘x’ and ‘y’ ∀x {1,….,m} and ∀y {1,….,n}

Now consider the bandwidth-pools BP(1) and BP(5). SS2, SS3 and SS4 cannot bid for these pools because of budget and bandwidth limitations. Hence these allocations are not possible. From Table 7, we finally determine that SS2 is allocated BP(2), SS3 could have been allocated BP(2) and BP(3) but preferably was allocated BP(3) as BP(3) had a higher bandwidth chunk which catered to the high unsatisfied bandwidth requirement of SS3. SS4 could have been allocated BP(2), BP(3) or BP(4) but it was allocated BP(4) as that particular bandwidth chunk best suits its needs. Hence the allocation of SS2 to BP(2), SS3 to BP(3) and SS4 to BP(4) is the best possible allocation abiding by the Bidding Rules. This allocation will generate a revenue of $0.98+$1.00+ $1.50 = $3.48.

On the other hand, SS1 and SS5 has the resources to bid for all bandwidth-pools and hence would choose to bid for the highest chunks of bandwidth namely BP(5) and BP(1). Thus 2 possible final combinations could arise, namely ss1 to BP91) and SS5 to BP(5) OR SS(1) to BP(5) and SS(5) to BP(1). Let us examine the revenues ( and generated by each of these combinations along with the already decided upon combinations of SS2 to BP(2), SS3 to BP(3) and SS4 to BP(4): = (SS1-BP(1),(SS2-BP(2),SS3-BP(3),SS4-BP(4)),SS5-BP(5)) = $4.79 + $3.48+ $3= $11.48 and = (SS1-BP(5),(SS2-BP(2),SS3-BP(3),SS4-BP(4)),SS5-BP(1)) = $5.23 + $3.48+ $3 = $11.78 Thus, we find >

However SS1 has higher demand requirements and budget hence is the probable winning combination, as SS1 will outbid SS5 for BP(5).

Thus is the winning combination that generates the highest revenue possible than any other combination (we have argued that other combinations will be eliminated as per the bidding rules) and it is the allocation that our scheme proposes. We have thus qualitatively and


37

numerically proven that our resource allocation scheme attains Nash Equilibrium.

3.3.2 No user can gain by changing its strategy (bid offer) unilaterally

We have already numerically shown in the previous section that the winning combination generates the maximum revenue. The SSs bidding strategies are restricted by their bandwidth requirement and budget. Hence SS2, SS3 and SS4 cannot bid for pool portions BP(1) and BP(5); SS2 cannot bid for BP(4) due to its lower bandwidth requirement though it has higher budget than SS4; SS3 can only bid for BP(3) because its demand requirement is higher than other bandwidth-pools namely BP(2) and for the same reason SS4 was best allocated BP(4). Again, SS1 and SS5 would maximize the revenue only by opting for BP(1) and BP(5). Any other combinations of SS and BPs will not lead to a higher revenue or more apt bandwidth need fulfillment. Thus no SS will gain by changing their strategy of bidding for other bandwidth-pool than what our resource allocation scheme spells out. We have thus proved the second required criteria for our scheme to have attained NE.

4. Simulation Set up and Results Our resource allocation scheme aims at providing applications with the bandwidth they require without confining them to the pre-determined bandwidth allocation that WiMAX decides for them based on their service class Table 8 tabulates the simulation parameters.

Table 8: Simulation parameters Simulation Parameters Parameter Value Traffic Type Constant Bit Rate

traffic Total available bandwidth 180 MB Bandwidth available in fixed pool for fixed bandwidth algorithm

180 MB

Bandwidth available in fixed pool for our scheme

120 MB

Bandwidth available in bidding pool for our scheme

60 MB

Number of SS in the system 25 Max. number of users per SS 18

Penalty for participating in repeated bidding (by SS)

$0.01

Penalty for participating in repeated bidding (by user)

$0.006

Demand of Gold Class users

1.5 MB

Demand of Silver Class users

1.5 MB

Demand of Bronze ( NRtPS and BE )users

1.5 MB

In addition, our scheme also maximizes the revenue earned by ISPs by offering this customized bandwidth allocation at a monetary cost. We compare our “bidding” resource allocation scheme with WiMAX’s “fixed” resource

allocation scheme where the entire available bandwidth is chopped up into four chunks for the four different traffic classes. WiMAX charges a flat rate from all users of all classes.

Figures 3(a) and 3(b) demonstrates the other important aspect of our scheme which is revenue maximization. The X-axis represents systems with increasing number of GOLD and SILVER users. The Y-axis denotes the revenue earned from these users. We see that WiMAX’s fixed allocation and flat-fee-for-all-users scheme leads to higher revenue as the number of users increase in the system but the “bidding” scheme leads to a much higher (as much as ~50%) revenue earned as it forces users to pay proportionally to their resource demand.

Figure 3(a): Revenue by varying number of gold user

Figure 3(b): Revenue earned by varying silver users

Figures 4(a) and 4(b) compares bandwidth allocation to UGS and RtPS traffic class in our bidding scheme and the fixed scheme of WiMAX. Our bidding scheme further classifies UGS and RtPS traffic into four priority classes based on the applications bandwidth and delay requirement. The bidding scheme takes into account these varying requirements and proportionally allocates bandwidth to these four different classes within the UGS and RtPS class (denoted by the brown bars) whereas WiMAX’s fixed allocation scheme does not make this distinction and allocates bandwidth equally to all applications within the UGS and RtPS class (as denoted by the green bars).

Figure 5(a) shows a distribution of bandwidth by the BS in a system comprising of 80% GOLD and SILVER users and 20% BRONZE users. The available bandwidth is 180 MB and the demand on the bandwidth amongst all the users is 250 MB.


38

Figure 4(a): Comparison of Bandwidth allocated to Gold users via WiMAX’s “Fixed” and our proposed “Bidding” Resource Allocation Scheme (Green bars denote the WiMAX scheme while the brown bars denote our Bidding Scheme)

Figure 4(b): Comparison of Bandwidth allocated to Silver users via WiMAX’s “Fixed” and our proposed “Bidding” Resource Allocation Scheme (Green bars denote the WiMAX scheme while the brown bars denote our Bidding Scheme).

Figure 5(b) shows a distribution of bandwidth by the BS in a system comprising of 50% GOLD and SILVER users and 50% BRONZE users. The available bandwidth is 180 MB and the demand on the bandwidth amongst all the users is 180 MB. Figure 5(c) shows a distribution of bandwidth by the BS in a system comprising of 20% GOLD and SILVER users and 80% BRONZE users. The available bandwidth is 180 MB and the demand on the bandwidth amongst all the users is 240 MB.

We showcase that even in such stringent bandwidth demand requirement, our scheme delivers bandwidth allocation in the most optimum fashion. The blue bars in Figure 5 denote the total bandwidth demand of the different traffic classes (marked in the X-axis), the green bars denote the bandwidth allocation as seen by WiMAX’s fixed scheme whereas the brown bars denote the bandwidth allocation as per our bidding scheme. As evident from Figure 5, the bandwidth demand of UGS and RtPS traffic is met more efficiently in our scheme as compared to WiMAX’s fixed scheme. In addition, the nRtPS and BE traffic doesn’t starve either. This is because, based on the traffic mix, the BS intelligently divides the available bandwidth into ratio of Fixed Pool: Bidding Pool depending on percentage of Gold, Silver and Bronze users. For e.g. If the system has 80% Gold and Silver users as in Figure 5(a), the BS divides the available bandwidth into

20:80 ratio of Fixed Pool: Bidding Pool. So that Gold and Silver users get as much resources as possible to satisfy their demands. However if 80% of users are Bronze then BS divides the available bandwidth into 80:20 ratio of Fixed Pool: Bidding Pool. Thus more resources are available for the Bronze users to meet their needs even after catering to the GOLD and SILVER users efficiently.

UGS RTPS NRTPS BE Figure 5(a) : Bandwidth allocated by WiMAX’s Fixed and our Bidding scheme to a system comprising of 80% Gold and Silver users.

UGS RTPS NRTPS BE Figure 5(b) : Bandwidth allocated by WiMAX’s Fixed and our Bidding scheme to a system comprising 50% Bronze users

UGS RTPS NRTPS BE Figure 5(c) : Bandwidth allocated by WiMAX’s Fixed and our Bidding scheme to a system comprising of 80% Bronze users.

We have experimented with several traffic mix and our scheme always emerges the winner in terms of resource allocation when compared to WiMAX’s fixed scheme.


39

Figure 6(a) and 6(b) provide enough evidence to support this argument. The figure shows that our scheme (indicated by Brown Bar) allocates higher resources to UGS and RtPS requirements as number of silver and Gold users increase in the system and hence performs better than WiMAX fixed scheme (indicated by green bar).

Figure 6(a): Resource allocated to UGS classes by WiMAX fixed scheme( Green Bar) and Bidding scheme ( Red Bar)

Figure 6(b): Resource allocated to RtPS classes by WiMAX fixed scheme( Green Bar) and Bidding scheme ( Red Bar)

We have also compared the performance of our scheme in satisfying BE - Figure7(b) and nRtPS - Figure7(a) requirements with increasing GOLD, Silver and Bronze users. As number of Gold and Silver users increase more resources are allocated to them and lesser to Bronze users. The simulation result in Figure 7 shows that existing WiMAX scheme performs slightly better than Bidding scheme when it comes to Bronze users. This is because Bidding scheme allocates more resources to Gold and Silver users than the WiMAX fixed scheme.

Figure 7(a): Resource allocated to NRtPS by WiMAX fixed scheme (Green Line) and Bidding scheme ( Red Line)

Figure 7(b): Resource allocated to BE by WiMAX fixed scheme (Green Line) and Bidding scheme ( Red Line)

5. Conclusion In this paper we have addressed the issue of broad traffic classification in WiMAX which we perceive inappropriate in today’s world of varied applications and their resource requirements. We suggest a finer grained traffic classification within the realms of WiMAX’s service class. We device a resource scheduling algorithm using a combination of WiMAX’s fixed resource allocation scheme and a Nash equilibrium based bidding scheme. We have framed rules and regulations to ensure that optimum resource allocation is achieved, without wasting any resources.

Revenue maximization is an off-spring of our resource allocation scheme by taking advantage of a user’s willingness to pay a higher price for extra resources. Simulation results prove that our “bidding” resource allocation scheme not only allocates resources optimally and more efficiently than WiMAX’s fixed resource allocation scheme but also maximizes the revenue earned by the ISPs for the same bandwidth allocation to users. Our future research entails mathematically formulating the bandwidth partitioning scheme into the fixed and the bidding pool based on the traffic mix to optimize the resource allocation.

References [1] ANSI/IEEE Std 802.16, “IEEE802.16 Standard:

Broadband Wireless Metropolitan Area Network,” 2000.

[2] Ha'ikel Yaiche and Ravi R. Mazumdar, “A Game Theoretic Framework for Bandwidth Allocation and Pricing in Broadband Networks,” in proceedings of IEEE/ACM Transactions on Networking, October 2000.

[3] Isto Kannisto, Timo Hamalainen, Jyrki Joutsensalo, Eero Wallenius and TeliaSonera,“Adaptive Scheduling Method for WiMAX Basestations” in Proceedings of European Wireless Conference, 2007.

[4] Jyrki Joutsensalo, Ari Viinikainen, Mika Wikström and Timo Hämäläinen , “Bandwidth allocation and pricing in multimode network,” in the Proceedings of IEEE Advanced Information Networking and Applications (AINA), 2006.


40

[5] Aymen Belghith, Loutfi Nuaymi and Patrick Maill´e, “Pricing of Real-Time Applications in WiMAX Systems,” in the Proceedings of IEEE Vehicular Technology Conference, Calgary, British Columbia, Canada, Fall 2008.

[6] Siamak Ayani and Jean Walrand, “Increasing Wireless Revenue with Service Differentiation,” in Proceedings of 3rd ACM Q2SWiNet, 2007.

[7] Richard Engelbrecht-Wiggans and Martin Shubik, Auctions, Bidding, and Contracting: Uses and Theory Studies in Game Theory and Mathematical Economics, New York University Press , 1983.

[8] Haili Song, Chen-Ching Liu and Jacques Lawarrée, “Nash Equilibrium Bidding Strategies in a Bilateral Electricity Market,” in the Proceedings of IEEE Transactions on Power Systems, February 2002.

[9] Ken Binmore, Game Theory: A Very Short Introduction, Oxford University Press, USA , 2007.

[10] Belghith, A. Nuaymi, L. ENST Bretagne, and Rennes, “Comparison of WiMAX Scheduling Algorithms and Proposals for the rtPS QoS Class”, in Proceedings of EW 2008-14th European Wireless Conference, June 2008.

Authors Profile Mahasweta Sarkar joined the department of Electrical and Computer Engineering at San Diego State University as an assistant Professor in August 2006, after receiving her Ph.D. from UCSD in December 2005. She received her Bachelor’s degree in Computer Science (Suma Cum Laude) from San Diego State University, San Diego in June, 2000. Prior to joining SDSU, she

worked as a Research Scientist at SPAWAR Systems Center, Point Loma, San Diego. Her research interests include protocol designs, flow control, scheduling and power management issues in Wireless Local Area Networks.

Padmasini Chandramouli is currently pursuing Master’s Degree in Electrical Engineering from San Diego State University. She received her Bachelor’s degree in Electrical Engineering from Nagarjuna University, India in April, 2003. Her areas of interest are scheduling in WiMAX, MAC sublayer design and VLSI

system design.


41

Relative Performance of Scheduling Heuristics in Heterogeneous Computing Environment

Ehsan Ullah Munir1, Shengfei Shi2, Zhaonian Zou2, Muhammad Wasif Nisar1, Kashif Ayyub1 and

Muhammad Waqas Anwar3

1Department of Computer Science, COMSATS Institute of Information Technology,

Quaid Avenue, Wah Cantt 47040, Pakistan [email protected]

2School of Computer Science and Technology, Harbin Institute of Technology,

Harbin, 150001, PR China

3Department of Computer Science, COMSATS Institute of Information Technology, Abbottabad 22060, Pakistan

Abstract: Heterogeneous computing (HC) environment consists of different resources connected with high-speed links to provide a variety of computational capabilities for computing-intensive applications having multifarious computational requirements. The problem of optimal assignment of tasks to machines in HC environment is proven to be NP-complete requiring use of heuristics to find the near optimal solution. In this work we conduct a performance study of task scheduling heuristics in HC environment. Overall we have implemented 16 heuristics, among them 7 are proposed in this paper. Based on experimental results we specify the circumstances under which one heuristic will outperform the others.

Keywords: Heterogeneous computing, Task scheduling, Performance evaluation, Task Partitioning heuristic

1. Introduction Heterogeneous computing (HC) environment consists of different resources connected with high-speed links to provide a variety of computational capabilities for computing-intensive applications having multifarious computational requirements [1]. In HC environment an application is decomposed into various tasks and each task is assigned to one of the machines, which is best suited for its execution to minimize the total execution time. Therefore, an efficient assignment scheme responsible for allocating the application tasks to the machines is needed; formally this problem is named task scheduling [7]. Developing such strategies is an important area of research and it has gained a lot of interest from researchers [3, 19, 20]. The problem of task scheduling has gained tremendous attention and has been extensively studied in other areas such as computational grids [8] and parallel programs [14].

The problem of an optimal assignment of tasks to machines is proven to be NP-complete requiring use of heuristics to find the near-optimal solution [4, 12]. Plethora of heuristics has been proposed for assignment of tasks to machines in HC environment [13, 15, 17, 18, 22]. Each heuristic has different underlying assumptions to produce near optimal solution however no work reports which heuristic should be used for a given set of tasks to be executed on different machines.

Provided with a set of tasks {t1, t2,…,tm}, a set of machines{m1, m2,…,mn} and expected time to compute of each task it on each machine jm , ETC(ti, mj)(1 ≤ i ≤ m,1 ≤ j ≤ n), in the current study we find out the task assignment strategy that gives the minimum makespan.

For task selection in heterogeneous environment different criteria can be used, e.g. minimum, maximum or average of expected execution time across all machines. In current work we propose a new heuristic based on task partitioning, which consider minimum (min), maximum (max), average (avg), median (med) and standard deviation (std) of expected execution time of task on different machines as selection criteria. We call each selection criterion a key. Each heuristic uses only one key. Scheduling process for the proposed heuristics works like this; all the tasks are sorted in decreasing order of their key, then these tasks are partitioned into k segments and after this scheduling is performed in each segment.

A large number of experiments were conducted on synthetic datasets; Coefficient of Variation (COV) based method was used for generating synthetic datasets, which provides greater control over spread of heterogeneity [2]. A comparison among existing heuristics is conducted and new heuristics are proposed. Extensive simulation results illustrate the circumstances when one heuristic would outperform other heuristics in terms of average makespan. To the best of our knowledge, there is no literature available that proposes for a given ETC which heuristic should be used; so this work is a first attempt towards this problem.

2. Related Work Many heuristics have been developed for task scheduling in heterogeneous computing environments. Min-min [9] gives the highest priority to the task for scheduling, which can be completed earliest. The idea behind Min-min heuristic is to finish each task as early as possible and hence, it schedules the tasks with the selection criterion of minimum earliest completion time. Max-min [9] heuristic is very similar to the Min-min, which gives the highest priority to the task


42

with the maximum earliest completion time for scheduling. The idea behind Max-min is that overlapping long running tasks with short-running ones. The Heaviest Task First (HTF) heuristic [23] computes each task’s minimum execution time on all machines and the task with the maximum execution time is selected. The selected task is the heaviest task among all tasks (note that the Max-Min algorithm selects the task with the latest minimum completion time, which may not be the heaviest one). Then this heaviest task is assigned to the machine on which this task has minimum completion time. Eight dynamic mapping heuristics are given and compared in [13], however the problem domain considered there involves priorities and multiple deadlines. In Segmented Min-min heuristic [22] the tasks are divided into four groups based on their minimum, maximum or average expected execution time, and then Min-min is applied on each group for scheduling. The Sufferage heuristic [17] is based on the idea that better mappings can be generated by assigning a task to machine that would suffer most in terms of expected completion time if that particular machine is not assigned to it. Sufferage assigns each task its priority according to its sufferage value. For each task, its sufferage value is equal to the difference between its best completion time and its second best completion time. The detailed procedure and its comparison with some widely used heuristics are presented in [17]. In [6] a new criterion to minimize completion time of non makespan machines is introduced. It is noted that although completion time of non makespan machine can be reduced but it can increase the overall system makespan as well. Dynamic mapping heuristics for class of independent tasks are studied in [17] and three new heuristics are proposed. Balanced Minimum Completion Time (BMCT), heuristic for scheduling independent tasks is given and compared in [18], which works in two steps: In first step it assign each task to machine which minimize the execution time, and then in second phase try to balance the load by swapping tasks in order to minimize completion time. In [5] the comparison of eleven heuristics is given and the Min-min heuristic is declared the best among all the other heuristics considered based on makespan criterion. Minimum standard deviation first heuristics is proposed in [16] where the task having the minimum standard deviation is scheduled first.

The work in this research is different from the related work that here different keys are used as a selection criterion suiting the ETC type. Moreover for any given ETC we provide the near optimal solution by using the heuristic best suited for a specific ETC type.

3. Problem Statement Let T = {t1, t2,…,tm} be a set of tasks, M = {m1, m2,…,mn} be a set of machines, and the expected time to compute (ETC) is a m × n matrix where the element ETCij represents the expected execution time of task ti on machine mj. For clarity, we denote ETCij by ETC(ti, mj) in the rest of the paper. Machine availability time, MAT(mj), is the earliest time machine mj can complete the execution of all the tasks

that have previously been assigned to it (based on the ETC entries for those tasks). The completion time (CT) of task ti on machine mj is equal to the execution time of ti on mj plus the machine availability time of mj i.e.

CT (ti, mj) = ETC(ti, mj) + MAT(mj) Makespan (MS) is equal to the maximum value of the completion time of all tasks i.e.

MS = max MAT(mj) for (1 ≤ j ≤ n) Provided with T, M and ETC our objective is to find the task assignment strategy that minimizes makespan.

4. Task Partitioning Heuristic In heterogeneous environment for task selection different criteria can be used, examples are minimum, maximum or average of expected execution time across all machines. In task partitioning heuristic we use minimum (min), maximum (max), average (avg), median (med) and standard deviation (std) of expected execution time of task on different machines as selection criteria; hereafter referred to as key. Given a set of tasks T = {t1, t2,…,tm}, a set of machines M = {m1, m2,…,mn}, expected time to compute (ETC) matrix then the working of proposed heuristic can be explained as follows: we compute the sorting key for each task (for each heuristic only one key will be used for sorting), then we sort the tasks in decreasing order of their sorting key. Next the tasks are partitioned into k disjoint equal sized groups. In last, tasks are scheduled in each group gx using the following procedure:

Procedure 1 a) For each task ti in a group gx find machine mj which completes the task at earliest.

b) If machine mj is available i.e. no task is assigned to machine then assign task to machine and remove it from list of tasks.

c) If there is already task tk assigned to machine i.e. machine mj is not available then compute the difference between the minimum earliest completion time (CT) and the second smallest earliest CT on all machines for ti and tk respectively.

1. If the difference value for ti is larger than that of tk then ti is assigned to machine mj.

2. If the difference value for ti is less than that of tk, then no changes to the assignment.

3. If the differences are equal, we compute the difference between the minimum earliest completion time and the third smallest earliest completion time for ti and tk respectively. And repeat 1-3. Every time if step 3 is selected, the difference between the minimum earliest completion time and the next earliest completion time (e.g. the fourth, the fifth…) for ti and tk are computed respectively. If all the differences are the same then the task is selected deterministically i.e. the oldest task is chosen.


43

Now the proposed Task partitioning algorithm can be summed up in the following steps:

Task Partitioning Heuristic 1. Compute the sorting key for each task:

Sub-policy1 (avg): Compute the average value of each row in ETC matrix

( ), / .i i jj

key ETC t m n= ∑

Sub-policy2 (min): Compute the minimum value of each row in ETC matrix

( )min , .i i jjkey ETC t m=

Sub-policy3 (max): Compute the maximum value of each row in ETC matrix

( )max , .i i jjkey ETC t m=

Sub-policy4 (med): Compute the median value of each row in ETC matrix

( ), .i i jjkey med ETC t m=

Sub-policy5 (std): Compute the standard deviation value of each row in ETC matrix

( ), .i i jjkey std ETC t m=

2. Sort the tasks in decreasing order of their sorting key (for each heuristic only one key will be used for sorting).

3. Partition the tasks evenly into k segments. 4. Apply the procedure 1 for scheduling each segment.

A scenario of ETC is given in Table 1 to describe the working of proposed heuristic. All machines are assumed to be idle for this case. Sorting key used for the algorithm is average (avg) i.e. tasks are sorted in decreasing order of their average value. Table 2 shows the task partitioning; tasks are partitioned into three segments which implies k = 3. Table 3 shows how the results are derived using procedure 1. Figure 1 gives the visual representation of task assignment for proposed heuristic.

Table 1: Scenario ETC matrix

Task no m1 m2 m3 m4 t1 17 19 31 17 t2 2 4 2 5 t3 18 11 12 7 t4 3 4 6 13 t5 4 2 2 3 t6 10 9 11 7 t7 13 26 28 10 t8 9 6 4 4 t9 10 13 8 5 t10 5 4 7 9 t11 7 9 6 13 t12 14 6 12 8 t13 14 8 12 20 t14 16 9 16 15 t15 18 11 5 7

Table 2: Task partitioning

Task no m1 m2 m3 m4 Avg t1 17 19 31 17 21.00 t7 13 26 28 10 19.25 t14 16 9 16 15 14.00 t13 14 8 12 20 13.50 t3 18 11 12 7 12.00 t15 18 11 5 7 10.25 t12 14 6 12 8 10.00 t6 10 9 11 7 9.25 t9 10 13 8 5 9.00 t11 7 9 6 13 8.75 t4 3 4 6 13 6.50 t10 5 4 7 9 6.25 t8 9 6 4 4 5.75 t2 2 4 2 5 3.25 t5 4 2 2 3 2.75

Table 3: Execution process of Procedure 1 on each group

execution process on group 1 1st pass min. CT difference t1→ m1 17 0 t14→ m2 9 6 t3→ m4 7 4 2nd pass min. CT difference t7→ m4 17 11 t13→ m3 12 5

execution process on group 2 1st pass min. CT difference t15→ m3 17 3 t12→ m2 15 9 2nd pass min. CT difference t6→ m2 24 0 t9→ m4 22 3 t11→ m3 23 1

execution process on group 3 1st pass min. CT Difference t4→ m1 20 8 2nd pass min. CT Difference t10→ m1 25 3 t8→ m4 26 1 3rd pass min. CT Difference t2→ m3 25 2 4th pass min. CT Difference t5→ m2 26 1

Figure 1. Visual representation of task assignment in task

partitioning heuristic

4.1 Heuristics Notation In task partitioning heuristic tasks are sorted based on average, minimum, maximum, median and standard


44

deviation, and each heuristic is named as TPAvg, TPMin, TPMax, TPMed and TPStd. The algorithms Segmented min-min (med) and Segmented min-min (std) are also implemented for the evaluation purpose. The naming conventions and source information for all existing and proposed heuristics are detailed in Table 4.

Table 4: Summary of compared heuristics

No Name Reference

No Name Reference

H1 TPAvg New H9 Smm-avg [22] H2 TPMin New H10 Smm-min [22] H3 TPMax New H11 Smm-max [22] H4 TPMed New H12 Smm-med New H5 TPStd New H13 Smm-std New H6 Min-min [9] H14 MCT [17] H7 Max-min [9] H15 minSD [16] H8 Sufferage [17] H16 HTF [23]

5. Experimental Results and Analysis

5.1 Dataset In the experiments, COV based ETC generation method is used to simulate different HC environments by changing the parameters µtask, Vtask and Vmachine, which represent the mean task execution time, the task heterogeneity, and the machine heterogeneity, respectively. The COV based method provides greater control over the spread of the execution time values than the common range-based method used previously [5, 10, 11, 21].

The COV-based ETC generation method works as follows [2]: First, a task vector, q, of expected execution times with the desired task heterogeneity is generated following gamma distribution with mean µtask and standard deviation µtask × Vtask. The input parameter µtask is used to set the average of the values in q. The input parameter Vtask is the desired coefficient of variation of the values in q. The value of Vtask quantifies task heterogeneity, and is larger for high task heterogeneity. Each element of the task vector q is then used to produce one row of the ETC matrix following gamma distribution with mean q[i] and standard deviation q[i] × Vmachine such that the desired coefficient of variation of values in each row is Vmachine, another input parameter. The value of Vmachine quantifies machine heterogeneity, and is larger for high machine heterogeneity.

5.2 Comparative Performance Evaluation The performance of the heuristic algorithm is evaluated by the average makespan of 1000 results on 1000 ETCs generated by the same parameters. In all the experiments, the size of ETCs is 512×16, the value of k = 3, the mean of task execution time µtask is 1000, and the task COV Vtask is in [ ]0.1, 2 while the machine COV Vmachine is in [0.1, 1.1].

The motivation behind choosing such heterogeneous ranges is that in real situation there is more variability across execution times for different tasks on a given machine than the execution time for a single task across different machines.

The range bar for the average makespan of each heuristic shows a 95% confidence interval for the corresponding average makespan. This interval represents the likelihood that makespans of task assignment for that type of heuristic fall within the specified range. That is, if another ETC matrix (of the same type) is generated, and the specified heuristic generates a task assignment, then the makespan of the task assignment would be within the given interval with 95% certainty. In our experiments we have also considered two metrics in comparison of heuristics. Such metrics have also been considered by [18]

• The number of best solutions (denoted by NB) is the number of times a particular method was the only one that produced the shortest makespan.

• The number of best solutions equal with another method (denoted by NEB), which counts those cases where a particular method produced the shortest makespan but at least one other method also achieved the same makespan. NEB is the complement to NB.

The proposed heuristics are compared with 11 existing heuristics. Experiments are performed with different ranges of task and machine heterogeneity.

In the first experiment we have fixed the value of Vtask = 2 and then increase the value of Vmachine from 0.1 to 1.1 with increment of 0.2 in each step. The results of NB and NEB are shown in the Table 5 (best values shown in bold). From the values we can see that for high values of Vmachine H16 is the best heuristic. And in all other cases one of the proposed heuristic H2 or H5 outperforms all other heuristics. Figure 2 gives the comparison of average makespan of the all heuristics considered.


45

Table 5: NB and NEB values table when fix Vtask = 2

Cov of tasks H1 H2 H3 H4 H5 H6 H7 H8 H9 H10 H11 H12 H13 H14 H15 H16 0.1 NB 86 197 169 78 245 0 0 96 0 0 0 0 0 0 0 4 NEB 97 27 48 92 29 0 2 18 0 0 0 0 0 0 0 2 0.3 NB 101 252 112 132 90 0 0 213 0 1 0 0 0 0 0 0 NEB 62 54 48 62 52 0 1 49 0 0 0 0 0 0 0 4 0.5 NB 101 352 98 106 65 0 0 92 0 1 1 1 1 0 0 19 NEB 105 84 104 103 99 0 1 90 1 0 1 1 0 0 0 10 0.7 NB 82 350 62 89 47 0 0 45 1 2 4 1 2 0 0 146 NEB 100 59 98 96 99 0 2 89 0 0 2 1 1 0 0 32 0.9 NB 60 199 43 62 44 0 0 11 5 2 2 4 0 0 0 381 NEB 103 78 115 103 110 0 14 94 1 0 2 0 1 2 0 90 1.1 NB 17 69 22 21 16 0 0 9 0 1 0 3 1 0 0 575 NEB 167 156 160 163 160 0 47 156 1 0 3 1 2 5 0 202

(a)

(b)

(c)

(d)

(e)

(f)

Figure 2. Average makespan of the heuristics when Vtask = 2 and Vmachine = (a) Vmachine= 0.1; (b) Vmachine = 0.3; (c) Vmachine = 0.5; (d) Vmachine = 0.7; (e) Vmachine = 0.9; (f) Vmachine = 1.1

In the second experiment we have fixed the value of Vtask = 1.1 and then increase the value of Vmachine from 0.1 to 1.1 with increment of 0.2 in each step. The results of NB and NEB are shown in the Table 6. From the values it is clear

that here in all the cases one of the proposed heuristic H2 or H5 is best. Figure 3 gives the comparison of average makespan of all the heuristics considered.

Table 6: NB and NEB values table when fix Vtask = 1.1



46

(a)

(b)

(c)

(d)

(e)

(f)


In the third experiment we have fixed the value of Vtask = 0.6 and then increase the value of Vmachine from 0.1 to 1.1 with increment of 0.2 in each step. The results of NB and NEB are shown in the Table 7. From the values

it is clear that here in all the cases proposed heuristic H5 outperforms all other heuristics. Figure 4 gives the comparison of average makespan of all the heuristics.

Table 7. NB and NEB values when fix Vtask = 0.6



47

(a)

(b)

(c)

(d)

(e)

(f)


In the fourth experiment we have fixed the value of Vtask = 0.1 and then increase the value of Vmachine from 0.1 to 1.1 with increment of 0.2 in each step. The results of NB and NEB are shown in the Table 8. From the values

it is clear that here in all the cases proposed heuristic H5 outperforms all other heuristics. Figure 5 gives the comparison of the average makespan of all the heuristics.

Table 8: NB and NEB values when fix Vtask = 0.1



48

(a)

(b)

(c)

(d)

(e)

(f)


From all these experiments we conclude that in most of circumstances one of the proposed heuristics H2 or H5 outperforms the existing heuristics in terms of average makespan. In the remaining cases H16 performs better.

5.3 Algorithm to find best heuristic

Based on the values of Vtask and Vmachine we divide ETC into three different regions. If the values of Vtask and Vmachine are high (here Vtask = 2 and 0.9 <= Vmachine <= 1.1) then ETC falls in the region 1, if either of them is medium (here Vtask = 1.1 or 0.3 <= Vmachine < =0.7) then it falls in region 2 and if either of them is low (here 0.1 <= Vtask <= 0.6 or 0.1 <= Vmachine <= 0.2) then it falls in region 3. Fig. 6 shows the three regions and best heuristic for each region.

COV of

Tasks

COV of Machines 0.1 0.3 0.5 0.7 0.9 1.1 2 H5 H2 H2 H2 H16 H16 1.1 H5 H2 H2 H2 H2 H2 0.6 H5 H5 H5 H5 H5 H5 0.1 H5 H5 H5 H5 H5 H5

Figure 6. Division of ETC in different regions

The procedure for finding a best heuristic is given below in Algorithm Best Heuristic, which suggests the best heuristic depending on ETC type.

Best heuristic Input: expected time to compute matrix (ETC) Output: best heuristic

Compute the Vtask and Vmachine if Vtask is high and Vmachine is high then ETC belongs to region1 if Vtask is medium or Vmachine is medium then ETC belongs to region2 if Vtask is low or Vmachine is low then ETC belongs to region3 end if switch(region) case region1: return H16 case region2: return H2 case region3: return H5 end switch

6. Conclusions Optimal assignment of tasks to machines in a HC environment has been proven to be a NP-complete problem. It requires the use of efficient heuristics to find near optimal solutions. In this paper, we have proposed, analyzed and implemented seven new heuristics. A comparison of the proposed heuristics with the existing heuristics was also performed in order to identify the circumstances in which one heuristic outperforms the others. The experimental results demonstrate that in most of the circumstances one of the proposed heuristics H2 or H5 outperforms all the existing heuristics. Based on these experimental results, we are also able to suggest, given an ETC, which heuristic should be used to achieve the minimum.

Acknowledgements The authors are thankful to COMSATS Institute of Information Technology for providing the support for this

Region 1

Region 2 Region 3


49

research. This work is supported by the Key Program of The National Natural Science Foundation of China under Grant No.60533110. A preliminary version of portions of this document appear in the proceedings of the Modelling, Computation and Optimization in Information Systems and Management Sciences (MCO’08).

References [1] Ali, S., Braun, T.D., Siegel, H.J., Maciejewski, A.A.,

Beck, N., Boloni, L., Maheswaran, M., Reuther, A.I., Robertson, J.P., Theys, M.D., Yao B., (2005). Characterizing source allocation heuristics for heterogeneous computing systems. In: A.R. Hurson (Ed.), Advances in Computers, vol. 63: Parallel, Distributed, and Pervasive Computing, Elsevier, Amsterdam, The Netherlands, p.91-128.

[2] Ali, S., Siegel, H.J., Maheswaran, M., Ali, S., Hensgen, D., (2000). Task Execution Time Modeling for Heterogeneous Computing Systems, Proceedings of the 9th Heterogeneous Computing Workshop, p.85–200.

[3] Barbulescu, L., Whitley, L. D., Howe, A. E., (2004). Leap Before You Look: An Effective Strategy in an Oversubscribed Scheduling Problem. Proceedings of the 19th National Conference on Artificial Intelligence, p.143–148.

[4] Baca, D. F., (1989). Allocating modules to processors in a distributed system, IEEE Transactions on Software Engineering 15 (11) 1427–1436.

[5] Braun T.D., Siegel H.J., Beck, N., Bölöni, L.L., Maheswaran, M., Reuther, A.I., Robertson, J.P., Theys, M.D., Yao, B., Hensgen, D., Freund, R.F., 2001. A comparison of eleven static heuristics for mapping a class of independent tasks onto heterogeneous distributed computing systems. Journal of Parallel and Distributed Computing 61(6) 810 – 837.

[6] Briceno, L.D., Oltikar, M., Siegel, H.J., Maciejewski, A.A., (2007). Study of an Iterative Technique to Minimize Completion Times of Non-Makespan Machines, Proceedings of the 17th Heterogeneous Computing Workshop.

[7] El-Rewini, H., Lewis, T. G., Ali, H. H., (1994). Task Scheduling in Parallel and Distributed Systems, PTR Prentice Hall, New Jersey, USA.

[8] Foster, I., Kesselman,C., (1998). The Grid: Blueprint for a New Computing Infrastructure, Morgan Kaufman Publishers, San Francisco, CA, USA.

[9] Freund R.F., Gherrity, M., Ambrosius, S., Campbell, M., Halderman, M., Hensgen, D., Keith, E., Kidd, T., Kussow, M., Lima, J.D., Mirabile, M., Moore, L.,

Rust, B., Siegel, H.J., (1998). “Scheduling Resources in Multi-User, Heterogeneous, Computing Environments with Smartnet”, Proceedings of the 7th Heterogeneous Computing Workshop, p.184-199.

[10] Ritchie, G., Levine, J., 2004. A hybrid ant algorithm for scheduling independent jobs in heterogeneous computing environments, Proceedings of the 23rd Workshop of the UK Planning and Scheduling Special Interest Group.

[11] Ritchie, G., Levine, J., 2003 A fast, effective local search for scheduling independent jobs in heterogeneous computing environments, Proceedings of the 22nd Workshop of the UK Planning and Scheduling Special Interest Group, p.178-183.

[12] Ibarra, O.H., Kim, C.E., 1977. Heuristic algorithms for scheduling independent tasks on non-identical processors, Journal of ACM 24 (2):280–289.

[13] Kim, J.K., Shivle, S., Siegel, H.J., Maciejewski, A.A., Braun, T.D., Schneider, M., Tideman S., Chitta, R., Dilmaghani, R.B., Joshi, R., Kaul, A., Sharma, A., Sripada, S., Vangari, P., Yellampalli, S.S., 2007. Dynamically mapping tasks with priorities and multiple deadlines in a heterogeneous environment. Journal of Parallel and Distributed Computing, 67(2):154-169.

[14] Kwok Y. K., Ahmad, I., 1999. Static scheduling algorithms for allocating directed task graphs to multiprocessors, ACM Computing Surveys, 31(4):406–471.

[15] Kwok, Y.K., Maciejewski, A.A., Siegel, H.J., Ahmad, I., Ghafoor, A., 2006. A semi-static approach to mapping dynamic iterative tasks onto heterogeneous computing system, Journal of Parallel and Distributed Computing, 66(1):77-98.

[16] Luo, P., K. Lu, Z.Z. Shi, 2007. A revisit of fast greedy heuristics for mapping a class of independent tasks onto heterogeneous computing systems, Journal of Parallel and Distributed Computing, 67 (6): 695-714.

[17] Maheswaran, M., Ali, S., Siegel, H. J., Hensgen, D., Freund, R. F., 1999. Dynamic Matching and Scheduling of a Class of Independent Tasks onto Heterogeneous Computing Systems, Proceedings of the 8th IEEE Heterogeneous Computing Workshop, p.30–44.

[18] Sakellariou, R., Zhao, H., 2004. A Hybrid Heuristic for Dag Scheduling on Heterogeneous Systems, Proceedings of the 13th Heterogeneous Computing Workshop.

[19] Shestak, V., Chong, E. K. P., Maciejewski, A. A., Siegel, H. J., Bemohamed, L., Wang I., Daley, R., 2005. Resource Allocation for Periodic Applications in


50

a Shipboard Environment. Proceedings of the 14th Heterogeneous Computing Workshop.

[20] Shivle, S., Siegel, H.J., Maciejewski, A.A., Sugavanam, P., Banka, T., Castain, R., Chindam, K., Dussinger, S., Pichumani, P., Satyasekaran, P., Saylor, W., Sendek, D., Sousa, J., Sridharan, J.,Velazco, J., 2006. Static allocation of resources to communicating subtasks in a heterogeneous ad hoc grid environment, Journal of Parallel and Distributed Computing, 66 (4): 600–611.

[21] Shivle, S., Sugavanam, P., Siegel, H.J., Maciejewski, A.A., Banka, T., Chindam, K., Dussinger, S., Kutruff, A., Penumarthy, P., Pichumani, P., Satyasekaran, P., Sendek, D., Smith, J., Sousa, J., Sridharan, J., Velazco, J., 2005. Mapping subtasks with multiple versions on an adhoc grid, Parallel Computing. Special Issue on Heterogeneous Computing. 31(7):671– 690.

[22] Wu, M.Y,. Shu, W., Zhnag, H., 2000. Segmented min-min: A Static Mapping Algorithm for Meta-Tasks on Heterogeneous Computing Systems. Proceedings of the 9th Heterogeneous Computing Workshop, p.375–385.

[23] Yarmolenko, V., J. Duato, D.K. Panda P. Sadayappan, 2000. Characterization and Enhancement of Static Mapping Heuristics for Heterogeneous Systems. International Conference on Parallel Processing, pp: 437-444.

Authors Profile

Ehsan Ullah Munir received his Ph.D. degree in Computer Software & Theory from Harbin Institute of technology Harbin, China in 2008. He completed his Masters in Computer Science from Barani Institute of Information Technology, Pakistan in 2001. He is currently an assistant professor in the department of Computer Science

at COMSATS Institute of Information Technology, Pakistan. His research interests are task scheduling algorithms in heterogeneous parallel and distributed computing.

Shengfei Shi received the B.E. degree in computer science from Harbin Institute of Technology, China, in 1995, his MS and PhD degrees in computer software and theory from Harbin Institute of Technology, China, in 2000 and 2006 respectively. He is currently an associate professor in the School of Computer Science and Technology at Harbin Institute of

Technology, China. His research interests include wireless sensor networks, mobile computing and data mining.

Zhaonian Zou is a PhD student in the Department of Computer Science and Technology, Harbin Institute of Technology, China. He received his bachelor degree and master degree from Jilin University in 2002 and

2005 respectively. His research interests include data mining, query processing, and task scheduling.

Muhammad Wasif Nisar is a PhD candidate at the School of Computer Science and Technology, Institute of Software, GUCAS China. He received his BSc in 1998 and MSc Computer Science in 2000 from University of Peshawar, Pakistan. His research interest includes Software Estimation, Software Process

Improvement, Distributed systems, Databases, and CMMI-based Project Management.

Kashif Ayyub is a MS(CS) student at the Iqra University, Islamabad Campus, Pakistan. His research interest includes NLP and OCR for Urdu Language. He is currently Lecturer in Computer Science Department, COMSATS Institute of Information Technology, Wah Cantt., Pakistan.

Muhammad Waqas Anwar received his Ph.D. degree in Computer Science from Harbin Institute of technology Harbin, China in 2008. He is currently an assistant professor in the department of Computer Science at COMSATS Institute of Information Technology, Abbottabad, Pakistan. His areas of research area NLP and

Computational Intelligence.


51

Solution for Green Computing

Rajguru P.V.1, Nayak S.K.2 More D.S.3

1Department of Computer Science and IT, Adarsh college, Hingoli (Maharashtra),India

[email protected]

2Head, Department of Computer Science Bahirji Smarak Mahavidyalaya, Basmathnagar Dist-Hingoli (Maharashtra), India

[email protected]

3Head, Department of Environmental Science Bahirji Smarak Mahavidyalaya, Basmathnagar Dist-Hingoli (Maharashtra), India

[email protected]

Abstract: Environmental and energy conservation issues

have taken center stage in the global business arena in recent years. The reality of rising energy costs and their impact on international affairs coupled with the increased concern over the global warming climate crisis and other environmental issues have shifted the social and economic consciousness of the business community.

Green Computing, or Green IT, is the practice of implementing policies and procedures that improve the efficiency of computing resources in such a way as to reduce the environmental impact of their utilization. Green Computing is founded on the “triple bottom line” principle which defines an enterprise’s success based on its economic, environmental and social performance. This philosophy follows that given that there is a finite amount of available natural resources, it is in the interest of the business community as a whole to decrease their dependence on those limited resources to ensure long-term economic viability. Just as the logging industry long ago learned that they need to plant a tree for each that they cut, today’s power consumption enterprises must maximize the conservation of energy until renewable forms become more readily available. This is often referred to as “sustainability” – that is, the ability of the planet to maintain a consistent level of resources to ensure the continuance of the existing level of society and com-mercial enterprise.

In this paper we discuss green computing approach and give green computing solutions that will help for ecofriendly environment.

Keywords: Computer System, Flat Panel Display, Energy Conservation, Energy Star.

1. Introduction Over the last fifteen years, computers have transformed the academic and administrative landscape. There are now over 100 computers on campus of Adarsh College, Hingoli. Personal computers (PC) operation alone may directly account for nearly 1, 50,000 Rs. approximately per year in Adarsh College Hingoli. Computers generate heat and require additional cooling which adds to energy costs. Thus, the overall energy cost of Adarsh College about personal computers is more likely around 2,25,000 Rs. Approximately. A meeting computer cooling need in summer (and winter) often compromises the efficient use of

building cooling and heating systems by requiring colder fan discharge temperatures. In the summer, these temperatures may satisfy computer lab cooling needs while overcooling

Figure 1. Green computing environment

other spaces. Adarsh College given commitment to energy conservation and the environmental stewardship, we must address the issue of responsible computer use by adopting conserving practices, annual savings of 40,000 Rs. approximately are possible.

2. Green Computing Approach

Figure 2. Green Keyboard


52

2.1 Green use Reducing the energy consumption of computers and other information systems as well as using them in an environmentally sound manner.

2.2 Green disposal Refurbishing and reusing old computers and properly recycling unwanted computers and other electronic equipment.

2.3 Green design Designing energy-efficient and environmentally sound components, computers, servers, cooling equipment, and data centers.

2.4 Green manufacturing Manufacturing electronic components, computers, and other associated subsystems with minimal impact on the environment. These four paths span a number of focus areas and activities, including -

• Design for environmental sustainability. • Energy-efficient computing. • Power management • Server virtualization. • Responsible disposal and recycling. • Green metrics, assessment tools, and methodology. • Environment-related risk mitigation. • Use of renewable energy sources and • Eco-labeling of IT products.

Figure 3. Green computing environment cycle

3. Energy Consumption of PC A typical desktop PC system is comprised of the computer itself (the CPU or the “box”), a monitor, and printer. Your CPU may require approximately 100 watts of electrical power. Add 50-150 watts for a 15-17 inch monitor, proportionately more for larger monitors. The power requirements of conventional laser printers can be as much as 100 watts or more when printing though much less if idling in a “sleep mode.” Inkjet printers use as little as 12 watts while printing and 5 watts while idling. How a user operates the computer also factors into energy costs. First let’s take the worst case scenario, continuous operation. Assuming you operate 200 watt PC system day and night

everyday, direct annual electrical cost would be over 1500 Rs. In contrast, if you operate your system just during normal business hours, say 40 hours per week, the direct annual energy cost would be about 1000 Rs plus, of course, the cost of providing additional cooling. Considering the tremendous benefits of computer use, neither of the above cost figures may seem like much, but think of what happens when these costs are multiplied by the many thousands of computers in use at CU. The energy waste Cost adds up quickly. Here are some tested suggestions that may make it possible for you to reduce your computer energy consumption by 80 percent or more while still retaining most or all productivity and other benefits of your computer system, including network connectivity.

4. Solutions for green computing

4.1 Energy efficiency Maximizing the power utilization of computing systems by reducing system usage during non-peak time periods.

4.2 Reducing Electronic Waste Physical technology components (keyboards, monitors, CPUs, etc.) are often not biodegradable and highly toxic. Several business and governmental directives have been enacted to promote the recycling of electronic components and several hardware manufacturers have developed biodegradable parts.

4.3 Employing thin clients

These systems utilize only basic computing functionality (and are sometimes even diskless), utilizing remote systems to perform its primary processing activity. Since antiquated systems can be used to perform this function, electronic waste is reduced. Alternatively, new thin client devices are now available that are designed with low power consumption.

4.4 Telecommuting

Providing the facilities necessary to allow employees the ability to work from home in order to reduce transportation emissions.

4.5 Remote Administration

Allowing administrators the ability to remotely access, monitor and repair systems significantly decreases the need for physical travel to remote offices and customer sites. As with telecommuting, this reduced travel eliminates unnecessary carbon emissions.

4.6 Green Power Generation

Many businesses have chosen to implement clean, renewable energy sources, such as solar and wind, to partially or completely power their business.

4.7 Green Computing Practices

You can take a giant step toward environmentally


53

responsible or “green” computing by conserving energy with your computer. But green computing involves other import steps as well. These pertain to paper use, toner cartridges, disposal of old computer equipment and purchasing decisions when considering new computer equipment.

4.8 Reducing Paper Waste

Rather than creating a paperless office, computer use has vastly increased paper consumption and paper waste. Here are some suggestions for reducing waste:

• Print as little as possible. • Review and modify documents on the screen and

use print. • Preview the document. • Minimize the number of hard copies and paper

drafts you make. • Instead of printing save information to disks.

4.9 Recycle Waste Paper • Buy and use recycled paper in your printers and

copiers. From an environmental point of view, the best recycled paper is 100 percent post consumer recycled content.

• Save e-mail whenever possible and avoid needless printing of e-mail messages.

• Use e-mail instead of faxes or send faxes directly from your computer to eliminate the need for a hard copy. When you must fax using hard copies, save paper using a “sticky” fax address note and not a cover sheet.

• On larger documents, use smaller font sizes (consistent with readability) to save paper.

• If your printer prints a test page whenever it is turned on, disable this unnecessary feature.

• Before recycling paper, which has print on only one side, set it aside for use as scrap paper or in printing drafts.

• When documents are printed or copied, use double sided printing and copying. If possible, use the multiple pages per sheet option on printer proper ties.

• When general information-type documents must be shared within an office, try circulating them instead of making an individual copy for each person. This can also be done easily by e-mail.

