Research Article Multi-Objective Approach for Energy-Aware...
Transcript of Research Article Multi-Objective Approach for Energy-Aware...
Hindawi Publishing CorporationThe Scientific World JournalVolume 2013 Article ID 350934 13 pageshttpdxdoiorg1011552013350934
Research ArticleMulti-Objective Approach for Energy-Aware WorkflowScheduling in Cloud Computing Environments
Sonia Yassa12 Rachid Chelouah1 Hubert Kadima1 and Bertrand Granado2
1 LRIS Laboratory EISTI Avenue du Parc 95011 Cergy-Pontoise France2 ETIS Laboratory CNRS UMR8051 University of Cergy-Pontoise ENSEA 6 Avenue du Ponceau 95014 Cergy-Pontoise France
Correspondence should be addressed to Sonia Yassa soniayassaeistieu
Received 6 August 2013 Accepted 12 September 2013
Academic Editors S H Rubin and A F Zobaa
Copyright copy 2013 Sonia Yassa et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited
We address the problem of scheduling workflow applications on heterogeneous computing systems like cloud computinginfrastructures In general the cloud workflow scheduling is a complex optimization problem which requires considering differentcriteria so as to meet a large number of QoS (Quality of Service) requirements Traditional research in workflow scheduling mainlyfocuses on the optimization constrained by time or cost without paying attention to energy consumptionThemain contribution ofthis study is to propose a new approach for multi-objective workflow scheduling in clouds and present the hybrid PSO algorithmto optimize the scheduling performance Our method is based on the Dynamic Voltage and Frequency Scaling (DVFS) techniqueto minimize energy consumption This technique allows processors to operate in different voltage supply levels by sacrificingclock frequencies This multiple voltage involves a compromise between the quality of schedules and energy Simulation resultson synthetic and real-world scientific applications highlight the robust performance of the proposed approach
1 Introduction
Cloud computing presents an interesting technology thatfacilitates the execution of scientific and commercial applica-tions It provides on demand flexible and scalable services tocustomers through a pay per use basis It can usually providethree kinds of services IaaS (Infrastructure as a Service) PaaS(Platform as a Service) and SaaS (Software as a Service)These services are offered with different service levels so as tomeet the needs of various customer groups Although manycloud services have a similar functionality (eg computingservices storage services network services etc) they differfrom each other by non-functional qualities termed QoS(Quality of Service) parameters such as service time servicecost service availability service energy consumption serviceutilization and so forth
These QoS parameters may be defined and proposed bydifferent SLAs (Service Level Agreements) An SLA specifiesthe QoS requirements of negotiated resources the minimumexpectations and limits that exist between consumers andproviders Applying such an SLA represents a binding con-tract Lack of such agreements can lead applications to move
away from the cloud and will compromise the future growthof cloud computing
Several scientific applications such as those of bioinfor-matics chemistry and astronomy contain a great number oftasks that have precedence constraints They can be definedby DAGs (Directed Acyclic Graph) These scientific work-flows typically involve complex data of different sizes and longterm computer simulations They need high computationpower and the availability of large infrastructures that gridand more recently cloud computing environments providewith different QoS levels
Due to the importance of workflow applications severalresearch projects have been conducted to develop work-flow management systems with scheduling algorithms Theprojects Condor Dagman [1] Gridbus toolkit [2] Iceni [3]Pegasus [4] and so forth are designed for grids whereascloudbus toolkit [5] SwinDeW-C [6] VGrADS [7] and soforth are developed for clouds These systems can be viewedas a type of platform service facilitating the automation ofscientific and commercial applications on the grid and cloudby masking their orchestrations and executions
2 The Scientific World Journal
In order to effectively schedule the tasks and data appli-cations on these cloud environments workflow manage-ment systems require more elaborated scheduling strategiesto meet QoS constraints and the precedence relationshipsbetween workflow tasks The study of workflow schedulingis becoming an important challenge in the area of cloudcomputing
The workflow scheduling in the cloud is a difficultproblem This problem is even more difficult when there areseveral factors to be considered namely (1) the various QoSrequirements of customers like service response time servicecost and so forth (2) the heterogeneity dynamicity andelasticity of cloud services (3) the various ways of combiningthese services to execute workflow tasks (4) the transfer oflarge volumes of data and so forth However the workflowscheduling problem is seen as a combinatorial problemwhere it is impossible to find the globally optimal solutionby using simple algorithms or rules It is well known as anNP-complete problem [8] and depends on the problem size
The workflow scheduling problem has been widely stud-ied in many previous works [9ndash12] Most of these workshave concentrated only on two QoS parameters namelythe deadline and budget In this paper we extend theseworks to handle multiple QoS requirements We address theQoS-based workflow scheduling which aims to minimizethe cost and total time execution of user applications asspecified in the SLA Furthermore the scheduler must alsobe able to schedule workflow tasks so as to maximize theprovider profits by minimizing energy consumption whilepreserving the users QoS preferences We achieve this byusing an iterative method called Multi-objective DiscreteParticle SwarmOptimization (MODPSO) combinedwith theDynamic Voltage and Frequency Scaling (DVFS) techniqueThis last one allows a compromise between system perfor-mance and energy consumption