4.10 Reusing and recycling

Adarsh College & institutions in Hingoli city generates many printer toner, ink jet cartridges and batteries a year. Instead of tossing these in the garbage, they can be recycled, saving resources and reducing pollution and solid waste. To recycle spent toner or ink jet cartridges (printer and some fax), deposit them at electronic dumping site. To recycle batteries, drop them off at any of the battery collection Shop. Computer diskettes may be inexpensive, but why keep buying more? Diskettes with outdated information on them can be reformatted and reused. When you are done with your diskettes, recycle them.

5 Recommendations Environmentally responsible computer use implies not buying new equipment unless there is a demonstrated need. Thus, before buying new equipment, consider the following questions:

• Do you really need a new computer or printer? • Can you meet your needs (with less expense and

environmental damage) by upgrading existing equipment?

• Can you find a solution in software rather than hardware?

• If you do need new equipment, buy efficient and buy green. Do research online and talk to the Buffalo Chip in the UMC about purchasing environmental and socially responsible equipment.

• Buy only “Energy Star” computers, monitors and printers. Flat panel monitors use about half of the electricity of a cathode-ray tube (CRT) display.

• Buy a monitor only as large as you really need. A 17 inch monitor uses 30 percent more energy than a 15 inch monitor when each is in an active mode.

• Buy ink jet printers, not laser printers. These use 80 to 90 percent less energy than laser printers and print quality can be excellent.

• Create network and share printer and other equipments.

• Consider leasing equipment as an alternative to purchasing. Leased equipment is typically refurbished or recycled, and packaging is reduced.

Of all these, “Energy Efficiency” provides the greatest potential for quick return on investment, ease of implementation, and financial justification. Several commercial solutions for improving computing energy efficiency have recently become available and we strongly recommend the adoption of such a solution not only for its environmental implications, but also for its common-sense reduction on IT infrastructure costs.

6. Conclusions The Research Paper contains useful techniques, guidelines described suitable for reducing energy consumption of any computer institute or a personnel user of computer. By following these simple guidelines lot of energy can be saved and thousands of rupees can be saved at village, taluka, district or any metro city ultimately saving the environment & conserving the energy.

References [1] “Energy Saving Guidelines for PCs”. www.colorado.edu /its/docs/energy.html “ [2] “The Computer for the 21st Century” Scientific

American, [3] www.adarshcollege.org [4] Adarsh College, Hingoli Accounts Department. [5] http://en.wikipedia.org/wiki/Green_computing


54

[6] "It's Easier to Say Green Than Be Green", Lave 96 Technology Review, Vol. 99 No.8, November/December 1996, pp. 68-69. [7 ] www.cosn.org/greencomputing. [8] “Rich Kaestner - CoSN Green Computing Project Director”. [9] http://www.ce.cmu.edu/GreenDesign [10] www.epa.gov/epaoswer/hazwaste/recycle/ecycling/ manage.html Authors Profile

Rajguru Prakash Vithoba received the M.C.M. from Dr. Babasaheb Ambedkar Marathwada University, Aurangabad and M.Phil. degree in Computer Science from Algappa University Karikudi in 2005 and 2008 respectively. During 2005-till today, he is working as Sr. Lecturer of Computer Science in Adarsh Education

Society’s, Hingoli.

S.K.Nayak received M.Sc. degree in Computer Science from S.R.T.M.U, Nanded. In 2000 he joined as lecturer in Computer Science at Bahirji Smarak Mahavidyalaya, Basmathnagar. From 2002 he is acting as a Head of Computer Science department. He is doing Ph.D. He attended many national and

international conferences, workshops and seminars. He is having 10 international publications. His areas of interest are ICT, Rural development, Bioinformatics.

More D.S. received the M.Sc. degree in Environmental Science from S.R.T.M.U, Nanded (Maharashtra) India in 2006. He is working as a Head, Department of Environmental Science in Bahirji Smarak Mahavidyalaya, Basmathnagar, Dist-Hingoli (Maharashtra), India.


55

BER Analysis of BPSK Signal in AWGN Channel

Dr. Mohammad Samir Modabbes

Aleppo University, Faculty of Electrical and Electronic Engineering, Department of Communications Engineering, Syria

[email protected]

Abstract: This paper presents extensive results of a experimental and simulation test of additive white Gaussian noise (AWGN) channel on binary phase-shift keying (BPSK) modulated signal. The bit error rate (BER) is measured and then simulation results are provided to verify the accuracy of error rate measurements. Results for all measurements and simulation are presented in curves. Results show that coherent detection of BPSK signal offers approximately 60-70% improvement in BER over the non-coherent detection.

Keywords: Wireless Communications, BPSK detection, noise,

BER analysis

1. Introduction Multipath propagation in wireless cellular systems leads to small-scale fading. Rayleigh distribution is a useful model for describing fading conditions for medium or large cellular systems. The Ricean distribution is another important fading model for describing small-scale fading. In [1] a precise BER expression for a BPSK signal corrupted by Ricean-faded co-channel interfering signals was obtained, their work focused on Ricean-faded co-channel interfering signals. In [2] The BER performance of BPSK orthogonal frequency division multiplexing (OFDM) system is analyzed in the cases of AWGN, but their work focused on OFDM system only. In [3] An exact expression for the bit error rate of a synchronous maximum-likelihood receiver used to detect a band limited BPSK signal in the presence of a similar BPSK signal, additive white Gaussian noise and imperfect carrier phase recovery is derived. It is shown that the receiver is more robust to carrier phase mismatch at high signal to-interference ratios (SIR) than low SIR.

2. Error performance for band-pass modulation

Band pass modulation is the process by which an information signal is converted to sinusoidal waveform [4, 5]:

)](sin[()()( 0 tttVtv ϕω += (1)

Where in digital systems the duration of carrier sinusoidal is called digital symbol. The probability of symbol error (Ps) is one of performance evaluation parameter of correct reception. The probability of incorrect detection of any receiver is termed the probability of symbol error Ps, but it’s often convenient to specify system performance by the probability of bit error Pb. A symbol error doesn’t mean that

all the bits within the symbol are in error in other words Pb is less than or equal to Ps. Thus ban-pass modulation can be defined as the process whereby the amplitude, frequency or phase of RF carrier or a combination of them is varied or will has M discrete values in accordance with the information to be transmitted. The efficiency of digital band-pass modulation technique depends on Eb/No parameter which can be expressed as the ratio of average signal power to average noise power (SNR) [1, 4]. 3. Probability of bit error for non-coherent

detection of band-pass modulation

When the receiver does not exploit phase reference of the carrier to detect the received signals, the process is called non-coherent detection. Thus a complexity reduction in non-coherent systems can be achieved at the expense of error performance. For non-coherent analysis we consider equally two band-pass modulated signals set (i=1, 2) at the input of the detector, Where:

)()()( tntstr ii += (2) r i (t) - the received signal s i (t) - the transmitted signal n (t) - the noise signal

The two band pass modulated signals can be defined by:

)cos(2)( 0 ϕω += tTEtsi Tt ≤≤0 (3)

E is the signal energy per bit; T is the bit duration. Considering the noise in the received signal n (t) is a white Gaussian noise we use the decision rule to choose the minimum error [5, 8]. This rule for the detector stated in terms of decision regions is whenever the received signal ri (t) is located in region1 , choose signal s1, when it is located in region2 choose signal s2 as in figure (1).

2/)( 210 aa +=γ represents the optimum threshold for minimizing the probability of making an incorrect decision given equally likely signals. For the case when s1is sent such that r (t) =s1 (t) + n (t), the output z2(T) is a Gaussian noise random variable only without signal component, so it has Rayleigh distribution. While z1 (T) has a Rician distribution since the input to envelope detector is a sinusoid plus noise.


56

We can compute Pb by integrating the area of either probability density functions p (z/ s1) or p (z/ s2):

Figure 1. the two conditional probability density

functions p (z / s1) and p (z / s2)

)4

exp(21

20

2

σApb −= (4)

Where TEA /2= , 20σ is the noise variance. We can

express the noise at the filter output as:

WN020 =σ (5)

W=R bits/s = 1/T is the filter bandwidth. Thus equation (5) becomes:

)2

exp(21

0NEp b

b −= (6)

Where 2/2TAEb = the signal energy per bit, N0 is is the noise power per hertz. Equation (6) gives the probability of bit error for non-coherent detection of FSK and is identical to the probability of bit error for non-coherent detection of OOK signals since envelope detector is used for detection [1, 7]. From equations (5), (6) it's clear that the error performance depends on the filter bandwidth and that Pb proportional to W.

4. Probability of bit error for non-coherent

detection of BPSK modulation

For standard PSK signal of the form:

)cos()( 11 ttAts ϕω += (7)

the digital modulation is carried in the angle of s(t) by φ1(t), which assumes discrete values from a set of M equally spaced points in [0, 2π] at the sample times T seconds apart. Thus the Nth message or baud is modulated by:

φ1 (NT) = 2πk/M; k=0, 1, 2… M-1 (8) Where each of M values of k is equally probable.

In BPSK there are no fixed decision regions in the signal space, but the decision is based on the phase difference between successively received signals. Thus the probability of bit error is equal to [5]:

)exp(21

0NE

p bb −= (9)

5. Probability of bit error for coherent

detection of BPSK For a coherent receiver, which compares the received wave with the un-modulated carrier, A cos (ω1t), and produces instantly the signed phase difference between the two points [6, 9], an M-ary symbol is transmitted in one baud by the value of k. Considering the external thermal noise in the received signal n(t) is a white Gaussian noise and is modeled in the usual Gaussian random process with uniform spectral density [10,11]. Hence,

n (t ) = n1 (t) cos ω1 t − n2 (t) sin ω2 t (10) Where n1 (t) and n2 (t) are stationary independent, zero mean Gaussian random processes with power 2σ

At a certain instant, the combined input signal at the detector is given by

s ′(t ) = s (t ) + n(t ) (11)

The detector examines the difference between the phase of the received signal and the reference phase and decides which symbol was transmitted. Assuming equal a priori symbol probabilities, for a proper decision we need to define decision thresholds by dividing the circle into regions, for example for the case of M = 8 at the instant of detection, if the phase of the received signal lies within the region, 0 ≤θ ≤ π/4 , we make the decision that the symbol, having been transmitted, corresponds to the value k = 1, but for BPSK (M=2), we make the decision that the symbol having been transmitted corresponds to the value k= 1, if the phase of the received signal lies within the region, 0 ≤θ ≤ π. 6. Measurement setup The tests were conducted using measurement system as shown in figure (2), which simulates the transmission medium over which digital communication takes place with frequency range from about 160 Hz to 70 kHz. The modulator produce BPSK signal with carrier frequency 2.4 kHz. A random pulse generator injects a variable amount of noise into the input signal and has frequencies in the range 75Hz to 600 Hz. Amplitude noise can be measured by determining the bit error rate (BER), which is the number of incorrect bits received in reference to the total number of bits transmitted. Error counter in error rate calculations block measures the number of incorrect bits received (error count), where The

0γ=0

region1

p(z/s1)

a1 a2

region2

p(z/s2)

Decision Line

Z(T)


57

transmitted and received data are compared bit by bit in the comparator, if the bit is not match an error pulse is generated, the errors are totalized in a counter over a fixed time period (106 ms is the time required for 128 data bits) generated by a one shot. Each time when the counter is reset a 106 ms one shot is triggered, and the error pulses from the XOR gate are totalized by the counter only during 106 ms frame. Then a display indicates how many errors occurred within the time interval.

Figure 2. BPSK measurement system A performance comparison of coherent and non-coherent BPSK detection in the presence of noise was made at:

v cutoff frequency of low pass filter 1.5 kHz

v total number of bits transmitted 128 bits

v SNR was calculated for 4Vp-p (2,828Vrms) input signal amplitude and variable amplitude of noise signal.

Figure (3) shows measurement results of BER versus SNR for the coherent and non-coherent detection of BPSK modulation at different noise amplitude.

0 5 10 15 20 250.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

SNR[dB]

BE

R

Coherent PSKNoncoherent PSK

Figure 3. Probability of bit error versus SNR for

measurement BPSK detection

7. Simulation results

Simulation on BPSK transmission channel with AWGN using SIMULINK was made. Figure (4) shows simulation

results of BER versus SNR for the coherent and non-coherent detection of BPSK modulation.

Figure (5) shows a comparison between simulation and measurement results of BER versus SNR for coherent and non-coherent detection of BPSK modulation.

0 5 10 15 20 250

0.05

0.1

0.15

0.2

0.25

0.3

0.35

SNR[dB]B

ER

Coherent PSKNoncoherent PSK

Figure 4. Probability of bit error versus SNR for

simulation BPSK detection

0 5 10 15 20 250

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

SNR[dB]

BE

R

Simulation coherentPSKSimulation noncoherent PSKMeasured coherentPSKMeasured noncoherentPSK

Figure 5. Probability of bit error versus SNR for measurement and simulation BPSK detection

8. Conclusions

Results show that coherent detection of BPSK offers approximately 60-70% improvement in BER over the non-coherent detection. Simulation results offer approximately 40-50% improvement over the measuring results. Measurement and simulation results led us to conclude that BPSK modulation has the best performance in coherent detection.

Comparator Error rate calculations

Coherent det.

Noncoherent Det.

Noise

BPSK Σ

NRZ

Amplifier

Switch


58

References [27] M.K. Simon & M.S. Alouini., 2000-"Digital

Communication over Fading Channels: A Unified Approach to Performance Analysis", John Wiley & Sons.

[28] Zheng Du, J. Cheng and C. Norman, 2007-"BER Analysis of BPSK Signals in Ricean-Faded Co-channel Interference" IEEE Transactions on communications, Vol. 55, No. 10.

[29] P. C. Weeraddana, N. Rajatheva and H. Minn., 2009 - " Probability of Error Analysis of BPSK OFDM Systems with Random Residual Frequency Offset" IEEE Transactions on communications, Vol. 57, No. 1.

[30] A. M. Rabiei, and N. C. Beaulieu., 2007-" Exact Error Probability of a Band-limited Single-Interferer Maximum Likelihood BPSK Receiver in AWGN" IEEE Transactions on Wireless communications, Vol. 6, No. 1.

[31] M. S. Modabbes, 2004- “Performance Analysis of Shift Keying Modulation in the Presence of Noise” - 1st International Conference on Information and Communication Technologies: From Theory to Applications, ICTTA’04, Proceedings, 19-23, pp:271 – 272.

[32] Soon H. Oh and Kwok H. Li, 2005-" BER Performance of BPSK Receivers Over Two-Wave with Diffuse Power Fading Channels", IEEE Transactions on Wireless communications, Vol. 4, No. 4.

[33] José María López-Villegas, Senior Member, IEEE, and Javier Jose Sieiro Cordoba, 2005 - " BPSK to ASK Signal Conversion Using Injection-Locked Oscillators-PartI: Theory", IEEE Transactions on Microwave Theory and Techniques, Vol. 53, No. 12.

[34] R. Kwan and C. Leung, 2002 - “Optimal Detection of a BPSK Signal Contaminated by Interference and Noise", IEEE COMMUNICATIONS LETTERS, Vol. 6, No. 6.

[35] H. Roelofs, R. Srinivasan and W. van Etten, 2003- “Performance estimation of M-ary PSK in co-channel Interference using fast simulation", IEE Proc.- communications, Vol. 150, No. 5.

[36] H. Meng, Y. L. Guan and S. Chen, 2005 - “Modeling and Analysis of Noise Effects on Broadband Power-Line Communications", IEEE Transactions on Power Delivery, Vol. 20, No. 2.

[37] A. M. Rabiei, and N. C. Beaulieu, 2007 “An Analytical Expression for the BER of an Individually Optimal Single Co-channel Interferer BPSK Receiver", IEEE Transactions on communications, Vol. 55, No. 1.

Author Profile Mohammad Samir Modabbes received the B.S. degree in Electronic Engineering from University of Aleppo in 1982. M.S.

and Ph.D. degrees in Communications Engineering from High Institute of Communications in Sankt Petersburg in 1988 and 1992, respectively. He is working as Associate Professor at the Faculty of Electrical and Electronic Engineering, Department of Communications Engineering, University of Aleppo Syria. Since 2006 he is working as Associate

Professor at Qassim University, College of Computer, Department of Computer Engineering in Saudi Arabia. His research interests are: Analysis and performance evaluation of Data transmission Systems, Wireless Communication Systems


59

The Use of Mobile Technology In Delivering E-Learning Contents: UNITEN experience

Mohd Hazli.M.Z, Mohd Zaliman Mohd Yusoff and Azlan Yusof

Universiti Tenaga Nasional, College of Information Technology, KM7 Jalan Kajang-Puchong, 43009 Kajang, Selangor, Malaysia.

{hazli, zaliman, azlany}@uniten.edu.my

Abstract: E-learning had become more affordable with cheaper but faster Internet connection. This has encouraged the development of mobile applications and integration of such applications into conventional e-learning system. This paper presents a working prototype of SMS-based mobile learning system which includes the delivery and dissemination of interactive learning contents to students using mobile phone at University Tenaga Nasional (UNITEN). Challenges and issues related to the deployment of SMS-based application are also discussed.

Keywords: e-Learning, mobile technology, short message service application

1. Introduction In today modern life, mobile communication activities are very popular among teenager and adolescence. This has resulted in the raise of many mobile applications including mobile learning [2]. In fact, mobile learning presents great potential as it offers unique learning experience which allows students to learn their lesson anywhere, anytime and very personalized. Moreover, it provides enriching, enlivening values as well as adding variety to conventional lessons or courses.

There is several benefits that mobile learning approach can offer. For example, mobile learning approach encourages both independent and collaborative learning experiences. Attewell [2] reported that many learners who taking part in mobile learning class enjoyed the opportunity of using mobile devices in their learning sessions. Besides, Mobile learning approach helps to combat resistance on the use of ICT and bridge the gap between mobile phone literacy and ICT literacy [9]. For example, it is reported in [6] that, post-participation, a number of learners of displaced young adults studying ESOL (English for Speakers of Other Languages) who had previously avoided using PCs actively continue working with their PCs on tasks such as writing letters. In fact, for some learners, their computer skills and confidence were enhanced to such an extent that they felt able to offer support and assistance to their peers. Besides that, several successful results of the benefit gained from mobile learning environment were also reported by researches in Sweden, Italy and UK [9].

While mobile learning is regarded as the new promising learning paradigm, adaptation of this learning approach into Malaysian community must be done carefully [10]. One of the potential drawbacks to the success of the implementation of mobile learning approach is the delivering technology [10]. Compared to develop country, Malaysia is still lacking

of the telecommunication infrastructure especially in the area of critical broadband services, blue tooth and WIFI technologies [10]. Thus, it is a challenge for Malaysian researchers to study the best suitable delivering method that suits Malaysian environment. Without the right setting, the potential of mobile learning technology cannot be utilized by Malaysian students. Suitable delivering method is important so that the progress of mobile learning could be accelerated by providing better learning opportunities for students. In the next section, results of a survey that was carried out to study the perspective amongst students with regard to mobile learning is presented. 2. The SMS based Mobile Learning System

(SMLS) A survey was conducted amongst students of Universiti Tenaga Nasional (UNITEN) to study their perspective on mobile devices usage in obtaining information related to their studies. This includes dissemination of learning contents such as information on lecture’s notes, quizzes, class information and other related student activities. There were 55 unpaid students of UNITEN involved in this study. They represent students from various degree levels (from foundation to final year students), different degree program (Engineering background as well as IT background students) including those with position in students’ club and society.

Results from the survey indicate that very significant percentage (83%) of respondents agrees that using SMS to obtain information related to their study and daily activities in university was more effective and convenient. Besides that, the coverage of SMS applications is wider compared to social networking or instant messenger as it has better penetration level in this country. In fact, the penetration level of cellular phone in Malaysian is more than 98%, compared to only 21% of broadband penetration [10]. Underpinning by the survey results, we develop a working prototype to incorporate mobile technology in delivering learning contents. The SMLS application was developed for both students and lecturers. For students, the SMLS application allows them to request various information related to their learning routines such as checking for class schedule, downloading assignment and obtaining their assessment marks.

As for lecturers, they can use the system to disseminate important announcements to student such as posting the postponement of class or giving the instruction for downloading notes or assignments. Moreover, the SMLS


60

application provides web-based interfaces for lecturers to update student information (e.g. students’ contact numbers and email addresses), class information and upload notes or assignments to SMLS server (see Figure 1).

Figure 1. Interface of SMLS

3. Design and Implementation

3.1 System Architecture Figure 2 below shows the architecture of the SMS technology used in SMLS prototype. The prototype is developed based on the current SMS technology. As to reduce the development time and concentrate more on the development of the learning contents, a third party SMS gateway service was deployed. The third party SMS gateway provides the application programming interface between the SMLS and the service provider’s SMS centre.

Internet

Internet

SMS Centre

SMS Centre SMS Gateway 32244

SMLS (SMS Application Server)

MobileService Provider

MobileService Provider

MobileDevice

MobileDevice

MobileDevice

Internet

Web Browser

Web Browser

Figure 2. SMLS system architecture Each SMS gateway has a unique number known as the shortcode. The shortcode is used by SMS Centre of mobile service provider to route the message to the respective SMS Gateway. In our context, the shortcode used is 32244. It

means, when a mobile phone user sent SMS to 32244, SMS centre of the respective mobile service provider will route the message to the specific SMS Gateway server. Based on the first keyword in the message, the Gateway will then make a HTTP GET request to SMLS. Upon receiving the request, SMLS server will verify the message and process the request as explain in the later part of this section. An example of SMS sent by a user is as below:

Figure 3. Sending SMS to SMLS

The first string which is “sms2u” is the keyword for SMLS application (see Figure 3). The Gateway will invoke the HTTP request using a GET method to SMLS server together with several other parameters for processing as below: http://smls.net/receive.php?from=0123456789&[email protected].&time=13+May+2010+17:03:52&msgid=10001&shortcode=32244 After processing the string, SMLS will return the result accordingly. Basically the result contains 4 parameters as describe in the Table 1 below:

Table 1: Return string parameters Sequence Parameter name Description

1 Type Message type 0 – normal text 1 – ringtone 2 – operator logo 4 – picture message

2 To The receiver’s mobile number. Must be in international format, For example: 60123456789

3 Text The content of the message, which must be URLencoded

4 Price 30 – RM0.30 An example of a return string is as below: 0,60123456789,CSEB364%20notes%20had%20been%20sent%20to%20adam%40yahoo.com.my.,30

To : 32244 SMS2U cseb364 notes [email protected]

Send Cancel


61

The above string contains message type code (0), the recipient phone number (0123456789), the result or information requested in url encoded and the premium charge (30) in cents. The Gateway then processes the string and forwards the message to the respective SMS Centre of the particular service provider. Figure 4 shows the example SMS sent by a user to 32244 and the reply from SMLS.

Figure 4. SMS replies from SMLS

3.2 Server Components of SMLS In general, SMLS comprises of two main components; the Information On Demand (IOD) module and the Bulk SMS module. The IOD module is used to handle information as per user request. This is a one-to-one service type. It means, if information is requested by a user, the information is returned or sent only to that particular user. In general, the server components of the SMLS are represented by Figure 5 below.

ApacheWeb

Server MySQLDatabase

MailServer

FTPServer

PHP Scripts

IODModules

BulkSMS

Modules

Applicationsscripts

Figure 5. SMLS server components

The Bulk SMS module is used when SMLS needs to send message to multiple users. For example, let assume that a lecturer needs to send announcement from his mobile phone. In this case, the lecturer can send SMS containing

special keywords together with the announcement to be made to SMLS as below:

Sms2u announce cseb324 class is postponed.

Upon receiving the request, SMLS will invoke the Bulk SMS module to send the announcement to phone numbers listed under CSEB324 class. The status of the operation will then send back to the lecturer’s cellular phone. Besides IOD and Bulk SMS module, there are also a database server (MySQL) and PHP modules, which control the flow and processes of user requests. They also act as the interfaces between the SMS gateway and other web services, like emails and FTP. So, upon receiving the HTTP GET request, the web server will execute and perform specific PHP scripts according to the keyword. In some cases, when necessary, the script will invoke the database server to establish a connection to the database or request the mail server to send email to an email address.

Figure 6 describes the flow of SMLS processes. For example, in the event of receiving a SMS, SMLS will first verify the message and then performs the required task. Every incoming SMS will be logged into an incoming SMS database table. Also, for every valid keyword, the required task will be performed (i.e. either make a query to database and get the result or invoke the sendmail function to send emails). Also, an acknowledgment will be sent through the SMS gateway, then to the SMS centre and eventually to the student’s mobile device for every valid SMS request.

S t a r t

S M S M e s s a g e R e c e iv e d

V a l id R e q u e s t?

P r o c e s s S M S s t r i n g

o b ta in d r e le v a n t

i n fo r m a t io n

R e t u r n e d e r r o r

m e s s a g e

F a ls e

T r u e

R e tu r n e dth e in fo r m a t io n

r e q u e s t e d

E n d

R e c o r d L o g

R e c o r d L o g

Figure 6. General flowchart of SMLS

3.3 System Functionality Table 2 below describes the tasks and application offered by SMLS system to students and lecturers. SMLS has been successfully implemented for student of CSEB324 Data

From : 32244 CSEB364 notes had been sent to [email protected].

Close Delete


62

Structure and Algorithms class during Semester 2 2009/2010 in UNITEN.

Table 2: SMLS functionalities Actor Task Description

Lecturer Announcement Send announcement to all students in a class. Announcement can be in the form of instruction to students or notification related to their class.

Student Request notes Student can request notes and the notes will then be emailed to their email address.

Check class information

Student can check class schedule, class announcement or any class related information.

Download quiz Student can request quiz paper to be emailed to their email.

Review question Student can request for review question for a chapter for refreshment before class.

Class synopsis. Student can request for class or chapter synopsis for refreshment before or after class.

Check marks Student can check their marks (quizzes, test or any assessment marks)

4. Issues and Challenges

Although SMLS works as planned, there are several important issues to be highlighted.

4.1 Reliability and performance issue The reliability of current implementation of the SMLS prototype architecture relies on several servers. However, these server are not owned by the university– they are belongs to the service provider (i.e. SMS Centre and the SMS gateway). As such, the performance of the SMLS prototype depends much on the performance and availability of both servers. During the trial, there were cases where the reply takes longer than 15 seconds and it was caused by the Gateway server. In some isolated cases, the reply even failed to reach the user. This is solely due to the network connection problem. Often, the problem was due to the connectivity failure between the third party SMS Gateway and the service provider SMS centre. Thus, to overcome the performance

and reliability issues, two solutions are proposed. First, university can choose a third party SMS gateway provider with better reliability and performance track record. The option is a simple but the recurring costs (i.e. maintenance, upgrading) of such gateway can be very expensive. A better option the problem is for university to setup its own SMS gateway server. Although the initial setup cost is expensive, the university can recoup its investment from the charged imposed by the service provider to the user.

4.2 Cost of SMS issue Since SMLS is using a third party SMS gateway, cost of sending IOD to user is charged at a premium rate with minimum of RM0.30 per message received. This amount is on top of standard SMS rate imposed by the respective service provider. Students, with financing constraint were a bit reluctant to use the system, unless they have no choice.

For an SMS massage, about 40% to 60% of the premium charges goes to the service provider. Whatever remains will be shared between the third party SMS gateway and the content provider (in this case the university). If the university owns the gateway server, almost half of the cost can be subsidized by the university. Furthermore, the mobile service provider could play their role in reducing premium charged for education purpose as part of their corporate social responsibility project.

4.3 Attestation and Privacy issue

Despite the ability of the SMLS prototype system to detect students’ mobile phone, the issue of user verification and authentication still posed a major challenge in MSLS environment. It is hard to develop a function that able to authenticate the user’s identity in the SMLS environment. Although some mobile phone such as PDA has the user authentication function using finger print to authenticate the authorized phone user, such mobile phone is seldom used by students. In most cases, the access to the mobile devices is open to everyone; thus, who has access to the phone can use the phone to send SMS and obtain privacy information such as student’s marks, or maybe, answer a quiz on behalf of other students.

In this pilot project, we just assume that the sender is the trustworthy student. However, in order for an SMLS system to be widely used and accepted, attestation framework has to be developed.

4.4 Copyright of the learning material

It is common that academic materials are usually copyrighted. In mobile learning environment where materials such as lecture notes and questions are disseminated through mobile phone without the copyright notice. Hence, the intellectual property of the material is not protected. As such, the students could easily reproduce the material and disseminate the material to other students. Therefore, serious consideration should be taken into account including looking for means to protect lecturers’ intellectual properties right for their learning materials.

4.5 Equal opportunity among students In mobile learning environment, the use of mobile


63

technology gadget is mandatory. In SMLS context, a student should have a mobile phone with SMS capability. Despite the mobile phone penetration in the country is high (almost 100%), there are still students who cannot afford to have it. Unlike personal computers, mobile phone is personal accessories and therefore not provided by the university. Thus, it creates unequal learning opportunity to those who cannot afford to buy the mobile phone. Serious actions should be taken to bridge the gap between students as the possession of a mobile phone is an important requirement if the SMLS were to be implemented in larger scale especially in the third world country.

4.6 University policy and acceptance We also look into the university policy on extra charges to be paid by the students. Since university is bound by rules and regulation of the Ministry of Higher Education, extra charges to the students have to be addressed carefully. This is to avoid any discrepancies with the current rules and regulation on charging the students. As a pilot project, SMLS is implemented only as an alternative and to complement the current web based technology. Students are not forced to use the system; instead, they opted to use the SMLS voluntarily as this brings new learning environment experience to students. 5. Conclusion With the intention to enrich student learning experience, SMLS had been developed. SMLS is an approach of implementing e-learning, using existing and inexpensive technology. Using SMLS, the learning process can be done in almost any place with a very minimal infrastructure – a hand phone. Further research and work need to be done to improve the system so that more complicated content such as multimedia content can be delivered. Besides that, more works should be done in the issues and challenges discuss in this paper in order to provide more conducive environment for students and lecturers to interact via SMLS. It is hoped that this project will not only generate new knowledge, but will also provide a new learning paradigm as it could improve students’ learning interest and eventually contribute to the improvement of their performance.

References [1] Albanese, M and Mitchell, S (1993).Problem-based

learning: a review of the literature on its outcomes and implementation issues. AcademicMedicine, 68: 52-81

[2] Attewell, J, 2005: Mobile technologies and learning A technology update and m-learning project summary: Learning and skills development agencies: http :// www.LSDA.org.uk, retrieved on 30 /10/2009

[3] Attewell, J and Savill-Smith, C (2003).M-learning and social inclusion - focusingon learners and learning. Proceedings of MLEARN 2003: Learning with Mobile Devices. London, UK: Learning and Skills Development Agency, 3-12

[4] Attewell J, Savill-Smith C (eds) (2004). Learning with mobile devices: research and development –a book

of papers. London: Learning and Skills Development Agency.At www.LSDA.org.uk/files/PDF/1440.pdf, accessed Nov 2009.

[5] Bacsich, P, Ash, C, Boniwell, K and Kaplan, L (1999). The Costs of Networked Learning. Sheffield, UK: Sheffield Hallam University. Available online at: www.shu.ac.uk/cnl/report1.html

[6] Corlett, D, Sharples, M, Chan, T and Bull, S (2004). A mobile learning organiser for university students. Proceedings of the 2nd International Workshop on Wireless and Mobile Technologies in Education. JungLi, Taiwan: IEEE Computer Society, 35-42

[7] Johnsen ET (2004).Telenor on the EDGE. 3G Americas at www.3gamericas.org/English/ PressReleases/PR200469682.html, accessed Nov 2009.

[8] Lonsdale, P, Baber, C, Sharples, M and Arvanitis, TN (2003). A context-awareness architecture for facilitating mobile learning. Proceedings of MLEARN 2003: Learningwith Mobile Devices. London, UK: Learning and Skills Development Agency, 79-85

[9] Lonsdale, P, Baber, C, Sharples, M,Byrne, W, Arvanitis, T, Brundell, P and Beale, H (2004). Context awareness for MOBIlearn: creating an engaging learning experience in an art museum. Proceedings of MLEARN 2004. Bracciano, Rome: LSDA

[10] Malaysian Communications and Multimedia Commission (SKMM), 2009 Brief Industry Trends Report: 2H 2008 http://www.skmm.gov.my, assesses April 2010

[11] NRPB (2004). Mobile Phones and Health 2004: Report by the Board of NRPB.Didcot, Oxfordshire: National Radiation Protection Board.

[12] Prensky M (2004). What Can You Learn from a Cell Phone? – Almost Anything.At www.marcprensky.com/writing/ Prensky-what_Can_You_Learn_From_a_Cell_Phone-FINAL.pdf, accessed November 2009.

Authors Profile Mohd Hazli M.Z obtained his BACHELOR OF Computer Science and M.Sc (Computer Science) from Universiti Teknologi Malaysia

in 1998 and 2001 respectively. Started his career as system engineer in an IT company in 1998 as system developer and joined the UNITEN in the year of 2001 as Lecturer in Software Engineering Department. He currently pursues his PhD in Software Testing. Mohd Z. M. Yusoff obtained his BSc and MSC in Computer Science from Universiti Kebangsaan Malaysia in 1996 and 1998 respectively. He started his career as a Lecturer at UNITEN in 1998 and has been appointed as a Principle Lecturer at UNITEN since 2008. His has produced and presented more than 60 papers for local and international conferences.

His research interest includes modeling and applying emotions in various domains including educational systems and software agents, modeling trust in computer forensic and integrating agent in knowledge discovery system.


64

Azlan Yusof obtained his BSc Computer and MSc (Computer Science) from Universiti Teknologi Malaysia in 2000 and 2002 respectively. Started his career in UNITEN in the year of 2003 as lecturer in Software Engineering department.


65

Automated Detection and Recognition of Face in a Crowded Scene

Chandrappa D N1, M Ravishankar2 and D R RameshBabu3

1 SJB Institue of Technology, Bangalore, India [email protected]

2 Professor ISE, Dayananda Sagar college of Engineering, Bangalore, India

[email protected]

3Professor CSE, Dayananda Sagar college of Engineering, Bangalore, India Abstract: Face detection and recognition play an important role in the process of biologic character identity. Biological character identification happens in two phases, first is the face detection and then recognition. Face detection is performed by feature extraction considering skin color model and performing skin segmentation and morphological operations in different environmental variations. It extends with extracting the skin color regions by skin likelihood and then the obtained face is cropped, scaled and stored in a specified location using the digital image processing techniques. For the purpose of face recognition, cropped face images are projected onto a feature space (face space) that best encodes the variation among known face images by using Eigenface method to identify the authenticated person.

Keywords: Face detection, Face recognition, Skin Segmentation, Skin-Color model

1. Introduction Automatic recognition of individuals based on their biometric characteristics has attracted researcher’s attention in recent years due to the advancement in image analysis methods and the emergence of significant commercial applications. There is an ever increasing demand for automated personal identification in a wide variety of applications ranging from low to high security such as: Human Computer Interaction (HCI), access control, surveillance, airport screening, smart card and security. Even though biometrics such as finger prints and iris scans deliver very reliable performance, the human face remains an attractive biometric because of its advantages over other forms of biometrics. The face detection algorithm, “Human Skin Color Clustering for Face Detection” [5] uses 2D color (chromatic) space (RGB) for detecting skin color pixels. Some color spaces have a property of separating the luminance component from chromatic component and with that at least partial independence of chromaticity and luminance is achieved. Such color spaces are for example YUV, RGB, HSI, etc. The original face detection algorithm [1] was developed to work best under standard daylight illumination. After the skin color classification is done for every pixel of the image, the skin region segmentation is performed. Unsuitable regions are then eliminated on the basis of geometric properties of

the face. Face Recognition using Eigen-faces [6] algorithm recognizes the person by comparing characteristics of the face to those of known individuals. Eigenface method [7] for effectiveness is evaluated by comparing false acceptance rates, false rejection rates and equal error rates, and verification operations on a large test set of facial images, which present typical difficulties when attempting recognition, such as strong variations in lighting conditions and changes in facial expression. The Proposed method introduces a Skin color model that helps to detect faces from different environmental variations and after this stage, model is proposed for Recognition the faces from crowded image. This paper is organized as follows: Section 2 discusses face detection using Skin color, Segmentation, Morphological operation, Region labeling, Euler’s test, Template matching and Recognition by using Eigen-faces. Section 3 discussing the Results of the proposed method and Section 4 concludes the paper.

2. Proposed Method In the proposed method, the goal is to detect the presence of faces in crowded image using Chrominance method and Recognition using Eigen vector algorithm. The system composed of two stages: Face Detection with feature extraction and Face Recognition.

2.1 Face Detection Face detection is achieved by detecting areas that indicate a human face. Face detection is based on using grayscale and color information and implementing an algorithm to detect faces independently of the background color of the scene.

2.1.1 Skin Color Model An RGB image when compared to other color models like YUV or HSI, has the disadvantage of not clearly separating the mere color (chroma) and the intensity of a pixel, which makes it very difficult to robustly distinguish skin colored regions, as it often has large intensity differences due to lightning and shadows. In the RGB space, the triple component (r,g,b) represents not only color but also luminance. Luminance may vary across a person's face due


66

to the ambient lighting and is not a reliable measure in separating skin from non-skin region.

In order to segment human skin regions from non-skin regions based on color there is a need for a reliable skin color model that is adaptable to people of different skin colors and to different lighting conditions. Luminance has to be removed from the color representation in the chromatic color space. Chromatic colors, also known as "pure" colors in the absence of luminance, are defined by a normalization process shown below:

r = R/(R+G+B) (1)

b = B/(R+G+B) (2)

Also color green is redundant after the normalization because r+g+b = 1.

Chromatic colors have been effectively used to segment color images in many applications. Although skin colors of different people appear to vary over a wide range as they differ much less in color than in brightness.

A total of 50 skin samples from 20 color images were used to determine the color distribution of human skin in chromatic color space. As the skin samples were extracted from color images, the skin samples were filtered using a low-pass filter to reduce the effect of noise in the samples. This skin color model can transform a color image into a grayscale image such that the gray value at each pixel shows the likelihood of the pixel belonging to the skin. A sample color image and its resulting skin-likelihood image are shown in Figure 1. All skin regions (like the face, the hands and the arms) were shown brighter than the non-skin region and Segmentation that region brighter than Skin Color Model.

2.1.2 Skin Segmentation Since the skin regions are brighter than the other parts of

the images, the skin regions can be segmented from the rest of the image through a thresholding process. Also, since people with different skins have different likelihood, an adaptive thresholding process is required to achieve the optimal threshold value for each case. The adaptive thresholding is based on the observation that stepping the threshold value down may intuitively increase the

segmented region. Using this technique of adaptive thresholding, many images yield good results; the skin-colored regions are effectively segmented from the non-skin colored regions. The skin segmented image of a color image resulting from this technique is shown in figure 2. The next stage in the process is Morphological operations that involves tools for extracting image components that are useful in the representation and description of region shape, such as boundaries and Skeletons using erosion and dilation.

Figure 2. Skin Segmented Image

2.1.3 Morphological Operations The basic morphological tools like filling, erosion and

dilation for investigation of binary image with ‘1’ representing skin pixels and ‘0’ representing non-skin pixels and morphological operations in order to separate the skin areas which are closely connected (when two faces are a bit overlap). The dilated binary image is multiplied with binary image from segmentation process to maintain the holes as shown in figure 3. The next stage to analyze the region to determine it is a face region or non-face regions and regions is labeling.

2.1.4 Region Labeling The binary image resulted after morphological operation

needs to be labeled such that each clustered group of pixels can be identified as a single region, in order for each region to be analyzed further to determine if it is a face region or not. That is instead of 1 and 0; each region is labeled as 1, 2, 3 and so on. Pixels with 0 values are unchanged. A function is used to show each different labeled region in different colors as shown in figure 4.

The next stage is to conduct Euler test for skin region and it should have at least one hole inside that region.

(a). The Original color Image

(b). The skin-likelihood image.

Figure 1. Sample color image and its resulting skin-likelihood image

(a). Eroded Image (b). Dilated Image

Figure 3. After Morphological operation


67

2.1.5 Euler Test

This stage is processed for each region at a time. The objective in this stage is to reject regions which have no holes as mentioned in the above region. There after analyze with several images, to decide that a skin region should have at least one hole inside that region. To determine the number of holes inside a region, we use the formula E = C – H (3) where E=Euler Number, C=Connected component, H=holes. The next stage is conduct to template matching to cross correlation between template face and grayscale region.

2.1.6 Template Matching This is the final stage of face detection where cross

correlation between template face and grayscale region is performed (obtained by multiplying the binary region with grayscale original image). The template face is an average frontal face of a person. This final stage involves the following processes: Firstly, width, height, orientation and centroid of binary region under consideration have to be computed. Then, the template face image is resized, rotated and its centroid placed on the centroid of the region in original grayscale image with only one region as shown in figure 6. The rotated template need to be cropped properly and the size need to be same with that of the region. Then cross correlation is calculated and threshold of 0.2 is set by experiment and the region which has less cross correlation value is rejected.

2.1.7 Final Detection Output

The rectangle box to mark the detected face as shown in figure. 7. Once we get the face regions and crop them, it becomes possible to then convert it to a gray image of size 92 x112 pixels and save each as individual images for further processing for face recognition.

2.2 Face Recognition Humans can recognize human faces with little difficulty

even in complex environments. Therefore, the human face is a natural and attractive biometric. Most approaches to automate human face recognition typically analyze face images automatically. Experimental results can thus be collected and approaches evaluated quickly by using eigenface method.

PCA approach [6] will treat every image of the training set as a vector in a very high dimensional space. The Eigen vectors of the covariance matrix of these vectors would incorporate the variation amongst the face images. Now each image in the training set would have its contribution to the Eigen vectors (variations). This can be displayed as an ‘eigenface’ representing its contribution in the variation between the images. In each eigenface some sort of facial variation can be seen which deviates from the original image. The face image to be recognized is projected on this face space. This yields the weights associated with the Eigen faces, which linearly approximate the face or can be used to reconstruct the face. These weights are compared with the weights of the known face images for it to be recognized as a known face used in the training set. The face image is then classified as one of the faces with minimum Euclidean distance.

Figure 4. Colored labeled region

Figure 5. After Euler’s test

Figure 6. Template Matching

(a) Face detected

(b) Cropped Image (c) Gray converted resized image

Figure 7. Final face detection


68

2.2.1 Calculating Eigen faces Let a face image Γ(x,y) be a two-dimensional N by N

array of intensity values. An image may also be considered as a vector of dimension 2N , so that a typical image of size 256 x 256 pixels becomes a vector of dimension 65,536, or equivalently, a point in 65,536-dimensional space. An ensemble of images, then, maps to a collection of points in this huge space.

Let the training set of face images be 1Γ , 2Γ , 3Γ , …, MΓ .

The average face of the set if defined by ∑=

Γ=ΨM

nnM 1

1.

Each face differs from the average by the vector Ψ−Γ=Φ nn . An example training set is shown in

Figure 8(a), with the average face Ψ .

This set of very large vectors is then subject to Principal Component Analysis, which seeks a set of M orthonormal vectors, nµ , which best describes the distribution of the

data. The kth vector, kµ is chosen such that

∑=

Φ=M

nn

Tkk M 1

2)(1µλ (4)

is a maximum, subject to

=

=otherwise

klk

Tl ,0

,1µµ

The vectors kµ and scalars kλ are the eigenvectors and eigenvalues, respectively, of the covariance matrix

∑=

=ΦΦ=M

n

TTnn AA

MC

1

1 (5)

where the matrix ]...[ 21 MA ΦΦΦ= . The matrix C,

however, is 2N by 2N , and determining the 2N eigenvectors and eigenvalues is an intractable task for typical image sizes. If the number of data points in the image space is less than the dimension of the space ( 2NM < ), there will be only 1−M , rather than 2N , meaningful eigenvectors.

Consider the eigenvectors nν of AAT such that

nnnT AA νλν =

The vectors determine linear combinations of the M training set face images to form the eigenfaces nµ :

MnA n

M

kknkn ,......,1,

1==Φ= ∑

=

ννµ (6)

With this analysis the calculations are greatly reduced, from the order of the number of pixels in the images ( 2N ) to the order of the number of images in the training set (M). In

practice, the training set of face images will be relatively small ( 2NM < ), and the calculations become quite manageable. The associated Eigen-values allow us to rank the Eigen vectors according to their usefulness in characterizing the variation among the images. The Training set, Normalized and Eigen faces are as shown below

3. Result and Discussion Results were evaluated using MATLAB environment for

different lighting conditions with uniform as well as non-uniform background. Crowded scene faces were detected by using skin color model algorithm for different stages are analysed experimentally in Figure1 to Figure 7. The experiment involves a set of 20 images tested in different conditions like light, Intensity, uniform and non-uniform background as shown in figure 9 – figure 12. Then the obtain face is cropped, scaled and stored in specified location for the Eigen faces Recognition.

(a). Original Image

(b). Detected Face Figure. 9: Very bright light condition (eg: sunlight)

under different background condition

(a)

Figure. 8 (a)Training set (b)Normalized Training set

(c ) Eigenfaces


69

Table No. 1 Comparison of Proposed method with Different Samples

4. Conclusion In this paper, the presented results could conclude that

detection algorithm based on only skin color and feature extraction considering skin color model, skin likelihood, performing skin segmentation and morphological operations. It works even with images taken on non-uniform background and different lightning and Intensity conditions. Face recognition is done by PCA-based eigenface method

approach, which has been widely cited as a prominent

application of PCA to pattern analysis. It is a simple and fast method that works reasonably well, though not very robust when dealing with some variations like changes in size. The main drawback of proposed method is that when the background color is similar to skin color, there is difficult to detect the skin areas in the image.

References [1] Ming Hu and Qiang Zhang , Zhiping Wang

“Application of Rough Sets to Image Pre-processing for Face Detection” IEEE International Conference on Information and Automation , June 20-23, 2008, Zhangjiajie, China.

[2] Sina jahanbin, hyohoon choi, “Automated facial feature detection and face recognition using gabor features on range and portrait images” IEEE transaction on 2008.

[3] Xiaogang Wang and Xiaoou,” Face Photo-Sketch Synthesis and Recognition” IEEE Trans. on Pattern analysis and machine intelligence

[4] Randazzo Vincenzo, Usai Lisa, “An Improvement of AdaBoost for Face-Detection with Motion and Color Information” IEEE 14th International Conference on Image Analysis and Processing (ICIAP) on 2007.

[5] Jure Kova C, Peter Peer, and Franc Solina “Human Skin Colour Clustering for Face Detection”.IEEE Region8, Eurocon 2003. pp.144-148. vol.2

[6] M. Turk and A. Pentland (1991). "Face recognition using eigenfaces". Proc. IEEE Conference on Computer Vision and Pattern Recognition. pp. 586–591.

[7] Thomas Heseltine1, Nick Pears, Jim Austin, Zezhi Chen “Face Recognition: A Comparison of Appearance-Based Approaches” VIIth Digital Image Computing: Techniques and Applications, Sun C., Talbot H., Ourselin S. and Adriaansen T. (Eds.), 10-12 Dec. 2003, Sydney.

(a). Original Image

(b). Detected Face

Figure. 10: Varying light intensity condition

(a). Original Image

(b). Detected Face

Figure. 11: Very bright light and uniform background condition

(a). Original Image

(b). Detected Face

Figure. 12: Very bright light and non-uniform background condition

Samples

Face Detection

Input faces

Faces Detected Efficiency

Sample1 14 13 92.8

Sample2 5 5 100

Sample3 8 7 80

Samples

Face Recognition

Input faces

Faces Recognized

Efficiency

Sample1 13 12 92.3

Sample2 5 5 100

Sample3 7 7 100


70

Authors Profile

Chandrappa D N is a research scholar. He is working as an Assistant Professor in Dept. of Electronics and communication Engineering at SJB Institute of Technology Bangalore-60. He has a B.E in degree in Electronics & Communication, M.E degrees in Power Electronics from University Visveswaraiah

College of Engineering and currently pursing PhD in Image processing. He has 10 years of teaching experience & has guided many UG and PG projects.

M Ravishankar received Ph.D degree in Computer Science from University of Mysore. He is working as a Professor in the department of Information Science and Engineering at Dayananda Sagar college of Engineering Bangalore and his research area includes Speech

processing , Image processing and computer vission.

D R RameshBabu received Ph.D degree in Computer Science from university of Mysore. He is working as a Professor in the department of Computer Science and Engineering at Dayananda Sagar college of Engineering Bangalore and his research area includes Digital Image processing.


71

Clustering Based Machine Learning Approach for Detecting Database Intrusions in RBAC Enabled

Databases

Udai Pratap Rao1, G. J. Sahani2, Dhiren R. Patel3

1 Dept. of Computer Engineering, S.V. National Institute of Technology Surat, Gujarat, INDIA [email protected]

2 Dept. of Computer Engineering, SVIT,Vadodara, Gujarat, INDIA

[email protected]

3 Dept. of Computer Science & Engineering, Indian Institute of Technology Gandhinagar, Ahmedabad, Gujarat, INDIA [email protected]

Abstract: Database security is an important issue of any organization. Data stored in databases is very sensitive and hence to be protected from unauthorized access and manipulations. Database management systems provide number of mechanism to stop unauthorized access to database. But, intelligent hackers are able to break the security of database systems. Most of the database systems are vulnerable or the environment in which database system resides may be vulnerable. People knowing such vulnerabilities can easily get access to database. Unauthorized suspicious activities can be trapped by database management systems. But, there are some authorized users who can violet the security constraints. Traditional database mechanisms are not sufficient to handle such attacks. Early Detections of any authorized or unauthorized access to database is very important for database recovery and to save the loss that can be occurred due to manipulation of data. There are number of database intrusion detection systems to detect intrusions in network systems, these IDSs cannot detect database intrusions. Very few IDS mechanism for databases has been proposed. Here we are proposing unsupervised machine learning approach for database intrusion detections in databases enabled with Role Based Access Control (RBAC) mechanism.

Keywords: Database Security, Clustering Technique, Malicious Transactions

1. Introduction Databases not only allow the efficient management and retrieval of huge amounts of data, but also they provide mechanisms that can be employed to ensure the integrity of the stored data. Data in these databases may range from credit card numbers to personal information like medical records. Unauthorized access or modification to such data results in big loss to customers. So, database security has become an important issue of most of the organizations. Recently number of database attack incidents has been occurred and number of customer records was stolen. Most of the attacks were encountered because of bad coding of database applications or exploiting database systems vulnerabilities. Web applications are the main sources of database attacks. Attackers may attack databases for several reasons and they may deduce newer techniques of database

attacks over a time. In today’s network environment is necessary to protect our data from attackers. Mainly database attacks are of two types: 1) intentional unauthorized attempts to access or destroy private data; 2) malicious actions executed by authorized users to cause loss or corruption of critical data. Although there are number of number of approaches available to detect unauthorized attempt to access data, attackers are succeeded in attacking the system because of the vulnerabilities. As database security mechanisms are not design to primarily detect intrusions, there are many cases where the execution of malicious sequences of SQL commands (transactions) cannot be detected. Therefore it becomes necessary to employ intrusion detection system [1]. In case a computer system is compromised, an early detection is the key for recovering lost or damaged data without much complexity. When an attacker or a malicious user updates the database, the resulting damage can spread very quickly to other parts of the database. Intrusion Detection System (IDS) provides good protections from attacks aimed at taking down access to the network, such as Distributed Denial of Service attacks and TCP SYN Flood attacks. But such systems cannot detect malicious database activity done by users. In recent years, researchers have proposed a variety of approaches for increasing the intrusion detection efficiency and accuracy [2]-[5]. But most of these efforts concentrated on detecting intrusions at the network or operating system level. But, there have been very few ID mechanisms specifically tailored to database systems. They are not capable of detecting malicious data corruptions. So, reasonable effort is required in area of database intrusion detection system. Intrusion detection systems determine the normal behavior of users accessing the database. Any deviation to such behavior is treated as intrusion. There are mainly two models of intrusion detection system, namely, anomaly detection and misuse detection. The anomaly detection model bases its decision on the profile of a user's normal behavior. It analyzes a user's current session and compares it with the profile representing his normal behavior. An alarm is raised if significant deviation is found


72

during the comparison of session data and user's profile. This type of system is well suited for the detection of previously unknown attacks. The main disadvantage is that, it may not be able to describe what the attack is and may sometimes have high false positive rate. In contrast, a misuse detection model takes decision based on comparison of user's session or commands with the rule or signature of attacks previously used by attackers. We are presenting unsupervised machine learning approach for database intrusion detections in databases enabled with role based access control (RBAC) mechanism. It means number of roles has been defined and assigned to users of database systems. Keeping database security in view, proper privileges are assigned to these roles. The rest of this paper is organized as follows. In section 2, we discuss related background. In section 3, a detailed overview about our approach is given. In section 4, analysis and result of our approach is presented. Finally in section 5 we conclude with the references at the end.

2. Related Work Application of machine learning techniques to database security is an emerging area of research. There are various approaches that use machine learning/data mining techniques to enhance the traditional security mechanisms of databases. Bertino et al. [6] have proposed a framework based on anomaly detection techniques to detect malicious behavior of database application programs. Association rule mining techniques are used to determine normal behavior of application programs. Query traces from database logs are used for this purpose. This scheme may suffer from high detection overhead in case of large number of distinct template queries. i.e. the number of association rules to be maintained will be large. DEMIDS is a misuse-detection system, tailored for relational database systems [7]. It uses audit log data to derive profiles describing typical patterns of accesses by database users. The main drawback of the approach presented as in [7] is a lack of implementation and experimentation. The approach has only been described theoretically, and no empirical evidence has been presented of its performance as a detection mechanism. Yi Hu and Brajendra Panda proposed a data mining approach [8] for intrusion detection in database systems. This approach determines the data dependencies among the data items in the database system. Read and write dependency rules are generated to detect intrusion. The approach is novel, but its scope is limited to detecting malicious behavior in user transactions. Within that as well, it is limited to user transactions that conform to the read-write patterns assumed by the authors. Also, the system is not able to detect malicious behavior in individual read-write commands. False alarm rate is may be more. It also does not hold good for different access roles. Sural et al. [9] have presented a approach for extracting dependency among attributes of database using weighted sequence mining. They have taken sensitivity of data items into consideration in the form of weights. Advantage of this approach is that more rules are

generated as compared to the approach presented in [8]. More rules generated reduce false alarms. But it is also not well suited approach for role based database access. Kamra et. al [10] have proposed a role based approach for detecting malicious behavior in RBAC (role based access control) administered databases. Classification technique is used to deduce role profiles of normal user behavior. An alarm is raised if roles estimated by classification for given user is different than the actual role of a user. The approach is well suited for databases which employ role based access control mechanism. It also addresses insider threats scenario directly. But limitation of this approach is that it is query-based approach and it cannot extract correlation among queries in the transaction.

3. Our Approach The approach we are presenting is a transaction level approach. Attributes referred together for read and write operations in transactions play important role in defining normal behavior of user’s activities.

For example consider the following transaction: Begin transaction select a1,a2,a3 from t1 where a1=25; update t2 set a4= a2+1.2(a3); End transaction

Where t1 and t2 are tables of the database and a1, a2, a3 are the attributes of table t1 and a4, a5 are the attributes of table t2 respectively. This example shows the correlation between the two queries of the transaction. It states that after issuing select query, the update query should also be issued by same user and in the same transaction. Approach presented in [10] can easily detect the attributes which are to be referred together, but it cannot detect the queries which are to be executed together. This example shows the correlation between the two queries of the transaction. It states that after issuing select query, the update query should also be issued by same user and in the same transaction. Our approach extracts this correlation among queries of the transaction. In this approach database log is read to extract the list of tables accessed by transaction and list of attributes read and written by transaction. The extracted information is represented in the form of following structure format: (Read, TB-Acc[ ], Attr-Acc[ ][ ], Write, TB-Acc[ ],Attr-Acc[ ][ ] ) Where Read and Write are binary fields while TB-Acc[ ] is binary vector of size equal to number of relations in database and Attr-Acc[ ][ ] is vector of N vectors and N is equal to the number of relations in the database. If transaction contains select query then Read is equal to 1 otherwise it is 0. Similarly, if transaction contains update or insert query Write is equal to 1 otherwise it is 0. Element TB-Acc[i]=1 if SQL command at hand access i-th table and 0 otherwise. Element Attr-Acc[i][j] = 1 if the SQL command at hand accesses the j- th attribute of the i-th table and 0 otherwise. Table 1 shows the representation of example transaction given above using this format.


73

Table 1: Representation of example transaction

Rd t1 t2 a1 a2 a3 a4 a5

1 1 0 1 1 1 0 0

Table 1: (Continued)

Wt t1 t2 a1 a2 a3 a4 a5

1 0 1 0 0 0 1 0

Where Rd=Read and Wt=Write

Values of fields of above structure will form the normal behavior of the transaction to be issued by user. Violation to such behavior will be detected as anomalous. The overall approach is depicted by figure 1.

Figure 1. Overview of the proposed approach

Information about the role of the users who had issued the transactions and the data items read written through these transactions is gathered from the database log. After gathering the history transaction from database log, it is preprocessed and stored as binary bits representing the items read and items written by the transactions in the form of structure presented above. Data generated in this form is form the dataset for clustering. Clustering forms the group of similar transactions. These groups represent the normal behavior of the users who have issued such transactions. It represents the role profile of the users who are authorized to issues such transactions. Once the role profiles are generated, next goal is to predict group of new incoming transactions. If the incoming transaction is found to be member of any of the cluster, then the transaction is considered as a valid transaction. If the incoming transaction is detected as an outlier, then it is considered as an invalid transaction and an alarm is generated. Valid transactions are fired on the database and are added to the database log. As history transactions are to be partitioned

into number of groups, we have used k-means clustering algorithm for clustering. K-means is the fastest among the partitioning clustering algorithms. Training tuples generated from database log has binary data fields. Therefore similarity measures of binary variables can be used for clustering such tuples. Similarity measure between two tuples for clustering algorithm of our approach is as follows.

ncount11

simm(t1,t2) =

ncount11 + ncount10 + ncount01

Where

ncount11 – count equals to number of similar binary fields of both the tuples t1 and t2 has value 1.

ncount10 – count equals to number of similar binary field of tuple t1 has value 1 and of tuple t2 has value 0.

ncount01 – count equals to number of similar binary field of tuple t1 has value 0 and of tuple t2 has value 1.

For example consider the following transactions:

Transaction tr1

Begin Transaction

select a1,a2 from t1;

update t2 set a4;

End Transaction

Corresponding bit pattern:

1 1 0 1 1 0 0 0 1 0 1 0 0 0 1 0

Transaction tr2

Begin Transaction

select a1,a3 from t1;

update t2 set a5;

End Transaction

Corresponding bit pattern:

1 1 0 1 0 1 0 0 1 0 1 0 0 0 0 1

ncount11 = 5

ncount10 = 2

ncount01 = 2

Similarly- similarity measure of tr1 and tr2 will be

simm(tr1,tr2) = 5/(5+2+2) = 55.5 %

Advantage of our unsupervised approach is that role information of the transactions need not have to log in database log. Behaviors of the users belonging to the same

Database Log (History Transactions)

Clustering (Learning Phase)

Preprocess (Read Items, Write Items)

Clusters (Role Profiles)

Current Session

User transaction

Comparison (Detection Phase)

Raise Alarm New DB Log

Update Outlier


74

role are grouped into same cluster. Approach is also well suited for the users with more than one role. Detection phase need to be generalized only.

4. Result and Analysis For verification of our approach, we generated number of database tables with number of attributes. We defined number of roles and generated number of transactions for these roles. Based on these transactions, we also generated large number of tuples as a training dataset. For detection, we generated number of valid as well as invalid transactions. We tested our approach by supplying valid as well as invalid transactions and our approach was detecting these transactions with full accuracy. We considered all the possible ways for generating valid and invalid transactions and we got the proper result for all the cases. Our approach is perfectly detecting correlations among commands of the transactions. We tested the approach by issuing the valid transactions by eliminating one of the SQL command from the transaction and it was detected as invalid transaction. When we issued the transactions with all the desired SQL commands, it was detected as valid transaction. Training time was also varying linearly with respect to number of training tuples as per the expectations. Figure 2 shows the nature of training time vs number of training tuples.

Figure 2. Training Time Vs Training Data

4. Conclusion In this paper we have proposed a new unsupervised machine learning approach of database intrusion detection for databases in which role based access control (RBAC) mechanism is enabled. It considers the correlations among the queries of the transaction and detects them accordingly. It does not require role information to be logged in database log. Clusters of transactions generated can also provide guidelines to the database administrator for role definitions.

References

[1] Fredrik Valeur, Darren Mutz, and Giovanni Vigna., “A learning-based approach to the detection of sql attacks,” In Proceedings of the International Conference on Detection of Intrusions and Malware,

and ulnerability Assessment (DIMVA), pages 123-140,2003.

[2] Lee, V. C.S., Stankovic, J. A., Son, S. H., “Intrusion Detection in Real-time Database Systems Via Time Signatures,” In Proceedings of the Sixth IEEE Real Time Technology and Applications Symposium, pages 121-128, 2000.

[3] Marco Vieira and Henrique Madeira, “Detection of Malicious Transactions in DBMS,” IEEE Proceedings- 11th Pacific Rim International Symposium on Dependable Computing, PP: 8, Dec 12-14, 2005.

[4] Ashish Kamra, Elisa Bertino, and Evimaria Terzi. , “Detecting anomalous access patterns in relational databases,” The International Journal on Very Large Data Bases (VLDB), 2008.

[5] Wai Lup LOW, Joseph LEE, Peter TEOH., “DIDAFIT: Detecting intrusions in databases through fingerprinting transactions,” ICEIS 2002 - Databases and Information Systems Integration, pages 121-127,2002.

[6] Elisa Bertino, Ashish Kamra, and James Early, “Profiling database application to detect sql injection attacks,” IEEE International Performance, Computing, and Communications Conference (IPCCC) 2007, pages 449–458, April 2007.

[7] C.Y. Chung, M. Gertz, and K. Levitt. , “DEMIDS: a misuse detection system for database systems,” In Integrity and Internal Control in Information Systems: Strategic Views on the Need for Control. IFIP TC11 WG11.5 Third Working Conference, pages 159-178, 2000.

[8] Yi Hu and Brajendra Panda, “A data mining approach for database intrusion detection,” In SAC ’04: Proceedings of the 2004 ACM symposium on applied computing, pages 711–716, New York, NY, USA, 2004.

[9] Abhinav Srivastava, Shamik Sural and A. K. Majumdar, “Database intrusion detection using weighted sequence mining,” Journal of Computers, Vol. 1, NO. 4, pages 8-12, JULY 2006.

[10] Elisa Bertino, Ashish Kamra and Evimaria Terzi, “Intrusion detection in rbac-administered databases,” In Proceedings of the Applied Computer Security Applications Conference (ACSAC), 2005.


75

Secure Password Authenticated Three Party Key Agreement Protocol

Pritiranjan Bijayasingh1 and Debasish Jena2

1Department of Computer Science and Engineering

International Institute of Information Technology, Bhubaneswar 751 013, India [email protected]

2Centre for IT Education

Biju Patnaik University of Technology Orissa, India

[email protected]

Abstract: In the past, several key agreement protocols are proposed on pre-shared password based mechanism. Due to development of communication technology, it is necessary to construct a secure end-to-end channel between clients. In this paper, an improved version of J. Kim et al proposed password-authenticated key exchange protocol has been proposed. The proposed scheme is secure against Denial of Service Attack, Perfect Forward Secrecy, Denning-Sacco Attack etc. Hence proposed scheme is much more superior to the previous scheme.

Keywords: Trusted third party, Kerberos, Denial of Services, and Key Exchange.

1. Introduction Secure communication between two users on a computer network is possible using either single key (conventional) encryption or public key encryption. In both systems, key establishment protocols are needed so the users can acquire keys to establish a secure channel. In single key systems, the users must acquire a shared communication key; in public-key systems, the users must acquire each others' public keys. In secure communications, one of the most important security services is user authentication. Before the communicating parties start a new connection, their identities should be verified. In the client-server model, password-based authentication is a favorable method for user authentication because of its easy-to-memorize property. As the passwords are easy to remember and selected from a small space, it allows a cryptanalyst to mount several attacks such as guessing attack, dictionary attack against it. Based on different cryptographic assumptions, various protocols have been proposed to achieve secure password-authenticated key exchange [7,9-12,19-20] for preventing these ever-present attacks. In the literature, most password-authenticated key exchange schemes assume that two parties share a common password. Two parties use their shared password to generate a secure common session key and perform key confirmation with regard to the session key. Most of these schemes consider authentication between a client and a server. In 2002, Byun et al. proposed a new password-authenticated key exchange protocol between two clients with different

passwords, which they call Client-to-Client Password-Authenticated Key Exchange (C2C-PAKE) protocol. In this protocol two clients pre-shared their passwords either with a single server (called a single-server setting) or respectively with two servers (called a cross-realm setting). However, in [13] Chen have shown the C2C-PAKE protocol in the cross-realm setting is not secure against dictionary attack from a malicious server in a different realm. In 2004, J. Kim et al, shown that the C2C-PAKE protocol is also vulnerable to the Denning-Sacco attack by an insider adversary. They have also modified the protocol to overcome the Denning-Sacco inside adversary attack.

In this paper, the weaknesses of the modified C2C-PAKE protocol are presented. Furthermore, the modified protocol, which repairs the problem of the modified C2C-PAKE protocol, is proposed.

The remaining of the paper is organized as follows. In section 2, a brief overview of the modified C2C-PAKE protocol is given along with its weaknesses. Then the improved version of the modified C2C-PAKE protocol is introduced in section 3. . Next, in Section 4, we briefly discuss the formal security notions of the proposed protocols. The paper concludes with the concluding remark and some directions for future work in section 5.

2. Review of Modified C2C-PAKE protocol In this section, the modified C2C-PAKE protocol in cross-realm setting is described with its weakness. 2.1 Modified C2C-PAKE Protocol 2.1.1 Computational Assumptions The scheme is based on numerical assumptions and computational assumptions. Let p, q be sufficiently large primes such that q|p − 1, and let G be a subgroup of Z∗

p of order q. During initialization step, a generator g ∈ G and hash function (H1, H2, H3, H4, H5) are published. All protocols throughout the paper are based on the discrete logarithm assumption (DLA) and Diffie-Hellman Protocol.


76

2.1.2. Protocol Description KDCA and KDCB are key distribution centers which store (Alice’s identity, ID(A), and password, pwa), and (Bob’s identity, ID(B), and password, pwb) respectively. K is a symmetric key shared between KDCA and KDCB. 1. Alice chooses x ∈R Z∗

p randomly, computes and sends Epwa(gx), ID(A) and ID(B) to KDCA. 2. KDCA obtains gx by decrypting Epwa(gx). KDCA selects r ∈R Z∗

p randomly and makes TicketB = EK(gx·r, gr, ID(A), ID(B), L). L is a lifetime of TicketB. Then KDCA sends TicketB, ID(A), ID(B) and L to Alice. 3. Upon receiving the message from KDCA, Alice forwards TicketB to Bob with ID(A). 4. Bob chooses y ∈R Z∗

p randomly, and computes Epwb(gy). Then he sends Epwb(gy), ID(A) and ID(B) to KDCB with TicketB. 5. KDCB obtains gx·r and gr by decrypting TicketB, selects r’ ∈R Z∗

p randomly and computes gx·r·r’ and gr·r’. Next KDCB

sends gx·r·r’ and gr·r’ to Bob. 6. Bob makes cs = H1(gx·y·r·r’ ) using gx·r·r’ and y. Then Bob chooses a random number a ∈R Z∗

p and computes Ecs(ga) and gr·r’·y. Finally, Bob sends Ecs(ga) and gr·r’·y to Alice. 7. Alice also can compute cs using gr·r’·y and x. Next, Alice selects b ∈R Z∗

p randomly and computes the session key sk = H2(gab) and Ecs(gb). Finally she sends Esk(ga) and Ecs(gb) for session key confirmation. 8. After receiving Esk(ga) and Ecs(gb), Bob gets gb by decrypting Ecs(gb) with cs and computes sk with gb and a. Bob verifies ga by decrypting Esk(ga) with sk. Bob sends Esk(gb) to Alice to confirm the session key. 9. Alice verifies gb by decrypting Esk(gb) with sk. 2.2 Cryptanalysis of Modified C2C-PAKE Protocol Let an outside attacker having knowledge of the whole cross-realm architecture comes in between client A and KDCA. In the first message transfer from A to KDCA, he may snoop the message and modify the IDB with some ID say, IDC which is a legitimate client in the other realm. Upon receiving the message from A, KDCA makes TicketC = EK(gx·r, gr, IDA, IDC, L) which is not the intended Ticket for client B. When KDCA sends the message to A in step-2, the attacker can again change the IDC to IDB. Hence, A doesn’t know what happened in between as she can’t decrypt the Ticket. In the same way the attacker may snoop the message in the step-4 and modify IDB to IDC. After decrypting the Ticket, KDCB assumes that client A wants to communicate with client C as he receives same IDC from both the message and the Ticket. He may discard Epwb(gy) as an redundant

message because, he can’t discard the TicketC as it is issued by KDCA. In step-5 KDCB sends the message to client C and the rest of the steps are executed normally. In this way, client A establishes the intended session key with client C not with the actual client B. In the, modified C2C-PAKE protocol there is no verifiable message where both A and B can verify their identities after step-4. In this way, the attacker may trap client C and communicate with client A pretending as client B. This causes identity mis-binding attack In step-2, the attacker may modify the Lifetime L of the ticket with a higher value than the value specified by KDCA. But, KDCB obtains the actual value of L. As client A has a high L value, she may request KDCB for service within the enhanced time period. But, KDCB is unable to provide the service after the actual lifetime L. This causes a Denial of Service (DOS) attack.

3. Proposed Protocol In our proposed protocol every host is registered with the TTP with different passwords. The hosts and the TTP agree upon a family of commutative one-way hash functions which is used for host authentication. One-way hashes of the passwords are being stored instead of storing the plaintext version of the passwords. One-way function is a function F such that for each x in the domain of F, it is easy to compute y=F(x), but given F(x) it is computationally infeasible to find any x. 3.1 Notations The following notations are used in this paper. Alice, Bob Honest Hosts TTP Trusted Third Party IDA, IDB Identities of Alice and Bob pwa, pwb Passwords of Alice and Bob EK(X) Encryption of plaintext X using key K DK(X) Decryption of plaintext X using key K SK Session Key between A and B H(pwa) One way hash of password of A g Generator of cyclic group p, q Large prime numbers A→B: M A sends message “M” to B TicketB Kerberos Ticket issued to A for service

from B sgnA( . ) Signature generated using the private

key of A K Shared Secret Key between TTP1 and

TTP2 3.2 Proposed Protocol Here we describe the steps involved in the protocol in detail. g, p and q are global parameters shared by protocol participants. 3.2.1 Single-Server Setting


77

Alice and Bob choose password pwa, pwb respectively and then transfer it to TTP through a secure channel. TTP stores (IDA, H(pwa)) and (IDB, H(pwb)) in its database. The following shows the steps:

i. Alice randomly chooses a number r, computes gr (mod p). The computed value is encrypted using the H(pwa) along with the IDs of the participating hosts. Alice then sends the calculated values to the server.

)](mod,,)[(: pgIDIDpwaHTA rBA→

ii. Server T then decrypts the received packet

using the H(pwa) of host A stored in its database to recover gr (mod p). Then server randomly chooses a number t and computes grt

(mod p) and encrypts the computed value concatenated with IDs of the two communicating entities, using the H(pwb) stored in its database. Then the computed value is sent to Bob by the server.

)](mod,,)[(: . pgIDIDpwbHBT trBA→

iii. After receiving the packet Bob decrypts it

using H(pwb) to get grt (mod p). He chooses a random number s, and computes an ephemeral key (Ke) as Ke=grts (mod p). Then he computes gs (mod p), concatenates it with IDA, encrypts the resulting value using H(pwb) and sends to the server T.

)](mod,)[(: pgIDpwbHTB sB→

iv. After receiving the packet, server decrypts it

using H(pwb) to recover gs (mod p). Then he computes gst(mod p), encrypts it using H(pwa) and sends it to Alice.

)](mod)[(: . pgpwaHAT ts→

v. Alice decrypts the received packet using

H(pwa) to get gst (mod p). Then she calculates the ephemeral key (Ke) as Ke=gstr (mod p). Alice chooses a random number a and computes RA=ga(mod p), encrypts RA using the ephemeral key Ke and sends it to Bob along with her own ID.

)(,: AKA REIDBAe

→

vi. Bob recovers RA by decrypting the received

packet using the ephemeral key Ke. He randomly chooses a number b as his private key, computes RB=gb (mod p) and encrypts RB using the ephemeral key Ke. Then he calculates the session key (SK) as SK=(RA)b (mod p)=gab

(mod p). Bob concatenates RA, RB and Alice’s ID, creates a signature of the result using his private key b to be verified by Alice. The signature is encrypted using the session key SK and sent to Alice along with the encrypted value of RB.

)),,((sgn),(: BAABSKBK RRIDEREABe

→

vii. Alice finds RB after decrypting with Ke. She computes the session key SK=(RB)a (mod p)=gba (mod p). After calculating the session key Alice verifies Bob’s signature. If the signature is verified, Alice concatenates Bob’s ID, RA and RB. Then she signs the result with her own private key, encrypts the signature with the session key SK and sends to Bob.

)),,((sgn: BABASK RRIDEBA→

viii. If Bob verifies Alice’s signature correctly, then he ensures that the same session key SK is occupied by both of them.

3.2.2 Dual-Server Setting Alice and Bob choose their distinct passwords pwa, pwb and registers themselves with TTP1 and TTP2 respectively using their passwords through a secure channel. TTP1 and TTP2 store (IDA, H(pwa)) and (IDB, H(pwb)) respectively in their own databases. We assume that TTP1 and TTP2 share a secret key K. The steps involved are described below:

i. A randomly chooses r and computes )(modpg r . Then she encrypts the computed value concatenated with IDs of the participating hosts using the H(pwa) and sends the calculated value to TTP1.

)](mod,,)[(:1 pgIDIDpwaHTTPA rBA→

ii. TTP1 obtains gr (mod p) after decrypting the

received packet by the H(pwa) of host A stored in its database. Then TTP1 randomly chooses a number t and computes grt (mod p) and gt (mod p). It prepares TicketB=EK(grt (mod p), gt (mod p), IDA, IDB, L). It concatenates IDs of the hosts with L and encrypts the resulting value with H(pwa). TTP1 sends TicketB to A along with the encrypted values of IDA, IDB and L. L is the lifetime of TicketB.

BBA TicketLIDIDpwaHATTP ],,,)[(:1 →

],,),(mod),(mod[ . LIDIDpgpgETicket BAttr

KB =

iii. After receiving the message, A forwards IDA and TicketB to TTP2.

BA TicketIDTTPA ,:2→

iv. Upon receiving the message, TTP2 decrypts

TicketB using shared key K to recover grt (mod p) and gt (mod p). It selects a random number t' and computes grtt' (mod p) and gtt' (mod p). Next, TTP2 concatenates the computed values with the IDs of both the hosts and encrypts the resulting value using H(pwb) and sends to B.

)](mod),(mod,,)[(: '.'..2 pgpgIDIDpwbHBTTP ttttr

BA→


78

v. B obtains grtt' (mod p), gtt' (mod p) after decrypting the message. He chooses s as random and finds an ephemeral key Ke= grtt's (mod p). He also randomly selects his private key as b and calculates his public key as RB=gb(mod p). Next, RB concatenated with the IDs of the hosts is encrypted using the ephemeral key Ke. Then he sends this encrypted value and gtt's (mod p) to A.

),,(),(mod: '..BBAK

stt RIDIDEpgABe

→

vi. After receiving the message A finds the ephemeral key Ke using r and gtt's (mod p). She also recovers RB from the message and chooses her private and corresponding public key as a, RA respectively, where RA= ga(mod p). She computes the session key SK=(RB)a (mod p)=gba (mod p). A concatenates RA, RB and IDB, creates a signature of the result using his private key a. The signature is encrypted using the session key SK and sent to B along with the encrypted value of RA.

)),,((sgn),(: BABASKAK RRIDEREBAe

→

vii. Upon receiving the message B obtains RA by

decrypting the message, computes the intended session key SK=(RA)b(mod p)=gab(mod p), and verifies A’s signature. If the signature is verified, B creates a signature exactly the same way as done by A and encrypts it using the session key SK. Then he sends the encrypted signature to A.

)),,((sgn: BAABSK RRIDEAB →

viii. Finally, A verifies the signature and if verified,

ensures that both of the hosts have the same session key SK.

4. Security Analysis of Proposed Protocol In this section, security of the proposed protocols is analysed. Our proposed protocols are secure against the type of attacks considered in [13,14] including Identity Mis-binding Attack and Denial of Service attack. 4.1. Identity Mis-binding Attack: Unlike the modified C2C-PAKE protocol, the IDs of the communicating entities are encrypted using the one-way hash value of the passwords in the proposed protocol. So, the adversary can’t change any of the IDs of the hosts. As a result, the proposed protocols are secure against identity mis-binding attack. 4.2. Denial of Service Attack: In the Dual-Server Setting of the proposed protocol, the lifetime L of the ticket issued by TTP1 is encrypted using key K in the TicketB as well as using H(pwa) to be transmitted to entity A. Hence, the value of L remains the same with A and TTP2 as specified by TTP1. As a result, the proposed protocol averts Denial of Service attack. 4.3. Perfect Forward Secrecy:

An adversary with pwa (or pwb) can easily compute gr by decrypting H(pwa)(gr). But these values do not help to compute Ke or SK in old sessions because session key generation is based on the Diffie-Hellman problem. Therefore the proposed protocol provides perfect forward secrecy.

4.4. Denning-Sacco Attack: Now we can show that our protocol is secure against Denning-Sacco attack. Like the original C2C-PAKE protocol, we also classify an adversary into two types. One is an Insider adversary and the other is an Outsider adversary. 4.4.1. In case of Outsider Adversary: Outsider adversary, with session keys Ke and SK can compute ga, gb and all conversations in the protocol. But he can not verify a candidate password pwa'(or pwb') of pwa (or pwb) since he can not get r (or s) which is a random secret value of A (or B). 4.4.2. In case of Insider Adversary with pwa: We are going to show that an adversary cannot mount a dictionary attack on pwb. To verify a candidate password pwb' of pwb, he must get gs. Since the value of s is a random number of B, he cannot compute valid gs. 4.4.3. In case of Insider Adversary with pwb: Similar to the case of insider adversary with pwa, he must get gr to verify a candidate password pwa' of pwa. Since the value of r is a random number of A, he cannot compute valid gr. 4.5 Dictionary Attack: In case of compromise of pwa or pwb, adversary can mount a dictionary attack if he gets gr or gs. However, he can not mount a dictionary attack as analyzed in Denning-Sacco attack. 4.6 On-line guessing attack, man in the middle attack and replay attack: It is the same as analyzed in the original C2C-PAKE protocol with regard to on-line guessing attack, man in the middle attack and replay attack. 4.7 Chen’s attack: Regarding Chen’s attack, there is no verifiable cipher text based on password in TicketB. So it is secure against the dictionary attack by a malicious TTP2. 5. CONCLUSION From the security analysis, we conclude that the proposed protocol meets all the security requirements defined in [13]. Furthermore, the protocol is secure against dictionary attack from a malicious server in a different realm.

References [1] Menezes A.,Oorschot P. van and Vanstone S.,

Handbook of Applied Cryptography, CRC Press, 1996 [2] Schneier Bruce., Applied Cryptography: Protocols and

Algorithms, John Wiley and Sons, 1994


79

[3] Stallings Williams., Cryptography and Network Security, 3rd Edition, Pearson Education, 2004

[4] B. A. Forouzan, Cryptography and Network Security, Tata McGraw Hill, Special Indian Edition, 2007

[5] W. Diffie and M. Hellman, “New Directions In Cryptography”. IEEE Transactions on Information Theory IT-11, pp. 644–654, November 1976

[6] Y. Her-Tyan and S. Hung-Min, “Simple Authenticated Key Agreement Protocol Resistant To Password Guessing Attacks”, ACM SIGOPS Operating Systems Review, vol. 36, no. 4, pp.14–22, October 2002

[7] M. Steiner, G. Tsudik, and M. Waidner, “Refinement And Extension Of Encrypted Key Exchange”. ACM Operating System Review, vol. 29, no. 3, pp. 22–30, 1995

[8] Y. Ding and P. Horster, “Undetectable On-Line Password Guessing Attacks”. ACM Operating System Review, vol. 29, no. 4, pp. 77–86, October1995

[9] C. L. Lin, H. M. Sun, and Hwang, “Three-Party Encrypted Key Exchange: Attacks And A Solution”. ACM Operating System Review, vol. 34, no. 4, pp. 12–20, October 2000

[10] M. Bellare, D. Pointcheval and P. Rogaway, “Authenticated key exchange secure against dictionary attacks,” in Eurocrypt 2000, LNCS 1807, pp. 139–155, Springer-Verlag, 2000.

[11] E. Bresson, O. Chevassut and D. Pointcheval, “New security results on encrypted key exchange,” in PKC 2004, LNCS 2947, pp. 145–158, Springer-Verlag, Mar. 2004.

[12] V. Boyko, P. MacKenzie and S. Patel, “Provably secure password-authenticated key exchange using Diffie-Hellman,” in Eurocrypt 2000, LNCS 1807, pp. 156–171, Springer-Verlag, May 2000.

[13] J. W. Byun, I. R. Jeong, D. H. Lee and C. S. Park, “Password-authenticated key exchange between clients with different passwords,” in ICICS’02, LNCS 2513, pp. 134–146, Springer-Verlag, Dec. 2002.

[14] L. Chen, “A Weakness of the Password-Autenticated Key Agreement between Clients with Different Passwords Scheme”. The document was being circulated for consideration at the 27th the SC27/WG2 meeting in Paris, France, 2003-10-20/24, 2003

[15] J. Kim, S. Kim, J. Kwak and D. Won, “Crypt-analysis and improvement of password authenticated key exchange scheme between clients with different passwords,” in ICCSA’04, LNCS 3043, pp. 895–902, Springer-Verlag, May 2004.

[16] D. Denning, G. Sacco, “Timestamps in key distribution protocols”. Communications of the ACM, Vol.24, No.8, pp. 533-536, 1981

[17] O. Goldreich and Y. Lindell, “Session-key generation using human memorable passwords only,” in Crypto 2001, LNCS 2139, pp. 408–432, Springer-Verlag, Aug. 2001.

[18] J. Katz, R. Ostrovsky and M. Yung, “Efficient password-authenticated key exchange using human-memorable passwords,” in Eurocrypt 2001, LNCS 2045, pp. 475–494, Springer-Verlag, May 2001.

[19] S. Jiang and G. Gong, “Password-based Key exchange With mutual authentication,” in SAC 2004, LNCS 3006, pp. 291-306, Springer-Verlag, 2004.

[20] S. Kulkarni, D. Jena, and S. K. Jena., “A novel secure key agreement protocol using trusted third party”. International Journal of Computer Science and Security, Volume (1): Issue (1), pp. 11–18, 2007.

Authors Profile

Pritiranjan Bijayasingh received the B.E. degree in Computer Science and Engineering from Balasore College of Engineering and Technology in 2005. He has joined Balasore College of Engineering and Technology as Lecturer since 26.08.2005. Now, he is persuing his M.Tech degree at

International Institute of Information Technology-Bhubaneswar, Orissa, India. His research area of interest is Information Security

Debasish Jena was born in 18th December, 1968. He received his B Tech degree in Computer Science and Engineering, his Management Degree and his MTech Degree in 1991, 1997 and 2002 respectively. He has joined Centre for IT Education as Assistant Professor since 01.02.2006. He has submitted his thesis for Ph.D. at NIT,

Rourkela on 5th April 2010. In addition to his responsibility, he was also IT, Consultant to Health Society, Govt. of Orissa for a period of 2 years from 2004 to 2006.His research areas of interest are Information Security, Web Engineering, Bio-Informatics and Database Engineering.


80

A Specific Replication Strategy of Middleware Resources in Dense MANETs

Anil Kumar1, Parveen Gupta2, Pankaj Verma3 & Vijay Lamba4

1C.S.E. Dept.,

H.C.T.M., Kaithal (INDIA). [email protected]

2C.S.E. Dept.,

APIIT, Panipat (INDIA) [email protected]

3C.S.E. Dept.,

H.C.T.M., Kaithal (INDIA) [email protected]

4E.C.E. Dept.,

H.C.T.M., Kaithal (INDIA)

Abstract: There are many dynamic services that are applicable on Adhoc Networks where many portable devices are scattered in a limited spatial regions. All mobile peers can autonomously cooperate without a fixed deployed network infrastructure such as shopping malls and airports. In this paper a Replication Strategy of middleware is proposed to manage, retrieve, and disseminate replicas of data/service components to cooperating nodes in a dense MANET. There are guidelines to enable optimistic lightweight resource replication capable of tolerating node exits/failures. Replication Strategy adopts original approximated solutions, specifically designed for dense MANET, which best suited for good scalability and limited overhead for dense MANET configuration (node identification and manager election), for replica distribution/retrieval, and for lazily-consistent replica degree maintenance.

Keywords: Replication, MANET, MANET Identification, Explored Node.

1. Introduction There are only few proposals which face the challenging issue of resource replication in mobile environments with the goal of increasing accessibility and effectiveness [1]. In particular, the paper proposes a middleware solution, called Replication Strategy in Dense MANETs), that transparently disseminates, retrieves, and manages replicas of common interest resources among cooperating nodes in dense MANETs [2]. Replication Strategy has the main goal of improving the availability of data and service components, and maintaining a desired replication degree for needed resources, independently of possible (and unpredictable) exits of replica-hosting nodes from the dense MANET. The paper presents how Replication Strategy addresses the primary challenging issues of dense MANETs, i.e., the determination of nodes belonging to the dense region, the sensing of nodes entering/exiting the dense MANET, and the dynamic election of replica managers responsible for orchestrating resource replication and maintenance. In this paper Replication Strategy protocols for dense MANET identification and manager election explained, extensive

simulation results about the Replication Strategy prototype, whose performance shows the effectiveness and the limited overhead of the proposed solution, by confirming the suitability of the application-level middleware approach even in conditions of high node mobility, notes about potential security issues, related work, and conclusions end the article.

2. REPLICATION STRATEGY SERVICES ON MANET

Replication Strategy provides replica distribution, replica retrieval, and replica degree maintenance facilities in a managed format as:

2.1 Replica Distribution The RD facility operates to transparently distribute resource replicas in the dense MANET [3] [4]. When a delegate enters a dense region, it communicates the descriptions of its shared resources, represented according to the Resource Description Framework (RDF) [5], to the replica manager, which autonomously decides the values of some parameters that influence the resource replication process.

2.2 Replica Retrieval The RR facility has the goal of retrieving resource replicas at provision time on the basis of the resource RDF-based descriptions [6], i.e., to dynamically determine the IP address of one suitable node hosting the requested resource and the unique name of the replica on that node. Resource retrieval is a hard task in MANETs, where a static infrastructure is not continuously available, thus preventing the usage of a fixed centralized lookup service known by all participants [8]. The usage of a single centralized repository with replica placement information is not viable [9]: i) a large number of requests could overwhelm the single point of centralization; ii) the repository would represent a single point of failure; and iii) the repository should be updated with strict consistency requirements not to hinder resource accessibility.

2.3 Replica Degree Maintenance Proactive RDM solutions available in the literature, such as [7], do not fit the provisioning environments addressed by Replication


81

Strategy. In fact, proactive RDM approaches usually require GPS-equipped wireless nodes that continuously monitor their mutual positions to foresee network exits, thus producing non-negligible network/computing overhead. Replication Strategy RDM, instead, implements a reactive solution with very low communication overhead by relaxing the constraint of anytime perfect consistency in the number of available replicas.

3. Replication Strategy on Dense MANET The Replication Strategy determines which nodes belong to the dense MANET and which ones among them have to play the role of replica managers.

3.3 Identification of Dense MANET Replication Strategy design a simple protocol where any node autonomously determines whether it belongs to the dense MANET or not. One node is in the dense MANET DM(n) only if the number of its neighbors, i.e., the nodes at single-hop distance, is greater than n. Each node autonomously discovers the number of its neighbors by exploiting simple single-hop broadcast discovery messages. By delving into finer details, at any time one Replication Strategy node can start the process of dense MANET identification/update; in the following, we will call that node the initiator. The initiator starts the original Replication Strategy protocol for dense MANET identification by broadcasting a discovery message that includes the number of neighbors (NoN) required to belong to the dense region and the identity of the sender. That number can be autonomously decided depending on the desired degree of connectivity redundancy: typically, a value between 10 and 20 ensures a sufficiently large set of alternative connectivity links. When receiving the discovery message, each node willing to participate replies by forwarding the message to its single-hop neighbors, if it has not already sent that message, and by updating a local list with IP addresses of detected neighbors.

3.4 Replica Manager Election The Replication Strategy middleware works to assign the manager role to a node located in a topologically central position. The protocol explores a manager candidate as a subset of nodes in the dense MANET, called Explored Nodes (ENs). Figure 1 shows a practical example of application of that guideline. The first step of the protocol considers node H: its farthest node is I, located at 4-hop distance; so, H is tagged with the value of that distance (H4 in the figure). Then, the Replication Strategy manager election protocol considers A, because it is the first node along the path from H to I: A's farthest node is I, at 3-hop distance (A3). At the next iteration, the protocol explores node D, which is chosen as replica manager by respecting the termination criteria described in the following of the section. Node D can reach any other node in the depicted dense MANET with a maximum path of two hops. Let us observe that Replication Strategy provides a simple way to react also to manager exits from the dense MANET. If the manager realizes it is going to exit, e.g., because its battery power is lower than a specified threshold, it delegates its role to the first neighbor node found with suitable battery charge and local memory. In the case the manager abruptly

fails, any resource delegate that senses the manager unavailability can trigger a new election procedure.

Figure 1. Replication Strategy exploring the sequence of ENs from H to A and A to D.

After having informally introduced the main guidelines of the protocol, let us now precisely specify how the manager election works. Each EN executes three operations: i) it determines the number of hops of the shortest paths connecting it to any farthest node in the dense MANET (the maximum of those hop numbers is called EN_value); ii) it identifies its neighbors located in the direction of its farthest nodes (forwarding_neighbors); and iii) it autonomously chooses the next EN among all the unexplored forwarding neighbors of already explored ENs with lowest associated values. To take possible device heterogeneity into account, Replication Strategy promotes the exploration only of ENs suitable to play the role of replica manager once elected. For instance, if a potential EN device has insufficient memory and too low battery life (if compared with configurable Replication Strategy thresholds), it is excluded from the manager election protocol. The protocol ends when either the Replication Strategy heuristic criterion determines there are no more promising nodes, or the current IN_value = Min_Int((worst explored EN_value)/2), where Min_Int(x) returns the least integer greater than x. Since Replication Strategy considers bi-directional links among MANET nodes, when the above equation is verified, it is easy to demonstrate that Replication Strategy has reached the optimal solution for the manager election. In particular, Replication Strategy combines two different strategies against manager assignment degradation, one proactive and one reactive, which operate, respectively, over large and medium time periods (Tp and Tr with Tp>>Tr). The proactive maintenance strategy establishes that the current manager always triggers a new manager election after Tp seconds. In addition to probabilistically improving the centrality of the manager position, the periodical re-execution of the election process contributes to distribute the burden of the role among different nodes, thus avoiding depleting the energy resources of a single participant. Moreover, let us rapidly observe that only the nodes located in the proximity of the dense MANET topology center have high probability to assume the manager role. Therefore, a target_EN_value can be easily determined equal to the EN_value of the current manager, thus speeding up manager election and reducing the protocol overhead. In addition to the above proactive degradation counteraction,


82

Replication Strategy exploits a reactive strategy that consists in repeating the farthest node determination at regular Tr periods, with the goal of understanding the current manager distance from the optimal placement. If the distance of manager farthest nodes, i.e., its new_EN_value, has increased if compared with the distance estimated at the moment of its election, i.e., its EN_value, Replication Strategy realizes that the manager has moved from the topologic center in a significant way; in that case, the manager itself triggers a new election process. The following subsections detail the adopted heuristics to limit the number of explored ENs and the exploited solution to determine, given a node, its farthest nodes in the dense MANET. Replication Strategy provides two tuning parameters that enable dense MANET administrators to trade between the quality of the manager election protocol and its performance. The first parameter, Desired_Accuracy, permits the initiator to tune the approximation considered acceptable for the election solution (see Figure 2). The second parameter, Max_Consecutive_Equal_Solutions, is introduced by observing that, when the Replication Strategy election protocol approaches the optimal solution, it often explores other candidate nodes without improving the current best EN_value. For each explored solution equal to the current best, Replication Strategy increases a counter; the counter resets when Replication Strategy finds a new solution outperforming the old best. The adopted heuristic stops the iterations when the counter reaches Max_Consecutive_Equal_Solutions. Figure 2 shows the pseudo-code of the Replication Strategy election protocol. explored_List = ø; forwarder_List = ø; best_Node = ø; best_Value = MAX; worst_Value = 0; unexplored_List = Initiator; while (unexplored_List != ø) { EN = Head(unexplored_List); EN_Value = Distance_From_Farthest(EN); explored_List = explored_List U EN; unexplored_List = unexplored_List – EN; forwarder_List = Get_Promising_Neighbors(EN); forwarder_List = forwarder_List – explored_List; if ((ENValue == Min_Int(worst_Value/2) || (ENValue <=worst_Value * desired_accuracy)) exit; if (EN_Value < best_Value) { best_Node = EN; best_Value = EN_Value; consecutive_Equal_Solutions = 0; unexplored_List = forwarder_List; } if (EN_Value > worst_Value) { worst_Value = EN_Value; if ((best_Value == Min_Int (worst_Value/2) || best_Value <= worst_Value*desired_accuracy)) exit; } if (EN_Value == best_Value)

{ consecutive_Equal_Solutions++; if (conecutive_Equal_Solutions == max_consecutive_equal_solutions) exit; unexplored_List = unexplored_List U forwarder_List; } } Print(best_Node)

Figure 2. Pseudo-code of the Replication Strategy manager election protocol.

Elected manager perform replica distribution, replica retrieval, and replica degree maintenance also.

4. Experimental Results For evaluation the approach, we have implemented a Replication Strategy by using NS2 simulator for dense MANET identification, manager election, and resource retrieval. The simulations have three main goals which are explained below

4.1 Network Overhead First, we have carefully evaluated the network overhead of the Replication Strategy protocols for dense MANET identification and manager election, to verify that the DMC proposals are lightweight enough for the addressed deployment scenario. We have tested the two protocols in different simulation environments, with a number of nodes ranging from 100 to 1100 (increasing of 100 nodes each step). For protocols we have measured the average number of messages sent by each participant, over a set of more than 1,000 simulations. The results reported in Figure 3 are normalized to the number of nodes actively participating in the protocol. The dense MANET identification protocol is designed to determine participant nodes by requiring only one local broadcast from each node reachable from the intiator. Given the dependence of manager election network overhead on the number of iterations, we have carefully investigated the convergence of the Replication Strategy protocol while varying the number of dense MANET nodes (see Figure 4). The results are average values on a set of simulations where the role of initiator is assigned to a different node at each simulation. The experimental results about the number of iterations have demonstrated the almost linear dependence on dense MANET diameter and have shown to be negligibly affected by the choice of the initiator node. Let us briefly observe that the non-monotonic growth of the overhead trace in Figure 3 is due to different factors. First, the dense MANET diameter only increases in correspondence with some threshold values of the number of participants. About the latency imposed by the Replication Strategy manager election protocol, let us briefly observe that it mainly depends on three factors: obviously, one is the number of dense MANET participants; the others are two Replication Strategy configurable timers that establish, respectively, the message delay for preventing broadcast storm in farthest node determination and the time interval waited by the current EN before delegating its role. All the presented results have been obtained by setting the former to 2.5s and the latter to 5s per hop of the diameter. These


83

values guarantee that the current EN passes node exploration responsibility only after having received most replies from farthest nodes, thus achieving its correct EN_value.

Figure 3. Number of iterations needed for the Replication Strategy manager election protocol.

Figure 4. Message sent/received per node in dense MANET identification and manager election.

4.2 Manager Election Inaccuracy We have measured its accuracy in assigning the manager role to a node close to the actual topology center of the dense MANET. We have run over 200 simulations in the most populated scenario of 1100 nodes and, for any simulation, we have measured the election inaccuracy defined as the hop distance between the manager chosen by the Replication Strategy protocol and the actual optimal solution.

Figure 5. Inaccuracy of the Replication Strategy manager election protocol

The results in Figure 5 are obtained by starting each election from a different initiator node. In more than 90% of the runs, the Replication Strategy protocol has identified either optimal solutions or quasi-optimal solutions at 1-hop distance from the actual optimum. The average inaccuracy is only 0.385 hops, which represents a largely acceptable value for the addressed application scenario.

4.3 Impact of Node Mobility on the Accuracy of the Replication Strategy Dense MANET Identification Protocol

To test the robustness of Replication Strategy solutions, we have evaluated the accuracy of the dense MANET identification protocol by varying the mobility characteristics of network nodes. Let us rapidly note that the manager election protocol executes for very limited time periods and re-starts its execution only after a long time interval; to a certain extent, it is assumable that it operates under static conditions. On the contrary, the dense MANET identification protocol should continuously work to maintain an almost consistent and updated view of the dense MANET participants, crucial for the effective working of Replication Strategy solutions. For this reason, the sub-section focuses on the behavior of our dense MANET identification protocol as a function of node mobility. Any pair of random movements of randomly chosen nodes occurs every M seconds, with M that varies from 10 to 60. Any other node movement not producing arrival/departure in/out the dense MANET does not affect at all the behavior of the Replication Strategy identification solution. Figure 6 reports the dense MANET identification inaccuracy, defined as the difference between the number of dense MANET participants determined by the Replication Strategy protocol and its actual value. The inaccuracy is reported as a function of the mobility period M and for different values of the time period used for Hello packets. Each point in the figure represents an average value obtained by capturing the state of the network over 30 different runs. The figure shows that the average inaccuracy is very limited and always within a range that is definitely acceptable for lazy consistent resource replication in dense MANETs. As expected, the inaccuracy grows when node mobility grows, for fixed values of the Hello message period.