The proposed approach is assessed by simulation runson a set of synthetic and real-world scientific applicationsSimulation results showed that this new multi-objectivealgorithm significantly improves the performance of relatedapproaches
The remainder of this paper is organized as followsSection 2 reviews several related works Section 3 presentsthe problemmodeling of theQoS basedworkflow schedulingSection 4 describes in detail our scheduling heuristic calledDVFS-MODPSO Section 5 shows an experimental evalu-ation of our heuristic Section 6 concludes the paper anddiscusses some future works
2 Related Work
The workflow scheduling problem in heterogeneous com-puting systems is an NP-hard optimization problem [8]meaning that the amount of computation needed to findoptimum solutions increases exponentially with the problemsize Previous works have proposed many heuristic andmeta-heuristic based approaches [13ndash16] to solve this prob-lem One of the most widely used heuristics for schedulingworkflow application is the Heterogeneous Earliest Finish
Time (HEFT) algorithm developed by Topcuoglu et al [17]HEFT is a static scheduling algorithm that attempts to min-imize execution time (makespan) It preserves the workflowprecedence constraints and produces a good schedule length
Most of these previous works have focused on mini-mizing the workflow execution time without consideringthe usersrsquo budget constraint However with the market-oriented business model in cloud computing environmentswhere users are billed for their consumption of resourcesseveral works that consider usersrsquo budget and deadline havebeen proposed [18ndash21] In [22] a study indicating howto schedule scientific workflow applications with budgetand deadline constraints onto computational grids usinggenetic algorithms is presented Authors in [6] proposedan improved cost-based scheduling algorithm for makingefficient scheduling of tasks to available resources in cloudIn [9] a particle swarm optimization (PSO) based heuristic isused to minimize the execution cost of scheduling workflowapplications to cloud resources
Besides makespan and cost energy consumption isbecoming more and more important in the cloud computingenvironments However cloud providers must adopt mea-sures not only to meet the userrsquo QoS requirements butalso to ensure that their profit margin is not dramaticallyreduced due to high energy consumptions The energyefficiency can conflict with the other QoS requirements(makespan cost) Incorporating the energy consumptioninto the workflow scheduling adds another layer of complex-ity Therefore recent works have concentrated on developingenergy-aware scheduling algorithms They have examinedvarious techniques such as dynamic power managementDynamic Voltage and Frequency Scaling (DVFS) or resourcehibernation [23ndash26] Authors in [27] presented an onlinedynamic power management strategy with many power-saving states They proposed a min-min based energy-awarescheduling algorithm to minimize energy consumption inheterogeneous computing systems In [26] a dynamic slackallocation technique which tries to use idle (slack) timeslots of processors to lower supply voltage (frequencyspeed)is presented These slack time slots occur due to earliercompletion andor dependencies of tasks Several DVS-basedapproaches for slack allocation have been proposed for bothindependent [28ndash35] and precedence-constrained [36ndash42]tasks In [43] an energy-aware scheduling algorithm anddetailed discussion of slack time computation are presentedThis scheduling algorithm reduces voltages during the com-munication phases between parallel tasks on homogeneousprocessors In [44] an Energy Conscious Scheduling heuris-tic (ECS) is proposed The heuristic is devised with relativesuperiority as a novel objective function which takes intoaccount energy andmakespan ECS is used to improve the bi-objective genetic algorithm proposed in [45] This latter hasbeen extended in [46] to a parallel model of their approach
All of these presented works have focused on optimizingeither a single or two objectives but none of them consider therelationships between several objectives namely the relation-ship between energy makespan and cost They do not takeinto account how each one of these criteria can affect othersTo deal with these misses we propose a Multi-Objective
The Scientific World Journal 3
Discrete Particle Swarm Optimization algorithm combinedwith DVFS technique (DVFS-MODPSO) to optimize allthree objectives at the same time Our new approach providesa set of solutions named Pareto solutions (ie non-dominatedsolutions) enabling the user to select the desired tradeoff
To the best of our knowledge none of the previ-ous scheduling approaches deal with the three-dimensionalmakespancostenergy optimization when tackling the prob-lem of scheduling workflow applications on heterogeneouscomputing environments such as the cloud computing oneswhich constitute our key novelties
3 Problem Modeling
In this section we describe our system model in a formalway Our ultimate goal is to distribute workflow tasks amongcloud services so as to optimize both the usersrsquo QoS criteriaand cloud providersrsquo profits by saving energy consumption oftheir servicesTherefore we first present the cloud computingmodel Then we describe our workflow model and the QoSparameters we deal We conclude this section by describingthe scheduling model formalized as a multi-objective opti-mization problem we solve