Figure 6. Accuracy of the Replication Strategy dense MANET identification protocol as a function of node mobility


84

However, even for relatively high values of the Hello period, the Replication Strategy identification inaccuracy is negligible for the addressed application scenario (on the average, always less than 1.7), for the whole range of node mobility that can be of interest for dense MANETs. Let us observe that this permits to set relatively high periods for Hello packets, while obtaining a low inaccuracy for the dense MANET identification, thus significantly reducing the message exchange overhead.

4.4. Replica Retrieval Message Overhead We have also performed a wide set of simulations to evaluate the message overhead of the Replication Strategy RR strategy, i.e., the overall number of search messages exchanged within the dense MANET, on the average, when a requester looks for the IP address of a node hosting a replica of the needed resource. To this purpose, we compare the Replication Strategy RR results on message traffic with the network overhead that would be generated by a trivial retrieval strategy based on flooding query messages to all nodes belonging to the dense MANET (in the following we call that strategy Query Flooding – QF). Figure 7 reports the Replication Strategy RR plot as a function of the rep_Hops parameter (the configurable number of IRPs disseminated in the dense MANET) in the most populated scenario of 1100 nodes. The figure shows the average number of search messages necessary to find the first IRP, normalized to the number of dense MANET participants. As expected, the number of search messages per node decreases as the number of disseminated IRPs grows. But, most important, even for a very limited number of IRPs, the Replication Strategy expanding ring RR strategy largely outperforms QF, by providing a suitable trade-off between RR complexity, provision-time RR traffic, and IRP distribution overhead.

Figure 7. Message overload for resource retrieval depending on the number of disseminated IRPs.

5. Conclusion Dense MANETs can benefit from the assumption of high node population to enable lazily consistent forms of replication for resources of common interest. This can significantly increase resource availability notwithstanding unpredictable node movements entry/exit dense MANETs. The Specific Replication Strategy project middleware

demonstrates how it is possible to control lightweight and effective replica management. The performance results obtained are encouraging further new Specific Replication Strategy related research activities. First, we are evaluating the impact of the proposed protocols on the performance of Replication Strategy-supported applications. In particular, we are carefully evaluating in which deployment conditions (dense MANET diameter and number of participants, node mobility, resource distributions/requests ratio, average replica size …) the traffic reduction due to the central position of the replica manager offsets the overhead incurred in its election. Finally, as already stated in previous parts of the paper, we are working on more sophisticated replica retrieval solutions, which also take into account simple security mechanisms for authorized resource access and incentive-based promotion of node collaboration.

References [1]. T. Qing, D.C. Cox, “Optimal Replication Algorithms

for Hierarchical Mobility Management in PCS Networks”, IEEE Wireless Communications and Networking Conf. (WCNC), year. 2002.

[2]. P. Bellavista, A. Corradi, E. Magistretti, “Lightweight Replication Middleware for data and Service Components in Dense MANETs”, 1st IEEE Int. Symp. on a World of Wireless, Mobile and Multimedia Networks (WoWMoM), June 2005.

[3]. M.J. Freedman, E. Freudenthal, D. Mazières, “Democratizing Content Publication with Coral”, 1st USENIX/ACM Symp. on Networked Systems Design and Implementation, Mar. 2004

[4]. C. Bettstetter, G. Resta, and P. Santi, “The Node Distribution of the Random Waypoint Mobility Model for Wireless Ad hoc Networks”, IEEE Transactions on Mobile Computing, Jul.-Sept. 2003.

[5]. S. Decker, P. Mitra, S. Melnik, “Framework for the Semantic Web: an RDF Tutorial”, IEEE Internet Computing, Vol. 4, No. 6, Nov.-Dec. 2000.

[6]. I. Aydin, C.-C. Shen, “Facilitating Match-Making Service in Ad Hoc and Sensor Networks using Pseudo Quorum”, 11th IEEE Int. Conf. Computer Communications and Networks, Oct. 2002.

[7]. M. Boulkenafed and V. Lssarny “A middleware service for mobile ad hoc data haring, enhancing data availability” in 4th ACM/IFIP/USENIX Middleware June 2003.

[8]. K. Chen, K. Nahrstedt, “An Integrated Data Lookup and Replication Scheme in Mobile Ad Hoc Networks”, SPIE Int. Symp. on the Convergence of Information Technologies and Communications (ITCom 2001), Aug. 2001.

[9]. M. Tamori, S. Ishihara, T. Watanabe, T. Mizuno, “A Replica Distribution Method with Consideration of the Positions of Mobile Hosts on Wireless Ad-hoc Networks”, 22nd Int. Conf. on Distributed Computing Systems Workshops, July 2002.


85

Reliable Soft-Checkpoint Based Fault Tolerance Approach for Mobile Distributed Systems

Surender Kumar1, R.K. Chauhan2 and Parveen Kumar3

1Haryana College of Technology & Management (H.C.T.M), Department of I.T

Ambala Road, Kaithal-136027, India [email protected]

2Kurukshetra University, Kurukshetra(K.U.K), Department of Computer Sc. & Applications

Kaithal-Pehowa Road, Kurukshetra-136119, India [email protected]

3Meerut Institute of Engg. & Tech. (M.I.E.T) , Department of Computer Sc. & Engineering, N.H. 58, Delhi-Roorkee Highway, Baghpat Road Bypass Crossing, Meerut-250005, India

[email protected]

Abstract: A checkpoint is a local state of a process saved on stable storage. In coordinated checkpointing, processes take their checkpoints in such a manner that the resulting global state is consistent. Mostly it follows two-phase commit structure. In the first phase, processes take tentative checkpoints and in the second phase, these tentative checkpoints are made permanent. In case of a single systems failure or reply negatively, processes rollback to last consistent checkpointed state. In mobile distributed systems process is considering as a MH without stable storage and there is wireless link between MH and MSS. So during checkpointing, an MH has to transfer a large amount of checkpoint data to its local MSS over the wireless network. Since MHs are prone to failure and during the failure of an MH these checkpoints become unnecessary. Therefore, transferring such unnecessary checkpoints and then rollback after failure, may waste a large amount computation power, bandwidth, time and energy. In this paper we propose a soft-checkpoint based checkpointing algorithm which are suitable for mobile distributed systems. Our soft-checkpoints are store on local MHs and have not any transferring cost. These soft- checkpoints are stored in stable only receiving the commit message or during failure of a node and discard locally, after receiving the abort message. Hence, in case of failure our soft- checkpoint has much less overhead as compare to tentative.

Keywords: Checkpointing, Coordinated checkpointing, uncoordinated Checkpointing, Fault Tolerance, Mobile Distributed Systems, Domino-free

1. Introduction Checkpointing and rollback recovery techniques for fault-tolerance in distributed systems have studied extensively in the literature. However, little attention has been devoted to fault–tolerance techniques for mobile distributed systems. Today, mobile user become common in distributed system due to availability, cost and mobile connectivity. Mobile distributed systems raise new issue such as mobility, low bandwidth of wireless channels, disconnections, limited battery power and lack of reliable stable storage of mobile nodes. Due to these unique features of mobile systems, it is

not appropriate to directly apply checkpointing protocol designed for fixed network distributed systems to mobile systems. Checkpointing algorithms are classified into two main categories: uncoordinated and coordinated. In uncoordinated approach each process takes its checkpoint independently without the knowledge of other process. This approach is simple, but suffers from domino-effect. In coordinated checkpointing approach, processes take checkpoints in such a manner that the resulting global state is consistent. So, this approach is domino-free [2], stores minimum number of checkpoints on the stable storage (maximum two). In a MDCS the checkpoint taken by a MH is need to be transferred to its current MSS due to lack of stable storage at MH level [2]. Most of the existing coordinated checkpointing protocols try to minimize the number of processes to take checkpoint and to make checkpoint algorithms non-blocking. These algorithms follows two-phase commit structure [1] - [5]. In the first phase, processes take tentative checkpoints and in the second phase, these are made permanent. The main advantage is that only one permanent checkpoint and at most one tentative checkpoint is required to be stored. In the case of a fault, processes rollback to last checkpointed state [1].

2. Related Work and Problem Formulation:

2.1 Related work The work presented in this paper shows the performance improvement over work reported in [1]-[5]. These algorithms either try to make the checkpointing algorithm either minimum-process or non-blocking or both minimum–process and non-blocking. All these algorithms follow two-phase commit structure. In the first phase, processes take temporary checkpoints when they receive the checkpoint request. These tentative checkpoints are store in the stable storage of MSS. In the second phase, if an MSS learns that all of its processes to


86

whom it sends checkpoint request, have taken their tentative checkpoints successfully, initiator MSS sends the commit message to all the participating node. However, in case of a single failure or reply aborted, initiator process broadcast the aborted message. After receiving commit request, processes convert their tentative checkpoints into permanent ones and on receiving abort message, every process that has already taken a temporary checkpoint must discard it. Later the checkpointing algorithm has to be restarted again from the previous consistent global state. As to take a checkpoint, an MH has to transfer a large amount of data to its local MSS over the wireless network. Since the wireless network has low bandwidth and the MHs have relatively low computation power. During failure theses transferred checkpoint become useless and discard later. These algorithms have higher checkpoint latency and recovery time as transferring such temporary checkpoints on stable storage may waste a large amount computation power, bandwidth, energy and time.

2.2 Problem formulation In mobile distributed system multiple MHs are connected with their local MSS through the wireless links. A process is considering as a MH without stable storage. During checkpointing, an MH has to transfer a large amount of checkpointed data (like variables, control information, register value and environments etc.) to its local MSS over the wireless network. So, it consume resources to transfer data and to rollback in case of any failure to its consistent state. If even a single MH fails to take a checkpoint all the checkpoint which is taken on MSS must be rollback. So, it increases the checkpoint latency, recovery time during and disk overhead during the failure. The objective of the present work is to design a checkpointing approach that is suitable for mobile computing environment. Checkpointing and recovery protocol for Mobile computing environment demands for efficient use of the limited resources of mobile environment i.e wireless bandwidth, battery power and memory etc. Consider a mobile environment, in which a MSS has 1000 MHs in their minimum set. During the first phase 999 MHs take their temporary checkpoints successfully and one MH fails to take checkpoint. In such case, checkpointing process must be aborted; MHs discard their temporary checkpoint and the system restarts its execution from a previous consistent global checkpoint saved on the stable storage during fault free operation. Observe that taking a temporary checkpoint in the stable storage and later discard it, affects the bandwidth utilization and waste the MHs limited battery power.

2.3 Basic idea The basic idea of the proposed scheme is to store the checkpoint as a soft-checkpoint in the memory, till the time it receives the hard checkpoint request from the initiator. The initiator asks all processes in minset[] whether they are agreed to soft-checkpoint or not. If one process reply abort, or fails to respond within a timeout period, then the initiator broadcast to aborted message and after receiving ABORT, process rollback to their previous consistent state locally. So this approach has not any transferring cost. If initiator know

that all processes takes their hard checkpoint successfully, then it forward commit message. This approach assumes that each site has its own local log, sufficient volatile memory and can therefore rollback or commit the transaction reliably.

3. Soft-checkpoint based Checkpointing Approach:

3.5 System model A mobile distributed system is a distributed system where some of the processes are running on mobile hosts (MHs)[7]. It consists of Static Hosts (SHs), Mobile Hosts(MHs) and the Mobile Support Stations(MSSs). So, the mobile distributed system can be considered as consisting of “n” MHs and “m” MSSs. The static network provides reliable, sequenced delivery of messages between any two MSSs, with arbitrary message latency. Similarly, the wireless network within a cell ensures FIFO delivery of messages between an MSS and a local MH. The links are FIFO in nature. An MH communicates with other nodes of system via special nodes called mobile support station (MSS).An MH can directly communicate with an MSS only if the MH is physically located within the cell serviced by MSS. A static node that has no support to MH can be considered as an MSS with no MH. A cell is a geographical area around an MSS in which it can support an MH .An MH can change its geographical position freely from one cell to another cell or even area covered by no cell .At any given instant of time an MH may logically belong to only one cell; its current cell defines the MH’s location and the MH is considered local to MSS providing wireless coverage in the cell. If an MH does not leave the cell, then every message sent to it from local MSS would receive in sequence in which they are sent.

3.6 The Soft checkpoint approach Therefore, in the present work we emphasize on to make computation faster, eliminating the overhead of taking temporary checkpoint on stable storage and finally utilize the available fast memory. In this context we proposed a soft-checkpoint based approach, in which processes takes their temporary checkpoints, as a soft-checkpoints on their main memory. The approach work as follows:

3.2.4 Action at the initiator Pj The initiator can be in one of four states during the checkpointing process: Initial state, Waiting state, Hard checkpointing state, and COMITTED/ABORTED state, as shown in state transition diagram of Figure[1]. Each state is followed by its previous state and work as follows. 1. When Pj initiates checkpointing algorithm:

{take_soft-checkpoint; Increment csn; set weight=1; Compute minset[]; send take soft-checkpoint request to all processes belongs to minset along with minset[], csn and weightj=weightj/2;} /*If Pi initiate its(x+1)th checkpoint then the set of processes on which Pi depends(directly or transitively) in its xth checkpoint is minimum set[8]. */

2. Initiator waits for response to the soft-checkpoint


87

request 3. If( time out/Failure)

{Broadcast Global_ ABORT; and go to step 5.} else if (weight ==1)//receives positive response from all {Sends take Hard checkpoint request message. old_csnj:= Soft checkpointj} else

go to step 2. 4. if(Receives ACK regarding taking Hard checkpoint

from all process in minset []) { Broadcast Global_COMMIT message; go to step 5.} else { Broadcast Global_ABORT message;}

5. END

3.2.5 Any at any Process Pi which belongs to minsetj[]

Process Pi which is the part of minsetj[] have the four states :Initial, Prepared to takes soft-checkpoint state, Prepared to take hard checkpoint state and COMMITTED/ABORTED state. Each state is followed by its previous state. The actions to be taken during these states are shown in Figure 2. Process Pi works in different conditions as follows: On the receipt of take Soft-checkpoint request:

• Take Soft-checkpoint; • Increment in soft_csn; • Sends reply to Initiator;

On sending computation message to Pj:

• Sends csni[i], minset with the message Upon receiving computation message from any process Pk: if (m.csn > csni[k]) { csni[k]:=m.csn; Increment csni[i];

Take Soft-checkpoint; Process the message; reset Recvi, Recvi[k]:=1;}

else if (m.csn= csni[j]) Process the message; update Recvi if necessary; else // m.csn<csni[j] Process the message; On receiving Global_ABORT message:

• Discard the soft checkpoint from the volatile memory; On receiving Global_COMMIT message:

• Discard old Hard checkpoint, if any; • Convert the Soft Checkpoint into Hard Checkpoint

Process takes the soft-checkpoint on happening of any below condition first:

• On the receipt of Take soft checkpoint request from initiator;

• Upon receiving the computation message, if (m.csn > csni[k]);

• If local time to take soft checkpoint expires

3.2.6 Increasing Reliability of Proposed Approach:

As soft-checkpoints are necessarily less reliable than hard checkpoint, because they are stored on volatile memory of an MHs. Hence, there is a grate need to convert theses soft -checkpoints in to hard (permanent) checkpoints. These soft

Figure 2. State transition diagram of process Pi

Figure 1. State transition diagram initiator process


88

checkpoints are converted in to hard, based on the following conditions:

• On receipt of the commit message from the initiator. • At the time of node disconnection and Handoff: in

mobile distributed system, a MH may get disconnected voluntarily and non-voluntary. In voluntarily disconnections, MH transfer its soft checkpoint in to the stable storage of MSS before disconnection. We call this disconnected checkpoint. In such case the MSS who receive the disconnected checkpoint, coordinate on the behalf of disconnected MH. Non-voluntary disconnections are treated as fault [8].

• On the basic of maxsoft: The number of soft checkpoint stored per hard checkpoint is called maxsoft, and it depends on the quality of service of current network [9].

4. Comparison with Hard Checkpointing Approach:

In this section we compare our soft checkpoint based approach with the hard checkpoint approach in different perspectives.

4.1 Checkpoint Latency Checkpointing latency is the time needed to save the checkpoint. It is observed that there is a big difference between the latency in soft and hard checkpointing based approach. This is due to fact that soft checkpoint based approach uses fast volatile memory and not have any transmission cost.

4.2 Transmission Cost Soft checkpoint based approach have not any transmission cost as the checkpoint are stored locally on volatile memory, instead through wireless link on stable storage of MSS.

Table 1: Soft Vs. Hard Checkpointing Approach Parameter Hard-based Soft-based Checkpoint Latency High Low Transmission Cost High Low Recovery Time High Low CPU Overhead High High Main Memory Requirement Low High Reliability High Low Efficiency Low High Portability High Low Additional Hardware Not

required Additional processors

Suitability For Large Systems

For Small Systems

Power Consume High Low 4.3 Recovery Time

It is observed that recovery time in soft checkpoint based approach is much less if the failure occurs before converting the soft checkpoint into hard checkpoint and it is comparable in the other case.

5. Conclusion In our proposed approach, a process in minset [ ] takes a soft checkpoint first and then soft checkpoint will be discarded, if it receives aborted message from the initiator. As soft checkpoints are saved on main memory of the mobile hosts [MHs], and then the soft checkpoint will be saved on the stable storage of MSS at a later time only if they receive the hard checkpoint request from the initiator. As a result soft checkpoint based approach requires low battery power of MHs, low checkpoint latency, low transmission cost, and low recovery time due to reduced disk accessed of MSS by the MHs. As soft checkpoint approach is less reliable, to make it reliable we transfer the soft checkpoint on stable storage by happen the conditions given in section 3.2.3.

References [1] Cao G. and Singhal M., “On coordinated

checkpointing in Distributed Systems,” IEEE Transactions on Parallel and Distributed Systems, vol. 9, no.12, pp. 1213-1225, Dec 1998.

[2] G. Cao and M. Singhal, “Mutable Checkpoints: A New Checkpointing Approach for Mobile Computing Systems,” In Proceedings of the IEEE Trans. Vol. 12, No. 2, pp. 157-172, Feb. 2001.

[3] Cao G. and Singhal M., “On the Impossibility Min-process Non-blocking Checkpointing and an Efficient Checkpointing Algorithm for Mobile Computing Systems,” In Proceedings of the International Conference on Parallel Processing, pp. 37-44, August 1998.

[4] Elnozahy E.N., Johnson D.B. and Zwaenepoel W., “The Performance of Consistent Checkpointing,” In Proceedings of the 11th Symposium on Reliable Distributed Systems, pp. 39-47, October 1992.

[5] R.Koo and S.Toueg, “Checkpointing and Roll-back Recovery for Distributed systems,” IEEE Transactions on Software Engineering, pages 23-31, January 1987.

[6] Elnozahy E.N., Alvisi L., Wang Y.M. and Johnson D.B., “A Survey of Rollback-Recovery Protocols in Message-Passing Systems,” ACM Computing Surveys, vol. 34, no. 3, pp. 375-408, 2002.

[7] Acharya A. and Badrinath B. R., “Checkpointing Distributed Applications on Mobile Computers,” In Proceedings of the 3rd International Conference on Parallel and Distributed Information Systems, pp. 73-80, September 1994.

[8] L. Kumar, M. Misra, R.C. Joshi, “Low overhead optimal checkpointing for mobile distributed systems Proceedings,” 19th IEEE International Conference on Data Engineering, pp 686 – 88, 2003.

[9] N. Naves and W.Fuchs, “Adaptive Recovery for Mobile Environments,” In Proceeding of the IEEE High-Assurance Systems Engineering Workshop, October 21-22, 1996, pp, 134-141. Also in Communications of the ACM, 1997, January, vol.40, no.1, pp.68-74.


89

Cryptography of a Gray Level Image Using a Modified Hill Cipher

V. U. K. Sastry1*, D. S. R. Murthy2, S. Durga Bhavani3

1Dept. of Computer Science & Engg., SNIST,

Hyderabad, India, [email protected]

2Dept. of Information Technology, SNIST,


3School of Information Technology, JNTUH,


Abstract: In this paper, we have used a modified Hill cipher for encrypting a Gray level image. Here, we have illustrated the process by considering a couple of examples. The security of the image is totally achieved, as the encrypted version of the original image, does not reveal any feature of the original image. Keywords: Cryptography, Cipher, Gray level image, Encrypted image, Modular arithmetic inverse.

1. Introduction The study of cryptography of gray level images [1 – 3] by using block ciphers has gained considerable impetus in the recent years. The transformation of an image from its original form to some other form, such that it cannot be deciphered what it is, is really an interesting one. In a recent investigation [4, 5], we have developed two large block ciphers by modifying the Hill cipher [3]. In these ciphers, the key is of size 512 bits and the plain text is of size 2048 bits. In one of the papers [6], the plain text matrix is multiplied by the key on one side and by its modular arithmetic inverse on the other side. From the cryptanalysis and the avalanche effect, we have noticed that the cipher is a strong one and it cannot be broken by any cryptanalytic attack. In the present paper, our objective is to develop a block cipher, and to use it for the cryptography of a gray level image. Here, we have taken a key containing 64 decimal numbers (as it was in [1]), and generated a key matrix of size 32 x 32 by extending the key in a special manner (discussed later), and applied it in the cryptography of a gray level image. In Section 2, we have developed a procedure for the cryptography of a gray level image. In Section 3, we have used an example and illustrated the process. Finally, in Section 4, we have drawn conclusions from the analysis.

2. Development of a Procedure for the Cryptography of a Gray Level Image Consider a gray level image whose gray level values can be represented in the form of a matrix given by P = [Pij], i = 1 to n, j = 1 to n. (2.1) Here, each Pij lies between 0 - 255. Let us choose a key k let it be represented in the form of a matrix given by K = [Kij], i = 1 to n, j = 1 to n, (2.2) where each Kij is in the interval [0, 255]. Let C = [Cij], i = 1 to n, j = 1 to n (2.3) be a matrix, obtained on encryption. The process of encryption and the process of decryption, which are quite suitable, for the problem on hand, are given in Fig. 1.


90

Figure1. The process of encryption and the process of decryption

Here, Mix ( ) is a function used for mixing thoroughly the decimal numbers (on converting them into binary bits) arising in the process of encryption at each stage of iteration. IMix ( ) is a function which represents the reverse process of Mix ( ). For a detailed discussion of these functions, and the algorithms involved in the processes of encryption and decryption, we refer to [1].

3. Illustration of the cryptography of the image Let us choose a key Q consisting of 64 numbers. This

can be written in the form of a matrix given by

The length of the secret key (which is to be transmitted) is 512 bits. On using this key, we can generate a new key E in the form

where U = QT, in which T denotes the transpose of a matrix, and R and S are obtained from Q and U as follows. On interchanging the 1st row and the 8th row of Q, the 2nd row and the 7th row of Q, etc., we get R. Similarly, we obtain S from U. Thus, we have

The size of this matrix is 16 x 16. This can be further

extended to a matrix L of size 32 x 32

where where H = ET, in which T denotes the transpose of a matrix, and F and G are obtained from E and H as follows. On interchanging the 1st row and the 16th row of E, the 2nd row and the 15th row of E, etc., we get F. Similarly, we obtain G from H. Thus, we have L. Here,

Now we obtain the key matrix K given by

where J is obtained from H by rotating, circularly, two rows in the downward direction.

The afore mentioned operations are performed for (1) enhancing the size of the key matrix to 32 x 32, and (2) obtaining the modular arithmetic inverse of K, in a trial

and error manner. The modular arithmetic inverse of K is obtained as


91

From (3.8) and (3.10), we can readily find that K K–1 mod 256 = K–1 K mod 256 = I. (3.15) Let us consider the image of a hand, which is given below.

Figure 2. Image of a Hand

This image can be represented in the form of a binary matrix P given by

where 1 denotes black and 0 denotes white.

On adopting the iterative procedure given in Fig. 1, we get the encrypted image C

On using (3.8), (3.10), (3.17), and the procedure for decryption (See Fig. 1.(b)), we get back the original binary image P, given by (3.16). From the matrix C, on connecting each 1 with its neighbouring 1, we get an image which is in a zigzag manner (See Fig. 3).

Figure 3. Encrypted image of the hand It is interesting to note that, the original image and the encrypted image differ totally, and the former one, exhibits all the features very clearly, while the later one does not reveal anything.

4. Conclusions In this analysis, we have made use of a modified Hill cipher for encrypting a binary image. Here we have illustrated the procedure by considering a pair of examples: (1) the image of a hand, and (2) the image of upper half of a person. Here, we have noticed that, the encrypted image is totally different from the original image, and the security of the image is completely enhanced, as no feature of the original image can be traced out in any way from the encrypted image. This analysis can be extended for the images of signatures and thumb impressions.

References [1] Hossam El-din H. Ahmed, Hamdy M. Kalash, and

Osama S. Farag Allah, “Encryption Efficiency Analysis and Security Evaluation of RC6 Block Cipher for


92

Digital Images”, International Journal of Computer, Information, and Systems Science, and Engineering, Vol. 1, No. 1, pp. 33 – 39, 2007.

[2] M. Zeghid, M. Machhout, L. Khriji, A. Baganne, and R. Tourki, “A Modified AES Based Algorithm for Image Encryption”, World Academy of Science, Engineering and Technology, Vol. 27, pp. 206 – 211, 2007.

[3] Bibhudendra Acharya, Saroj Kumar Panigrahy, Sarat Kumar Patra, and Ganapati Panda, “Image Encryption Using Advanced Hill Cipher Algorithm”, International Journal of Recent Trends in Engineering, Vol. 1, No. 1, May 2009.

[4] V. U. K. Sastry, D. S. R. Murthy, S. Durga Bhavani, “A Block Cipher Involving a Key Applied on Both the Sides of the Plain Text”, International Journal of Computer and Network Security (IJCNS), Vol. 1, No. 1, pp. 27 – 30, Oct. 2009.

[5] V. U. K. Sastry, D. S. R. Murthy, S. Durga Bhavani, “A Block Cipher Having a Key on One Side of the Plain Text Matrix and its Inverse on the Other Side”, Accepted for Publication in International Journal of

Computer Theory and Engineering (IJCTE), Vol. 2, No 5, Oct 2010.

[6] William Stallings, Cryptography and Network Security, Principles and Practice, Third Edition, Pearson, 2003.


93

Proxy Blind Signature based on ECDLP

Asis Kumar Tripathy1, Ipsita Patra2 and Debasish Jena3

1,2 Department of Computer Science and Engineering International Institute of Information Technology, Bhubaneswar 751 013, India

[email protected] , [email protected]

3Center For IT Education Biju Pattanaik University of Technology,Bhubaneswar, India

[email protected]

Abstract: Proxy blind signature is the combination of properties and behavior of two most common digital signature schemes i.e. proxy signature and blind signature. In this scheme proxy signer generates the blind signature on behalf of the original signer without knowing the content of the message. Proxy blind signature can be used in various applications such as e-voting, e-payment, mobile agent communication. In this paper cryptanalysis of DLP based proxy blind signature scheme with low computation by Aung et al. has been done. An efficient proxy blind signature scheme based on ECDLP has been proposed.

Keywords: Proxy Signature ,Blind signature,Proxy blind signature,ECDLP.

1. Introduction

1.1 Digital Signature A digital code that can be attached to an electronically transmitted that uniquely identifies the sender. Digital signatures are especially important for electronic commerce and are a key component of most authentication schemes. To be effective, digital signatures must be unforgeable. There are a number of different encryption techniques to guarantee this level of security.

1.2 Proxy signature The proxy signature scheme is a kind of digital signature scheme . In the proxy signature scheme , one user called the original signer ,can delegate his/her signing capability to another user called the proxy signer. This is similar to a person delegating his/her seal to another person in the real world.

1.3 Blind Signature The signer cannot determine which transformed message received for signing corresponds with which digital signature, even though the signer knows that such a correspondence must exist.

1.4 Proxy Blind Signature A proxy blind signature scheme is a digital signature scheme which combines the properties of proxy signature and blind signature schemes. A proxy blind signature scheme is a protocol played by two parties in which a user obtains a proxy signer’s signature for a desired message and the proxy signer learns nothing about the message. With such properties, the proxy blind signature scheme is useful in several applications such as e-voting, e-payment and mobile agent environments. In a proxy blind signature scheme, the proxy signer is allowed to generate a blind signature on behalf of the original signer.

1.5 Properties of the Proxy Blind Signature Scheme

In addition to the properties of Digital signature and proxy blind signature should satisfy the following properties. 1.Distinguish-ability: The proxy signature must be distinguishable from the normal signature. 2. Non-repudiation: Neither the original signer nor the proxy signer can sign message instead of the other party. Both the original signer and the proxy signer can not deny their signatures against anyone. 3. Unforgeability: Only a designated proxy signer can generate a valid proxy signature for the original signer (even the original signer cannot do it). 4. Verifiability: The receiver of the signature should be able to verify the proxy signature in a similar way to the verification of the original signature. 5. Identifiability: Anyone can determine the identity of the corresponding proxy signer from a proxy signature. 6. Prevention of misuse: It should be confident that proxy key pair should be used only for creating proxy signature, which conforms to delegation information. In case of any misuse of proxy key pair, the responsibility of proxy signer should be determined explicitly. 7. Unlinkability: When the signature is revealed, the proxy signer can not identify the association between the message and the blind signature he generated. When the signature is verified, the signer knows neither the message nor the signature associated with the signature scheme. In this paper, cryptanalysis of Aung et al[12] has been done and an efficient proxy digital signature based on ECDLP has


94

been proposed. The proposed scheme satisfies all the properties of a proxy blind signature scheme. The rest of this paper is organized as follows. In Section 2,some related work are discussed. Overview Of Aung et al.’s DLP Based Proxy Blind Signature scheme with Low Computation has been done in section 3 . In Section 4, Cryptanalysis of Aung et al.’s scheme has been done. An introduction to ECC is being described in section 5. In Section 6, Proposed scheme is being described. Security analysis of the proposed scheme has been done in section 7. In Section 8, Efficiency of the proposed scheme is being compared with the previous schemes and concluding remarks are being described in Section 9.

2. Related Work D. Chaum [4]introduced the concept of a blind signature scheme in 1982. In 1996 Mambo et al [2] introduced the concept of proxy signature. The two types of scheme: proxy unprotected (proxy and original signer both can generate a valid proxy signature) and proxy protected (only proxy can generate a valid proxy signature) ensures among other things, non-repudiation and unforgeability . The first proxy blind signature was proposed by Lin and Jan [1] in 2000. Recently Tan et al. [7] introduced a proxy blind signature scheme, which ensures security properties of the schemes, viz., the blind signature schemes and the proxy signature schemes. The scheme is based on Schnorr blind signature scheme Lee et al.[3] showed that a strong proxy signature scheme should have properties of strong unforgeability, verifiability, strong identifiability, strong nonrepudiation and prevention of misuse. Hung-Min Sun and Bin Tsan Hsieh [8] show that Tan et al.[6] schemes do not satisfy the unforgeability and unlinkability properties. In addition, they also point out that Lal and Awasthi [7] scheme does not possess the unlinkability property either. In 2004, Xue and Cao [9] showed there exists one weakness in Tan et al. scheme [5] and Lal et al. scheme [7] since the proxy signer can get the link between the blind message and the signature or plaintext with great probability. Xue and Cao introduced concept of strong unlinkability and they also proposed a proxy blind signature scheme. In 2007 Li et al.[10] proposed a proxy blind signature scheme using verifiable self-certified public key, and their scheme is more efficient than schemes published earlier. Recently, Xuang Yang and Zhaoping Yu[11] proposed new scheme and showed their scheme is more efficient than Li et al.[10]. In 2009 Aung et al.[12] proposed a new proxy blind signature scheme which satisfied all the security requirements of both the blind signature scheme and the proxy signature scheme.

3. Overview Of Aung et al.’s DLP Based Proxy Blind Signature scheme With Low Computation In this section,DLP Based Proxy Blind Signature scheme With Low Computation has been discussed.

3.1 Proxy Delegation Phase

Original signer selects random number and computes:

(1) (2)

sends along with the warrant to the proxy

signer. And then proxy signer checks:

(3) If it is correct, P accepts it and computes proxy signature secret key as follow:

(4) Note: responding proxy public key

3.2 Blind Signing Phase

Proxy signer selects random number and computes:

(5) and then sends to signature asker . To obtain the blind signature of message m, original signer randomly choose two random numbers and computes:

(6) (7)

(8) If =0 then has to select new tuple Otherwise sends to . After receiving proxy signer computes :

(9) and sends the sign message to .

3.3 Extraction Phase

While receiving , computes:

(10) Finally the signature of message is .


95

3.4 Verification Phase

The recipient of the signature can verify the proxy blind signature by checking whether

(11) Where If it is true, the verifier accepts it as a valid proxy blind signature, otherwise rejects.

Verifiability

The verifier can verify the proxy blind signature by checking

holds.

4. Cryptanalysis Of the Aung et al.’s Scheme The scheme doesn’t satisfy the property of verifiability as mentioned in Aung et al.[12]’s scheme. The verification steps are not correct.

5. Elliptic Curve Cryptography

5.1 Elliptic Curve over Finite Field

The elliptic curve operations defined on real numbers are slow and inaccurate due to round-off error. Cryptographic operations need to be faster and accurate. To make operations on elliptic curve accurate and more efficient, ECC is defined over two finite fields— prime field and binary field . The field is chosen with finitely large number of points suited for cryptographic operations.

5.1.1 EC over Prime Field

The equation of the elliptic curve on a prime field is

where, . Here, the elements of the finite field are integers between 0 and . All the operations such as addition, subtraction, division, and multiplication involves integers between 0 and . The prime number p is chosen such that there is finitely large number of points on the elliptic curve to make the cryptosystem secure. Standards for Efficient Cryptography (SEC) specifies curves with p ranging between 112-521 bits . The algebraic rules for point addition and point doubling can be adapted for elliptic curves over .

5.1.2 Point Addition Consider two distinct points J and K such that and Let where,

then,

(12) where, s is the slope of the line through J and K. If , then where, O is the point at infinity. If then, ; point doubling equations are used. Also,

5.1.3 Point Doubling Consider a point such that where, . Let

where, . Then,

(13)

where, s is the tangent at point and is one of the parameters chosen with the elliptic curve. If = 0 then, where, O is the point at infinity. 6. Proposed Scheme In this section, we propose an efficient proxy blind signature scheme based on ECC.The proposed scheme is divided into five phases: system parameters, proxy delegation, blind signing, signature extraction and signature verification.

6.1 System Parameters

The entities involved are three parties.We denote the x-coordinate of a point on the elliptic curve by . The scheme is constructed as follows. We make conventions that lowercases denote the elements in and capital letters denote the points in the curve . O: the original signer P: the proxy signer A: the signature asker : the original signer O's secret key : the original signer O's public key, : the proxy signer P's secret key : the proxy signer P's public key, : the designated proxy warrant which contains the identities information of the original signer and the proxy


96

signer, message type to be signed by the proxy signer, the delegation limits of authority, valid periods of delegation, etc. h(.) a secure one-way hash function. || the concatenation of strings.

6.2 Proxy Delegation Phase Original Signer randomly chooses , 1 < < n ; and computes: (14) (15) sends along with the warrant to the proxy signer. And then proxy signer checks: (16) If the equation holds, P computes proxy signature secret as follow: (17) and corresponding proxy public key 6.3 Blind Signing Phase Proxy Signer randomly choose , where 1 < < n and computes: (18) and then sends to signature asker A. To obtain the blind signature of message m, original signer O randomly choose two random numbers u,v and computes: (19) (20) (21) If then A has to select new tuple (u,v),Otherwise O sends e to P. After receiving e proxy signer P computes : (22) 6.4 Extraction Phase

While receiving , computes: (23) Finally the signature of message is 6.5 Verification Phase

The recipient of the signature can verify the proxy blind signature by checking whether (24) where If it is true, the verifier accepts it as a valid proxy blind signature, otherwise rejects.

7. Security Analysis of the proposed scheme

The Scheme has stronger security property because ECDLP is more difficult than DLP. While maintaining the security, the scheme requires less data size and less computations, so it is efficient. In this subsection , we show that our scheme satisfies all the security requirements of a strong proxy blind signature. Distinguish-ability: On the one hand, the proxy blind signature contains the warrant . On the other hand, anyone can verify the validity of the proxy blind signature, so he can easily distinguish the proxy blind signature from the normal signature. Nonrepudiation: The original signer does not obtain the proxy signer’s secret key and proxy signer does not obtain original signer’s secret key . Thus, neither the original signer nor the proxy signer can sign in place of the other party. At the same time, through the valid proxy blind signature, the verifier can confirm that the signature of the message has been entitled by the original signer, because the verifier must use the original signer’s public key during the verification. Likewise, the proxy signer cannot repudiate the signature. The scheme offers nonrepudiation property. Unforgeability: An adversary (including the original signer and the receiver) wants to impersonate the proxy signer to sign the message m. He can intercept the delegation information but he cannot obtain the proxy signature secret key . From Equation (15), we know that only the proxy signer holds the proxy signature secret key . Because of 1 < < n , the adversary can obtain the proper proxy signature secret key by guessing it with at most a probability .That is, anyone else (even the original signer and the receiver) can forge the proxy blind signature successfully with a probability . Verifiability: The proposed scheme satisfies the property of verifiability. The verifier can verify the proxy blind signature by checking holds. This is because:


97

= = Identifiability: The proxy blind signature contains the warrant . Moreover, in the verification equation which includes the original signer O’s public key and the proxy signer P’s public key . Hence, anyone can determine the identity of the corresponding proxy signer from a proxy signature. Prevention of misuse: The proposed scheme can prevent proxy key pair misuse because the warrant mw includes original signer and proxy signer identities information, message type to be signed by the proxy signer, delegation period, etc. With the proxy key, the proxy signer cannot sign messages that have not been authorized by the original signer. Unlinkability: During generation of the signature , the proxy signer has the view of transcripts .Since are specified by the original signer for all the signatures under the same delegation condition. The proxy unlinkability holds if and only if there is no conjunction between and . This is obvious from Equations (16)-(21). The value is only included in Equation (17) and connected to through Equation (18). For this, one must be able to compute R which is masked with two random numbers. Similarly, and may be associated with the signature through Equation (19) and (20) respectively. They fail again due to the random numbers. Even they are combined, the number of unknowns is still more than that of the equations. So, the proposed scheme provides indeed the proxy blindness property. 8. Efficiency of the proposed scheme The proposed scheme has been proved to be correct. It also satisfies the properties of proxy blind signature scheme, i.e, Distinguish-ability, Nonrepudiation,

Verifiability, unforgeability, Identifiability, Prevention of misuse and Unlinkability. The proposed scheme is compared with the known proxy blind signature scheme which are given below. Table 1: Comparative statement among proxy blind signature schemes Scheme Basis 80-Bit Security Strength interger fractorization problem IFP 1024-bit modulus discrete logarithm problem DLP |p| = 1024, |q| = 160 Proposed schemes ECDLP m= 163 bits 9. Conclusion We analyzed that Aung et al.’s DLP based proxy blind signature scheme with low computation does not satisfy the verifiability property of proxy blind signature scheme .Compared with Aung et al’s scheme , we present a more efficient and secure proxy blind signature scheme to overcome the pointed out drawback of the Aung et al’s scheme.We proved that our scheme is more efficient and secure than the previous schemes.

References [1] W. D. Lin, and J. K. Jan, "A security personal learning tools using a proxy blind signature scheme", Proc. Of International Conference on Chinese Language Computing, 2000, pp.273- 277. [2] M. Mambo, K. Usuda, and E. Okamoto, "Proxy signatures: Delegation of the power to sign messages", IEICE Transaction on Fundamentals, E79-A (1996), pp.1338-1353. [3] B. Lee, H. Kim, and K. Kim,"Strong proxy signature and its application", Australasian Conference on Information Security and Privacy(ACISP 2001), LNCS2119, Springer- Verlag, Sydney, 2001, pp.603-608. [4] D Chaum," Blind signature for untraceable payments, Advances in C ryptology", proceeding of CRYPTO 82, Springer- Verlag, New York, 1983, pp.199-203. [5] Z. W. Tan, Z. J. Liu, and C. M. Tang, "A proxy blind signature scheme based on DLP", Journal of Software, Vol14, No11, 2003, pp.1931-1935. [6] Tan Z. Liu Z. Tang C. "Digital proxy blind signature

schemes based on DLP and ECDLP". In MM Research Preprints, No. 21, MMRC, AMSS,Academia, Beijing, 2002, 212- 217. [7] Lal S. Awasthi A K. Proxy" Blind Signature Scheme". Journal of Information Science and Engineering. Cryptology ePrint Archive, Report 2003/072. Available at http://eprint.iacr.org/.


98

[8] Hung-Min Sun, Bin-Tsan Hsieh "On the Security of Some Proxy Blind Signature Schemes". Australasian InformationvSecurityWorkshop(AIS W2 004),Dunedin: Australian Computer Society press, 2004,75-78. [9] Q. S. Xue, and Z. F. Cao, "A new proxy blind signature scheme with warrant", IEEE Conference on Cybernetics and Intelligent Systems (CIS and RAM 2004), Singapore, 2004, pp.1385-1390. [10] J.G. Li, and S. H.Wang, "New Efficient Proxy Blind Signature Scheme Using Verifiable Self-certified Public Key", International Journal of Network Security, Vol.4, No.2, 2007, pp.193-200. [11] Xuan Yang, Zhaoping Yu , "Efficient Proxy Blind Signature Scheme based on DLP", International Conference on Embedded Software and Systems ( ICESS2008). [12] Aung Nway Oo and Nilar "The DLP based proxy blind signature scheme with low computation", 2009 Fifth International Joint Conference on INC, IMS and IDC. Authors Profile

Asis Kumar Tripathy received the B.E. degree in Information Technology from bialasore College of Engineering and Technology in 2006. He has joned the same college as Lecturer since 26.12.2006. Now, he is persuing his M.Tech degree at International Institute of Information Technology-Bhubaneswar,

Orissa, India. His research area of interest is Information Security

Ipsita Patra received the MCA degree from National Institute of Technology ,Rourkela . Now, she is persuing her M.Tech degree at International Institute of Information Technology-Bhubaneswar, Orissa, India. Her research area of interest are Information Security and Network security.

Debasish Jena was born in 18th December, 1968. He received his B Tech degree in Computer Science and Engineering, his Management Degree and his MTech Degree in 1991,1997 and 2002 respectively. He has joined Centre for IT Education as Assistant Professor since 01.02.2006. He has submitted his

thesis for Ph.D. at NIT, Rourkela on 5th April 2010. .His research areas of interest are Information Security , Web Engineering, Bio-Informatics and Database Engineering.


99

Comparative Study of Continuous Density Hidden Markov Models (CDHMM) and Artificial Neural

Network (ANN) In Infant’s Cry Classification System

Yousra Abdulaziz Mohammed, Sharifah Mumtazah Syed Ahmad

College of IT

Universiti Tenaga Nasional (UNITEN) Kajang, Malaysia

[email protected], [email protected] Abstract: This paper describes a comparative study between continuous Density hidden Markov model (CDHMM) and artificial neural network (ANN) on an automatic infant’s cries classification system which main task is to classify and differentiate between pain and non-pain cries belonging to infants. In this study, Mel Frequency Cepstral Coefficient (MFCC) and Linear Prediction Cepstral Coefficients (LPCC) are extracted from the audio samples of infant’s cries and are fed into the classification modules. Two well-known recognition engines, ANN and CDHMM, are conducted and compared. The ANN system (a feedforwaed multilayer perceptron network with backpropagation using scaled conjugate gradient learning algorithm) is applied. The novel continuous Hidden Markov Model classification system is trained based on Baum –Welch algorithm on a pair of local feature vectors. After optimizing system’s parameters by performing some preliminary experiments, CDHMM gives the best identification rate at 96.1%, which is much better than 79% of ANN whereby in general the system that are based on MFCC features performs better than the one that utilizes LPCC features.

Keywords: Artificial Neural Networks, Continuos Density

Hidden Markov Model; Mel Frequency Cepstral Coefficient, Linear Prediction Cepstral Coefficints, Infant Pain Cry Classification

1. Introduction Infants often use cries as communication tool to express their physical, emotional and psychological states and needs [1]. An infant may cry for a variety of reasons, and many scientists believe that there are different types of cries which reflects different states and needs of infants, thus it is possible to analyze and classify infant cries for clinical diagnosis purposes.

A number of research work related to this line have been reported, whereby many of which are based on Artificial Neural Network (ANN) classification techniques. Petroni and Malowany [4] for example, have used three different varieties of supervised ANN technique which include a simple feed-forward, a recurrent neural network (RNN) and a time-delay neural network (TDNN) in their infant cry classification system. In their study, they have attempted to to recognize and classify three categories of cry, namely ‘pain’, ‘fear’ and ‘hunger’ and the results demonstrated that

the highest classification rate was achieved by using feed-forward neural network. Another research work carried out by Cano [8] used the Kohonen's self organizing maps (SOM) which is basically a variety of unsupervised ANN technique to classify different infant cries. A hybrid approach that combines Fuzzy Logic and neural network has also been applied in the similar domain [7].

Apart from the traditional ANN approach, other infant cry classification technique studied is Support Vector Machine (SVM) which has been reported by Barajas and Reyes [2]. Here, a set of Mel Frequency Cepstral Coefficients (MFCC) was extracted from the audio samples as the input features. On the other hand, Orozco and Garcia [6], use the linear prediction technique to extract the acoustic features from the cry samples of which are then fed into a feedforward neural network recognition module.

Hidden Markov Model is based on double stochastic processes, whereby the first process produces a set of observations which in turns can be used indirectly to reveal another hidden process that describes the states evolution [13]. This technique has been used extensively to analyze audio signals such as for biomedical signal processing [17] and speech recognition [18]. The prime objective of this paper is to compare the performance of an automatic infant’s cry classification system applying two different classification techniques, Artificial Neural Networks and continuous Hidden Markov Model.

Here, a series of observable feature vector is used to reveal the cry model hence assists in its classification. First, the paper describes the overall architecture of an automatic recognition system which main task is to differentiate between an infant ‘pain’ cries from ‘non-pain’ cries. The performance of both systems is compared in terms of recognition accuracy, classification error rate and F-measure under the use of two different acoustic features, namely Mel Frequency Cepstral Coefficient (MFCC) and Linear Prediction Cepstral Coefficients (LPCC). Separate phases of system training and system testing are carried out on two different sample sets of infant cries recorded from a group of babies which ranges from newborns up to 12 months old.


100

The organization of the rest of this paper is as follows. Section 2 describes the nature of datasets used in the study. Section 3 presents an overall architecture of the recognition system, in terms of the features extracted and the classification techniques. On the other hand, Section 4 and Section 5, details out the descriptions of the experimental set-ups and the results of the findings respectively. Section 5 concludes the paper and highlights potential area for future works.

2. Experimental Data Sets The infant cry corpus collected is a set of 150 pain samples and 30 non-pain samples recorded from a random time interval. The babies selected for recording varies from newborns up to 12 month old, a mixture of both healthy males and females. The records are then sampled at 16000 Hertz. It is important to highlight here that the pain cry episodes are the result of the pain stimulus carried out during routine immunization at a local pediatric clinic, in Darnah, Libya. Recordings resulting from anger, or hunger were considered as non-pain utterances were recorded at quiet rooms at various infants’ home. Recordings were made on a digital player at a sample rate of 8000 Hertz and 4-bit resolution with microphone placed between 10 to 30 centimeters away from the infant’s mouth. The audio signals were then transferred for analysis to a sound editor and then re-sampled at 16000 with 16 bit resolution [3, 4, 5]. The final digital recordings were stored as WAV files.

3. System Overview In this paper, we consider two recognition engines for

infant’s cry classification system. The first is an artificial neural network (ANN) which is a very popular pattern matching in the field of speech recognition. Feedforward multilayer perceptron (MLP) network with a backpropagation learning algorithm is the most well known of ANN. The other technique is a continuous density hidden Markov model (CDHMM). The overall system is as depicted in Figure 1 below:

Figure 1. Infant Cry Recognition System Architecture

From Figure 1, we can say that infant cry recognition system implies three main tasks,

• Signal Preprocessing • Feature Extraction • Pattern Matching

3.1 Pre-Processing The first step in the proposed system is the pre-processing step which requires the removal of the ‘silent’ periods from the recorded sample. Recordings with cry units lasting at least 1 second from the moment of stimulus event were used for the study [4, 6]. The cry units are defined as the duration of the vocalization only during expiration [5].

The audio recordings were then divided further into segments of exactly 1 second length, where each represents a pre-processed cry segments as recommended by [2, 4, 6, 8]. Before these one second segments can be used for feature extraction, a process called pre-emphasis is applied. Pre-emphasis aims at reducing the high spectral dynamic range, and is accomplished by passing the signal through an FIR filter whose transfer function is given by. (1) A typical value for the pre-emphasis parameter ‘ ’ is usually 0.97. Consequently the output is formed as follow:

(2)

where s(n) is the input signal and y(n) is the output signal from the first order FIR filter. Every segment of 1 second is divided thereafter in frames of 50-milliseconds with successive frames overlapping by 50% from each other. The next step is to use a window function on each individual frame in order to minimize discontinuities at the beginning and end of each frame. Typically the window function used is the Hamming window and has the following form:

, (3)

Given the above window function and assuming that there are N samples in each frame, we will obtain the following signal after windowing.

, 0≤ ≤ (N-1) (4)

Figure 2. Hamming Window

Off the 150 pain and 30 non-pain recording samples, 625 and 256 one second cry segments were obtained respectively. Off these 881 cry segments, 700 were used for system training and 181 were used for system testing. It is important to use a separate set of cry segments for training and testing purposes in order to avoid obtaining biased testing results.


101

3.2 Feature Extraction 3.2.1 Mel-Frequency Cepstral Coefficients (MFCC)

MFCCs are one of the more popular parameter used by researchers in the acoustic research domain. It has the benefit that it is capable of capturing the important characteristics of audio signals. Cepstral analysis calculates the inverse Fourier transform of the logarithm of the power spectrum of the cry signal, the calculation of the mel cepstral coefficients is illustrated in Figure 3.

Figure 3. Extraction of MFCC from Audio Signals

The cry signal must be divided into overlapping blocks first, (in our experiment the blocks are in Hamming windows) which is then transformed into its power spectrum. Because human perception of the frequency contents of sounds does not follow a linear scale. They are approximately linear with logarithmic frequency beyond about 1000 Hz. The mel frequency warping is most conveniently done by utilizing a filter bank with filters centered according to mel frequencies as shown in Figure 4.

Figure 4. Mel Spaced Filter Banks

The mapping is usually done using an approximation (where fmel is the perceived frequency in mels),

)700

1log(2595)( fffmel +×= (5)

These vectors were normalized so that all the values within a given 1 second sample would lie between ±1 in order to decrease their dynamic range [4].

3.2.2 Linear Prediction Cepstral Coefficients (LPCC)

LPCC is Linear Predicted Coefficients (LPC) in the cepstrum domain. The basis of linear prediction analysis is that a given speech sample can be approximated with a linear combination of the past p speech samples [9]. This can be calculated either by the autocorrelation or covariance methods directly from the windowed portion of audio signal [10]. In this study the LPC coefficients were calculated using the autocorrelation method that uses Levinson-Durbin recursion algorithm. Hence, LPCC can be derived from LPC using the recursion as follows:

)0(0 rc = (6)

where r is derived from the LPC autocorrelation matrix

kmk

m

kmm ac

nkac −

−

=∑+=

1

1

)( for Pm <<1 (7)

kmk

m

pmkm ac

nkc −

−

−=∑=

1

)( for Pm > (8)

where p is the so called prediction order, ma represents the thm LPC coefficient and is the number of LPCC’s

needed to be calculated. Whereas mc is the thm LPCC. 3.2.3 DELTA coefficients

It has been proved that system performance may be enhanced by adding time derivatives to the static parameters [10]. The first order derivatives are referred to as delta features and can be calculated as shown in formula 9,

(9) where d, is the delta coefficient at time t, computed in terms of the corresponding static coefficients , to , and is the size of delta window.

3.3 Patter Matching 3.3.1 Artificial Neural Network (ANN) Approach

Neural Networks are defined as systems which has the capability to model highly complex nonlinear problems and composed of many simple processing elements, that operate in parallel and whose function is determined by the network's structure, the strength of its connections, and the processing carried out by the processing elements or nodes. The feedforward multilayer perceptron (MLP) network architecture using a backpropagation learning algorithm is one of the most popular neural networks. It consists of at least three layers of neurons: an input layer, one or more hidden layers and an output layer. Processing elements or neurons in the input layer only act as buffers for distributing the input signal xi to neurons in the hidden layer. The hidden and output layers have a non-linear activation function. A backpropagation is a supervised learning algorithm to calculate the change of weights in the network. In the forward pass, the weights are fixed and the input vector is propagated through the network to produce an output. An output error is calculated from the difference between actual output and the target. This is propagated backwards through the network to make changes to the weights [19]. For this work, a feed-forward multilayer perceptron using full connections between adjacent layers was trained and tested with input patterns described above in a supervised manner with scaled conjugate gradient back-propagation learning algorithm since it has shown good results in classifying infant’s cries than other NN algorithms [3, 20]. The number of computations in each iteration is significantly reduced mainly because no line search is required. Different feature sets were used in order to


102

determine the set that results in optimum recognition rate. Sets of 12 MFCC, 12MFCC+1ST derivative, 12 MFCC+1st & 2nd derivative, 20 MFCC, 16 LPCC and 16 LPCC+ 1st derivative were used. Two frame length was used, 50ms and 100 ms in order to determine the right choice that gives the best results. Two architectures were investigated in this study. First, with one hidden layer then with two hidden layers. The number of hidden neurons was varied to obtain optimum performance. The number of neurons in the input layer is decided by the number of elements in the feature vector. Output layer has two neurons each for one cry class The activation function used in all layers in this work is hyperbolic tangent sigmoid transfer function ‘TANSIG’. Training stops when any of these conditions occur:

• The maximum number of epochs (repetitions) is reached. we established 500 epochs at maximum because above this value, the convergence line do not have any significant change.

• The networks were trained until the performance is minimized to the goal i.e. mean squared error is less than 0.00001.

Figure 5. MLP Neural Network

3.3.2 Hidden Markov Model (HMM) approach The continuous HMM is chosen over the discrete counterparts since it avoids losing of critical signal information during discrete symbol quantization process and that it provides for better modeling of continuous signal representation such as the audio cry signal [12]. However, computational complexity when using the CHMMs is more than the computational complexity when using DHMMs. It normally takes more time in the training phase[11] An HMM is specified by the following:

• N, the number of states in the HMM; • ), the prior probability of state si being the first state of a state sequence. The collection of π1 forms the vector π = { π1, ….., πN}; • , the transition coefficients gives the probability of going from state si immediately to state sj. The collection of a ij forms the transition matrix A. • Emission probability of a certain observation o, when the model is in state . The observation o can be either discrete or continuous [11]. However, in this study a continuous HMM is applied whereby continuous observations o | = indicates the probability density function (pdf) over the observation space for the model being in state .

For the continuous HMM, the observations are continuous and the output likelihood of a HMM for a given observation

vector, can be expressed as a weighted sum of M multivariate Gaussian probability densities [15] as given by the equation

for 1≤ j ≤ N (10)

where ot is a d-dimensional feature vector, M is the number of Gaussian mixture components m is the mixture index (1≤ m ≤ M), Cjm is the mixture weight for the mth component, which satisfies the constraint • Cjm > 0 and

• ∑ ==M

m jmC1

1 , 1≤ j ≤ N , 1≤ m ≤ M

where N is the number of states, j is the state index, μjm is the mean vectors of the mth Gaussian probability densities function, and א is the most efficient density functions widely used without loss of generality[11, 15] which is defined as, (11)

In the training phase, and in order to derive the models for both ‘pain’ and’non-pain’ cries (i.e. to derive λpain and λnon-pain models respectively) we first have to make a rough guess about the parameters of an HMM, and based on these initial parameters more accurate parameters can be found by applying the Baum-Welch algorithm. This mainly requires for the learning problem of HMM as highlighted by Rabiner in his HMM tutorial [13]. The re-estimation procedure is sensitive to the selection of initial parameters. The model topology is specified by an initial transition matrix. The state means and variances can be initialized by clustering the training data into as many clusters as there are states in the model with the K-means clustering algorithm and estimating the initial parameters from these clusters [16]. The basic idea behind the Baum-Welch algorithm (also known as Forward-Backward algorithm) is to iteratively re-estimate the parameters of a model, and to obtain a new model with a better set of parameters ¸ which satisfies the following criterion for the observation sequence

, (12) where the given parameters are¸ . By setting, ¸ at the end of every iteration and re-estimating a better parameter set, the probability of can be improved until some threshold is reached. The re-estimation procedure is guaranteed to find in a local optimum. The flow chart of the training procedure is shown in Figure 6,


103

Figure 6. Training Flow Chart

Once the system training is completed, system testing is carried out to investigate the accuracy of the recognition system. The classification is done with a maximum-likelihood classifier, that is the model with the highest probability with respect to the observations sequence, i.e., the one that maximizes P(Model | Observations) will be the natural choice,

(13)

This is called the maximum likelihood (ML) estimate. The best model in maximum likelihood sense is therefore the one that is most probable to generate the given observations. In this study, separate untrained samples from each class where fed into the HMM classifier and were compared against the trained ‘pain’ and ‘non-pain’ model. This testing process mainly requires for evaluation problem of HMM as highlighted by Rabiner [13]. The classification follows the following algorithm:

If P(test_sample| λpain) > P(test_sample| λnon-pain)

Then test_sample is classified ‘pain’

Else

test_sample is classified ‘non_pain’ (14)

4. Experimental Results A total of 700 cry segments of 1 second duration is used

during system testing. Each feature vector is extracted at 50ms windows with 75% overlap between adjacent frames. The size of the input data frame used was 50ms and 100ms in order to determine which would yield the best results for this application. These settings were used in all experiments reported in this paper which compare between the performance of both types of infant cry recognition systems described above utilizing 12 MFCCs (with 26 filter banks) and 16LPCCs (with 12th order LPC) acoustic features. The effect of the dynamic coefficients for both used features was also investigated. After optimizing system’s parameters by performing some preliminary experiments, it was found that a fully connected (an ergodic) HMM topology with five states and eight

Gaussian mixture per state is the best choice to model the cry utterances because it scored the best identification rate. On the other hand, a fully connected feedforward MLP NN trained with backpropagation with Scaled Conjugate Gradient algorithm, with one hidden layer and five hidden neurons has given the best recognition rates. The best results for both systems were obtained using 50 ms frame length. The results of training and testing datasets are evaluated using standard performance measures defined as follows:

• The classification accuracy which is calculated by taking the percentage number of correctly classified test samples.

System Accuracy (%) = %100×T

xcorrect (15)

where xcorrect is the total number of correctly classified test samples and T is the overall total number of test samples.

• The classification error rate (CER) which is defined as follows [14]:

(16) • F- measure which is defined as follows [14],

(17)

(18)

(19) F-measure varies from 0 to 1, with a higher F-measure indicating better performance [14]. where tp is the number of pain test samples successfully classified, fp is the number of misclassified non-pain test samples, fn is the number of misclassified pain test samples, and finally tn the number of correctly classified non-pain test samples. Both systems performed optimally with 20 MFCC’s 50 ms window. For the NN based system the hierarchy of one hidden layer having 5 hidden nodes showed to be the best, while an ergodic HMM with 5 states and 8 Gaussians per state has resulted in the best recognition rates. The optimum recognition rates obtained were 96.1% for HMM trained with 20 MFCC, whereas for ANN the highest recognition rate was 79% using 20 MFCC also. For both systems trained with LPCC’s, the best recognition rate obtained for HMM was 78.5% using 16 LPCC+DEL, 10 Gaussians , 5 states and 50 ms whereas for ANN was 70.2% using 16 LPCC, 5 states, 8 Gaussians per state and 50 ms window.

Table I, Figures 7 and 8 summarizes a comparison between performance of both systems using different performance metrics for the optimum results obtained with both MFCC and LPCC features.


104

Table 1: Overall System Accuracy

Features

Performance Measures

MFCC LPCC

ANN HMM ANN HMM

F-measure 0.83 0.97 0.75 0.84

CER% 20.99 3.87 29.83 21.55

Accuracy% 79 96.1 70.2 78.5

1 2 3

0

10

20

30

40

50

60

70

80

90

100

F-measure CER Accuracy

Comparison between NN & HMM using MFCC

ANNHMM

FIGURE 7. COMPARISON BETWEEN ANN AND HMM

USING MFCC

1 2 3

0

10

20

30

40

50

60

70

80

90

F-measure CER Accuracy

Com parison between NN & HMM us ing LPCC

ANNHMM

Figure 8. Comparison between ANN and HMM using LPCC

5. Conclusion and Future Work In this paper we applied both of ANN and CDHMM to

identify infant pain and non-pain cries . From the obtained results, it is clear that HMM has been proved to be the superior classification technique in the task of recognition and discrimination between infants’ pain and non-pain cries with 96.1% classification rate with 20 MFCC’s extracted at 50ms. From our experiments, the optimal parameters of CHMM developed for this work were 5 states , 8 Gaussians per state, whereas the ANN architecture that yielded the highest recognition rates was with one hidden layer and 5 hidden nodes. The results also show that in general the system accuracy performs better with MFCC's rather than with LPCC’s features.

References [1] J.E. Drummond, M.L. McBride, “The Development of

Mother’s Understanding of Infant Crying,”, Clinical Nursing Research, Vol 2, pp 396 – 441, 1993.

[2] Sandra E. Barajas-Montiel, Carlos A. Reyes-Garcia, “Identifying Pain and Hunger in Infant Cry with Classifiers Ensembles”, Proceedings of the 2005 International Conference on Computational Intelligence for Modeling, Control and Automation, and International Conference on Intelligent Agents,

Web Technologies and Internet Commerce, Vol. 2, pp 770 – 775, 2005

[3] Yousra Abdulaziz, Sharrifah Mumtazah Syed Ahmad, “Infant Cry Recognition System: A Comparison of System Performance based on Mel Frequency and Linear Prediction Cepstral Coefficients”, Proceedings of International Conference on Information Retrieval and Knowledge Management (CAMP 10), Selangor, Malaysia, PP 260-263, March, 2010.

[4] M. Petroni, A.S. Malowany, C.C. Johnston, B.J. Stevens, “A Comparison of Neural Network Architectures for the Classification of Three Types of Infant Cry Vocalizations”, Proceedings of the IEEE 17th Annual Conference in Engineering in Medicine and Biology Society, Vol. 1, pp 821 – 822, 1995

[5] H.E. Baeck, M.N. Souza, ”Study of Acoustic Features of Newborn Cries that Correlate with the context”, Proceedings of the IEEE 23rd Annual Conference in Engineering in Medicine and Biology , Vol. 3, pp 2174 – 2177, 2001.

[6] Jose Orozco Garcia, Carlos A. Reyes Garcia, “Mel-Frequency Cepstrum Coefficients Extraction from Infant Cry for Classification of Normal and pathological cry with Feed- forward Neural Networks”, proceedings of international joint conference on neural Networks, Vol. 4, pp 3140-3145, 2003

[7] Cano-Ortiz, D. I. Escobedo-Becerro,” Classificacion de Unidades de Llanto Infantil Mediante el Mapa Auto- Organizado de Koheen”, I Taller AIRENE sobre Reconocimiento de Patrones con Redes Neuronales, Universidad Católica del Norte, Chile, pp. 24-29, 1999.

[8] Suaste, I., Reyes, O.F., Diaz, A., Reyes, C.A., “Implementation of a Linguistic Fuzzy Relational Neural Network for Detecting Pathologies by Infant Cry Recognition,” Advances of Artificial Intelligence - IBERAMIA,Vol. 3315, pp. 953-962, 2004

[9] Jose Orozco, Carlos A. Reyes-Garcia, “Implementation and Analysis of Training Algorithms for the Classification of Infant Cry with Feed-forward Neural Networks”, in IEEE International Symp. Intelligent Signal Processing, pp. 271 – 276, Puebla, Mexico, 2003.

[10] Eddie Wong, Sridha Sridharan, “Comparison of Linear Prediction Cepstrum Coefficients and Mel-Frequency Cepstrum Coefficientsfor Language Identification”, in Proc. Int. Symp. Intelligent Multimedia, Video and Speech Processing, Hong Kong May 24, 2001.

[11] Hesham Tolba,” Comparative Experiments to Evaluate the Use of a CHMM-Based Speaker Identification Engine for Arabic Spontaneous Speech”, 2nd IEEE International Conference on Computer Science and Information Technology, ICCSIT, pp. 241 – 245, 2009.

[12] Joseph Picone, “Continuous Speech Recognition Using Hidden Markov Models”, IEEE ASSP magazine, pp.26-41, 1991.

[13] L.R. Rabiner, "A tutorial on hidden markov models and selected applications in speech recognition".


105

Proceedings of the IEEE, Vol. 77 (2, pp. 257-286), 1989.

[14] Yang Liu, Elizabeth Shriberg, “Comparing Evaluation Metrics For Sentence Boundary Detection”, in Proc. IEEE Int. Conf. Vol. 4, pp185-188, 2007.

[15] Jun Cai, Ghazi Bouselmi, Yves Laprie, Jean-Paul Haton, “Efficient likelihood evaluation and dynamic Gaussian selection for HMM-based speech recognition”, Computer Speech and Language, vol.23, pp. 47–164, 2009.

[16] B. Resch, “Hidden Markov Models, A tutorial of the course computational intelligence”, Signal Processing and Speech Communication Laboratory.

[17] Dror Lederman, Arnon Cohen, Ehud Zmora, “On the use of Hidden Markov Models in Infants' Cry Classification”, 22nd IEEE Convention, pp 350 – 352, 2002.

[18] Mohamad Adnan Al-Alaoui, Lina Al-Kanj, Jimmy Azar, Elias Yaacoub, “Speech Recognition Using Artificial Neural Networks And Hidden Markov Models”, The 3rd International Conference on Mobile and Computer Aided Learning, IMCL Conference, Amman, Jordan, 16-18 April 2008.

[19] Sawit Kasuriya, Chai Wutiwiwatchai, Chularat Tanprasert, “Comparative Study of Continuous Hidden Markov Models (CHMM) and Artificial Neural Network (ANN) on Speaker Identification System”, International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems (IJUFKS), Vol. 9, Issue: 6, pp. 673-683, 2001.

[20] José Orozco, Carlos A. Reyes García, "Detecting Pathologies from Infant Cry Applying Scaled Conjugate Gradient Neural Networks”, Proceedings of European Symposium on Artificial Neural Networks, pp 349 – 354, 2003.


106

Scalable ACC-DCT Based Video Compression Using Up/Down Sampling Method

1G.Suresh, 2P.Epsiba, 3Dr.M.Rajaram, 4Dr.S.N.Sivanandam

1Research Scholar, CAHCET, Vellore, India.

[email protected]

2Lecturer, Department of ECE, CAHCET, Vellore, India. [email protected]

3Professor & Head, Department of EEE, GCE, Thirunelveli, India.

[email protected]

4Professor & Head, Department of CSE, PSG Tech, Coimbatore, India.

Abstract: In this paper, we propose a Scalable ACC-DCT based video compression using UP/DOWN sampling approach which tends to hard exploit the pertinent temporal redundancy in the video frames to improve compression efficiency with less processing complexity. Generally, video signal has high temporal redundancies due to the high correlation between successive frames. Actually, this redundancy has not been exposed enough by current video compression techniques. Our model consists on 3D to 2D transformation of the video frames that allows exploring the temporal redundancy of the video using 2D transforms and avoiding the computationally demanding motion compensation step. This transformation turns the spatial temporal correlation of the video into high spatial correlation. Indeed, this technique transforms each group of pictures (GOP) to one picture (Accordion Representation) eventually with high spatial correlation. This model is also incorporated with up/down sampling method (SVC) which is based on a combination of the forward and backward type discrete cosine transform (DCT) coefficients. As this kernel has various symmetries for efficient computation, a fast algorithm of DCT-based Scalability concept is also proposed. For further improvement of the scalable performance, an adaptive filtering method is introduced, which applies different weighting parameters to DCT coefficients. Thus, the decorrelation of the resulting pictures by the DCT makes efficient energy compaction, and therefore produces a high video compression ratio. Many experimental tests had been conducted to prove the method efficiency especially in high bit rate and with slow motion video. The proposed method seems to be well suitable for video surveillance applications and for embedded video compression systems. Keywords: SVC, Group of Pictures (GOP), ACC-DCT, Up/Down Sampling, ACC-DCT. 1. Introduction The main objective of video coding in most video applications is to reduce the amount of video data for storing or transmission purposes without affecting the visual quality. The desired video performances depend on applications requirements, in terms of quality, disks capacity and bandwidth. For portable digital video applications, highly-integrated real-time video compression and decompression solutions are more and more required. Actually, motion estimation based encoders are the most

widely used in video compression. Such encoder exploits inter frame correlation to provide more efficient compression. However, Motion estimation process is computationally intensive; its real time implementation is difficult and costly [1][2]. This is why motion-based video coding standard MPEG[12] was primarily developed for stored video applications, where the encoding process is typically carried out off-line on powerful computers. So it is less appropriate to be implemented as a real-time compression process for a portable recording or communication device (video surveillance camera and fully digital video cameras). In these applications, efficient low cost/complexity implementation is the most critical issue. Thus, researches turned towards the design of new coders more adapted to new video applications requirements. This led some researchers to look for the exploitation of 3D transforms in order to exploit temporal redundancy. Coder based on 3D transform produces video compression ratio which is close to the motion estimation based coding one with less complex processing [3][4][5][6]. The 3d transform based video compression methods treat the redundancies in the 3D video signal in the same way, which can reduce the efficiency of these methods as pixel's values variation in spatial or temporal dimensions is not uniform and so, redundancy has not the same pertinence. Often the temporal redundancies are more relevant than spatial one [3]. It is possible to achieve more efficient compression by exploiting more and more the redundancies in the temporal domain; this is the basic purpose of the proposed method. The proposed method consists on projecting temporal redundancy of each group of pictures into spatial domain to be combined with spatial redundancy in one representation with high spatial correlation. The obtained representation will be compressed as still image with JPEG coder. The rest of the paper is organized as follows: Section 2 gives an overview of basic definition of three dimensional DCT. Section 3 gives the basics of the proposed method and the modifications made to improve the compression ratio and also reduce the complexity. Experimental results were discussed in section 4. The section 5 concludes this paper with a short summary.


107

2. Definitions 2.1. Three dimensional DCT The discrete cosine transform (DCT)[4][7] has energy packing efficiency close to that of the optimal Karhunen-Loeve transform. In addition, it is signal independent and can be computed efficiently by fast algorithms. For these reasons, the DCT is widely used in image and video compression. Since the common three-dimensional DCT kernel is separable, the 3D DCT is usually obtained by applying the one-dimensional DCT along each of the three dimensions. Thus, the N ×N ×N 3D DCT can be defined as

3. Proposed Method The fundamental idea is to represent a video sequence with highly correlated form. Thus we need to expose both spatial and temporal redundancy in video signal. The video cube is the input of our encoder, which is a number of frames. This video cube will decomposed in to temporal frames which will be gathered into one 2D frame. The next step consists of coding the obtained frame. Normally, the variation of the 3D video signal is much less in the temporal domain than the spatial domain, the pixels in 3D video signal are more correlated in temporal domain[3]. For a single pixel model is denoted as p(x,y,t) where p is pixel value; x,y are pixel spatial coordinates; t is video instance at time. The following assumption will be the basis of the proposed model where we will try to put pixels- which have a very high temporal correlation in spatial adjacency. P(x,y,t)-p(x,y,t+1)<p(x,y,t)-p(x+1,y,t) (3) To exploit the succeeding assumption the temporal decomposition of the 3D video signal will be carried out and the temporal, spatial decomposition of one 8x8x8 video cube [8][9][10] is presented in the Figure.1. Thus the Accordion representation (Spatial Adjacency) is obtained from the basis assumption.

Figure 1. Temporal and spatial decomposition of one 8x8x8

video cube.

Accordion representation is formed by collecting the video cube pixels which have the same column rank and these frames have a stronger correlation compare to spatial frames. To improve correlation in the representation we reverse the direction of event frames. This tends to put in spatial adjacency that the pixels having the same coordinate in the different frames of the video cube. The following example i.e., Figure.2 clearly projecting the Accordion representation also minimizes the distance between the pixel correlated in the source.

Figure 2. Accordion Representation Example Continuation of the Accordion Representation a new concept is originated from scalable video coding (SVC)[16][17][18]19[20] technique; up/down sampling method using the DCT has a large degree of symmetries for efficient computation. Thus, a fast algorithm of the up/down sampling method is also included in our proposed method. For a performance improvement, an adaptive filtering method DCT up/down sampling is applied[13][14][15], which applies different weighting parameters to each DCT coefficient. Then we have to introduce quantization model and Entropy coding (RLE/Huffman) techniques for further performance improvement of the proposed system. A overall constructional details of the proposed model is explained in Figure.3.

Figure 3. Complete Constructional Details of the proposed

Model


108

3.1. Algorithm 1. Decomposition of the video in GOPs (Group of frames/pictures) 2. (a) Spatial Adjacency separation (Accordion Representation) of the GOP For x=0: (L * N)-1 do For y=0:(H-1) do If ((x/N)mod 2)!=0 then n=(N-1)-(x mod N) else n=x mod N end if IACC (x,y)=In (x/N, y) With n= ((x/N) mod2)(N-1)+1-2((x/N) mod 2)(x mod N) (b) For n=0:N-1 do For x=0:L-1 do For y=0:H-1 do

If(x mod 2)! =0 then X ACC= (N-1)-n(x*N) else X ACC=n(x*N) end if In(x,y)=IACC(X ACC, y) end for end for end for

with XACC=(( x/N) mod2)(N-1)+n(1-2(x/N) mod2))+x 3. Decomposition of the resulting frame into 8x8 blocks. 4. Introduce down sampling filter/ up sampling filter with DCT. 5. Quantization of the obtained coefficients. 6. ZigZag coding of the obtained coefficient. 7. Entropy (Huffman) coding of the coefficients. 4. Experimental Results This section verifies the performance of the proposed low complex scalable ACC-DCT based video compression model. We summarize the experimental results with some analysis and comments. By understanding the performance of the proposed method with different GOP value the best compression rate is obtained with GOP=8. Here Figure.4 GUI model is created to integrate the encoder and decoder sections.Figure.5 shows the progress of frame separation from video sequence. Encoder model is shown in Figure.6 and Figure.7 is the example Accordion representation for one GOP video cube. Figure.8: GUI for Decoding and Validation Process. Figure.9: GUI for Reconstructed output validation, then history of entire simulation is specified as ans. Finally Figure.10 shows the plot between Frame number Vs PSNR(dB), Figure.11 finds the orientation flow estimation of the sample sequence and compared the strength of our proposed model with other leading standards shown in Figure.11.

Figure 4. GUI Model

Figure 5. Frame Separation Model

Figure 6. Encoding Model


109

Figure 7. Accordion Representation example (Hall Monitor)

Figure 8. GUI for Decoding and Validation Process

Figure 9. GUI for Reconstructed output validation ans = hObject: 4.0011

eventdata: [] handles: [1x1 struct] q: 1 str1: 'frame' str2: '.bmp' Bitstream: {[51782x1 double]} Bitst: 51782 j1: 2 f: 1 filename_1: '1.bmp' Image1: [120x960 double] row: 244 col: 356 out: [120x960 double] Enc: [120x960 double] r: 120 c: 960 Input_filesize: 921600 i: 120 j: 960 QEnc: [120x960 double] ZQEnc: [1x115200 double] Level: 8 Speed: 0 xC: {[1x115200 double]} y: [51782x1 double] Res: [2x4 double] cs: 4 cc: 51782 dd: 51782 Compresed_file_size: 51782 Comp_RATIO: 71.1908 enctime: 345.6888

0 10 20 30 40 50 60 70 80 90 10031

32

33

34

35

36

37

38

Frame Number

PS

NR

(dB

)

PROPOSED METHODMPEG4

Figure 10. Frame number Vs PSNR(dB)(Hall Monitor)


110

Figure 11. Orientation Flow Estimation

(a)

(b) Figure 12. Motion analysis of the video sample (a) Speed

analysis (b) Histogram plot

Motion analysis like vector movement, frame block variation with time, speed analysis and histogram analysis etc are plotted and presented in figure12. To conclude the performance betterment of our proposed method PSNR value is calculated for different bit rate, and then compared with other popular methods shown in figure 13. For comparison with representative lossy compression algorithms, we have two performance measures were used as described below in equ 4 and equ 5.

Figure 13. comparison response between bit rate Vs PSNR

(dB) for different standard

Where Tc is the time consumed to compress one frame. TABLE I show compression ratio and speed for total of ninety six input video sequences. The proposed scheme shows a bit rate savings of an average of 3.19% over other schemes and the average compression ratio is effectively increased compared to other methods.

Table 1: Comparison of Compression Ratio and Speed Methods

CR% Average worst

CS(MB/sec)

Previous

66.96 40.96

37.56

Proposed

71.1908 41.47

36.18


111

5. Conclusion In this paper, we successfully extended and implemented a Scalable ACC-DCT based video compression using up/down sampling algorithm on MATLAB and provided experimental results to show that our method is better than the existing methods. We not only improved the coding efficiency in the proposed encoding algorithm but also it reduces complexity. As discussed in the experimental section, proposed method provides benefits of rate-PSNR performance at the good quality of base layer and low quality of enhancement layer. When SVC coding scenario meets these circumstances, proposed method should be useful. With the apparent gains in compression efficiency we foresee that the proposed method could open new horizons in video compression domain; it strongly exploits temporal redundancy with the minimum of processing complexity which facilitates its implementation in video embedded systems. It presents some useful functions and features which can be exploited in some domains as video surveillance. In high bit rate, it gives the best compromise between quality and complexity. It provides better performance than MJPEG and MJPEG2000 almost in different bit rate values. Over 2000kb/s bit rate values; our compression method performance becomes comparable to the MPEG 4 standard especially for low motion sequences.

Additionally, a further development of this model could be to combine ‘Accordion representation’ with other transformations such as wavelet Transformation and also aims to estimate the degraded model of the video compressed image through using the neural network. We can improve the quality of the compressed image, using neural network’s predictive ability strongly and high fault-tolerance improving the quantization parameter prediction in the video encoding, completing video image reconstruction.

References [1] E. Q. L. X. Zhou and Y. Chen, .Implementation of

h.264 decoder on general purpose processors with media instructions,. in SPIE Conf. on Image and Video Communications and Processing, (Santa Clara, CA), pp. 224.235, Jan 2003.

[2] M. B. T. Q. N. A. Molino, F. Vacca, .Low complexity video codec for mobile video conferencing,. in Eur. Signal Processing Conf. (EUSIPCO),(Vienna, Austria), pp. 665.668, Sept 2004.

[3] S. B. Gokturk and A. M. Aaron, .Applying 3d methods to video for compression,. in Digital Video Processing (EE392J) Projects Winter Quarter, 2002.

[4] T. Fryza, Compression of Video Signals by 3D-DCT Transform. Diploma thesis, Institute of Radio Electronics, FEKT Brno University of Technology, Czech Republic, 2002.

[5] G. M.P. Servais, .Video compression using the three dimensional discrete cosine transform,. in Proc.COMSIG, pp. 27.32, 1997.

[6] R. A.Burg, .A 3d-dct real-time video compression system for low complexity single chip vlsi

implementation,. in the Mobile Multimedia Conf. (MoMuC), 2000.

[7] A. N. N. T. R. K.R., .Discrete cosine transforms,. in IEEE transactions on computing, pp. 90.93, 1974.

[8] T.Fryza and S.Hanus, .Video signals transparency in consequence of 3d-dct transform,. in Radioelektronika 2003 Conference Proceedings, (Brno,Czech Republic), pp. 127.130, 2003.

[9] N. Boinovi and J. Konrad, .Motion analysis in 3d dct domain and its application to video coding, vol.20, pp. 510.528, 2005.

[10] E. Y. Lam and J. W. Goodman, .A mathematical analysis of the dct coef_cient distributions for images,. vol. 9, pp. 1661.1666, 2000.

[11] Information Technology—Coding of Moving Pictures and Associated Audio for Digital Storage Media at up about 1.5 Mbit/s: Video, ISO/IEC 13818-2 (Mpeg2-Video), 1993.

[12] Mpeg-4 Video Verification Model 8.0 ISO/IEC JTC1/SC29/WG11, MPEG97/N1796, 1997.

[13] Joint Scalable Video Model JSVM-5, Joint Video Team (JVT) of ISO/IEC MPEG & ITU-T VCEG, JVT-S202, Geneva, Switzerland, 2006.

[14] S. Sun, Direct Interpolation for Upsampling in Extended Spatial Scalability, Joint Video Team (JVT) of ISO/IEC MPEG & ITU-T VCEG, JVT-P012, Poznan, Poland, 2005.

[15] S. Sun, J. Reichel, E. Francois, H. Schwarz, M. Wien, and G. J. Sullivan, Unified Solution for Spatial Scalability Joint Video Team (JVT) of ISO/IEC MPEG & ITU-T VCEG, JVT-R018, Bangkok, Thailand, 2006.

[16] Y. Vatis, B. Edler, D. T. Nguyen, and J. Ostermann, “Motion and aliasing compensated prediction using a two-dimensional non-separable adaptive wiener interpolation filter,” in Proc. Int. Conf. Image Processing, Sep. 2005, pp. 894–897.

[17] A. Segall and S. Lei, Adaptive Upsampling for Spatially Scalable Coding, Joint Video Team (JVT) of ISO/IEC MPEG & ITU-T VCEG, JVT-O010, Busan, Korea, 2005.

[18] A. Segall, Study of Upsampling/Down-Sampling for Spatial Scalability Joint Video Team (JVT) of ISO/IEC MPEG & ITU-T VCEG, JVTQ083, Nice, France, 2005.

[19] A. Segall, Upsampling/Down-Sampling for Spatial Scalability, Joint Video Team (JVT) of ISO/IEC MPEG & ITU-T VCEG, JVT-R070, Bangkok, Thailand, 2006.

[20] G. J. Sullivan, Resampling Filters for SVC Upsampling, Joint Video Team (JVT) of ISO/IEC MPEG&ITU-T VCEG, JVT-R066, Bangkok, Thailand, 2006.

[21] Jun Ho Cho, Tae Gyoung Ahn, and Jae Hun Lee “ Lossless Image /Video Compression with Adaptive Plane Split” IEEE conference,2009.

Authors Profile G.Suresh received the B.E Degree in ECE from Priyadarshini Engineering College affiliated to Madras University in 2000, and the M.E degree from Anna University, Chennai, in 2004. Currently he registered for Ph.D and doing research under anna university, Chennai. His research focuses on video processing, and


112

compression techniques, Reconfigurable and adaptive logic design, and high performance clock distribution techniques. P.Epsiba received the B.E degree in ECE from Ranipettai Engineering College affiliated to Madras University in 2002, and the M.E degree from Anna University, Chennai, in 2008. Currently she is working as a lecturer in C. Abdul Hakeem College of Engineering and Technology. Her research focuses on video compression and filtering techniques, Reconfigurable and adaptive logic design, and high performance clock distribution techniques. Dr.M.Rajaram working as Professor and Head of the department of EEE in GCE, Thirunelveli. He is having more than twenty five years of teaching and research experience. His research focuses on networks security, image and video processing, FPGA architectures and power electronics. Dr.S.N.Sivanandam working as Professor and Head of the department of CSE in PSG Tech, Coimbatore. He is having more than thirty two years of teaching and research experience. His research focuses on Bioinformatics, computer Communication, video compression and adaptive logic design.


113

Study the Effect of Node Distribution on QoS Multicast Framework in Mobile Ad hoc Networks

Mohammed Saghir

Computer Department, Hodeidah University, Yemen Information Center, Hodeidah Governorate, Yemen

[email protected]

Abstract: The distribution of nodes in mobile ad hoc networks affects connectivity, network capacity, group member and route length. In this paper, the effect of node distribution on the QoS multicast framework (FQM) is studied. In order to do this, extensive simulations are performed for FQM with two well-known node placement models: The Random placement model and Uniform placement model. The performance of FQM with the placement models is studied under different node mobility and node density. The analysis of simulation results have shown that there is a difference between the performance of FQM-Uniform and FQM-Random when mobility is zero. This difference is decreased while mobility increased.

Key words: MANET, RANDOM, UNIFORM.

1. Introduction Wireless networking and multimedia applications are growing in importance rapidly. The motivation for supporting QoS multicasting in MANET is the fact that multimedia applications are becoming important for group communication. Among types of wireless networks, MANET provides flexible communication with low cost. All communications are done over wireless media without the help of wired base stations. The environment for MANETs is very volatile so connections can be dropped at any moment. Distant nodes communicate over multiple hops and nodes must cooperate with each other to provide routing. The challenges in MANET are attributed to mobility of intermediate nodes, absence of routing infrastructure, low bandwidth and computational capacity of mobile nodes. The network traffic is distributed through multi-hop paths and the construction of these paths is affected by node distribution. In addition, the node distribution affects connectivity, network capacity, group member and route length in mobile ad hoc networks. Moreover, if mobile nodes are not distributed uniformly, some area of ad hoc network may not be covered. In some cases, mobile nodes are located in the middle of simulation area and as a result, they have higher average connectivity degree than nodes at border [1].

This paper is structured as follows: Section 2 gives an overview on the previous work whereas Section 3 describes the QoS multicast framework FQM and defines the uniform and random placement models. In Section 4, the simulation results of implementing FQM with two different placement models are presented. Finally, Section 5 concludes this paper and makes mention of future work.

2. Literature Review It has been proven that the capacity of network does not increase with its size when nodes are stationary [2]. On the other hand, it has been proven that the mobility increases the capacity of mobile ad hoc networks [3]. The performances of some placement modes are studied in previous work. In [4], a new algorithm (a Node Placement Algorithm for Realistic Topologies-NPART) was proposed to create realistic network topologies. This algorithm provides realistic topology with different input data. The NPART algorithm is compared with the uniform placement model. The result of simulation shows that NPART algorithm reflects the properties for user initiated network. The researchers in [5] modified the Particle Swarm Optimization (PSO) using genetic algorithm to optimize node density and improve QoS in sensing coverage. The nodes in simulation area are divided into stationary nodes and mobile nodes and the study focuses on how to optimize mobile nodes distribution to improve QoS in sensing coverage in sensor network.

The properties of the random waypoint mobility model (RWP) are studied in [6], [7], [8], [9], [10], [11], [12] and the bugs that might occur when using this model is highlighted. The researchers concluded that the nodes distributions after long time of simulation are different from the initial nodes distributions. In addition, the random waypoint model and Brownian-like model are studied in [8] and concluded that the concentration of nodes in the center of simulation area in RWP model is based on the choice of mobility parameters. Moreover, the effect of RWP model on the node distribution is studied in square and circle area [13]. The behavior of mobile nodes in RWP mobility model is outlined and analyzed. Some parameters for distribution in the RWP model to accurate the movement of mobile nodes in a square area are defined.

The previous studies are focused on studying and updating the node placement models and mobility models while in this paper, we study the effect of node placement models on the performance of FQM framework as the node distribution affects the group member and construction of multi-hop paths communication in the QoS multicast routing protocols.


114

3. The FQM with two Placement Models

3.1 The FQM QoS multicast framework Multicast routing is more efficient in MANETs because it is inherently ready for multicast due to their broadcast nature that avoids duplicate transmission. Packets are only

multiplexed when it is necessary to reach two or more destinations on disjoint paths. This advantage conserves bandwidth and network resources [14]. In our previous work [15], we propose a cross-layer Framework FQM to support QoS multicast applications for MANETs. Figure 1 gives an overview on the cross-layer framework FQM.

Figure 1. An Overview on the Functionalities of the Cross-Layer Framework While Receiving Flows

The first component of the framework is a new and efficient QoS multicast routing protocol (QMR). The QMR protocol is used to find and maintain the paths that meet the QoS

requirements. Second, a distributed admission control which used to prevent nodes from being overloaded by rejecting the request for new flows that will affect the ongoing flows.


115

Third, an efficient way to estimate the available bandwidth and provides the information of the available bandwidth for other QoS schemes. Fourth, a source based admission control which used to prevent new sources from a affecting the ongoing sources if there is not enough available bandwidth for sending to the all members in the multicast group. Fifth, a cross-layer design with many QoS scheme: classifier, shaper, dynamic rate control and priority queue. These schemes work together to support real-time applications.

The traffic is classified and processed based on its priority. Control packets and real-time packets will bypass the shaper and be sent directly to the interface queue at MAC layer. The best-effort packets should be regulated based on the dynamic rate control. In terms of queue priority, control packets and real-time packets have higher priority than best-effort packets. All components in the framework are cooperating to provide the required level of services.

3.2 The Uniform Placement Model In Uniform placement model, the simulation area is divided into a number of cells based on the number of mobile nodes in the simulation area. Within each cell, a node is placed randomly. The uniform placement model is set to create topologies with an equaled average node degree and this improves connectivity.

3.3 The Random Placement Model In the Random placement model, mobile nodes are placed randomly in a given area according to a probability distribution of mobile nodes. Based on the probability distribution, the mobile nodes density is different from one area to another in the simulation area of ad hoc networks. Actually, the Random distribution is suitable as it reflects the real behavior of mobile nodes in ad hoc networks.

4. Performance Evaluation We have conducted experiments using GLOMOSIM [16] to study the effect of node distribution on the QoS multicast framework (FQM). The main concern of these experiments is to study the effect of the random placement model and the uniform placement model on FQM framework while supporting QoS multicast applications. This simulation was run using a MANET with different number of nodes moving over a rectangular 1000 m × 1000 m area for over 900 seconds of simulation time. Nodes in the simulation move according to the Random Waypoint mobility model provided by GLOMOSIM. Mobility speed is ranged from 0-20 m/s and the pause time is 0 s. We used one multicast source sending to 15 multicast destinations in all experiments (assuming all destinations were interested to receive from the source node). The radio transmission range is 250 M and the channel capacity is 2Mbit/s. Each data point in this simulation represents the average result of ten runs with different initial seeds.

The performance of FQM with uniform and random placement models is studied through the following

performance metrics: • Packet delivery ratio: the average of the ratio

between the number of data packets received and the number of data packets that should have been received at each destination. This metric indicates the reliability of the proposed framework.

• Control overhead: number of transmitted control packet (request, reply, acknowledgment) per data packet delivered. Control packets are counted at each hop. The available bandwidth in MANETs is limited so it is very sensitive to the control overhead.

• Average latency: the average end-to-end delivery delay is computed by subtracting packet generation time at the source node from the packet arrival time at each destination. The multimedia applications are very sensitive to the packet delay; if the packet takes long time to arrive at destinations, it will be useless.

• Jitter: it is defined as a variation in the latency of received packets. It is determined by calculating the standard deviation of the latency [17]. This is an important metric for multimedia applications and should be kept to a minimum value.

• Group Reliability: it is defined as the ratio of number of packets received at 95% of destination and number of packets should be received. This means that the packet is considered to be received only if it is received by 95% of the number of multicast group.

4.1 The Performance of FQM under Different Mobility In this section, we study the performance of FQM with the uniform and random placement models under different mobility.

4.1.1 Packet Delivery Ratio (PDR)

10

20

30

40

50

60

70

80

90

100

0 5 10 15 20Mobility (m/s)

PD

R

FQM-Random FQM-Uniform

Figure2. Performance of PDR vs. mobility.

In a stationary network, the nodes always remain either in the range of each other or out of range. When mobility is increased, the positions of mobile nodes are changed. The performance of PDR vs. increasing mobility is given in Figure2. The PDR in FQM-Uniform is significantly higher


116

than that in FQM-Random. In FQM-Uniform, the mobile nodes are distributed uniformly so the degree of neighbor nodes is equaled and as a result the traffic load is balanced through intermediate nodes.

In FQM-Random, the mobile nodes are distributed randomly so the degrees of nodes are different from one node to another and as a result, the traffic load may congest through some intermediate nodes. When mobility is increased, the distributions of nodes are affected and as a result the difference of PDR between FQM-Uniform and FQM-Random are decreased.

4.1.2 Control Overhead (OH)

00.030.060.090.120.150.180.210.240.270.3

0.330.360.39

0 5 10 15 20

Mobility (m/s)

OH

per

Pack

et D

eliv

ery


Figure3. Performance of OH vs. mobility.

Figure 3 shows the Control OH vs. increasing mobility. The results show that control OH for FQM-Uniform is lower than FQM-Random when mobile nodes are static; this is because the number of data packets that received in FQM-Uniform was higher than that received in FQM-Random. As mobility increased, the number of data packets that received in FQM-Uniform decreased and this affect the percentage of control OH per packet. As a result of this, the differences between FQM-Uniform and FQM-Random are decreased.

4.1.3 Average latency (AL)

0

10

20

30

40

50

60

70

80

90

100

0 5 10 15 20

Mobility (m/s)

AL

(ms)


Figure4. the Performance of AL vs. mobility.

In wireless networks such as IEEE 802.11, the mobile nodes share the same channel and use the contention mechanism to capture the channel so the bandwidth availability affected by the number of nodes in the surrounding and this increases network delay and jitter. The results in Figures 4

show that AL for FQM-Uniform is relatively lower than FQM-Random. In FQM-Uniform, the traffic is distributed through different paths while in FQM-Random the traffic is congested and as a result, the AL is increased. In addition, when mobility is increased, the uniform distribution of nodes is changed and as a result, the AL in FQM-Uniform is increased.

4.1.1 Jitter

0

10

20

30

40

50

60

70

80

90

100

0 5 10 15 20

Mobility (m/s)Ji

tter

(m

s)


Figure 5. Performance of jitter vs. mobility.

Jitter occurs due to temporally lack of wireless connections and scheduling issues on the link layer [18]. The number of hops in the path is affected by the node distribution and as a result, the jitter is affected. Frequently changing routes could increase the jitter since the time for selecting forward nodes and the delay variation between the old and new routes increase the jitter. Figure 5 gives an overview on the performance of jitter vs. increasing mobility. The results show that the jitter for FQM-Uniform is relatively less than FQM-Random when mobile nodes are static. When mobility increased, the uniform node distribution is affected and as a result the differences between FQM-Uniform and FQM-Random in jitter are decreased.

4.1.5 Group Reliability (GR) The Group Reliability vs. increasing mobility is given in Figure 6. The GR for FQM-Uniform is higher than that in FQM-Random when mobility is zero. As mobility is increased, the difference between FQM-Uniform and FQM-Random in group reliability is decreased as discussed in section 4.1.1.

0

10

20

30

40

50

60

70

80

90

100

0 5 10 15 20

Mobility (m/s)

GR


Figure 6. Performance of GR vs. mobility.


117

4.2 The Performance of FQM under Different Nodes Density

In this section we study the performance of FQM with the uniform and random placement models under different nodes density with no mobility to focus on the effect of node density.

In a denser network, the probability of mobile nodes to sense the activities of its neighbor nodes increases so the packet collision that is coming due to hidden terminals is reduced [19]. In addition, when the network density increases, the number of connection increases so packets can finds paths to arrive at destinations.

In sparser network, the connectivity will be small so the number of delivered data packets is few due to lack of routes [20]. On the other side, when the network density is very high, the interference between nodes increases and this increases the collisions and reduces the channel access which leads to drop data packets.

4.2.1 Packet Delivery Ratio (PDR) The packet delivery ratio as a function of network density for FQM-Uniform and FQM-Random is given in Figure 7. The figure shows that there is a difference between PDR FQM-Uniform and PDR FQM-Random. The difference is increased when the node density is increased. This is because in FQM-Uniform, the mobile nodes uniformly distributed so when node density is increased, the number of paths increased and traffic load is balanced and as a result the PDR increased.

In FQM-Random, when node density is increased with random distribution, the traffic is congested and the available bandwidth is reduced as a result of increase contention and collision between mobile nodes.

10

20

30

40

50

60

70

80

90

100

100 150 200 250Number of nodes

PDR


Figure 7. PDR vs. network density

4.2.2 Average Latency (AL) The number of nodes in the neighborhood affects the available bandwidth and this increases the network delay and jitter. Figure 8 reflects the difference between the AL in FQM-Uniform and FQM-Random. The AL in FQM-Random is slightly increased while network density increased.

0

30

60

90

120

150

180

210

240

270

300

100 150 200 250

Number of nodes

AL

(ms)


Figure 8. AL vs. network density

4.2.3 Jitter The performance of FQM as a function of network density is described in Figure 9. The results describe the difference between the jitter in FQM-Random and FQM-Uniform.

0

40

80

120

160

200

240

280

320

360

400


Jitt

er (m

s)FQM-Random FQM-Uniform

Figure 9. Jitter vs. network density

4.2.4 Group Reliability (GR) Figure 10 shows the performance of the group reliability vs. increasing network density. From the Figure, there is a difference between GR in FQM-Uniform and FQM-Random and this is because PDR in FQM-Uniform is higher than PDR in FQM-Random as discussed before in Section 4.2.1.

0

10

20

30

40

50

60

70

80

90

100


GR


Figure 10. GR vs. network density


118

4.3 The Performance of FQM under High Density and High Mobility

The effect of mobility with low node density is discussed in details in section 4.1. In this Section, the effect of high mobility with high density on the performance of FQM with uniform and random placement models is studied. The node density is 100 mobile nodes and mobility is 20 m/s.

4.3.1. Packet Delivery Ratio (PDR) The packet delivery ratio as a function of mobility with high density for FQM-Uniform and FQM-Random is given in Figure 11. The figure shows that the difference between PDR for FQM-Uniform and FQM-Random is decreased while mobility increased even the node density is high. This is because the uniformly distribution is affected with high mobility and as a result, the traffic is congested and available bandwidth is reduced.

10

20

30

40

50

60

70

80

90

100

0 20Mobility (m/s)

PDR


Figure 11. PDR as a Function of mobility with high density

4.3.2. Average Latency (AL) The number of nodes in the neighborhood affects the available bandwidth and this increases the network delay and jitter. Figure 12 shows that the difference between the AL in FQM-Uniform and FQM-Random is decreased while mobility is increased even the node density is high.

0

30

60

90

120

150

180

210

240

270

300

0 20Mobility (m/s)

AL

(ms)


Figure 12. AL as a Function of mobility with high density

4.3.3. Jitter The results in Figure 13 reflect that the difference between the jitter in FQM-Random and FQM-Uniform is decreased

while mobility is increased; this is because the average latency is changed while mobility is increased.

0

40

80

120

160

200

240

280

320

360

400

0 20Mobility (m/s)

Jitte

r (m

s)


Figure 13. Jitter as a Function of mobility with high density

4.3.4. Group Reliability (GR) The performance of the group reliability with high mobility and high node density is described in Figure 14. The Figure shows that the difference between GR in FQM-Uniform and FQM-Random is decreased while mobility is increased; this is because the uniformly distribution is affected with high mobility and as a result, the traffic is congested and available bandwidth is reduced as discussed before in Section 4.3.1.

0

10

20

30

40

50

60

70

80

90

100

0 20Mobility (m/s)

GR


Figure 14. GR as a Function of mobility with high density

5. Conclusion and Future work.

In this paper, we have studied the performance of FQM framework with two placement models under different node mobility and node density. From the results, the performance of the QoS multicast framework FQM with Uniform placement model (FQM-Uniform) is better than the performance of FQM with Random placement model (FQM-Random) when the mobility of nodes is zero. Although the Uniform placement model is superior Random placement model, the Random placement model is more suitable to reflect the real behavior of nodes in mobile ad hoc networks.


119

This is because the mobility is the main characteristic of mobile ad hoc network. Moreover, the Uniform placement model is suitable in some applications of sensor networks where static sensors are used.

The analysis of simulation results shows that the mobility model has the most effect on the performance of the FQM QoS multicast as it changes the distribution of mobile nodes and as a result, it affects the group member and network capacity. In future work, we intend to study the performance of the FQM QoS multicast framework with different mobility models.

References

[1] C. Bettstetter and C. Wagner, “The Spatial Node Distribution of the Random Waypoint Model,” In Preceding the First German Workshop Mobile Ad Hoc Networks, 2002.

[2] P.gupat and P. Kumer, “the capacity of ad hoc networks”, IEEE transaction information theory, V 64, no 2, pp388-404.

[3] M. grossglauser and D.Tse, “mobility increase capacity of ad hoc networks”, In the proceeding of the IEEE INFOCOM, pp1360-1369,2001.

[4] B. Milic and M. Malek, “NPART - Node Placement Algorithm for Realistic Topologies in Wireless Multi-hop Network Simulation”, In Proceedings of the 2nd International Conference on Simulation Tools and Techniques , 2009

[5] P. Song, J. Li, K. Li and L. Sui, “Researching on Optimal Distribution of Mobile Nodes in Wireless Sensor Networks being Deployed Randomly,” International Conference on Computer Science and Information Technology, 2008.

[6] C. Bettstetter, “Mobility Modeling in Wireless Networks: Categorization, Smooth Movement, and order Effects,” ACM Mobile Comp. and Comm. Rev., vol. 5, no. 3, 2001.

[7] C. Bettstetter, H. Hartenstein, and X. Perez-Costa, “Stochastic Properties of the Random Waypoint Mobility Model,” ACM/Kluwer Wireless Networks, 2004.

[8] D. M. Blough, G. Resta, and P. Santi, “A Statistical Analysis of the Long-Run Node Spatial Distribution in Mobile Ad Hoc Networks,” In the Proceeding ACM Int’l Workshop Modeling, Analysis, and Simulations of Wireless and Mobile Systems (MSWiM), 2002.

[9] T. Camp, J. Boleng, and V. Davies, “A Survey of Mobility Models for Ad Hoc Network Research,” Wireless Comm. & Mobile Computing (WCMC), vol. 2, no. 5, pp. 483-502, 2002.

[10] E. Royer, P. Melliar-Smith, and L. Moser, “An Analysis of the Optimum Node Density for Ad Hoc Mobile Networks,” In the Proceeding of the IEEE Int’l Conf. Comm. (ICC), 2001.

[11] J. Song and L. Miller, “Empirical Analysis of the Mobility Factor for the Random Waypoint Model,” In Proceeding of the OPNETWORK, 2002.

[12] J. Yoon, M. Liu, and B. Noble, “Random Waypoint Considered Harmful,” In the Proceeding of the IEEE INFOCOM, 2003.

[13] C. Bettstetter, G. Resta and P. Santi, “The Node Distribution of the Random Waypoint Mobility Model for Wireless Ad Hoc Networks,” IEEE Transaction on Mobile Computing, Vol, 2, No. 3, JULY-SEPTEMBER, 2003.

[14] M. Hasana and L. Hoda, Multicast Routing in Mobile Ad Hoc Networks: Kluwer Academic Publishers, 2004.

[15] M. Saghir, T. C. Wan, and R. Budiarto, “A New Cross-Layer Framework for QoS Multicast Applications in Mobile Ad hoc Networks,” International Journal of Computer Science and Network Security, (IJCSNS), vol. 6, pp. 142-151, 2006.

[16] http://pcl.cs.ucla.edu/projects/glomosim. [17] K. Farkas, D. Budke, B. Plattner, O. Wellnitz, and L.

Wolf, “QoS Extensions to Mobile Ad Hoc Routing Supporting Real-Time Applications,” In the proceeding of the 4th ACS/IEEE International Conference on Computer Systems and Applications, Dubai-UAE, 2006.

[18] O. Farkasa, M. Dickb, X. Gub, M. Bussec, W. Effelsbergc, Y. Rebahid, D. Sisalemd, D. Grigorase, K. Stefanidisf, and D. Serpanosf, “Real-time service provisioning for mobile and wireless networks,” Computer Communications, vol. 29, pp. 540-550, 2006.

[19] C. Lin, H. Dong, U. Madhow, A. Gersho, “Supporting real-time speech on wireless ad hoc networks: inter-packet redundancy, path diversity, and multiple description coding,” In Proceedings of the 2nd ACM international workshop on Wireless mobile applications and services on WLAN hotspots, USA, 2004.

[20] A. NILSSON, “Performance Analysis of Traffic Load and Node Density in Ad hoc Networks,” In proceeding the Fifth European Wireless Conference, Spain, 2004.

Author Profile

Mohammed Saghir received his B.S from Technology University, Iraq in 1998, M.Sc. from Al Al-Bayt University, Jordan in 2004 and his Ph.D from University Sains Malaysia 2008 in Computer Science. He is working as a lecturer in Hodeidah University, Yemen. His current research interests include, Mobile Ad hoc networks, QoS multicast routing in MANETs, WiMax.


120

ICAR: Intelligent Congestion Avoidance Routing for Open Networks

Dr. P.K.Suri1 and Kavita Taneja2

1Dean, Faculty of Sciences, Dean, Faculty of Engineering, Professor, Deptt. Of Comp. .Sci.& Appl. Kurukshetra University, Kurukshetra, Haryana ,India

[email protected]

2Assistant Professor, M.M. Inst. of Comp. Tech. & Business Mgmt., Maharishi Markandeshwar University, Mullana, Haryana, India.

[email protected]

Abstract: MANETs (Mobile Ad-hoc Networks) idiosyncrasies have emerged as a promising networking theme that allows groups of Nomadic Nodes (NN) to organize themselves in order to form an immediate network and that too in the absence of any central entities for configuration or routing (e.g. DHCP server). MANETs are instant distributed systems, where any node can become a source or destination, is host as well as router and are considered as peers. Hence, routes in MANETs are often multi-hop in nature with several intermediate nodes that forward the data of their peers. Congestion Control is an important design criterion for efficient mobile computing. The theme of any mobile computing is that the “ infrastructure less” devices or NN must be capable of locating each other & communicating automatically in dynamic, energy efficient, fault tolerant, latency proof and secure framework. Despite of lot of recent research, MANETs lack load-balancing capabilities, and thus, they fail to provide expected performance especially in the case of a large volume of traffic. In this paper, we present a simple but very effective method to support Quality of Service (QoS), by the use of load-balancing and congestion avoidance routing scheme. This intelligent approach allows NN to select next hop based on traffic sizes to balance the load over the network. A simulation study is presented to demonstrate the effectiveness of the proposed scheme and the results shows that the proposed design aims to maximize network performance in terms of reduced packet loss rate and increased throughput.

Keywords: Mobile Ad Hoc Network (MANET), Nomadic Node (NN), Intelligent Congestion Avoidance Routing (ICAR).

1. Introduction Recent advances in wireless communication technology and mobile computing have lead to virtual Big Bang of mobile technology resulting in an explosive growth in the number of NN intended to communicate via an anytime anywhere network. A Mobile Ad Hoc Network (MANET) is the promising answer, a multihop wireless network [2] in which all NNs (Figure1) such as personal computers, personal digital assistants and wireless phones are wildly mobile and communicating. MANET provides an effective method for constructing an instant network on the airs that uses a distributed control system by which all the NNs communicate with each other without a central authority. But facilities come as a package with challenges. Due to the limited transmission range of wireless network interfaces [1], [3], [4], multiple hops may be needed to exchange data between devices in the network. Others may include

dynamic topology, device discovery, limited bandwidth, limited resources, limited physical security, infrastructure-less and self operated, poor transmission quality and topology maintenance characteristic of open networks. It is evident that existing solutions are not fully equipped to tame these challenges to a reasonable level [5], [6]. Hence, Computing in such a commercialized hostile environment governed by mobility (shown in Figure 1) must be intelligent to ensure high packet delivery ratio with minimum loss of bandwidth and overhead and hence QoS.

Figure 1. A Pure MANET Existing routing protocols [11]-[12] tend to neglect the issue of congested areas in the open networks and hence are handicap to balance the traffic load over the NNs creating highly congested regions in which the data packets suffered a long buffering time and the NNs experienced a highly contended access to the medium. On the other hand, the proposed Intelligent Congestion Avoidance Routing (ICAR) protocol avoids the creation of such regions by selecting routes based on traffic metric and not the shortest path.

Rest of the paper is as follows. Section 2 discusses MANET applications and major challenges encountered due to NNs in open networks. Section 3 summarizes related work including existing routing protocols and their pitfalls. Section 4 gives the protocol design of ICAR. We present implementation and performance evaluation of ICAR in Section 5 and finally article is concluded in Section 6.


121

2. MANET: Applications and Challenges MANET invokes popular communication architectures to offer to NNs the opportunity to share information and services, in the absence of any infrastructure. Today the commercial aspect of MANET enables cars to exchange information about road conditions, traffic jams or emergency situations leading to improved road safety, efficiency and driving comfort for the customer. The beauty lies in the diversity of the MANET applications [7], [9] ranging from military, law enforcement, national security (global war), to disaster relief, and rescue mission like crisis management during natural disasters to user oriented brighter side. Commercially, MANET can expand the capabilities of mobile phones, communication in exhibitions, conferences, sale presentations, restaurants, etc. There are several challenges existing in the field of MANETs and to list a few one can start with the autonomous behavior of each node leading to no centralized administration [8]. For example, the frequent change in network topology due to the NNs causes a great deal of control information to flow onto the network. The small capacity of batteries and the bandwidth limitation of wireless channels are major factors [1], [3], [10]. Moreover, data access focused at a single point may incur impossibility of communication and make quality of service worse. This becomes a serious consideration, especially with recent trends to transferring huge data including video.

3. Related Work Congestion control is a key problem in mobile ad-hoc networks. The standard TCP congestion control mechanism is not able to handle the special properties of a shared wireless multihop channel well. In particular the frequent changes of the network topology ruled by NNs and the shared nature of the wireless channel pose significant challenges. Many approaches have been proposed to overcome these difficulties. The problems in using conventional routing protocols in MANET [12] are as follows: • Existing protocols cannot cope with frequent and

unpredictably changing topological connectivity [11]. • Conventional routing protocol could place heavy

computational burden on mobile computers in terms of battery power and network bandwidth.

• Inefficient convergence characteristics. • Wireless media has limited range unlike wired media. One of the issues with routing in open networks concerns whether nodes should keep track of routes to all possible destinations or instead keep track of only those destinations that are of immediate interest [18]. Routing protocols are classified broadly into two main categories as proactive routing protocols and reactive routing protocols [12]. Proactive routing protocols are derived from legacy Internet. Proactive protocols that keep track of routes for all destinations in the ad hoc have the advantage that communications with arbitrary destinations experience minimal initial delay from the point of view of the application. When the application starts, a route can be immediately selected from the routing table. Such protocols are called proactive because they store route information

even before it is needed. They are also called table driven because routes are available as parts of a well-maintained table. The popular proactive routing protocols [12], [15] are Destination-Sequenced Distance Vector (DSDV) and Global State Routing (GSR). Since the network topology is dynamic, when a link goes down, all paths that use that link are broken and have to be repaired. If no applications are using these paths, then the effort gone in to repair may be considered wasted that cause scarce bandwidth resources to be wasted and may lead to further congestion at intermediate network points. These Protocols are applicable only for low mobility networks and are scalable in the number of flows and number of nodes but are not scalable in the frequency of topology change. In contrast, reactive routing protocols establish the route to a destination only when requested. To overcome the wasted work in maintaining routes not required, on-demand, or reactive protocols have been designed. Reactive routing protocols save the overhead of maintaining unused routes at each node, but the latency for many applications will drastically increase. Most applications are likely to suffer a long delay when they start because a route to the destination will have to be acquired before the communication can begin. Reactive protocols namely Ad Hoc On-Demand Distance Vector (AODV) [13] and Dynamic Source Routing (DSR) [14] are suitable for networks with high mobility and relatively small number of flows. Shortest path algorithm used in the existing routing protocols does not provide optimal results when the primary route is congested. Congestion aware routing protocols [16]-[17] were therefore proposed to that effect due to the fact that besides route failures, network congestion is the other important cause of packet loss in MANETs.

4. ICAR: Protocol Design The policy of design for mobile networks is that if two

or more routes for NNs communication exist then simply the path with the minimum traffic metric associated is selected. Objectives include maximizing network performance from the application point of view while minimizing the cost of network itself in accordance with its capacity. In this paper we develop a new Routing framework that is service oriented for MANET which will cover following issues: • Intelligent routing algorithm ICAR that encompasses

an area with broadband wireless coverage and also can dynamically cope with congestion and path load.

• Supporting NNs that can act as routers for other network devices, i.e., each node is an implicit router further extending the network.

ICAR protocol operation is illustrated in Figure 2. The service required by the device A resides at NN ‘H’. The NN

‘K’ can act as a route to NN ‘A’, helping it to reach the service at NN ‘H’. Additionally, NNs ‘K’ & ‘H’ can share the data about other services with the NN ‘A’. Suppose NN ‘A’ requires a service which is being offered by NN ‘H’ at a distance of 2 hops (as path length is a secondary issue here) but the most congestion free route is to be selected. So ‘A’ broadcasts a request HELO looking for the service through a lesser loaded path. Alongside service request, ‘A’ also advertise its own services. All neighbors NNs receiving the broadcast update their service directories. NN ‘K’ has


122

information on the required service available with NN ‘H’; therefore on receiving ‘A’s’ request it will act as pseudo destination and unicast a response route ACK to ‘A’ containing route information of ‘H’ in addition to services offered by ‘H’ and ‘K’ itself. Here one additional field is proposed in the standard packet header. This field is called the Traffic Field, and it is initialized to zero by the source NN before broadcasting a route discovery packet. Every intermediate NN that receives the route discovery packet calculates its current total traffic based on its total neighbor NNs at one instance and adds it to the value of the Traffic Field on the incoming packet. The result of the addition is assigned to the Traffic Field before the NN rebroadcasts the packet. When generating the route discovery Reply, besides copying the route discovery packet’s node list to the reply packet, the target NN copies the value of the Traffic Field from the route discovery packet to the reply packet. When the source NN receives the Reply, it is in fact receiving the path to the intended destination and the Traffic Metric associated with that path. To allow the source NN to obtain more than one path, the destination NN must reply to all route discovery requests it receives, and the source of the route discovery must wait for an interval of time after starting the route discovery process. When the NN obtains more than one path for next hop, it simply selects the path with the minimum Traffic metric associated with it. Also ICAR prohibit intermediate NNs from replying to route discoveries using their cached paths that guarantee the utilization of nearly current Traffic information. When a NN sends a HELO Message, it includes its current Local Traffic metric on the message. Every NN maintains a list of its neighboring NNs and their local traffic. When the NN receives a HELO Message from a neighbor, it checks its list of neighbors. If the neighbor is already on the list, it updates the neighbor’s Local Traffic. Otherwise, it adds the new neighbor to the list. ICAR defines the removal of listed neighbor if it fails to receive “3-consecutive HELO messages.” The Local Traffic in the neighbor list results in the Regional Traffic of the NN. Such data sharing with next hop NN is dirty write in and is likely to be successful, since the availability of information about required service ‘A’ makes it highly probable that other related services residing on ‘H’ and even ‘K’ might interest ‘A’. So a discovered route can be cached for further references but only for a span of time that is a function of the mobility parameter. On receiving the response from ‘K’, ‘A’ updates its service directory and routing table with the information on ‘H’ and services of ‘K’ and the path least congested from A to H. If more than one NNs reply with the response, then only the response having the route information with the least congestion will be considered. If none of the NNs has information about service required by ‘A’, they will re-broadcast. This message is rejected by the NNs that have already received the request during the first rebroadcast by employing the (Sequence ID, Broadcast ID) in the request message. Also, if the more than 2 hops are involved, backward pointers are setup and routing tables are updated on each intermediate NN. Once the congestion free route discovery is made the NNs along the route may be requested to volunteer as pseudo destinations through a chain of unicast trusted pseudo destinations request.

C

E

F

G

H

A

K

B

REQ REQ

REQ

REQ REQ

REQ

REQ

RACK

RACK

RACK

RACK

RACK

Routing Table

Routing Table

Routing Table

Routing Table

Routing Table

Routing Table

Routing Table

Routing Table

Figure 2. ICAR Operation

5. Implementation and Performance Study To evaluate the effectiveness of ICAR, we have

simulated the scheme and compare it to the Implicit Source Routing (ISR) [19] protocol. ISR is a competitive routing protocol based on shortest path. We compare the ICAR to ISR using the ns-2 network simulator [20], which includes a mobility extension that was ported from CMU’s Monarch Group’s mobility extension to ns. CMU’s Monarch mobility extension to ns-2 allows the simulation of multi-hop ad hoc wireless networks. The extension includes functionalities to simulate NN’s movements, and to transmit and receive on wireless channels with a realistic radio propagation model. For the ISR simulation, we used the latest version available from the VINT project that comes with ns-2. That version includes DSR with a full implementation of the Implicit Source Routing (ISR) technique. The parameter values used by both protocols are summarized in Table 1. Constant Bit Rate (CBR) traffic was used in our simulation. The source and destination devices were randomly selected, and each simulation shows the results of 32 connections. To vary the size of the traffic, we divided the 32 connections into four groups of eight connections each. The size of the CBR packets for the first group was 128 bytes, second was 256 bytes, third was 512 bytes and fourth group was 1024 bytes. We modeled a 4 packets/sec send rate for all the groups. The mobility patterns in our simulation followed the random model in which each node starts at a random location, chooses a new location in a rectangular space (1500 m x 300 m) randomly, and starts its trip to the new location at a randomly chosen speed (uniformly distributed between 0–20 m/sec). After reaching its new location, NN pauses for a period of time (wait time) and then starts a new trip to a new location. We varied the mobility of the NNs by varying the wait time values. The results we present in this paper are based on 50 simulation runs (50 nodes) with run length of 500 seconds.


123

Table 1: Values used in simulation

Parameters ICAR ISR Send buffer size 64 64 Routing table size 64 30 Cache on off HELO interval dynamic N.A. Lost HELO allowed 03 N.A. WAIT 600 ms N.A. Interface queue size 50 50 Ten runs of different traffic and mobility scenarios are

averaged to generate each data point. However, identical traffic and mobility scenarios were used for both protocols. We used three performance metrics to compare our schema to ISR. The first metric is the Packet Delivery Ratio, which is defined as the percentage of data packets delivered to their destination NN of those sent by the source NN. The second metric is the Routing Overhead of both protocols, which is defined as the number of routing packets “transmitted” per each data packet “delivered.” On multi-hop routes, each transmission of the routing packets is counted as one transmission. We chose not to include the forwarding information carried in each data packet in our calculation of the overhead because the size is the same for both protocols.

Figure 3. Packet Delivery Ratio

The third metric used is the Average End-to-End Delay of the data packets. We used eleven pause time values (0, 50, 100, 150, 200, 250, 300, 350, 400, 450, and 500 s) to differ the mobility level (with 0 s pause time meaning continually moving nodes and 500 s representing stationary nodes).

ICAR had a better delivery ratio than ISR (see Figure 3). For the data packets sent, ICAR delivered an average of 19.65% higher than ISR. The difference in the delivery ratio between both protocols is significant at the high level of mobility where the pause times are 0 s and 50 s. With pause times 0 s and 50 s, ICAR delivered 26.21% and 20.12% higher than ISR, respectively. At its best performance, ISR did not even deliver 50% of data packets sent whereas; ICAR delivered 70% of the data packets that were sent. These results are due to fact that ICAR distributed the traffic among the NNs and in a way to avoid the creation of highly congested areas. We found ISR concentrate the traffic through centrally located NNs because it allows the NNs to reply to path discoveries from their cached routes that caused their interface queues to overflow and suffer from high drop rates leading to packet collision and eventually

their loss. ICAR also outperformed ISR in the routing overhead metric (see Figure 4). ISR incurred an average of 11.42 overhead packets per data packet higher than ICAR. We found that the higher the level of mobility, the higher the difference in overhead between ICAR and ISR. At the highest level of mobility, ISR incurred about 38.53 of overhead packets per data packet, whereas ICAR incurred about 1.04 overhead packets per data packet. At the lowest level of mobility, ISR incurred about 2.29 of overhead packets per data packet, while ICAR incurred about 0.42 overhead packets per data packet. The reason for such high overhead is the additional Route discoveries incurred by ISR through its salvaging process because of the congested network. We found that the data packets experienced delay times long enough to invalidate, due to the mobility, the routes of these packets. For those packets to be salvaged, ISR initiates the Route Discovery process to find alternative routes. ICAR surpassed ISR in the average end-to-end delay metric (see Figure 5). The average end-to-end delay for ICAR was 2.59 s while it was 6.11 s for ISR. Generally, the average end-to-end delay of ICAR was about three seconds less than that of ISR. The difference is significant at the highest level of mobility, where the average end-to-end delays for ICAR and ISR were 3.44 s and 8.70 s, respectively.

Figure 4. Overhead

Figure 5. Average End-to-End Delay

The reason for this is that ISR does not balance the traffic over the NNs, it created highly congested regions in which the data packets suffered a long buffering time and the NNs experienced a highly contended access to the medium. On

0

12

34

5

67

89

10

0 100 200 300 400 500 600

Pause Time (seconds)

ISR ICAR

z

Ave

rage

End

-to-E

nd D

elay

(sec

onds

)

0

10

20

30

40

50

60

70

80

0 100 200 300 400 500 600


P A C K E T D E L. R A T I o

ISR

ICAR

0

5

10

15

20

25

30

35

40

45

0 100 200 300 400 500 600


ISR ICAR

Ove

rhea

d pa

cket

s pe

r dat

a Pa

cket


124

the other hand, ICAR avoided the creation of such regions by selecting routes based on Traffic metric and not the shortest path.

6. Conclusion and Future Work In this paper, we presented the ICAR protocol keeping in view the nature of MANETs, to route the traffic through least congested paths. We have made an effort to take the discovery and delivery of services to the level of routing, ICAR is compared to a shortest path based routing protocol, namely ISR. ICAR distributes the load on a large network area, thus increasing the spatial reuse. The simulation study has confirmed the advantages of ICAR over ISR. We intend to explore the application of the idea introduced in ICAR to other routing protocols in our future study.

References [1] Mario Gerla, “From Battlefields To Urban Grids: New

Research Challenges in Ad Hoc Wireless Networks,” Pervasive and Mobile Computing, 1(1), pp.77-93, March 2005.

[2] Marco Conti, “Multihop Ad Hoc Networking: The Theory,” Communications Magazine, IEEE, Vol. 45, pp. 78-86, 2008.

[3] R. Ramanathan and J. Redi, “A Brief Overview of Ad Hoc Networks: Challenges and Directions,” IEEE Communications Magazine, 40(5), pp. 20--22, 2002.

[4] S. Hariri, “Autonomic Computing: Research Challenges and Opportunities,” CAIP Seminar, February 2004.

[5] T. Camp, J. Boleng, and V. Davies, “A Survey of Mobility for Ad Hoc Network Research,” Wireless Commun. Mobile Comput. (WCMC), pp. 483--502, 2002.

[6] Al Hanbali , A. A. Kherani , R. Groenevelt , P. Nain , E. Altman, “Impact of Mobility on the Performance of Relaying in Ad Hoc Networks,” Extended version, Computer Networks: The International Journal of Computer and Telecommunications Networking, 51(14), pp.4112-4130, October, 2007

[7] S. Hadim, J. Al-Jaroodi, and N. Mohamed, “Middleware Issues and Approaches for Mobile Ad Hoc Networks,” In Proccedings of 3rd IEEE Consumer Communications and Networking Conference, CCNC 2006, pp. 431--436, 8--10 Jan. 2006.

[8] C. Bettstetter, "On the Connectivity of Ad Hoc Networks," The Computer Journal, 47(4), pp. 432--447, April 2004.

[9] Konrad Lorincz , et al., “Sensor Networks for Emergency Response: Challenges and Opportunities,” IEEE Pervasive Computing, 3(4), pp.16-23, October 2004

[10] Lu Yan , Xinrong Zhou, “On Designing Peer-to-Peer Systems over Wireless Networks,” International Journal of Ad Hoc and Ubiquitous Computing, 3(4), pp.245-254, June 2008.

[11] Michael Gerharz , Christian de Waal , Peter Martini, and Paul James, “Strategies for Finding Stable Paths in Mobile Wireless Ad Hoc Networks,” In Proceedings of the 28th Annual IEEE International Conference on Local Computer Networks, pp.130, October 20-24, 2003

[12] Mina Masoudifar, “A Review and Performance Comparison of QoS Multicast Routing Protocols for MANETs, Ad Hoc Networks,” 7(6), pp.1150-1155, August, 2009.

[13] C. Perkins , E. Belding-Royer , S. Das, “Ad hoc On-Demand Distance Vector (AODV) Routing,” RFC Editor, 2003.

[14] D. Johnson, “The Dynamic Source Routing (DSR) Protocol for Mobile Ad Hoc Networks for IPv4,” RFC4728, February 2007.

[15] T. Clausen and P. Jacquet, "Optimized Link State Routing Protocol (OLSR)," IETF Mobile Ad Hoc Networks (MANET) Working Group Charter, October 2003, www.ietf.org/rfc/rfc3626.txt.

[16] X. Gao, X. Zhang, D. Shi, F. Zou, and W. Zhu, “Contention and Queue-Aware Routing Protocol for Mobile Ad Hoc Networks,” WiCOM, September 2007.

[17] D.A. Tran and H. Raghavendra, “Congestion Adaptive Routing in Mobile Ad Hoc Networks,” IEEE Transactions on Parallel Distributed Systems, 17(11), pp. 1294-1305, November 2006.

[18] Hui Xu , J. J. Garcia-Luna-Aceves, “Neighborhood Tracking for Mobile Ad Hoc Networks,” Computer Networks: The International Journal of Computer and Telecommunications Networking, 53 (10), p.1683-1696, July, 2009.

[19] Yih-Chun Hu and D. B. Johnson, “Implicit Source Routes for On Demand Ad Hoc Network Routing”, In Proceedings of the ACM Symposium on Mobile Ad Hoc Networking & Computing, MobiHoc, Long Beach, California, USA, Oct. 2001.

[20] VINT Project. The Network Simulator - ns-2, http://www.isi.edu/nsnam/ns/, 2002

Authors Profile

Dr. P. K. Suri received his Ph.D. degree from Faculty of Engineering, Kurukshetra University, Kurukshetra, India and Master’s degree from Indian Institute of Technology, Roorkee (formerly known as Roorkee University), India. Presently, he is Dean, Faculty of Science, Kurukshetra

University and is working as Professor in the Department of Computer Science & Applications, Kurukshetra University, Kurukshetra, India since Oct. 1993. He has earlier worked as Reader, Computer Sc. & Applications, at Bhopal University, Bhopal from 1985-90. He has supervised six Ph.D.’s in Computer Science and thirteen students are working under his supervision. He has more than 110 publications in International / National Journals and Conferences. He is recipient of ‘THE GEORGE OOMAN MEMORIAL PRIZE' for the year 1991-92 and a RESEARCH AWARD –“The Certificate of Merit – 2000” for the paper entitled ESMD – An Expert System for Medical Diagnosis from INSTITUTION OF ENGINEERS, INDIA. His teaching and research activities include Simulation and Modeling, SQA, Software Reliability, Software testing & Software Engineering processes, Temporal Databases, Ad hoc Networks, Grid Computing and Biomechanics.

Kavita Taneja has obtained M.Phil(CS) from Alagappa University, Tamil Nadu and Master of Computer Applications from Kurukshetra University, Kurukshetra, Haryana , India. Presently, she is working as Assistant Professor in M.C.A. at M.M.I.C.T & B.M., M.M. University, Mullana, Haryana, India. She is pursuing Ph. D in Computer Science

andapplications from Kurukshetra University. She has published and presented over 10 papers in National /International Journals/Conferences and has bagged BEST PAPER AWARD, 2007 at International Conference for the paper entitled “Dynamic Traffic -Conscious Routing for MANETs” at DIT, Dehradun. She is supervising five M.Phil scholars in Computer Science. Her teaching and research activities include Simulation and Modeling and Mobile Ad hoc Networks.


125

Using Word Distance Based Measurement for Cross-lingual Plagiarism Detection

Moslem mohammadi1, Morteza analouei2

1Payam Nour University of Miandoab, Miandoab, Iran

[email protected]

2department of Computer Science and Engineering , Iran University of Science and Technology, Tehran, Iran

[email protected]

Abstract: - a large number of articles and papers in various languages are available due to the expanding of information technology and information retrieval. Some authors translate these papers from original languages to other languages so that they claim as their own papers. In this paper a method for detection of exactly translated papers from English to Persian language is proposed. Document classification is performed by corner classification neural network (CC4) method to limit the investigated documents, before the main process is commenced. For bi-lingual text processing, language unification is necessary that is performed by a bi-lingual dictionary. After language unification, the suspicious text fragments are rephrased according to the original language. Rephrasing is performed like crystallization process. The similarity is computed based on equivalent word place distance in English and Persian; in fact, the paragraphs in a suspicious text are probed in the original text. Proposed system is tested in two conditions, first with rephrasing and adopted distance based similarity and other with tradition method. By comparing the outcomes in two experiments distinguished that proposed method’s discriminability is good.

Keywords: plagiarism detection, document classification, bi-lingual text processing

1. Introduction and motivation At first, there are several questions that will be discussed. They are

i. What is the research? ii. What is the plagiarism?

iii. How an essay is criticized?; and iv. What is the scientific misconduct?

These questions and instances of like are ambiguous questions that don’t have a clear answer or they share overlapped scopes. These terms are defined in table 1. Some of the paradigms mentioned, are in the academic dishonesty domain and some of them are not. As mentioned above, these fields have shared overlapped scopes and the border line between correct works and incorrect works must be cleared. Plagiarism is prevalent due to spread enlargement digital information especially text information on the internet and other devices. The simple definition of plagiarism is appropriating another author’s writing or

compilation. The plagiarism may be done through manipulation of resource, such as text, film, speech and other literary compilation or artifacts, and through manner, which Maurer[1] listed four broader categories of plagiarism as follows:

i. Accidental: due to lack of plagiarism knowledge. ii. Unintentional: probably the initiation of the same

idea at same time. iii. Intentional: a deliberate act of copying someone

work without any credit or reference. iv. Self-plagiarism: republished self published work.

jalali et al.[2] believe that the first two cases occur in non-English countries since authors are not competent to use English and to avoid plagiarism. They have suggested that plagiarism should be avoided by employing a combination of measures such as explicit warning, using plagiarism detection software, disseminating knowledge and improving the academic and writing skills. In this paper we will discuss about text plagiarism that can be occurred in same language (monolingual) or other ones(cross-lingual). In first one, a person copies, rewrites or rephrases another's text in a similar language and plagiarism detection in same cases can be performed bye tradition document similarity algorithms. But in second, a thief author translates from source language to target language. For detection; we need extra knowledge bases such as a dictionary, wordnet and word correlation table. Actually; plagiarism detection is time consuming because there are many large text corpora that must be compared with suspected document, though most of them are irrelevant. For this reason if we eliminate irrelevant documents from comparison corpora, we’ll decrease an indispensable time spontaneously. This is realized by classification methods that will be explained later. The paper is organized as follow: in next section the related works is reviewed and document classification is discussed in section 3. Thereafter in section 4 the proposed approach will be explained then datasets and experimental results are analyzed.


126

Table 1: plagiarism related concepts

1 Research Research can be defined as the search for knowledge or any systematic investigation to establish facts.

2 Plagiarism Use or close imitation of the language and thoughts of another author and the representation of them as one's own original work.

3 Criticism Is the judgment (using analysis and evaluation) of the merits and faults of the actions or work of another individual.

4 Assemblage built primarily and explicitly from existing texts in order to solve a writing or communication problem in a new context

5 Essay mill An essay mill (or paper mill) is a ghostwriting service that sells essays and other homework writing to university and college students

6 Ghostwriter A ghostwriter is a professional writer who is paid to write books, articles, stories, reports, or other texts that are officially credited to another person.

7 Scientific misconduct Scientific misconduct is the violation of the standard codes of scholarly conduct and ethical behavior in professional scientific research.( fabrication, plagiarism, ghostwriting)

2. Related works The root of plagiarism detection can be found in the document similarity. Huang et al [3] proposed a sense based similarity measure for cross-lingual documents that uses senses for document representation and adopted fuzzy set functions for resemblance calculating. Uzuner et al. [4] identify plagiarism when works are paraphrased using by syntactic information related to creative aspect of writing such as ‘sentence-initial and -final phrase structure’, ‘semantic class of verb’ and ‘syntactic class of verb’. Barron et al.[5] proposed the cross-lingual plagiarism analysis with a probabilistic method which calculates a bilingual statistical dictionary on the basis of the IBM-1 model in English and Spanish plagiarized examples. Their proposal calculates the probabilistic association between two terms in two different languages. In [6] an approach based on support vector machine(SVM) classifier has been proposed to determine similarity between English and Chinese text. Subsequently, evaluating semantic similarity among texts measured by means of a language-neutral clustering technique based on Self-Organizing Maps (SOM). Selamat et al [7] have used a Growing Hierarchical Self-Organized Map (GHSOM) to detect translated English documents to Arabic texts. Gustafan et al.[8] have proposed a method, relies on pre-computed word-correlation factors for determining the sentence-to-sentence similarity which handle various plagiarism techniques based on substitution, addition, and deletion of words in sentences. Consequently, the degree of resemblance of any two documents to detect the plagiarized copy is computed based on sentence similarity and the visual representation of same sentences is generated.

3. Document classification Document classification has an important place on text mining tasks because that it limits the amount of texts must be processed. There are several methods for information classification such as K- Nearest Neighbors (KNN), Naïve Bayes, Support Vector Machine (SVM)[9], Corner Classification Neural Network (CC4)[10], ensemble classifiers[11] and etc. In this system completion of the document classification are performed by CC4 that is a kind of three-layered feed forward neural network [12] like figure

1. The X vector is a dictionary includes all words of documents; in other words, it is the feature set of documents that must be classified. To save time, after stopwords are eliminated, we only use 20 frequently appeared words in document for classification. Each entry of X vector is 1 if its corresponding word is in 20 frequently appeared words of document, otherwise it is 0. The Y vector represents the class tags and here the k value is the number of classes that in our system it is 3. There are distinct classifiers for each language. The non iterative training phase is one of important benefit of these networks. Incremental training is the other benefit of these networks that appropriate these networks for classifying enormous documents. In fact, adding the new training data to extant CC4 network is very easy and simple. Furthermore, when the documents are almost of the same size, CC4 neural network is an effective document classification algorithm.

Figure 1. CC4 structure

4. Proposed method Bi-lingual Plagiarism can be supposed as a translation problem. Clumsy translators or plagiarists are using a literal translation. Thus in this case, sentence length in both texts very probability are equal but innate features and structures of either one are different. Automatic detection of awkward translation seems easy. The professional plagiarist can change sentences sequence or split longer sentences in two or more sentences or do vice versa. For these reasons and other motivations that will be discussed, we used a distance based method. Figure 2 is the schematic view of proposed


127

method. There are three main tasks in proposed system. They are

i. Document classification: just as explained in previous section, the classification determines what documents must be processed.

ii. Document representation (translation): documents in different language must be uniformed that will be illustrated in next section.

iii. Similarity calculation: this part of system uses a word distance base method.

Figure 2. Architect of proposed system

4.1. Document representation In this research we are dealing with two kinds of documents in the Persian and English languages which the processing phase for similarity detection is accomplished in Persian. The Persian documents after eliminating the non-verbal stopwords[13], are represented in vector space model [14]. In this adopted model, each paragraph is placed in a vector. For plagiarism detection from English documents to Persian, the unification of their language must be performed. Whereas each word in the English text can have several meanings and equivalences in Persian, each vector of English words is changed into jagged matrix of the Persian words according to the bilingual dictionary. The schematic description of details comes in figure 3.

Persian paragraph: Tà t1 t2 t3 … ti … t

n

English paragraph: English

(W)

Persian equivalence (TW)

W1 11p

21p

31p

W2 12p

22p

W3 13p 2

3p 33p 4

3p 53p

... … … … Wj 1

jp 2jp 3

jp 4jp 5

jp 6jp 7

jp

… … … … Wm 1

mp 2mp 3

mp 4mp

Figure 3. Document representation

4.2. Similarity evaluating As mentioned before, each vector of English words is changed into jagged matrix of the Persian words, the tradition vector space methods such as euclidean, cosine, dice, jaccard similarity measures[15] that use the inner product of vectors can’t be used for similarity calculation because one side of problem is vector and other side is matrix. Let t1, t2, · · ·, tn be fragments conforming a suspicious text T in Persian, and let w1, w2, · · · ,wm be a collection of W original fragments in English. According to the structural differences between investigated languages, the smallest piece of text which must be processed is a Paragraph. The sentence length in English and Persian is different [16], and one sentence in English during its translation may be rendered in two or more sentences. Furthermore, the structural differences exist. For example, the verb placement in the English sentences is in the middle of sentence but in the Persian sentences is in the end of sentence. Moreover, the sentence boundary detection is an extra task that can be ignored. At first, the jagged matrix of TW is created according to meaning of including words of W like figure 3. The main aim is computing similarity between T and TW. For this purpose, each word in T is probed in the TW matrix and a vector (indi) based on indices of matching words such as (1) is created.

}{ jii TWtjind ∈= | (1)

The similarity measurement is inspired from crystallization. The crystallization process consists of two major events, nucleation and crystal growth1. The main idea is creation of the fragments of text like crystal. indi vectors are equivalent the solvent and words on them are solute. For crystal growth we need a lot of nucleus that here we use the words in T which their indi vectors contain only one word. After define nucleuses, we should describe crystal growth process that perform based on T vector. Relation (2) describes it.

=

otherwise nucleus a index tonearest in W equal havet don' tif0

nucleus is t

i

ii

i

ifindCS

(2)

CS is the resembled vector to the T vector that contains nearest equvalence to the T words. The similarity amount of this Resembled vector must be measured. For this, we use an adoptive euclidean distance[17] . Euclidean distance is computed for two vectors like relation (3), but in adoptive form we want to influence similarity with word situation distance in CS vector.

21

22i1i21 d -d),(

= ∑

i

DDL (3)

If most of a fragment's words of text in T are neighbors in CS, there are most probable that these two fragments are same. For computing amount of similarity we use the relation (4).

Classifier

English documents

Persian document

Sorted documents

Classifier

Dictionary

Translated

Similarity evaluating

Rate of similarity


128

∑= +

−+

=T

i pii

slCSCSWTsim

1 1 )])([1(

1),( (4)

T and W vectors already are illustrated. CSi shows existence of ti word of T in W text fragment and in the case of existence, it shows the location of that word in W. sl is the average sentence length, where if distance of two words is equal with sl, the similarity amount equals with 0.5. p is the slope of curve, which the great value express sharp slope and diminishes the effect of two far words in similarity evaluating. Figure (4) shows the relation (4) behavior with various numbers for |CSi+1- CSi|(distance of two words). There are several details in computation and implementation such as treatment with words with CSi zero value that here these details have not been explained.

0

0.2

0.4

0.6

0.8

1

1.2

1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49

distance

effe

ct o

n si

mila

rity

Figure 4. effect of distance on similarity with sl= 15 and

p=8 This fact is important that how far the neighbor words in original paragraph are from each other in the suspicious paragraph. Suppose tk and tk+1 are adjacent in T and wi and wi+d are their equivalents in W. if d or in other words the distance of two words were closed to each other, we can consider wi and wi+d in an identical sentence. By generalizing this fact, the W and T fragments are equal if most of W words equivalences are neighbor in T. in other side of coin; we restate T words by illustrated reconstruction method in CS. If CS contents come from different part of text or various W, we can not find a unique same for T. Total similarity between two documents with T and W fragments is computed by relation (5) that there Pdoc and Edoc are respectively Persian and corresponding English document.

∑∑=i j

ji

EdocPdocWTsim

EdocPdocsimilarity),min(

),(),( (5)

5. Experimental result Dataset gathering in bi-lingual research is the main and time consuming challenge especially if you are among the beginners in that domain. In order to evaluate the proposed approach we used a parallel corpus containing 200 documents in Persian and English that were collected and translated from internet then converted into .txt format. This corpus contains documents in English and exact translate of them in Persian from various genre such as social short

stories, computer science texts in scope of computer networks and text mining. Furthermore, 100 Persian and 100 English documents without corresponding translation in corpus were added in same genre. These added documents were used for false positive evaluation. For classification task, we supposed three types of documents so called “story”, “network” and “text”. When there is no lack of reports on measurement of semantic similarity of concepts, evaluation of the performance of semantic similarity measure between cross-lingual documents has not yet been standardized because there is no universally recognized benchmark[3] . However, a simple thumb of rule is that a document should be very similar to its high-quality translation. With this assumption, the evaluation criterions adopted in this paper, are the rate of successful match between the Persian documents and their parallel English documents (true positive results) and furthermore maximum similarity rate of documents that don’t have corresponding translation in corpus (false positive results). Persian document (DPi) is classified by Persian classifier and all English documents from same class for using in similarity computation are extracted. The most similar document along similarity rate of it for each Persian document is determined. This is rational that we expect the similarity coefficient in unparallel document would be closed to 0 and in 200 parallel documents near to 1. For final result the mean of all similarities in parallel and unparallel groups are calculated as the table (2) shows.

Table 2: mean of similarities in parallel and unparallel documents with sl= 15 and p=8

Tf based method Proposed method Parallel 0.76 0.81 unparallel 0.35 0.09

First column in table (2) represents tradition method results in which similarity of two documents is computed based on words occurrences of a document in other one. In other words this is a term frequency based method, also in this case rephrasing is not performed. Considering, tf based method has high precision in parallel documents matching, but its error rate in unparallel documents are excessive. Second column shows proposed method results that in this method, term frequency doesn't have any influence on similarity, and similarity is computed based on distance. As results discloses, similarity rate for each group are meaningful.

6. Conclusion Plagiarism is a challenging problem in research domain and academic comportment. For distinguishing plagiarist from the researcher, automatic detection of plagiarism is necessary. In this paper we developed a system for detection of exact translation from English to Persian. In bi-lingual plagiarism, in which system dealing with two different languages that each word in one language has several equivalences in the other. Thus selecting a suitable meaning is essential. In proposed method, this is realized by rephrasing task that inspired from crystallization. Structural


129

differences between investigated languages are another challenge in bi-lingual tasks. That this problem is solved by using a distance based similarity measurement. Obtained outcomes are very promising to detect exact translated texts. Using all validate words in similarity computation is One of the best properties of proposed method

7. References [1] H. Maurer, F. Kappe, B. Zaka, ‘Plagiarism-a survey’,

Journal of Universal Computer Science, 2006, 12, (8), pp. 1050-1084

[2] M. Jalalian, L. Latiff, ‘Medical researchers in non-English countries and concerns about unintentional plagiarism’, Journal of Medical Ethics and History of Medicine, 2009, 2, (2), pp. 1-2

[3] H.H. Huang, H.C. Yang, Y.H. Kuo, ‘A Sense Based Similarity Measure for Cross-Lingual Documents’. Proc. Intelligent Systems Design and Applications, 2008. ISDA'08. Eighth International Conference on2008 .

[4] O. Uzuner, B. Katz, T. Nahnsen, ‘Using syntactic information to identify plagiarism’. Proc. Proceedings of the 2nd Workshop on Building Educational Applications Using NLP2005 pp.37-44

[5] A. Barron-Cedeno, P. Rosso, D. Pinto, A. Juan, ‘On cross-lingual plagiarism analysis using a statistical model’, Proc. of PAN-08, 2008

[6] C.H. Lee, C.H. Wu, H.C. Yang, ‘A Platform Framework for Cross-Lingual Text Relatedness Evaluation and Plagiarism Detection’. Proc. Innovative Computing Information and Control, 2008. ICICIC'08. 3rd International Conference on 2008 pp. 303-307.

[7] A. Selamat, H.H. Ismail, ‘Finding English and translated Arabic documents similarities using GHSOM’2008 pp. 460-465

[8] N. Gustafson, M.S. Pera, Y.K. Ng, ‘Nowhere to hide: Finding plagiarized documents based on sentence similarity’. Proc. IEEE/WIC/ACM International Conference on Web Intelligence and Intelligent Agent Technology, 2008. WI-IAT'082008 pp.690-696

[9] D. Zhang, W.S. Lee, ‘Question classification using support vector machines’. Proc. Proceedings of the 26th annual international ACM SIGIR conference on Research and development in informaion retrieval2003

[10] C. Enhong, Z. Zhenya, A. Kazuyuki, W. Xu-fa, ‘An extended corner classification neural network based document classification approach’, Journal of Software, 2002, 13, (5), pp. 871-878

[11] M. Mohammadi, H. Alizadeh, B. Minaei-Bidgoli, ‘Neural Network Ensembles Using Clustering Ensemble and Genetic Algorithm’. Proc. Convergence and Hybrid Information Technology, 2008. ICCIT'08. Third International Conference on2008 pp. 761-766

[12] M. Mohammadi, B. Minaei-Bidgoli, ‘Using CC4 neural networks for Persian document classification’. Proc. 2nd Iran Data Mining Conference(IDMC008) Tehran 2008

[13] K. Sheykh Esmaili, A. Rostami,: ‘A list of persian stopwords’. Proc. Technical Report No. 2006-03,

Semantic Web Research Laboratory, Sharif University of Technology, tehran - iran 2006.

[14] G. Salton, A. Wong, C.S. Yang, ‘A vector space model for information retrieval’, Journal of the American Society for information Science, 1975, 18, (11), pp. 613-620

[15] H.F. Ma, Q. He, Z.Z. Shi, ‘Geodesic distance based approach for sentence similarity computation’. Proc. Machine Learning and Cybernetics, 2008 International Conference on2008 pp. 2552-2557

[16] C.D. Manning, H. Schütze, ‘Foundations of statistical natural language processing’ , MIT Press, 2002.

[17] E. Greengrass, ‘Information retrieval: A survey’, University of Maryland, Baltimore County, 2000.


130

Experimental Comparison of Fiber Optic Link Impact on Shift Keying Modulations

Dr. Mohammad Samir Modabbes

Aleppo University, Faculty of Electrical and Electronic Engineering,

Department of Communications Engineering, Syria [email protected]

Abstract: In this paper an experimental comparison of fiber optic communication link impact on shift keying modulated signals in presence of external noise is analyzed. The relation of BER versus SNR for each type of shift keying modulation is shown. A comparison measurement of voltage loss through an optical fiber link with different lengths is calculated. Results show that phase shift keying (PSK) modulation offers the advantages of being more immune to light scattering and absorption losses of fiber optic link and external noise than other shift keying modulations and is preferred to use in optical fiber transmission.

Keywords: BPSK modulation, Fiber optic Communications,

noise, BER analysis

1. Introduction Fiber optics is widely used today and is becoming more common in everyday life. Its greatest use is in the field of communications for voice, video and data signals transmission through small flexible threads of glass. These fiber optic cables far exceed the information capacity of coaxial cable. They are also smaller and lighter in weight than conventional copper systems and are immune to electromagnetic interference and crosstalk [1, 2]. There are two main factors to consider when transmitting signals through optical fibers: signal to noise ratio (SNR) and the bit error rate (BER).

2. Objective Both absorption and scattering in optical fiber are depending on light wavelength and the nature of signals traveling through it, and are specified by manufacturers in decibels per kilometer.. In [3] the performance of eight-channel 10.7 Gb/s systems was computed using advanced modulation formats, four non return-to-zero (NRZ)-type and three return-to-zero (RZ)-type formats with different phase characteristics. In [4] a novel phase shift keying technique was proposed that uses optical delay modulation for fiber-optic radio links. Using only a 2x1 switch and a delay line, this technique enables modulation of a millimeter-wave carrier at bit rates of several gigabits per second or higher, where high-speed devices are not needed. But only binary phase shift keying was experimentally demonstrated. Therefore studying optical fibers effect on different signals traveling through it is of a great importance. In [5] an exact analysis of BER performance of generalized hierarchical PSK constellations under imperfect phase or

timing synchronization over additive white Gaussian noise channels is made.

3. Methodology Fiber attenuation in dB per km for different types of shift keying modulated signals is determined using phototransistor for measuring relative light power.

Light falls on phototransistor controls its photo current, which is proportional to the relative light power, by measuring light power (voltage or current) detected by Phototransistor for two different lengths of optical fibers we can calculate the optical power ratio between the two lengths of that cable. Then by dividing the power ratio between the two cables by the length deference we can calculate the optical power loss in dB/km as following:

Power loss = )/())/log(10( 1221 LLpp − [dB/km] (1) Where: P1 output power of first optical fiber P2 output power of second optical fiber L1 length of first optical fiber L2 length of second optical fiber

The effect of optical power loss and amplitude noise for each type of shift keying modulated signals can be measured by determining the relation of BER versus SNR. BER is the number of incorrect bits received in reference to the total number of bits transmitted:

BER = incorrect received bits / total transmitted bits (2)

SNR is the ratio of input signal amplitude (Vinput ) to noise signal amplitude (Vnoise ) in decibels [1,5]:

SNR = 20 log (Vinput / Vnoise) [dB] (3)

4. Experimental Setup The experimental measurements were conducted on telephone channel simulator, in lab environment, using experimental board No. AS91025 from LABVOLT company, with glass fibers (1 & 5m) and input signal frequency about half kilohertz as shown in figure (1), real measurements environment require long fibers (several kilometers) and high frequencies (tens of Giga hertz), but even though the results on the experimental board can be


131

verified and expanded to cover the real fiber transmission system. Modulator produces different types of shift keying modulated signals (ASK, OOK, FSK, PSK) with carrier frequency 2.4 kHz (approximately five times the highest frequency of the baseband signal). Then modulated signals are transmitted by the fiber optic transmitter (FOT) and received by the fiber optic receiver (FOR). FOT has an infrared LED light source with a peak wavelength 820 nm and with typical spectral bandwidth 45 nm (50% less than peak wavelength). As it's known light source speed affects the bandwidth of a fiber optic system. The greater the bandwidth requirement, the greater the need to turn the light source on and off more quickly. Therefore light source speed is defined in terms of rise time (tr) and the following equation approximates the maximum bandwidth (Bwmax) [6, 7]:

Bwmax = 0.35/tr [Hz] (4)

Figure 1. Telephone channel simulator

LED has a typical rise time of 3 ns, which means that the FOT maximum bandwidth is approximately 120 MHz A random pulse generator (RPG) injects a variable amount of noise into the received signal after FOR stage and has frequencies up to 600 Hz. Error counter measures the number of incorrect bits received (error count), where The transmitted and received data are compared bit by bit in a comparator (XOR gate), if the bit is not match an error pulse is generated, the errors are totalized in a counter over a fixed time period or frame (106 ms is the time required for 128 data bits) generated by one shot. Each time when the counter is reset a 106 ms one shot is triggered, and the error pulses from the XOR gate are totalized by the counter only during 106 ms frame. A performance comparison of shift keying modulation transmission through a glass fiber optic link with graded index (62.5/125 μm) was made in the presence of noise at cutoff frequency of low pass filter 1.5 kHz; total number of bits transmitted 128 bits; SNR was calculated for 4Vp-p (2.828 Vrms) input signal amplitude and variable amplitude of noise signal.

5. Results and discussions 5.1 Voltage loss

Table (1) shows the measured received light voltage for different modulated signals detected by phototransistor and transmitted by different length of glass fibers:

Table 1: Measured received light voltage Fiber type The received light voltage [mV]

ASK OOK FSK PSK Glass

fiber(1m) 4.46 4.66 4.26 4.46

Glass fiber(5m)

4.44 4.65 4.25 4.45

Table (2) shows the calculated optical voltage loss in dB/m for different modulated signals as in equation (5):

Table 2: Calculated optical voltage loss

Loss = )/())/log(20( 1221 LLvv − [dB/m] (5) By comparing these results we conclude that: the received light voltage and voltage loss are affected by fiber length and fiber material. OOK and PSK modulations have minimum voltage loss (~4dB/km); while the worst ASK modulation has maximum voltage loss (~9dB/km).

5.2 BER vs. SNR

The graphics below show BER vs. SNR for the coherent detection of different types of shift keying modulation for different noise amplitude using different length of glass cables:

6 8 10 12 14 16 180

0.1

0.2

0.3

0.4

0.5

SNR[dB]

BE

R

ASK-1m glassOOK-1m glassFSK-1m glassPSK-1m glass

Figure 2. BER versus SNR using 1m glass fiber

Fiber type

Loss dB/m ASK OOK FSK PSK

Glass fiber

9.7x10-3 4.6x10-3 5.1x10-3 4.8x10-3

INPUT

MOD FOT

FOR

Fiber

DEMOD OUTPUT COMP

RPG

Carrier

ERROR


132

6 8 10 12 14 16 180

0.1

0.2

0.3

0.4

0.5

0.6

SNR[dB]

BER

ASK-5m glassOOK-5m glassFSK-5m glassPSK-5m glass

Figure 3. BER versus SNR using 5m glass fiber

By comparing all graphics in figures (2 & 3) we conclude that attenuation increases with fiber length, ASK modulation has the worst performance, OOK and FSK modulations have steady attenuation level in all SNR range, while in PSK modulation attenuation degreases remarkably as SNR ratio increases.

6. CONCLUSIONS An experimental comparison of fiber optic communication link impact on shift keying modulated signals (ASK, OOK, FSK, PSK) in the presence of external noise is analyzed. The experimental measurements were conducted on telephone channel simulator using board No. AS91025 from LABVOLT Company. Fiber attenuation is determined using phototransistor for measuring relative light power. Measurements results show that:

1. Losses due to light scattering and absorption can be determined by comparing different lengths of identical fiber.

2. Voltage Loss increases with fiber length and affected by wavelength and fiber material

3. ASK modulation has maximum loss in glass fiber; while FSK & OOK modulations have less and equal value.

4. For ASK and OOK at a given signal to noise ratio as the difference between the two levels of the carrier for “1” and “0” states increases the BER decreases which improves noise immunity and attenuation

5. PSK modulation has the best performance than other digital modulations in glass fiber, because attenuation degreases to minimum values as SNR ratio increases.

These results led us to conclude that PSK modulation offers the advantages of being more immune to light scattering and absorption losses of fiber optic and external noise than other shift keying modulations and is preferred to use in data transmission systems with optical fiber links.

References [1] Bernard Sklar., 2001- “Digital Communications

Fundamentals and Applications”, Prentice- Hall, New Jersey.

[2] Harold B. Killen., 1991-”Fiber Optic Communications”, Prentice-Hall, New Jersey

[3] Yihong M., S. Lobanov and S. Raghavan., 2007- "Impact of Modulation Format on the Performance of Fiber Optic Communication Systems with Transmitter Based Electronic Dispersion Compensation", Optical Society of America

[4] Y. Doi, S. Fukushima, T. Ohno, Y. Matsuoka, and H. Takeuchi., 2000- "Phase Shift Keying Using Optical Delay Modulation for Millimeter-Wave Fiber-Optic Radio Links", Journal of light wave Technology, Vol. 18, No. 3

[5] P. K. Vitthaladevuni, and M.S. Alouini, 2005- “Effect of Imperfect Phase and Timing Synchronization on the

Bit-Error Rate Performance of PSK Modulations”, IEEE Transactions on communications, Vol. 53 No. 7 [6] H. Meng, Y. L. Guan, and S. Chen., 2005- “Modeling

and Analysis of Noise Effects on Broadband Power-Line Communications", IEEE Transactions on Power Delivery, Vol. 20 No. 2

[7] A. Demir, 2007- “Nonlinear Phase Noise in Optical Fiber-Communication Systems”, Journal of Lightwave Technology, Vol. 25, No. 8

Author Profile

Mohammad Samir Modabbes received the B.S. degree in Electronic Engineering from University of Aleppo in 1982. M.S. and Ph.D. degrees in Communications Engineering from High Institute of Communications in Sankt Petersburg in 1988 and 1992, respectively. He is working as Associate Professor at the Faculty of Electrical and Electronic Engineering, Department of

Communications Engineering, University of Aleppo Syria. Since 2006 he is working as Associate Professor at Qassim University, College of Computer, Department of Computer Engineering in Saudi Arabia. His research interests are: Analysis and performance evaluation of Data transmission Systems, Wireless Communication Systems


133

A New Cluster-based Routing Protocol with Maximum Lifetime for Wireless Sensor Networks

Jalil Jabari Lotf1, seyed hossein hosseini nazhad ghazani2, Rasim M. Alguliev3 Institute of Information Technology of Azerbaijan National Academy of Science, Baku, Azerbaijan

[email protected], [email protected], 3 [email protected]

Abstract: In recent years, one of the main concerns in wireless sensor networks has been to develop a routing protocol which has an efficient and effective energy yield. Since the ability of sensor nodes is limited, the energy retaining and a long life network is an important issue in wireless sensor networks. In the present paper, the routing protocol MLCH, which has the characteristics of hierarchical routing and self-configuration, has been proposed. MLCH corrects the current routing protocols according to radio beams and the number of cluster members in several directions. In this method, the clusters will be distributed evenly in the network. The efficiency of the suggested algorithm was measured by simulations and comparing the resulted outcomes with the previous methods’ resulted outcomes. The results show that the suggested algorithm has a high efficiency.

Keywords: Wireless sensor Network, Routing, clustering.

1. Introduction A sensor network is composed of a lot of sensors which are accumulated to examine a state in an environment. As the name suggests, in sensor networks, the nodes have the responsibility of sensing and sending the received information through the network to a Base station or sink. Also the nodes usually have limited energy resources which are not changeable or rechargeable after they finish, because of the unavailability of the nodes or the dominant conditions. Thus, the important issue in the design of routing algorithms in these networks is that they are optimal in energy usage. In order to decrease the energy consumption appropriately, we need the data distribution techniques with an efficient energy yield. According to [1], there are three principle methods for data distribution: medium storage, local storage and external storage. In medium or local storage, the data are kept inside the network and the questions are sent to the nodes which store the desired data. In this paper we will study the external storage in which we need to store the data in a constant center out of the network. The life span of a sensor network can be defined as the distant between the time when the network starts to work and the time when the first or the last sensor loses its energy. Clustering is one of the energy yield techniques to increase the lifetime of the sensor network [2, 3]. Increasing the lifetime of the sensor is often accompanied by the data combination [4]. Each cluster chooses a node as a head-node, and sends the collected data first to the head node and then to the base station. The head-nodes can combine the data in sensors to minimize the amount of data and sends them to the base station. And when the size of the network

increases, the clusters can be organized hierarchically. MLCH is a protocol based on clustering, which minimizes the waste of energy in wireless sensor networks. This new algorithm creates the clusters according to radio range and their even distribution throughout the network. The proposed method, tries to improve the clusters’ topology in order to minimize the energy consumption. In section 2, some of the previous works on routing protocols based on wireless sensor networks will be summarized. In section 3, the suggested algorithm has been explained in detail. Section 4 explains the experimental outcomes and poses the system efficiency and section 5 will finish this paper. In this section we will conclude our discussion and suggest works to be done in the future. .

2. Previous works LEACH [5] is a protocol based on clustering, which uses the random rolling of the head-nodes in order to distribute the energy burden equally between sensor nodes in a network. After the clusters are created, the head-nodes spread the timing algorithm TDMA which identifies the order in which the members of clusters being sent to the head-node. Each node sends the data to head-node in a unique period of time. When the last node sends the data, a random head-node is chosen for the next period. TEEN [6] is a protocol based on protocol LEACH with two limitations. First: as soon as the absolute amount of the data exceeds a certain amount HT, the node which receives it, should activate the sender and report this. Second: when the changes in the received amount is higher than a specific amount ST, the node is forced to reactivate the sender and report the received data. The node will report the data, when the received amount is higher than HT or the change in the amount of the received data is higher than ST. PEGASIS [7] is one of chain protocols based on protocol LEACH. It is similar to multi-layer chain protocols which creates some chains from the top of nodes. Each head-node sends and receives its data, only using one neighbor and the collected data moves from each head-node to the other one. Finally, only a certain head-node sends them to the base station. HEED [8] is a distributed clustering method which uses the same methods as the previous algorithms. The only difference is that it uses several parameters instead of one parameter to form its sub-networks. The head-node selection in this algorithm is based on the amount of sending power


134

Steady-state phase Setup Phase

Time

in each node which is used as a primary parameter. Also the second parameter is used to improve the communication inside each cluster and identifying that if a node is located in the communicative range of several head-nodes, which one will be chosen as its head-node. MLCH is a protocol based on a clustering which tries to minimize waste of energy in sensor networks. The key characteristics of MLCH are as follows:

1. Network lifetime increases 2. It has self- configuration 3. The head-nodes are distributed evenly 4. Sending is hierarchical.

Since the energy consumption has a direct relation with the square root of the distance, MLCH tries to cluster the sensors in a way that the sensors’ density is less and the distance between them in the clusters is not big. In this protocol, the head-nodes are chosen in a way that they can cover all the area and the head-nodes can send their data to the base station hierarchically. Thus the network lifetime is increased and the head-nodes are distributed evenly and the load will be balanced.

3. The proposed algorithm Here we will discuss the problems related to one of the most common hierarchical routing protocols for wireless sensor networks and then will introduce the suggested algorithm.

3.7 The problems with LEACH LEACH uses a self- configuration method and decreases the energy consumption a great deal. However, it has some deficiencies too. First, LEACH doesn’t study the distribution of head-nodes because it chooses the head-nodes randomly and it doesn’t have any control over the evenly distribution of them in the environment. Figure (1) is an example, showing the head-nodes in the environment which has been created by LEACH.

Figure 1. 5 random choices of head-nodes in LEACH

The black nodes show the head-nodes. All of them are distributed in the top right corner. So the nodes in the left side consume a lot of energy to establish a connection with those head-nodes. The number of nodes in each cluster has an unequal distribution. In figure 1, the head-node 1, doesn’t include more than 5 clusters. While the head-node 5, includes 50 members. So its energy will be finished very

quickly. The nodes which are located farther than head-node 5 also will consume a lot of energy to send the data. After all data have been collected from the members by the head-node, it sends them directly to the base station. It is clear that those head-nodes which are far from the base station will have problems with the direct sending of the data and they cannot maintain the needed energy.

3.8 Clustering with maximum lifetime To solve the above problems in LEACH, the clustering method and the kind of connections between the head-nodes and base station should be changed and modified. Basically we use MLCH to solve these two problems. The structure of clustering in MLCH avoids the unequal distribution of the members of the clusters. Also the hierarchical routing pattern avoids the long term direct connection between the head-nodes and the base station. Our proposed algorithm includes a setup phase and a steady-state phase (figure 2). Each period includes setup phase and steady-state phase. According to the algorithm, in setup phase, the head-nodes are distributed evenly and in steady-state phase, first the clusters are created and then the sensor nodes send the sensed data to the cluster’s head-node which they belong to. Each of the head-nodes combines the received data from the cluster nodes with its own data and sends it to the base station. Whenever the head-node in a cluster stops working, “steady-state” will be finished and thus the system will return to the setup phase. The system will continue this cycling until all the sensor nodes present in the network stop working (or the simulation time is finished).

Figure 2. Round of sensor network operation

3.2.7 Setup phases Node number 1, sends a “Hello” message to its neighbors. The TTL of the “Hello” message is set in a way that we need only to recollect the neighbors of only one node. Also the radio range has been adjusted on a certain sending range and this limits the cluster area conditions. (e.g. a circle with A meter radiant). This message is to inform the nodes which are located in the area concerned. “I am a head-node in the layer X”. All the nodes which are smaller than the radiant specified (e.g. B meter radiant) will receive and record this message. Furthermore, such nodes never spread their received message because we emphasize that only one head-node is present in one radio range. Then we have the nodes which are located in the distance between the sense radiant and radio range (in the assumptions we have B to A meter) and they receive this message. Each of these nodes which are chosen from among the nodes located in a distance less than the sense radiant and have the maximum ability will be

1

2

3

4

5


135

chosen as a head-node in the next layer and they will be the new heads in the next layer. Also for each of these head-nodes, a successor will be chosen to start working when the first one has some problems or can’t work well and this is done to increase the error durability in the network. After repeating these steps, we have distributed our own head-nodes evenly, as is shown in figure 3.

Figure 3. even distribution of head-nodes in the

environment

3.2.8 Steady-state phase This phase includes three steps: installation step, permanence step, and sending step. In the previous phase we have chosen our head-nodes so in the installation step they are already distributed evenly. Now we should create our clusters. Each head-node can choose sensors within a radio range to be its cluster’s member. Then the TDMA time window will be timed for each member of the cluster in each period. In this step, each node will turn on the receiver as like as the LEACH. Then the head-node will spread a message which includes the timing window TDMA. Each member of the cluster will have the timing window which belongs to it. And the amount of remaining ability of each sensor (to choose a head-node for next period) along with the received data in the defined timing window will be reported to the head-node. Permanence step: The clusters have been created and the TDMA timing has been done, then we can transfer the data. With this assumption that the nodes always have some data to be sent, they send their energy and their received data in their allocated window to the head-node. The cluster nodes adjust their sent energy dynamically and according to the time span of sending the message and the size of the sent messages. In permanence step, only the head-node will always turn on its receiver. While the cluster members will turn on their receivers, only at the time allocated. Sending step: Regarding the fact that most part of the sent energy is equal with the square root of the distance, when a head-node (e.g. A) wants to send its data to the base station, first it does some calculations. It calculates the function d(x) for all other head-nodes first. The function d(x) is the

distance of the head-node A from the base station through a head-node like X.

( ) 2sin

2kxxA ddxD −− += x∈{all other headnodes} (1)

Then the minimum amount of these functions will be chosen and it will be compared with the square root of the distance of head-node A from the base station.

2sin))(( kAdxDMin −≤ (2)

If we find a medium head-node (like B) which consumes less energy, then the head-node A sends its data to that medium head-node. If this is not done, it will send the data to base station like LEACH. When the data reaches to the head-node B, the above mentioned algorithm will be repeated. This process will continue until the data reach the base station. This sending method will improve the energy consumption. Because first if we have a lot of head-nodes and want to send our data to the base station through multi-hop method, the cluster head which is nearer to the base station will lose its energy sooner and this causes the head-nodes to stop working. Second if we want to send the data directly as the LEACH, then if a cluster head is farther than the base station, it loses its energy sooner than the other head-nodes and it stops working. But in our protocol we use a combination of these which causes to consume much less energy than the other routing protocols.

4. Simulation and analysis

4.1 The simulation environment The following amounts are the parameters used in the simulations. The network size is 100*100 m2 with 100 sensor nodes. The number of members in one cluster is 10, 15 or 20. We will use a simple radio model with a transfer speed of 40 kbps. Each size of the data is equal to 36 bytes. The network topology is created with a random even distribution.

4.2 Efficiency measurement and analysis The main operations of sensor nodes in this system have been divided into some periods. Each period includes a setup phase and a steady-state phase. In setup phase and according to clustering algorithms, the network is divided into several clusters and the routing is based on the cost and only regarding the head-nodes and the base station the data collection will be done (figure 3). In data sending phase (static state), the sensor nodes will produce the sensed data and send them to the head-node of the cluster which they belong to. Each head-node combines the data received from the cluster nodes with its own data and sends them to the base station according to the rout created by the steady-state phase.

5-1

4-2

4-1

3-1

3-2

2-1

2-2 1-1 3-3

4-3

One Two Three Four

Five


136

First Round

Cluster formation using clustering

algorithms

Cluster-head selection

Yes

No

Go to steady phase

Cluster-head

Create Cluster and Node senses data

Node forward sensed data to cluster-head

Cluster-head aggregates data

Yes

No

Cluster-head sends aggregated

data to sink

Go to setup

Figure 4. The process of sensor network setup phase

Figure 5. The process of sensor network “steady-state” phase

To measure the proposed algorithm, 6 different tests according to the table (1) have been done.

Table 1: The characteristics of tests done

Test number

Number of Sensor nodes

The time duration for simulation

(seconds)

Number of members

1 100 1000 10 2 100 1000 15 3 100 1000 20 3 200 1000 10 5 200 1000 15 6 200 1000 20

The results of the experiments are shown in figures (6) to (11). The show results are the averages for 50 times of simulations for each one of the experiments.

0

20

40

60

80

100

120

1 95 189 283 377 471 565 659 753 847 941

LEACHMLCH

Figure 6. Comparing Protocols LEACH and MLCH Number of active nodes:100 , Number of members:10

0

20

40

60

80

100

120

1 148 295 442 589 736 883

LEACH

MLCH


0

20

40

60

80

100

120

1 100 199 298 397 496 595 694 793 892 991

MLCH

LEACH



137

0

50

100

150

200

250

1 95 189 283 377 471 565 659 753 847 941

LEACHMLCH


0

50

100

150

200

250

1 95 189 283 377 471 565 659 753 847 941

LEACHMLCH


0

50

100

150

200

250

1 95 189 283 377 471 565 659 753 847 941

LEACHMLCH


5. Conclusions and works in the future Here we suggested a new protocol based on clustering with maximum lifetime for wireless sensor networks. MLCH improves LEACH by using a very equally distributed cluster and decreasing the unequal topology of the clusters. MLCH uses the radio range to create a cluster in a certain environment. And also it modifies the connection distance of the head-nodes with the base station through a

hierarchical tree. In the future, we want to solve the problems dealing with the simultaneous sending of the nodes’ data. Creating thousands of nodes and sending the data simultaneously is difficult and one of the resolutions can be to use CSMA/CA instead of TDMA. Also the yield parameters’ characterization can be an alternative resolution. Finding the yield parameters, such as the number of members in each cluster and the radio range for MLCH is also important.

References [1] S. Ratnasamy, D. Estrin, R. Govindan, B. Karp, L. Yin

S. Shenker, and F. Yu. “Data-centric storage in sensornets”. Proceedings of the ACM First Workshop on Hot Topics in Networks, 2001.

[2] W. Choi, P. Shah and S. K. Das,“A Framework for Energy-Saving Data Gathering Using Two-Phase Clustering in Wireless Sensor Networks”, Proceedings of the Mobile and Ubiquitous Systems, Boston, MA, August 2004.

[3] O. Younis and S. Fahmy, “Distributed Clustering in Ad-hoc Sensor Networks: A Hybrid, Energy-Efficient Approach”, Proceedings of the IEEE INFOCOM 2004, Hong Kong, China, March, 2004.

[4] D. Hall, “Mathematical Techniques in Multisensor Data fusion,” Artech House, Boston, MA, 1992.

[5] W. Heinzelman, A. Chandrakasan, and H. Balakrishnan. “Energy-Efficient Communication Protocols for Wireless Microsensor Networks (LEACH)”. Proceedings of the 33rd Hawaii International Conference on Systems Science- Volume 8, pp. 3005-3014, January 04-07, 2000

[6] A. Manjeshwar and D. Agrawal, “TEEN: a Routing Protocol for Enhanced Efficient in Wireless Sensor Networks,” Proceedings of the 15th International Parallel and Distributed Processing Symposium, pp. 2009-2015, 2001.

[7] Lindesy and C. Raghavendra, “PEGASIS: Power-Efficient Gathering in Sensor Information System,” Proceedings of the IEEE Aerospace Conference, pp. 1-6, March 2002.

[8] Ossama Younis and Sonia Fahmy, “Distributed Clustering for Scalable, Long-Lived Sensor Networks”, Proceedings of the 9th Annual International Conference on Mobile Computing and Networking, ACM MobiCom, San Diego, CA, September 2003.


138

Design of Scheduling Algorithm for Ad hoc Networks with Capability of Differential Service

Seyed Hossein Hosseini Nazhad Ghazani1, Jalil Jabari Lotf 2,Mahsa Khayyatzadeh3, R.M. Alguliev4

1Institute of Information Technology of ANAS,

Azerbaijan Republic [email protected]



3Departmant of Electrical Engineering, Urmia University

Urmia, Iran [email protected]



Abstract: A mobile ad hoc network (MANET) is a collection of mobile nodes that can communicate with each other without using any fixed infrastructure. To support multimedia applications such as video and voice MANETs require an efficient routing protocol and quality of service (QoS) mechanism. Again Ad-hoc network is mainly constrained by the bandwidth. The proposed scheduling in this research adopting the Contention Window size of flows and also services the flows by select the fixed or minimally mobile nodes that provide network backbone access. The novelty of this model is that it is robust to mobility and variances in channel capacity and imposes no control message overhead on the network, again in calculate CW size, attention to collision rate and flow’s allocated QoS.

Keywords: Ad Hoc, Scheduling, QoS, Routing.

1. Introduction A MANET2 is essentially a special form of distributed system with the features like no fixed topology, no fixed connectivity, varying link capacity, without any central control and is constrained by the lack of resources. Due to the above items routing in such networks experiences link failure more often. Hence, a routing protocol supports QoS requires to consider the reasons of link failure to improve its performance. Many routing schemes and frameworks have been pro-posed to provide QoS support for ad hoc networks [1, 2, 3, 4, 5]. Among them, INSIGNIA [1] uses an in-band signaling protocol for distribution of QoS information.

SWAN [2] improves INSIGNIA by introducing an Additive Increase Multiplicative Decrease (AIMD)-based rate control algorithm. Both [3] and [4] utilize a distance vector protocol to collect end-to-end QoS information via either flooding or hop-by-hop propagation. CEDAR [5] proposes a core-extraction distributed routing algorithm that

maintains a self-organizing routing infrastructure, called the “core”.

These approaches do not consider the contentious nature of the MAC layer and the neighbor interference on multi-hop paths. This leads to inaccurate path quality prediction for real-time flows. Additionally, most of the work does not consider the fact that a newly admitted flow may disrupt the quality of service received by ongoing real-time traffic flows. Recently, other work has proposed the performance improvement of MAC protocols and the support of service differentiation. Many of these approaches specifically target IEEE 802.11 [6]. For example, studies in [1,6,12,16] propose to tune the contention windows sizes or the inter-frame spacing values to improve network throughput, while studies in [1, 4, 14, 23, 31] propose priority-based scheduling to provide service differentiation. To support both types of applications in ad hoc networks, effective QoS-aware protocols must be used to allocate resources to flows and provide guarantees to real-time traffic in the presence of best effort traffic [15]

2. QOS FRAMEWORK [16]

2.1 Targeted network modification This framework targets the support of real-time traffic in large-scale mobile networks. In these environments, fixed wireless routers may be placed to serve as a network backbone. The majority of proposed routing protocols place no preference on the selection of paths with regard to node mobility, i.e., highly mobile nodes have the same likelihood of inclusion on a communication path as stationary nodes. While best-effort traffic may be more tolerant to these events, the quality of real-time traffic will be significantly degraded and is likely to become unacceptable. The utilization of fixed wireless routers in these networks will


139

greatly improve the quality of real-time traffic by the elimination of intermediate link breaks. Figure 2 illustrates an example network where real-time and best-effort traffic utilize different routes [16].

Figure 1. Functionality of the framework at IP and MAC layers.

Figure 2. An example of the routes for different traffic.

2.2 Call setup for real-time traffic When a real-time flow is requested, a call setup process is needed to acquire a valid transmission path with satisfied QoS requirement. Call setup also enables effective admission control when the network utilization is saturated. This requires accurate estimation of channel utilization and prediction of flow quality, i.e., throughput or transmission delay.

The proposed QoS approach is based on model-based resource estimation mechanism, called MBRP [19]. By modeling the node backoff behavior of the MAC protocol and analyzing the channel utilization, MBRP provides both per-flow and aggregated system wide throughput and delay[16]. Call setup Call setup process based on the modified AODV routing protocol, which can be divided into a Request and a Reply phase. In the request phase, the source node sends Route Request messages (RREQ) for the new flow. The RREQ packet reaches the destination if a path with the needed quality exists. During the reply phase, the destination node sends a Route Reply message (RREP) along the reverse path to the source node [16].

2.3 Prioritized medium access Communication in ad hoc networks occurs in a distributed fashion. There is no centralized point that can provide resource coordination for the network; every node is responsible for its own traffic and is unaware of other traffic [16].

In Ad hoc networks, priority scheduling algorithm is based on IEEE 802.11[6].Currently, there are several

approaches that propose to provide service differentiation based on 802.11, by either assigning different minimum contention window sizes ( minCW ) , Arbitrary Inter Frame Spacing (AIFS), or back-off ratios, to different types of traffic. These approaches can all provide differentiation; however, the parameters are typically statically assigned and cannot adapt to the dynamic traffic environment. This reduces the usage efficiency of the network [16].

We propose an adaptive scheme to address trade-off. The basic idea is that, because the state of ad hoc networks can vary greatly due to mobility and channel interference, it is advantageous to adjust the back-off behavior according to the current channel condition.

To achieve service differentiation, as well as to adapt to the current network usage, we combine the collision rate and current QoS of flow with the exponential back-off mechanism in IEEE802.11. To do it, classifies flows into three types: delay-sensitive flows, bandwidth sensitive flows and best effort flows. The delay-sensitive flows, such as conversational audio/video conferencing, require that packets arrive at the destination within a certain delay bound. The bandwidth-sensitive flows, such as on-demand multimedia retrieval, require a certain throughput. The best effort flows, such as file transfer, can adapt to changes in bandwidth and delay. Due to the different requirements of flows, each type of flows has its own contention window adaptation rule [15]. 1) Delay-Sensitive Flows: For a delay-sensitive flow, the dominant QoS requirement is end-to-end packet delay. To control delay, the end-to-end delay requirement d must be broken down into per-hop delay requirements. Each hop locally limits packet delay below its per-hop requirement to maintain the aggregated end-to-end delay below d. For this paper, each node is assigned with the same per-hop delay requirement, d/ m, where m is the hop count of the flow.

)

)(

1(*)()1(

md

nDmd

anCWnCW−

+=+ (1)

Where the superscript n represents the thn update iteration, D denotes the actual peak packet delay at the node during a update period and α is a small positive constant (α=0.1) [15].

2) Bandwidth-Sensitive Flows: For a bandwidth sensitive

flow, the dominant QoS requirement is throughput, which requires that at each node along the flow’s route, the packet arrival rate of the flow should match the packet departure rate of the flow.

))(()()1( nQqnCWnCW −+=+ β (2) where q is a threshold value of the queue length that is

smaller than the maximum capacity of the queue, Q represents the actual queue length and β is a positive constant(β=1). If Q is larger than q, the algorithm decreases CW to increase the packet departure rate to decrease queue length. If Q is smaller than q, the algorithm increases CW to decrease the packet departure rate and free up resources for other flows. As the queue size varies around the threshold


140

value q, the average throughput of the flow matches its requirement [15].

3) Best Effort Flows: Best effort flows are tolerant to

changes in service levels and do not have any hard requirements about bandwidth or packet delay. The purpose of updating the contention window size of best effort flows is to prevent best effort flows from congesting the network and degrading the service level of real-time flows.

)))((1()()1( nFfnCWnCW −+×=+ γ (3) Where f is a congestion threshold for idle channel time,

F is the actual idle channel time and γ is a positive constant (γ=0.1) [15]

When the average idle channel time F is smaller than the threshold value f, the network is considered congested and the contention window size of the best effort traffic is increased to avoid decreasing the service level of real-time traffic. On the other hand, if the network is lightly loaded so that the idle channel time is larger than f, the contention window size of best effort traffic is decreased so that the idle bandwidth can be utilized [15].

Additional to combine the collision rate with the exponential back-off mechanisms, we use the follow algorithm [16]:

TimeSlotCWpricolRrRandoffBack _*]min*)*2(,0[ +=− (4)

where colR denotes the collision rate between a station’s two successful frame transmissions, and pri is a variable associated with the priority level of the traffic. By applying Eq. (4), traffic with different priority levels will have different back-off behavior when collisions occur. Also traffic with same priority levels will have different back-off behavior when collisions occur, with attention to flow current status. Specifically, after a collision occurs, low priority traffic will back-off for longer, and subsequently high priority traffic will have a better chance of accessing the channel.

3. MODEL VALIDATION In this section, we study the behavior of a station with a Markov model, and we obtain the stationary probabilityτ that the station transmits a packet in a generic (i.e., randomly chosen) slot time. This probability does not depend on the access mechanism (i.e., Basic or RTS/ CTS) employed. Consider a fixed number m of contending stations. In saturation conditions, each station has immediately a packet available for transmission. With attention to Markov model (Figure 3), the probability τ is: [18]

(5)

Figure 3. Markov Chain model for the back-off window size.

In Eq.(5) W denotes the Contention Window size of flow. With attention to Eq.(5), probability that the station with low CW versus the station with big CW obtain the channel and transmits a packet, is high. Using the above three contention window adaptation algorithms and Eq.(4), ensures that flows dynamically adjust their contention parameters to meet their own QoS needs with attention to collision rate.

4. CONCLUSION Using the above three contention window adaptation algorithms (1,2,3) and Eq.(4), ensures that real-time flows dynamically adjust their contention parameters to meet their own QoS needs with attention to collision rate. A real-time flow that did not get its required QoS in the past due to competition from other flows decreases its contention window size so that statistically it will have a higher chance to obtain the channel in the future (Eq.(5)). A best effort flow, on the other hand, increases its contention window size when the network is considered busy and hence releases the channel to the real-time flows.

The novelty of this model is that it is robust to mobility and variances in channel capacity and imposes no control message overhead on the network and in calculates CW size, attention to collision rate and flow’s current QoS.

References [1] S. Lee, G.-S. Ahn, X. Zhang, and A. T. Campbell.

“INSIGNIA: An IP-Based Quality of Service Framework for Mobile Ad Hoc Networks”. Journal of Parallel and Distributed Computing, Special issue on Wireless and Mobile Computing and Communications, 60:374–406, 2000.

[2] G.-S. Ahn, A. Campbell, A. Veres, and L.-H. Sun. “Supporting Service Differentiation for Real-Time and Best-Effort Traffic in Stateless Wireless Ad Hoc Networks (SWAN)”. IEEE Transactions on Mobile Computing, vol 1, pp.192–207, July- Septemeber 2002.

[3] S. Chen and K. Nahrstedt. “Distributed Quality-of-Service Routing in Ad-Hoc Networks”. IEEE Journal of Selected Areas in Communications, vol 17, pp 1454-1465, August 1999.

[4] T. Chen, M. Gerla, and J. Tsai. “QoS Routing Performance in a Multi-hop, Wireless Network”. In

∑ −=++

=1

0 )2(12)(

mi

ippWWpτ


141

Proceedings of the IEEE ICUPC’97,vol 2, pp 557—61 , San Diego, CA, October 1997.

[5] P. Sinha, R. Sivakumar, and V. Bharghavan. CEDAR: “A Core-Extraction Distributed Ad Hoc Routing Algorithm”. In Proceedings of the IEEE Conference on Computer Communications (INFOCOM), pages 202–209, New York, NY, 1999.

[6] IEEE Computer Society.”IEEE Standard for Wireless LANMedium Access Control and Physical Layer Specification”. Novermber 1999.

[7] I. Ada and C. Castelluccia. “Differentiation Mechanisms for IEEE 802.11”. In Proceedings of the IEEE Conference on Computer Communications (INFOCOM), Anchorage, Alaska, April 2001.

[8] F. Cal´i, M. Conti, and E. Gregori. “Tuning of the IEEE 802.11 Protocol to Achieve a Theoretical Throughput Limit”. IEEE/ACM Transactions on Networking, vol 8, pp 785-799, December 2000.

[9] T. S. Ho and K. C. Chen. “Performance Evaluation and Enhancement of CSMA/CA MAC Protocol for 802.11 Wireless LANs”. In Proceedings of the IEEE PIMRC,vol 18, pp 535-547, Taipei, Taiwan, October 1996.

[10] H. Kim and J. C. Hou. “Improving Protocol Capacity with Model-based Frame Scheduling in IEEE 802.11-operated WLANs”. In Proceedings of the Ninth Annual International Conference on Mobile Computing and Networking (Mobi COM’03), pages 190–204, San Diego, CA, September 2003.

[11] A. Banchs, X. Perez-Costa, and D. Qiao. “Providing Throughput Guarantees in IEEE 802.11e Wireless LANs”. In Proceedings of the th18 International Teletraffic Congress (ITC-18), Berlin, Germany, September 2003.

[12] V. Kanodia, C. Li, A. Sabharwal, B. Sadeghi, and E. Knightly. “Distributed Multi-Hop Scheduling and Medium Access with Delay and Throughput Constraints”. In Proceedings of the Seventh Annual International Conference on Mobile Computing and Networking (MobiCOM’01), Rome, Italy, July 2001.

[13] R. Rozovsky and P. Kumar. ”SEEDEX: A MAC Protocol for Ad Hoc Networks”. In Proceedings of the

nd2 __ACM International Symposium on Mobile Ad Hoc Networking and Computing (MobiHoc’01), pp 67-75, Long Beach, CA, October 2001.

[14] A. Veres, A. T. Campbell, M. Barry, and L.-H. Sun. “Supporting Service Differentiation in Wireless Packet Networks Using Distributed Control”. IEEE Journal of Selected Areas in Communications, vol 19, pp 2081, October 2001.

[15] Yaling Yang and Robin Kravets. "Distributed QoS Guarantees for Realtime Traffic in Ad Hoc Networks".Technical Report UIUCDCSR-2004-2446, June 2004.

[16] Yuan Sun, Elizabeth M. Belding-Royer, Xia Gao and James Kempf. "Real-time Traffic Support in Large-Scale Mobile Ad hoc Networks" Proc. of BroadWIM 2004, San Jose, CA, October 2004.

[17] Sally Floyd and Van Jacobson. “Random early detection gateways for congestion avoidance”.

EEE/ACM Transactions on Networking, vol 1, pp:397–413, 1993.

[18] G.Bianchi, "Performance Analysis of IEEE 802.11 Distributed Coordination Function", In IEEE Journal on Selected Areas in Communication, Vol 18, pp 535-547, March 2000.

[19] Y. Sun, X. Gao, E. M. Belding-Royer, and J. Kempf. Modelbased Resource Prediction for Multi-hop Wireless Networks. In Proceedings of the st1 IEEE International Conference on Mobile Ad-hoc and Sensor Systems (MASS), Ft. Lauderdale, FL, October 2004.


142

An Optimization Problem for Evaluation of Image Segmentation Methods

Javad Alikhani koupaei, Marjan Abdechiri

Department of Mathematics, Payamenour University, Isfahan, Iran.

[email protected] [email protected]

Abstract: Image segmenting is one of the most important steps in movie and image processing and the machine vision applications. The evaluating methods of image segmenting that recently introduced. In this paper, we proposed a new formulation for the evaluation of image segmentation methods. In this strategy using probabilistic model that utilize the information of pixels (mean and variance) in each region to balance the under-segmentation and over-segmentation. Using this mechanism dynamically set the correlation of pixels in the each region using a probabilistic model, then the evaluation of image segmentation methods introduce for an optimization problem. For solving this problem (evaluation of image segmentation methods) use the novel Imperialist Competitive Algorithm (ICA) that was recently introduced has a good performance in some optimization problems. In this paper a new Imperialist Competitive Algorithm is using chaotic map (CICA2) is proposed. In the proposed algorithm, the chaotic map is used to adapt the radius of colonies movement towards imperialist’s position to enhance the escaping capability from a local optima trap. Some famous benchmarks used to test proposed metric performance. Simulation results show this strategy can improve the performance of the unsupervised evaluation segmentation significantly. Keywords: Image segmentation, Imperialist Competitive Algorithm, Segmentation Evaluation.

1. Introduction Image segmentation is used to partition an image into separate regions for analysis and understanding image. Different methods have been introduced for segmenting image. There are two main approaches in image segmentation: region segmentation and boundary detection. We consider region-based image segmentation methods, because it has better results for texture images but there is no appropriate scale for evaluating these algorithms yet. The most usual evaluating method is the visual one in which the user visually observes different segmenting method at hand. Being time-consuming and gaining different results by users is disadvantages of this method. In supervised method, different segmented images are compared and evaluated with a ground truth image which has been made by the experts or different users. This method is the best method because of its high evaluating precision. Up to now most researches has been one on the supervised methods. In spite of their simplicity and low cost this method don’t have a proper efficiency because of miscue resulted from user improper choosing and spending a long time to examine different existing segmenting methods and also the

need of having the main segmented image of the intended image at hand. Unsupervised method does not require comparison with a manually- segmented reference image, has received little attention. The key advantage of unsupervised segmentation evaluation ability to evaluate segmentations independently of a manually-segmented reference images. This metric is good for processing real-time systems. The evaluating unsupervised which are given up to now, are base on the features of the image in locality area and the number of areas and the number of pixels in each region. In this paper, we examine a new scales for evaluating segmenting with and unsupervised methods. In this paper, we formulated the evaluation of image segmentation methods for an optimization problem. For solving this problem used ICA algorithm. So far, different evolutionary algorithms have been proposed for optimization which among them, we can point to a search algorithms were initially proposed by Holland, his colleagues and his students at the University of Michigan. These search algorithms which are based on nature and mimic the mechanism of natural selection were known as Genetic Algorithms (GAs) [1,2]. Particle Swarm Optimization algorithm proposed by Kennedy and Eberhart [3,4], in 1995. Simulated Annealing [5] and Cultural Evolutionary algorithm (CE), developed by Reynalds and Jin [5], in the early 1990s etc. The ant colony optimization algorithm (ACO), is a probabilistic technique for solving computational problems that can be reduced to finding good paths through graphs. This algorithm is a member of ant colony algorithms family, in swarm intelligence methods. Initially proposed by Marco Dorigo in 1992 in his PhD thesis [6][7] , the first algorithm was aiming to search for an optimal path in a graph, based on the behavior of ants seeking a path between their colony and a source of food. Differential evolution (DE) is an optimization algorithm. The DE method is originally due to Storn and Price [8][9] and works on multidimensional real-valued functions which are not necessarily continuous or differentiable. Recently, a new algorithm [10], in 2007, which has inspired not natural phenomenon, but of course from a socio-human from phenomenon. This algorithm has looked at imperialism process as a stage of human's socio-political evolution. The Imperialist Competitive Algorithm makes relation between humans and social sciences on one hand, and technical and mathematical sciences on the other hand, having a completely new viewpoint about the optimization topic. In the ICA algorithm, the colonies move towards the imperialist country with a random radius of movement. In


143

[11] CICA algorithm has been proposed that improved performance of ICA algorithm by the chaotic maps are used to adapt the angle of colonies movement towards imperialist’s position to enhance the escaping capability from a local optima trap. The ICA algorithm is used for Neural Network Learning based on Chaotic Imperialist Competitive Algorithm [12]. In this paper, we have proposed a new formulation for the evaluation of image segmentation methods that solved with Imperialist Competitive Algorithm. We introduce in this paper a study of unsupervised evaluation criteria that enable the quantification of the quality of an image segmentation result. This evaluation metric computes some statistics for each region in a segmentation result. Suggested scales engage in evaluation methods of segmenting by extracting image features in spatial domain. This method evaluate by evolutionary algorithm (ICA). These methods compare considering the segmented images and the main image. For this comparative study, we use two database composed of 200 images segmented. We will explain the suggested methods afterwards. This article is organized as follows: In Section 2, provides an introduction of the unsupervised evaluation criteria and highlight the most relevant ones and related work. In section 3, we introduced the Imperialist Competitive Algorithm (ICA). In section 4, described proposed algorithm and definition of chaotic radius in the movement of colonies toward the imperialist. In section 5, we present unsupervised evaluation methods and optimization problem. In Section 6, comparing results and show role correlation metric and our evaluation finally, in Section 7 we present a summary of our work and provide pointers to further work.

2. Related Work Unsupervised method does not require comparison with a manually-segmented reference image, has received little attention and it is quantitative and objective. Supervised evaluation methods, evaluate segmentation algorithms by comparing the resulting segmented image against a manually segmented reference image, which is often referred to as ground-truth. The degree of similarity between the human and machine segmented images determines the quality of the segmented image. One benefit of supervised methods over unsupervised methods is that the direct comparison between a segmented image and a reference image is believed to provide a finer resolution of evaluation. Unsupervised method also known as stand-alone evaluation methods or empirical goodness methods [13].

Table 1. Classification of evaluation methods. Class Details

Analytic methods Methods attempt to characterize an algorithm itself in terms of principles, requirements, complexity etc.

Empirical goodness methods

Computing a “goodness” metric on the segmented image without a priori knowledge [14].

Empirical discrepancy methods

A measure of discrepancy between the segmented image output by an algorithm [15].

Region Differencing Computing the degree of overlap of the cluster associated with each pixel in one segmentation [16][17][18].

Boundary matching

Matching boundaries between the segmentations, and computing some summary statistic of match quality [16][19][20].

Information-based

Formulate the problem as that of evaluating an affinity function that gives the probability of two pixels belonging to the same segment [16][21][22][23].

Unsupervised methods instead evaluate a segmented image based on how well it matches a set of features of segmented images as idealistic by humans. For solving these problem we need to use unsupervised methods so unsupervised evaluation suitable for online segmentation in real-time systems, where a wide variety of images, whose contents are not known beforehand, need to be processed. We for evaluation segmented image need to original image and some of segmented images. There are two major problems with segmentation: under-segmentation and over-segmentation [24][25] are shown in Figure1. We need to minimize the under- or over-segmentation as much as possible.

Figure 1. a) A ground truth image. b) Under-segmented image. c) Over-segmented image.

In the case of under-segmentation, full segmentation has not been achieved, i.e. there are two or more regions that appear as one. In the case of over-segmentation, a region that would be ideally present as one part is now, split into two or more parts. These problems, though important, are not easy to resolve. Recently a large number of unsupervised evaluation methods have been proposed. Without any a priori knowledge, most of evaluation criteria compute some statistics on each region or class in the segmentation result. We consider region-based image segmentation methods. Most of these methods consider factors such as region uniformity, inter-region heterogeneity, region contrast, line contrast, line connectivity, texture, and shape measures [26]. An evaluation methods has been proposed by Liu and Yang (1994) [27], that it is compute the average squared color error of the segments, penalizing over-segmentation by weighting proportional to the square root of the number of segments. It requires no user-defined parameters and is independent of the contents and type of image. The evaluation function:

Where N is the number of segments, is the number of pixels in segment j, and is the squared color error of region j. The F evaluation function has a very strong bias towards under-segmentation (segmentations with very few

a

b c


144

regions) and penalizing over-segmentation by weighting proportional to the square root of the number of segments. This metric is independent of the type of image. F is bias towards under-segmentation, so An evaluation methods has been proposed by Borsotti et al.(1998), that extended F by penalizing segmentations that have many small regions of the same size. Borsotti improved upon Liu and Yang’s method, and improved F' by decreasing the bias towards both over-segmentation and under-segmentation. Proposing a modified quantitative evaluation (Q) [28], where

The variance was given more influence in Q by dividing by the logarithm of the region size, and Q is penalized strongly by when there are a large number of

segments. So Q is the less biased towards both under segmentation and over-segmentation. More recently, Zhang et al.(2004), proposed the evaluation function E, an information theoretic and the minimum description length principle (MDL). This segmentation evaluation function instead of using squared color error they use region entropy as its measure of intra-region uniformity that measures the entropy of pixel intensities within each region [29]. To prevent a bias towards over-segmentation, they define the layout entropy of the object features of all pixels in image where any two pixels in the same region have the same object feature. Pal and Bhandari also proposed an entropy-based segmentation evaluation measure for intra-region uniformity based on the second-order local entropy. Weszka and Rosenfeld proposed such a criterion with thresholding that measures the effect of noise to evaluate some threshold images. Based on the same idea of intra-region uniformity, Levine and Nazif also defined criterion LEV1 that computes the uniformity of a region characteristic based on the variance of this characteristic. Complementary to the intra-region uniformity, Levine and Nazif defined a disparity measurement between two regions to evaluate the dissimilarity of regions in a segmentation result. We compare our proposed method against the evaluation functions of F, E and Q .

3. Introduction of Imperialist Competitive Algorithm (ICA)

In this section, we introduce ICA algorithm and chaos theory.

3.1. Imperialist Competitive Algorithm (ICA) Imperialist Competitive Algorithm (ICA) is a new evolutionary algorithm in the Evolutionary Computation field based on the human's socio-political evolution. The algorithm starts with an initial random population called countries. Some of the best countries in the population selected to be the imperialists and the rest form the colonies of these imperialists. In an N dimensional optimization problem, a country is a array. This array defined as below

The cost of a country is found by evaluating the cost function f at the variables . Then

The algorithm starts with N initial countries and the best of them (countries with minimum cost) chosen as the imperialists. The remaining countries are colonies that each belong to an empire. The initial colonies belong to imperialists in convenience with their powers. To distribute the colonies among imperialists proportionally, the normalized cost of an imperialist is defined as follow

Where, is the cost of nth imperialist and is its normalized cost. Each imperialist that has more cost value, will have less normalized cost value. Having the normalized cost, the power of each imperialist is computed as below and based on that the colonies distributed among the imperialist countries.

On the other hand, the normalized power of an imperialist is assessed by its colonies. Then, the initial number of colonies of an empire will be

Where, is initial number of colonies of nth empire and is the number of all colonies.

To distribute the colonies among imperialist, of the colonies is selected randomly and assigned to their imperialist. The imperialist countries absorb the colonies towards themselves using the absorption policy. The absorption policy shown in Fig.2, makes the main core of this algorithm and causes the countries move towards to their minimum optima. The imperialists absorb these colonies towards themselves with respect to their power that described in (8). The total power of each imperialist is determined by the power of its both parts, the empire power plus percents of its average colonies power.

Where is the total cost of the nth empire and is a positive number which is considered to be less than one.

In the absorption policy, the colony moves towards the imperialist by x unit. The direction of movement is the vector from colony to imperialist, as shown in Fig.2, in this figure, the distance between the imperialist and colony shown by d and x is a random variable with uniform distribution. Where is greater than 1 and is near to 2. So, a proper choice can be . In our implementation is

respectively.


145

(10)

In ICA algorithm, to search different points around the imperialist, a random amount of deviation is added to the direction of colony movement towards the imperialist. In Fig. 2, this deflection angle is shown as , which is chosen randomly and with an uniform distribution. While moving toward the imperialist countries, a colony may reach to a better position, so the colony position changes according to position of the imperialist.

Figure2. Moving colonies toward their imperialist

In this algorithm, the imperialistic competition has an important role. During the imperialistic competition, the weak empires will lose their power and their colonies. To model this competition, firstly we compute the probability of possessing all the colonies by each empire considering the total cost of empire.

Where, is the total cost of nth empire and is the normalized total cost of nth empire. Having the normalized total cost, the possession probability of each empire is computed as below

after a while all the empires except the most powerful one will collapse and all the colonies will be under the control of this unique empire.

4. Proposed Algorithm In this paper, we have proposed a new Imperialist Competitive Algorithm using the chaos theory (Chaotic Imperialist Competitive Algorithm CICA2). The primary ICA algorithm uses a local search mechanism as like as many evolutionary algorithms. Therefore, the primary ICA may fall into local minimum trap during the search process and it is possible to get far from the global optimum. To solve this problem we increased the exploration ability of the ICA algorithm, using a chaotic behavior in the colony movement towards the imperialist’s position. So it is intended to improve the global convergence of the ICA and to prevent it to stick on a local solution. 4.1. Definition of chaotic radius in the movement of colonies towards the imperialist

In this paper, to enhance the global exploration capability, the chaotic maps are incorporated into ICA to enhance the ability of escaping from a local optimum. The radius of movement is changed in a chaotic way during the search process. Adding this chaotic behavior in the imperialist algorithm absorption policy we make the conditions proper for the algorithm to escape from local peaks. Chaos variables are usually generated by the some well-known chaotic maps [30],[31]. Eq.(13), shows the mentioned chaotic maps for adjusting parameter (radius of colonies movement towards the imperialist’s position) in the proposed algorithm.

Where, is a control parameter. is a chaotic variable in kth iteration which belongs to interval of (0,1). During the search process, no value of is repeated. The CICA algorithm is summarized in Fig.3.

Figure2. The AICA algorithm.

Figure3. The CICA2 algorithm.

5. Unsupervised Image Segmentation and CICA2 algorithm

As mentioned before in algorithms the evaluations are bias towards the under-segmentation or over-segmentation. At first, we compute correlation and then present a new metric for evaluation of segmented images. In this method, we extract the statistical information about the image from the each region to provide an adaptive evaluation. We proposed a probabilistic model [32]-[35], to decrease error of evaluation. The probabilistic model P(x) that we use here is a Gaussian distribution model. The joint probability distribution of pixels given by the product of the marginal probabilities of the countries:

Where

The average, µ, and the standard deviation, , for the pixels of each region is approximated as below:

(1) Initialize the empires and their colonies positions randomly. (2) Compute the adaptive x (colonies movement radius towards the imperialist’s position) using the probabilistic model. (3) Compute the total cost of all empires (Related to the power of both the imperialist and its colonies). (4) Pick the weakest colony (colonies) from the weakest empire and give it (them) to the empire that has the most likelihood to possess it (Imperialistic competition). (5) Eliminate the powerless empires. (6) If there is just one empire, then stop else continue. (7) Check the termination conditions.


146

Using this probabilistic model, the density of pixels is computed in each region. If the pixels density in the current region is more than the previous region, then with 75% the previous correlation of the evaluation of the pixels will be decrease and with 25% the mentioned correlation will be increase.

, is the current correlation of pixel. , is the correlation of previous region and α is the constant value of decreasing and increasing the correlation of evaluation. The value of α is 0.5. Otherwise, if the pixels density in the current region is less than the previous region, then with 75% the previous correlation of the evaluation of the pixels will be increased and with 25% the mentioned correlation will be decreased.

If the pixels density in the current region is more than the previous region, it means that may be the pixels are in a good region. In Eq. (18), depending on the density of the pixels distribution, we set the correlation of region so that each pixel can escape from the dense area with 25% and with 75% the pixel is in its region with a decreasing correlation.

Figure 4. A sequence from formulating the evaluation of

Image Segmentation Methods for an optimization problem. In Eq. (19), if the pixels density in the current region is less than the previous region, each pixel with possibility of 25% is in its region with a decreasing correlation and with 75% the pixel is in its region with an increasing correlation. This way, provides a more efficient evaluation in all over the image. A good segmentation evaluation should maximize the uniformity of pixels within each segmented region, and minimize the uniformity across the regions. We propose a new function for evaluation of image segmentation for this function we need region-based segmentation. We compute variances of the R, G and B (variances color for image segmented in K-means) pixels of the region.

is variances intensity of pixels in region that you see in Eq. (20), and N is number of regions.

Where in the Eq.(21). Means of region original image is and means of region in gray-level segmented

image is .

We formulated the Evaluation of Image Segmentation Methods for an optimization problem and solved this problem with ICA algorithms. This method has a good precision for evaluating segmented image. This fitness function computes derivation of regions in segmented image. This Algorithm is show in Fig.5. This fitness function and ICA algorithm evaluate image segmentation algorithms. In this paper, ICA algorithm used for minimization the problem. We test this metric and is show result in next section.

Figure 5. The CICA2 algorithm for evaluation of image segmentation methods.

6. Experimental Result We empirically studied the evaluation methods F, Q, E and CICA2 algorithm on the segmentation results from two different segmentation algorithms, the Edge Detection and Image Segmentation (EDISON). It developed by the Robust Image Understanding Laboratory at Rutgers University. We used EDISON to generate images that vary in the number of regions in the segmentation to see how the evaluation methods are affected by the number of regions. The second segmentation algorithm is canny that is available in Berkeley dataset. We use these two segmentation methods to do a preliminary study on the effectiveness of these quantitative evaluation methods on different segmentation parameterizations and segmentation techniques. We use two dataset Berkeley with 1000 images and 1200 images, for computing error in evaluation. In this section, we analyze the previously presented unsupervised evaluation criteria. We describe experimental results to evaluate CICA2 algorithm and results from four evaluation methods are examined and compared. We compute the effectiveness of F, Q, E and CICA2 based on their accuracy with evaluations provided by a small group of human evaluators. In our first set of experiments, we vary the total number of regions in the segmentation (using EDISON to generate the segmentations) to study the sensitivity of these an objective evaluation methods to the number of regions in the segmentation. With an increase in the number of regions, the segmented images clearly look better to the observer, since more details are preserved. However, more regions do not necessarily

(1) Eq.(21) is cost function. (1) Initialize the empires and their colonies positions randomly. (2) Compute the total cost of all empires (Related to the power of both the imperialist and its colonies). (4) Pick the weakest colony (colonies) from the weakest empire and give it (them) to the empire that has the most likelihood to possess it

(Imperialistic competition). (5) Eliminate the powerless empires. (6) If there is just one empire, then stop else continue. (7) Check the termination conditions.


147

make a better segmentation, since over-segmentation can occur and it is a problem for evaluations.

The proposed algorithm is a good algorithm for evaluation of segmented image because this method has a controller for under-segmentation and over-segmentation. The corr has important role in evaluation so error of evaluation is the less. The effectiveness is described by accuracy, which is defined as the percentage of the number of times the evaluation measure correctly matches human evaluation result divided by the total number of comparisons in the experiment. We compute the effectiveness of F, Q, E and CICA2 algorithm based on their accuracy with evaluations provided on four dataset that is shown in Table.2.

Table 2: Accuracy (%) of the evaluation measures

Accuracy (%) F Q E CICA2 algorithm

1)Image Segmentation (EDISON)

%73.3 %76.22 %74.81 %80.09

2)Berkeley dataset

(canny)

%64.3 %68.22 %63.81 %75.85

3)1200 images- Berkeley

%71.01 %73.35 %75.50 %84.63

4)1000 images- Berkeley

%62.43 %68.6 %71.32 %83.43

The results, given in Table 2, once again demonstrate the bias of many of the evaluation methods towards under-segmentation. F and E, achieve low accuracy in this experiment. On the other hand, those measures that are more balanced or less biased towards under-segmentation, i.e. Q and CICA2 algorithm, achieve higher accuracy. Overall, CICA2 algorithm performs best here. We evaluate an image with four unsupervised evaluation and can see that CICA2 algorithm is better than F. and CICA2 algorithm no sensitive to under-segmentation and over-segmentation. It computes correlation and pixel density for each region and control error of under-segmentation and over-segmentation in evaluation.

Figure 6. Run-time for four metrics for evaluation 100 image (second).

In Fig.6, we are shown that run-time for evaluation 100 images in CICA2 algorithm is better than E, F and Q.

0 10 20 30 40 50 60 70 80 90 1000

20

40

60

80

100

120

Generation

Cos

t

Error of evaluation of segmented images

data1data2

Figure 7. Cost of evaluation of segmented images datasets

3,4.

In Fig.7, we are shown that error for evaluation 100 images in CICA2 algorithm is near the zero.

7. Conclusion And Future Work In this paper, we present an optimization method that objectively evaluate image segmentation. In this paper, we proposed a new formulation for the evaluation of image segmentation methods. In this paper using probabilistic model that utilize the information of pixels (mean and variance) in each region to balance the under-segmentation and over-segmentation. Using this mechanism dynamically set the correlation of pixels in the each region using a probabilistic model, then the evaluation of image segmentation methods introduce for an optimization problem. We first present four segmentation evaluation methodologies, and discuss the advantages and shortcomings of each type of unsupervised evaluation, among others. Subjective and supervised evaluations have their disadvantages. For example tedious to produce and can vary widely from one human to another and time-consuming. Unsupervised segmentation evaluation methods offer the unique advantage that they are purely objective and do not require a manually-segmented reference image and those embedded in real-time systems. We have demonstrated via our preliminary experiments that our unsupervised segmentation evaluation measure, CICA2 algorithm, improves upon previously defined evaluation measures in several ways. In particular, F has a very strong bias towards images with very few regions and thus do not perform well. Q outperforms F but still disagrees with our human evaluators more often than E did. The correlation and density in each region are important components in obtaining our results. Coding evaluation problem and present a new cost function and solving a optimization problem is interesting directions for future research.


148

References [1]H. M¨uhlenbein, M. Schomisch and J. Born, "The

Parallel Genetic Algorithm as Function Optimizer", Proceedings of The Fourth International Conference on Genetic Algorithms, University of California, San diego, pp. 270-278,1991.

[2]J.H. Holland. "ECHO: Explorations of Evolution in a Miniature World", In J.D. Farmer and J. Doyne, editors, Proceedings of the Second Conference on Artificial Life, 1990.

[3]J. Kennedy and R.C. Eberhart, "Particle swarm optimization", in: Proceedings of IEEE International Conference on Neural Networks, Piscataway: IEEE, pp. 1942–1948, 1995.

[4]X. Yang, J. Yuan, J. Yuan and H. Mao," A modified particle swarm optimizer with dynamic adaptation", Applied Mathematics and Computation, 189 (2): pp. 1205-1213, 2007.

[5] X. Jin and R.G. Reynolds, "Using Knowledge-Based Evolutionary Computation to Solve Nonlinear Constraint Optimization Problems: A Cultural Algorithm Approach", In Proceedings of the IEEE Congress on Evolutionary Computation, 3: pp. 1672–1678, 1999.

[6]A. Colorni, M. Dorigo et V. Maniezzo, "Distributed Optimization by Ant Colonies", actes de la première conférence européenne sur la vie artificial, Paris, France, Elsevier Publishing, pp.134-142, 1991.

[7]M. Dorigo, "Optimization, Learning and Natural Algorithms", PhD thesis, Politecnico di Milano, Italie, 1992.

[8] R. Storn and K. Price,."Differential evolution - a simple and efficient heuristic for global optimization over continuous spaces", Journal of Global Optimization, 11: 341–359, 1997.

[9]R.Storn, "On the usage of differential evolution for function optimization", Biennial Conference of the North American Fuzzy Information Processing Society (NAFIPS), pp. 519–523, 1996.

[10]E. Atashpaz-Gargari and C. Lucas, "Imperialist Competitive Algorithm: An Algorithm for Optimization Inspired by Imperialistic Competition", IEEE Congress on Evolutionary Computation (CEC 2007), pp. 4661-4667, 2007.

[11]H. Bahrami, K. Faez, M. Abdechiri, "Imperialist Competitive Algorithm using Chaos Theory for Optimization", UKSim-AMSS 12th International Conference on Computer Modeling and Simulation, 2010.

[12] M. Abdechiri, K. Faez and H. Bahrami, "Neural Network Learning based on Chaotic Imperialist Competitive Algorithm", The 2nd International Workshop on Intelligent System and Applications (ISA2010), 2010.

[13]H. Zhang, J. E. Fritts and S. A. Golman, "An Entropy - based Objective Evaluation method for image segmentation", SPIE Electronic Imaging – Storage and Retrieval Methods and Applications for Multimedia, pp.38-49, 2004.

[14]M.D. Levine, A. Nazif, "Dynamic measurement of computer generated image segmentations", IEEE

Transactions on Pattern Analysis and Machine Intelligence 7: pp. 155–164, 1985.

[15]W.A. Yasnoff, J.K. Mui and J.W. Bacus, "Error measures for scene segmentation", Pattern Recognition 9:pp.217–231, 1977.

[16]D. Martin, "An Empirical Approach to Grouping and Segmentation", PhD dissertation, Univ. of California, Berkeley, 2002.

[17]D. Martin, C. Fowlkes, D. Tal, and J. Malik, "A Database of Human Segmented Natural Images and Its Application to Evaluating Segmentation Algorithms and Measuring Ecological Statistics", Proc. Int’l Conf. Computer Vision, 2001.

[18]H.I. Christensen and P.J. Phillips, "Empirical Evaluation Methods in Computer Vision" , eds. World Scientific Publishing, July 2002.

[19]J. Freixenet, X. Munoz, D. Raba, J. Marti, and X. Cuff, "Yet Another Survey on Image Segmentation: Region and Boundary Information Integration", Proc. European Conf. Computer Vision, pp. 408-422, 2002.

[20]Q. Huang and B. Dom, "Quantitative Methods of Evaluating Image Segmentation", Proc. IEEE Int’l Conf. Image Processing, pp. 53-56, 1995.

[21]C. Fowlkes, D. Martin, and J. Malik, "Learning Affinity Functions for Image Segmentation", Proc. IEEE Conf. Computer Vision and Pattern Recognition, 2: pp. 54-61, 2003.

[22]M. Meila, "Comparing Clusterings by the Variation of Information", Proc. Conf. Learning Theory, 2003.

[23] M.R. Everingham, H. Muller, and B. Thomas, "Evaluating Image Segmentation Algorithms Using the Pareto Front, " Proc. European Conf. Computer Vision, 4: pp. 34-48, 2002.

[24]H. Zhang, Jason E. Fritts and S. A. Goldman , "Image Segmentation Evaluation: A Survey of Unsupervised Methods", Computer Vision and Image Understanding (CVIU), 110(2): pp. 260-280, 2008.

[25]S. Chabrier, B. Emile, H. Laurent, C. Rosenberger and P. Marche, "Unsupervised evaluation of image segmentation application multispectral images", in: Proceedings of the 17th international conference on pattern recognition, 2004.

[26]Zhang, Y.J, "A survey on evaluation mehods for image segmentation", Pattern Recognition 29:pp. 1335–1346, 1996.

[27]J.Liu, Y.-H. Yang, "Multi-resolution color image segmentation", IEEE Transactions on Pattern Analysis and Machine Intelligence 16 (7):pp. 689–700, 1994.

[28] M. Borsotti, P. Campadelli, R. Schettini, "Quantitative evaluation of color image segmentation results", Pattern Recognition Letters 19 (8):pp. 741–747,1998.

[29] Y. J. Zhang and J. J. Gerbrands, "Objective and quantitative segmentation evaluation and comparison". Signal Processing 39:pp. 43–54, 1994.

[30]WM. Zheng, Kneading plane of the circle map. Chaos, Solitons & Fractals, 4:1221, 1994.

[31]HG. Schuster, "Deterministic chaos: an introduction", 2nd revised ed. Weinheim, Federal Republic of Germany: Physick-Verlag GmnH; 1988.


149

[32]A. Papoulis, "Probability, Random Variables and Stochastic Processes", McGraw-Hill, 1965.

[33]R. C. Smith and P. Cheeseman, "On the Representation and Estimation of Spatial Uncertainty", the International Journal of Robotics Research,5(4), 1986.

[34]T.K. Paul and H. Iba, "Linear and Combinatorial Optimizations by Estimation of Distribution Algorithms", 9th MPS Symposium on Evolutionary Computation, IPSJ, Japan, 2002.

[35]Y. Bar-Shalom, X. Rong Li and T. Kirubarajan, "Estimation with Applications to Tracking and Navigation", John Wiley & Sons, 2001.

Authors Profile

Javad Alikhani koupaei the B.S. degrees received the in mathematics (application in computer), University of Esfahan, dec, 1993 and M.S. degrees in mathematics, Tarbiat modarres University, Tehran, Iran, february 1998, he study in "A new recursive algorithm for a gussian quadrature formula

via orthogonal polynomials", He is faculty at the Department of Mathematics, Payamenour University, Isfahan, Iran

Marjan Abdechiri received her B.S. degree in computer Engineering from Najafabad University, Isfahan, Iran, 2007. He is now pursuing the Artificial intelligence Engineering of Qazvin University, Her research interests are computational intelligence and image processing, machine learning, evolutionary computation.


150

Segmentation of MR Brain tumor using Parallel ACO J.Jaya1 and K.Thanushkodi 2

1 Research Scholar, Anna University,Chennai,Tamil Nadu,India. [email protected]

2Director, Akshaya College of Engg, & Tech.Coimbatore,Tamil Nadu,India

Abstract: One of the most complex tasks in digital image processing is image segmentation. This paper proposes a novel image segmentation algorithm that uses a biologically inspired technique based on ant colony optimization (ACO) algorithm. It has been applied to solve many optimization problems with good discretion, parallel, robustness and positive feedback. The proposed new meta-heuristic algorithm operates on the image pixel data and a region/neighborhood map to form a context in which they can merge. Hence, we segment the MR brain image using ant colony optimization algorithm. Compared to traditional metaheuristic segmentation methods, the proposed method has advantages that it can effectively segment the fine details. The suggested image segmentation strategy is tested on a set of real time MR Brain images.

Keywords: MRI, Segmentation, Ant colony optimization, & Registration.

5. Introduction The field known as biomedical analysis has evolved

considerably over the last couple of decades. The widespread availability of suitable detectors has aided the rapid development of new technologies for the monitoring and diagnosis, as well as treatment, of patients. Over the last century technology has advanced from the discovery of x-rays to a variety of imaging tools such as magnetic resonance imaging, computed tomography, positron emission tomography and ultrasonography. The recent revolution in medical imaging resulting from techniques such as computed tomography (CT) and magnetic resonance imaging (MRI) can provide detailed information about disease and can identify many pathologic conditions, giving an accurate diagnosis. Furthermore, new techniques are helping to advance fundamental biomedical research. Medical imaging is an essential tool for improving the diagnoses, understanding and treatment of a large variety of diseases [1, 7, 8].

The extraordinary growth experimented by the medical image processing field in the last years, has motivated the development of many algorithms and software packages for image processing. There are also some efforts to develop software that could be easy to modify by other researchers, for example the Medical Imaging Interaction Toolkit, is intended to fill the gap between algorithms and the final user, providing interaction capabilities to construct clinical applications. Therefore, many software packages for visualization and analysis of medical data are available for

the research community. Among them we can find commercial and non commercial packages. Usually, the former ones are focused on specific applications or just on visualization being more robust and stable, whereas the latter ones often offer more features to the end user. 2. Existing Methods

Several methods have been proposed to segment brain MR images. The methods are either based on the intrinsic structure of the data or based on statistical frameworks. Structural based methods rely on apparent spatial regularities of image structures such as edges, and regions. However, the performance is not satisfying when the images are with noise, artifacts, and local variations, which is often the case in real data. Instead, statistical based methods use a probability model to classify the voxels into different tissue types based on the intensity distribution of the image. Methods based on statistical frameworks can be further divided into non-parametric methods or parametric methods.. In non-parametric methods, the density model of the prior relies entirely on the data itself, i.e. the K-nearest-neighbors (K-NN) method. Non-parametric methods are adaptive, but its limitation is the need of a large amount of labeled training data. In contrast, non-parametric methods rely only on the explicit functional form of intensity density function of the MR image [10, 11]. 3. Proposed Method

Figure1. Block Diagram of Proposed work 3.1 Image Acquisition

The development of intra-operative imaging systems has contributed to improving the course of intracranial neurosurgical procedures. Among these systems, the 0.5T intra-operative magnetic resonance scanner of the Kovai

Image Acquisition

Preprocessing

Enhancement

Segmentation


151

Medical Center and Hospital (KMCH, Signa SP, GE Medical Systems) offers the possibility to acquire 256*256*58(0.86mm, 0.86mm, 2.5 mm) T1 weighted images with the fast spin echo protocol (TR=400,TE=16 ms, FOV=220*220 mm) in 3 minutes and 40 seconds. The quality of every 256*256 slice acquired intra-operatively is fairly similar to images acquired with a 1.5 T conventional scanner, but the major drawback of the intra-operative image is that the slice remains thick (2.5 mm).

Images of a patient obtained by MRI scan is displayed as an array of pixels (a two dimensional unit based on the matrix size and the field of view) and stored in Mat lab 7.0.Here, grayscale or intensity images are displayed of default size 256 x 256.The following figure displayed a MRI brain image obtained in Mat lab 7.0. The brain MR images are stored in the database in JPEG format. Fig 2 shows the image acquisition.

Figure 2. Image Acquisition

3.2 Preprocessing

Preprocessing functions involve those operations that are normally required prior to the main data analysis and extraction of information, and are generally grouped as radiometric or geometric corrections. Radiometric corrections include correcting the data for sensor irregularities and unwanted sensor or atmospheric noise, removal of non-brain voxels and Converting the data so they accurately represent the reflected or emitted radiation measured by the sensor [9].

In this work tracking algorithm is implemented to remove film artifacts. The high intensity value of film artifacts are removed from MRI brain image. During the removal of film artifacts, the image consists of salt and pepper noise.

Figure 2 a) Before Preprocessing

Figure 2 b) After Preprocessing

3.3 Enhancement Image enhancement methods improve the visual

appearance of Magnetic Resonance Image (MRI). The role of enhancement technique is removal of high frequency components from the images. This part is used to enhance the smoothness towards piecewise-homogeneous region and reduces the edge-blurring effect. Conventional Enhancement techniques such as low pass filter, Median filter, Gabor Filter, Gaussian Filter, Prewitt edge-finding filter, Normalization Method are employable for this work.

This proposed system describes the information of enhancement using weighted median filter for removing high frequency components such as impulsive noise, salt and pepper noise, etc. and obtained high PSNR and ASNR values. PSNR = 0.924,ASNR =.929

Figure 3. Enhancement

Figure3.Performance Evaluation of Enhancement Stage 3.5 Segmentation

Segmentation is the initial step for any image analysis. There is a task for segmentation of brain MRI images. That is to obtain the locations of suspicious areas to assist radiologists for diagnose. Image segmentation has been approached from a wide variety of perspectives. Region-based approach, morphological operation, multiscale analysis, fuzzy approaches and stochastic approaches have been used for MRI image segmentation but with some limitations. Local thresholding is used by setting threshold


152

values for sub-images. It requires selection of a window size and threshold parameters. Wu et al. presented that the threshold for a pixel is set as the mean value plus the Root Mean Square (RMS) noise value multiplied by a selected coefficient in a selected square region around the thresholded pixel. Kallergi et al. compared local thresholding and region growing methods. It showed that the local thresholding method has greater stability, but is more dependent on parameter selection. Woods et al. used local thresholding by subtracting the average intensity of a 15×15 window from the processed pixel. Then, region growing is performed to group pixels into objects. Comparing with the multi-tolerance region growing algorithm and the active contour model, it showed that the speed of the algorithm is more than an order of magnitude faster than the other two.

Edge detection is a traditional method for segmentation. Many operators, Roberts gradient, Sobel gradient, Prewitt gradient and Laplacian operator, were published in the literature. Some mathematical morphological operations such as erosion, top-hat transformation and complicated morphological filters and multi-structure elements can also be used. It is good in dealing with geometrically analytic aspects of image analysis problems. Stochastic approaches have also been used to segment calcifications. Stochastic and Bayesian methods have provided a general framework to model images and to express prior knowledge. Markov Random Field (MRF) model was used to deal with the spatial relations between the labels obtained in an iterative segmentation process. The process-assigning pixel labels iteratively.

4. Proposed parallel Ant Colony System (ACS) Ant Colony Optimization (ACO) is a population-based

approach first designed by Marco Dorigo and coworkers, inspired by the foraging behavior of ant colonies [13, 14, 15]. ndividuals ants are simple insects with limited memory and capable of performing simple ctions.However, the collective behavior of ants provides intelligent solutions to problems such as finding the shortest paths from the nest to a food source. Ants foraging for food lay down quantities of a volatile chemical substance named pheromone, marking their path that it follows. Ants smell pheromone and decide to follow the path with a high probability and thereby reinforce it with a further quantity of pheromone [12, 16]. The probability that an ant chooses a path increases with the number of ants choosing the path at previous times and with the strength of the pheromone concentration laid on it.

In this work, the labels created from the MRF method and the posterior energy function values for each pixel are stored in a solution matrix. The goal of this method is to find out the optimum label of the image that minimizes the posterior energy function value. Initially assign the values of number of iterations (N), number of ants (K), initial pheromone value (T0). [Hint: we are using N=20, K=10, T0=0.001].

4.1 Pheromone Initialization For each ant assign the initial pheromone value T0.

And for each ant select a random pixel from the image which has not been selected previously. To find out the pixels is been selected or not, a flag value is assigned for each pixel. Initially the flag value is assigned as 0, once the pixel is selected the flag is changed to 1. This procedure is followed for all the ants. For each ant a separate column for pheromone and flag values are allocated in the solution matrix.

4.2 Local Pheromone Update Update the pheromone values for all the randomly

selected pixels using the following equation: Tnew = (1 – ρ) * Told + ρ * T0 , where Told and Tnew are the old and new pheromone values, and ρ is rate of pheromone evaporation parameter in local update, ranges from [0,1] i.e., 0 < ρ < 1. Calculate the posterior energy function value for all the selected pixels by the ants from the solution matrix. The ant, which generates this local minimum value, is selected and whose pheromone is updated using the following equation:

Tnew = (1 – α) * Told + α * ∆Told, where Told and Tnew are the old and new pheromone values, and α is rate of pheromone evaporation parameter in global update called as track’s relative importance, ranges from [0,1] i.e., 0 < α < 1, and ∆ is equal to ( 1 / Gmin). For the remaining ants their pheromone is updated as: Tnew = (1 – α) * Told, here, the ∆ is assumed as 0. Thus the pheromones are updated globally. This procedure is repeated for all the image pixels. At the final iteration, the Gmin has the optimum label of the image. To further enhance the value, this entire procedure can be repeated for any number of times. In our implementation, we are using 20 numbers of iterations.

Figure 4. Segmented MRI

5. Conclusion

In this paper, a meta heuristic based image segmentation approach was presented. The multi-agent algorithm takes advantage of various models to represent the agent’s problem solving expertise within a knowledge base and perform an inference mechanism that governs the agent’s behavior in choosing the direction of their proceeding steps. This paper details an image segmentation method based on the ant colony optimization algorithm. The improved accuracy rate according to the experimental results is due to better characterization of natural brain structure.


153

Experiments on both real and synthetic MR images show that the segmentation result of the proposed method has higher accuracy compared to existing algorithms.

References [1] André Collignon, Dirk Vandermeulen, Paul Suetens,

Guy Marchal,” 3D multi-modality medical image registration using feature space clustering “,SpringerLink, Volume 905pp.193-204,Berlin (1995).

[2] Alexis Roche, Gregoire Malandain, Nicholas Ayache, Sylvain prima, “Towards a better comprehension Medical Image Registration”, Medical Image Computing and Computer-Assisted Intervention-(MICCAI’99) ,Volume 1679,pp. 555-566, (1999).

[3] Aaron Lefohn, Joshua Cates, Ross Whitaker,: ”Interactive GPU-Based level sets for 3D Brain Tumor Segmentation”,April 16,2003.

[4] Ahmed Saad, Ben Smith,Ghassan Simultaneous Segmentation, Kinetic Parameter Estimation and Uncertainty Visualization of Dynamic PET Images”,( MICCAI) pp-500-510 ( 2007).

[5] Albert K.W.Law,F.K.Law,Francis H.Y.Chan,:”A Fast Deformable Region Model for Brain Tumor Boundary Extraction”,IEEE ,Oct 23, USA,(2002).

[6] Amini L, Soltanian-Zadeh H, Lucas.C,:“Automated Segmentation of Brain Structure from MRI”, Proc. Intl.Soc.Mag.Reson.Med.11(2003).

[7] Ceylan.C, Van der Heide U.A, Bol G.H, Lagendijk .J.J.W,Kotte A.N.T.J,”Assessment of rigid multi-modality image registration consistency using the multiple subvolume registration(MSR) method”, Physics in Medicine Biology, pp. 101-108,(2005).

[8] Darryl de cunha, Leila Eadie, Benjamin Adams, David Hawkes, “Medical Ultrasound Image similarity measurement by human visual system(HVS) Modelling”, spingerlink, volume 2525,pp.143-164 ,berlin, (2002).

[9] Dirk-Jan Kroon,” Multimodality non-rigid demon algorithm image registration “, Robust Non-Rigid Point Matching, Volume 14, pp. 120-126, (2008).

[10] Duan Haibin, Wang Daobo, Zhu Jiaqiang and Huang Xianghua, “Development on ant colony algorithm theory and its application,” Control and Decision, Vol. 19, pp. 12-16, (2004).

[11] Dorigo.M, Di Caro G, Gambardella L M, “Ant algorithms for discrete optimization”, Artificial Life, 1999, Vol.5, No.2.

[12] ErikDam,Marcoloog,Marloes Letteboer,:”Integrating Automatic and Interactive Brain tumor Segmentation”, Proceedings of the 17th International Conference on Pattern Recognition (ICPR’04), IEEE Computer Society.

[13] H. He, Y. Chen, “Artificial life for image segmentation,” International Journal of Pattern Recognition and Artificial Intelligence 15 (6), 2001, pp. 989-1003.

[14] J. Liu, Y.Y. Tang, “Adaptive image segmentation with distributed behavior-based agents,” IEEE

Transactions Pattern Analysis and Machine Intelligence 21, June 1999, pp. 544-551.

[15] John Ashburner , Karl J. Friston,” Rigid body registration “, The welcome dept of image neuro science , 12 queen square, London.

[16] M. Dorigo, Mauro Birattari and Thomas Stiitzle, “Ant Colony Optimization artificial Ants as a Computational Intelligence Technique”, Computational Intelligence magazine, Nov. 2006, pp.28-39.

[17] P.S. Shelokar, V.K. JayRaman and B.D. Kulkami, “An ant colony clustering approach for clustering”, analytical chemical act 509, 2004, pp.187- 195.

Acknowledgement The author wishes to thank Dr.M.Karnan, for his guideness in this area and also extend wishes to thank Doctor Pankaj Metha for his Suggestion on tumor recognition with his knowledge and experience in imaging area. The MRI Image data obtained from KMCH Hospital, Coimbatore, Tamil Nadu, and India.


154

Session Based Load Optimization Techniques to Enhance Security & Efficiency of E-commerce

Transactions

R.K. Pateriya1, J.L. Rana2 and S.C. Shrivastava3

1Associate Professor , Department of Information Technology, Maulana Azad National Institute of Technology (MANIT) Bhopal, India

[email protected]

2Professor , Department of Computer Science and Engineering, Maulana Azad National Institute of Technology (MANIT) Bhopal, India

[email protected]

3 Professor , Department of Electronics Engineering, Maulana Azad National Institute of Technology (MANIT) Bhopal, India

[email protected] Abstract: Today internet based e-commerce has become a trend and business necessity. Secure Socket layer (SSL) is the world standard for cyber security. A SSL session contains temporally and logically related request sequences from the same client. In e-commerce session integrity is a critical metric. Overload on server can lead e-commerce applications to considerable revenue losses, response times may grow to unacceptable levels and as a result the server may saturate or even crash .Session based admission control techniques is able to control the server load . The purpose of this paper is to review about various session based admission control techniques to avoid server overload. Overload control is a critical goal so that a system can remain operational even when the incoming request rate is several times greater than the system capacity and this admission control mechanism based on session will maximize the number of sessions completed successfully, allowing e-commerce sites to increase the number of transactions completed, generating higher benefits and optimizes performance.

Keywords: Admission control, Application servers, Overload control, Service differentiation

1. Introduction E-Commerce is a growing phenomenon as consumers gain experience and comfort with shopping on the Internet .Most of e-commerce web sites applications are session-based. Access to a web service occurs in the form of a session consisting of a sequence of individual requests. Placing an order through the web site involves further requests relating to selecting a product, providing shipping information, arranging payment agreement and finally receiving a confirmation. So for a customer trying to place an order or a retailer trying to make a sale, the real measure of a web server performance is its ability to process the entire sequence of requests needed to complete a transaction. The higher the number of sessions completed the higher the amount of revenue that is likely to be generated as discussed in [3]. Sessions that are broken or delayed at some critical stages, like checkout and shipping, could mean loss of revenue to the web site. Security between network nodes

over the Internet is traditionally provided using Secure Socket Layer (SSL). It is commonly used for secure http connections where credit card information is going to be sent along a network and this gives e-commerce the confidence it needs to allow on-line banking and shopping.. SSL provides and encrypted bi-directional data stream, data is encrypted at the sender's end and decrypted at the receiver's end. It can perform mutual authentication of both the sender and receiver of messages and ensure message confidentiality. This process involves certificates that are configured on both sides the connection. Although providing these security capabilities does not introduce a new degree of complexity in web application structure, it increases remarkably the computation time needed to serve a connection, due to the use of cryptographic techniques, becoming a CPU intensive workload.

Two problems are typically encountered with deploying e-commerce Web sites presented in [1,2]. First is overload, where the volume of requests for content at a site temporarily exceeds the capacity for serving them and renders the site unusable. Second is responsiveness, where the lack of adequate response time leads to lowered usage of a site and subsequently, reduced revenues. During overload conditions, the service’s response times may grow to unacceptable levels, and exhaustion of resources may cause the service to behave erratically or even crash causing denial of services. For this reason, overload prevention in these applications is a critical issue. Several mechanisms have been proposed in [1,3,4] to deal with overload, such as admission control, request scheduling, service differentiation, service degradation.

Request scheduling refers to the order in which concurrent requests should be served. A well known form is queuing theory (SRPT) that shortest remaining processing time first scheduling minimizes queuing time .Better scheduling can always be complementary to any other mechanism .Service differentiation is based on differentiating classes of customers so that response times of preferred clients do not suffer in the presence of overload. Service degradation is based on avoiding refusing clients as


155

a response to overload but reducing the service offered to clients for example in the form on providing smaller content.

Admission control generally requires two components knowing the load that a particular job will generate on a system, and knowing the capacity of that system. By keeping the maximum amount of load just below the system capacity, overload is prevented and peak throughput is achieved discussed in [1,2]. The goal of overload control is to prevent service performance from degrading in an uncontrolled fashion under heavy load, it is often desirable to shed load.

The rest of the paper is organized as follows: Section II Overview of Session based admission control (SBAC) techniques. Section III presents SSL connection differentiation and admission control. Sections IV CPU Utilization-Based Implementation of SBAC Mechanism. Section V Adaptive admission control technique. Section VI Comparative study of SBAC techniques. Section VII Conclusion.

2. Overview of SBAC Techniques The Admission control is based on reducing the amount of work the server accepts when it is faced with overload. For example, admission control on per request basis may lead to a large number of broken or incomplete sessions when the system overloaded Sessions have distinguishable features from individual requests that complicate the overload control. Session-based workload gives a new interesting angle to revisit and re-valuate the definition of web server performance. It proposes to measure a server throughput as a number of successfully completed sessions. The reason, for failure of admission control techniques that work on a per request basis discussed in [1,2], is because it leads to large number of broken or incomplete sessions when the system is overloaded, hence cause revenue loss. Session integrity is a critical metric in e-commerce. Research in admission control, can be roughly categorized under two broad approaches presented in [3,5]: First is reducing the amount of work required when faced with overload, and second is differentiating classes of customers so that response times of preferred clients do not suffer in the presence of overload. This paper focuses on first approach by reducing amount of work for admission control. There are two desirable, but somewhat contradictory, properties for an admission control mechanism: stability and responsiveness discussed in [1,2]. In case, when the server receives an occasional burst of new traffic, while still being under a manageable load, the stability, which takes into account some load history, is a desirable property for admission control mechanism. It helps to maximize server throughput and to avoid unnecessary rejection of newly arrived sessions. However, if a server's load during previous time intervals is consistently high, and exceeds its capacity, the responsiveness is very important: The admission control policy should be switched on as soon as possible, to control and reject newly arriving traffic. There is a trade-off between these two desirable properties for an admission control mechanism.

The following two values help to check an admission

control goodness discussed in [1,2]: First is the percentage of aborted requests, which server can determine based on the client side closed connections. Aborted requests indicate that the level of service is unsatisfactory. Typically, aborted requests lead to aborted sessions, and could serve as a good warning sign of degrading server performance; second is a percentage of connection refused messages sent by a server, in the case of full listen queue. Refused connections are the dangerous warning sign of an overloaded server and its inevitable poor session performance. If both of these values are zero then it reveals that an admission control mechanism uses an adequate admission control function to cope with current workload and traffic rate. Good admission control strategy which minimizes a percentage of aborted requests and refused connections (ideally to 0) and maximizes the achievable server throughput .Now in the following sections we are going to discuss three techniques of session based admission control and their comparative study

3. SSL Connection Differentiation and Admission Control

SSL connections is 7 times lower than when using normal connections. Based on the SSL connection differentiation, a session-based adaptive admission control mechanism is implemented in [3,4] .This mechanism allows the server to avoid throughput degradation and response time increments occurred during overload conditions. The server differentiates full SSL connections from resumed SSL connections and limits the acceptance of full SSL connections to the maximum number possible without overloading the available resources, while it accepts all the resumed SSL connections. In this way, this admission control mechanism maximizes the number of sessions completed successfully, allowing e-commerce sites based on SSL to increase the number of transactions completed, thus generating higher benefit . The SSL protocol fundamentally has two phases of operation SSL handshake and SSL record protocol. Two different SSL handshake types can be distinguished discussed in [3,4 ,9]: The full SSL handshake and the resumed SSL handshake. Most of the computation time required when using SSL is spent during the SSL handshake phase, which features the negotiation between the client and the server to establish a SSL connection. The full SSL handshake is negotiated when a client establishes a new SSL connection with the server, and requires the complete negotiation of the SSL handshake, including parts that need a lot of computation time to be accomplished The SSL resumed handshake is negotiated when a client establishes a new HTTP connection with the server but resumes an existing SSL connection, the SSL session ID is reused hence part of the SSL handshake negotiation can be avoided, reducing considerably the computation time for performing a resumed SSL handshake Note their is big difference between the time to negotiate a full SSL handshake compared to the time to negotiate a resumed SSL handshake i.e. (175 ms vs. 2 ms) given in [3,4].


156

A session oriented adaptive mechanism discussed in [3,4] performs admission control based on SSL connection differentiation continuously monitors incoming secure connections to the server, performs online measurements distinguishing new SSL connections from resumed SSL connections and decides which incoming SSL connections are accepted and hence maximizing the number of sessions successfully completed . This maximum depends on the available processors for the server and the computational demand required by the accepted resumed connections.

Following are the definition of variable used in calculation taken from [3,4]] :K sampling interval (currently defined as 2 s), O(k) :defined as the number of resumed SSL connections that arrive to the server during that interval. CTO: The average computation time entailed by a resumed SSL connection .CTN: the average computation time entailed by a new SSL connection . N(k): defined as the maximum number of new SSL connections that can be accepted by the server during that interval without overloading. A(k): defined as the number of processors allocated to the server. Admission control mechanism periodically calculates, at the beginning of every sampling interval k maximum number of new connection allowed. Since resumed SSL connections have preference with respect to new SSL connections all resumed SSL connections are accepted.(O(k) · CTO) is the computation time required by the already accepted resumed SSL connections .Hence maximum number of new connection allowed: N(k)=(k*A(k)–O(k)*CTO)/CTN (1)

Figure 1. Completed sessions by the original Tomcat with different numbers of processors

Figure 2. Completed sessions with overload control with

different numbers of processors

Above figure are taken from [3,4] shows considerable

improvement using overload control mechanism as compared to without overload control policy and maximizes number of completed session

4. CPU Utilization Based SBAC Mechanism

A simple implementation of session based admission control is based on server CPU utilization presented in [1,2] . It measures and predicts the server utilization, rejects new sessions when the server becomes critically loaded and sends an explicit message of rejection to the client of a rejected session.

U_ac (admission control) is threshold which establishes the critical server utilization level to switch on the admission control policy);T1; T2; :::; Ti a sequence of time intervals used for making a decision whether to admit (or to reject) new sessions during the next time interval, this sequence is defined by the ac-interval length; F_ac is an ac-function used to evaluate the predicted utilization and distinguish two different values for server utilization ; U_measured(i) is a measured server utilization during;Ti{ the i-th ac-interval}; U_predicted( i+1) is a predicted utilization computed using a given ac-function f_ac after ac-interval Ti and before a new ac-interval Ti+1 begins,. U_predicted(i+1)=f_ac(i+1) (2) f_ac(1)=U_ac (3) f_ac(i+1)=(1-k)*f_ac(i)+k*U_measured(i) (4) where k is a damping coefficient between 0 and 1, and it is called ac-weight coefficient. which cover the space between ac-stable and ac-responsive policies. A web server with an admission control mechanism re-evaluates its admission strategy on intervals T1; T2; :::; Ti; ::: boundaries. Web server behavior for the next time interval Ti+1 is defined in [1,2] in the following way :.If ( U_predicted(i+1) > U_ac) then any new session arrived during Ti+1 will be rejected, and web server will process only requests belonging to already accepted sessions or if ( U_predicted(i+1) < U_ac ) then web server during Ti+1 is functioning in a usual mode: processing requests from both new and already accepted sessions. It is the simplest techniques but are not giving very satisfactory results. 5. Adaptive Admission Control Technique

Predictive admission control strategy and Hybrid

admission policies as discussed in[1,2] allow the design of a powerful admission control mechanism which tunes and adjusts itself for better performance across different workload types and different traffic loads[2].

Predictive admission control strategy evaluates the observed workload and makes its prediction for the load in the nearest future. It consistently shows the best performance results for different workloads and different traffic patterns. For workloads with short average session length, predictive strategy is the only strategy which provides both highest server throughput in completed sessions and no (or, practically no) aborted sessions.


157

Hybrid admission control strategy which tunes itself to be more responsive or more stable on a basis of observed quality of service. It successfully combines most attractive features of both responsive and stable policies. It improves performance results for workloads with medium to long average session length. 6. Comparative study of SBAC Techniques

CPU utilization based implementation presented in [1,2] is the simplest implementation of session based admission control but can break under certain rates and not work properly, reason is that the decision ,whether to admit or reject new sessions, is made at the boundaries of ac-intervals and this decision can not be changed until the next ac-interval. However, in presence of a very high load, the number of accepted new sessions may be much greater than a server capacity, and it inevitably leads to aborted sessions and poor session completion characteristics

Hybrid admission control strategy covered in [2] which tunes itself to be more responsive or more stable on a basis of observed quality of service. It successfully combines most attractive features of both ac-responsive and ac-stable policies. It improves performance results for workloads with medium to long average session length.

Predictive admission control strategy also covered in [2] which estimates the number of new sessions a server can accept and still guarantee processing of all the future session requests. This adaptive strategy evaluates the observed workload and makes its prediction for the load in the nearest future. It consistently shows the best performance results for different workloads and different traffic patterns. For workloads with short average session length, predictive strategy is the only strategy which provides both: highest server throughput in completed sessions and no (or, practically no) aborted sessions. Session-based adaptive overload control mechanism based on SSL connections differentiation and admission control presented in [3,4] prioritizes resumed connections maximize the number of sessions completed and also limits dynamically the number of new SSL connections accepted depending on the available resources and the number of resumed SSL connections accepted, in order to avoid server overload. 7. Conclusion

SSL is commonly used for secure http connections where

sensitive information is going to be sent along networks. SSL session integrity is a critical metric in e-commerce. Overload can lead e-commerce applications to considerable revenue losses or may cause response times to grow to unacceptable levels hence overload control is a critical goal. To meet this goal either apply predictive or hybrid overload control strategy based on session length which tunes itself for giving better performance according to different workload or an alternative approach is to apply SSL connection differentiation and admission control technique which prioritizes resumed SSL session over new session for overload control. These admission control mechanism based on

session will maximizes the number of sessions completed successfully and allow e-commerce sites to increase the number of transactions completed, therefore help in enhancing security and performance.

References [1] L. Cherkasova, P. Phaal “Session Based Admission

Control: a Mechanism for Improving the Performance of an Overloaded Web Server.” HP Laboratories Report No. HPL-98-119, June, 1998.

[2] L. Cherkasova, P. Phaal, “Session-based admission control: A mechanism for peak load management of commercial websites” IEEE Transactions on Computers LI (6), pp. 669–685 ,2002.

[3] Jordi Guitart, David Carrera, Vicenç Beltran, Jordi Torres and Eduard Ayguade “Session-Based Adaptive Overload Control for Secure Dynamic Web Applications” In Proceeding of International conf on Parallel Processing (ICPP) , pp. 341-349, 2005.

[4] Jordi Guitart , Vicenc Beltran , David Carrera , Jordi Torres , Eduard Ayguade “Designing an overload control strategy for secure e-commerce applications” LI (XV), pp. 4492-4510 , 2007.

[5] M. Harchol-Balter, B. Schroeder, N. Bansal, M. Agrawal, “Size-based scheduling to improve web performance” ACM Transactions on Computer Systems , XXI (II) , pp. 207–233 , 2003

[6] D. Mosberger, T. Jin “A tool for measuring web server performance” Workshop on Internet Server Performance (WISP’98) in conjunction with SIGMETRICS’98 Madison, Wisconsin, USA, pp59–67 , 1998 .

[7] S. Elnikety, E. Nahum, J. Tracey, W. Zwaenepoel, “A method for transparent admission control and request scheduling in e-commerce web sites” 13th International Conference on World Wide Web (WWW’04), New York, USA, pp. 276–286, 2004

[8] B. Urgaonkar, P. Shenoy, “Cataclysm: Handling extreme overloads in internet services” Tech. Rep. TR03-40, Department of Computer Science, University of Massachusetts, USA, December 2003

[9] H. Chen, P. Mohapatra, “Overload control in QoS-aware webservers” Elsevier journal Computer Networks XLII (I) , pp.119–133, 2003

[10] A.O. Freier, P. Karlton, C. Kocher, “The SSL Protocol.Version 3.0” November 1996. Available: http://wp.netscape.com/eng/ssl3/ssl-toc.htm

R K Pateriya M.Tech & B.E. in Computer Science & Engg. and working as Associate Professor in Information Technology Department of MANIT Bhopal . Total 17 Years Teaching Experience ( PG & UG ).


158

Dr. J. L . Rana Professor & Head of Computer Science & Engg deptt. in MANIT Bhopal .He has received his PhD from IIT Mumbai & M.S. from USA (Huwaii) .He has Guided Six PhD.

Dr. S. C. Shrivastava Professor & Head of Electronics Engg. department of MANIT Bhopal. He has Guided three PhD , 36 M.Tech and presented nine papers in international & twenty papers in national conference in India.

vol2 no6

Documents

Transcript of vol2 no6