31 Cloud Computing Model The cloud computing systemused in this work is a set of resources offered by a cloudprovider to run client applications Our cloud model isinspired by the model described in [45] We assume that thecloud is hosted in data centers composed of heterogeneousmachines These data centers deliver a variety of serviceshosted on thousands of IT servers which are made availableas subscription-based services in a pay-as-you-go model Inour model the cloud computing system consists of a set of119875 = 119901
1 1199012 119901
119898 heterogeneous processors which are
fully interconnected The processors have varied processingcapability delivered at different processing prices (see ec ofTable 1) Each processor 119901
119895isin 119875 is DVFS-enabled that is it
can operate with different VSLs (Voltage Scaling Level iedifferent clock frequencies) For each processor 119901
119895isin 119875
a set 119881119895of V VSLs is randomly and uniformly distributed
among three different sets of VSLs (Table 1) We consider thatprocessors consume energy during periods of inactivity thatis when a processor is idling it is assumed that the lowestvoltage is supplied [44] Because clock frequency transitionoverheads usually take a negligible amount of time theseoverheads are not considered in this paper and the inclusionof such an overheadwill have no bearing on the overall modelof the proposed study
Additionally each processor 119901119895isin 119875 has a set of links
119871119901119895= 119897119895 1199011 119897119895 1199012 119897119895 119901119896 1 le 119896 le 119898 where 119897
119895119894isin 119877+ is
the available bandwidthmdashmeasured in Mega bits per second(Mbps)mdashin the link between processors 119901
119895and 119901
119894 with 119897
119895119895=
1 We assume that a message can be transmitted from oneprocessor to another while a task is executed on the recipientprocessor Finally communication between tasks executed onthe same processor is neglected Table 1 shows DVFS levelsRelative speeds (RSpeed) and execution costs (ec) for threeprocessor classes (TURIONMT-34 OMAP PENTIUMM)
32 Workflow Application Model We model a cloud work-flow application as a Directed Acyclic Graph (DAG) denotedas 119866(119881 119864) The set of nodes 119881 = 119879
1 119879
119899 represents
the tasks in the workflow application the set of arcs denotesprecedence constraints and the controldata dependenciesbetween tasks An arc is in the form of 119889
119894119895= (119879119894 119879119895) isin 119864
where119879119894is called the parent task of119879
119895119879119895is the child task of119879
119894
119889119894119895is the data produced by119879
119894and consumed by119879
119895We assume
that a child task cannot be executed until all of its parent taskshave been completed In a given task graph a task with noparent is referred as an entry task and one without any childis called an exit task Since our algorithm involves only oneentry and one exit tasks we add two dummy tasks 119879entry and119879exit which have zero execution time to the beginning andthe end of the workflow respectively These dummy tasks areconnected with zero-weight arcs to the actual entry and exittasks respectively
We assume that each task 119879119894isin 119881 has an associated
basic execution time which is an independent value for eachmachine We denote 119908
119894119895 the basic computation time of a
task 119879119894on a compute resource 119901
119895at maximum speed and
voltage (ie it corresponds to Level 1 in Table 1) The averageexecution time of the task 119879
119894is defined as
119908119894=
119898
sum119895=1
119908119894119895
119898 (1)
Real computation time119908119903119894119895119896
of the task 119879119894on machine 119901
119895
using relative execution speed 119904119896119895is defined as
119908119903119894119895119896=119908119894119895
119904119896119895
(2)
We also assume that every edge (119879119894 119879119895) isin 119864 is associated
with value tr119894119895 representing the time needed to transfer data
from 119879119894to 119879119895 The transfer time can be calculated according
to the bandwidth 119897119909119910
between the resources executing thesetasks (119901
119909and 119901
119910resp) as follows
tr119894119895=119889119894119895
119897119909119910
(3)
However a communication time is only required whentwo tasks are assigned to different processors That is thecommunication time when tasks are assigned to the sameprocessor can be neglected that is 0
In general the execution costs (ec) and transmission costs(trc) are inversely proportional to the execution times andtransmission times respectively
We define pred (119879119894) as the set of all predecessors of 119879
119894and
succ (119879119894) as the set of all successors of 119879
119894 An ancestor of node
119879119894is any node 119879
119896that is contained in pred (119879
119894) or any node
119879119895that is also an ancestor of any node 119879
119896contained in pred
(119879119896)The Earliest Start Time and the Earliest Finish Time of a
task 119879119894on a processor 119901
119895are represented as EST (119879
119894 119901119895) and
EFT (119879119894 119901119895) respectively EST (119879
119894) and EFT (119879
119894) represent the
earliest start and finish times on any processor respectively
4 The Scientific World Journal
Table 1 Voltage Scaling Levels and relative speeds of processors
VSLP1 TURIONMT-34 P2 OMAP P3 PENTIUMM
Volt Freq R Speed Volt Freq R Speed Volt Freq R Speed(V) (Ghz) () (V) (Ghz) () (V) (Ghz) ()
1 12 18 100 135 06 100 148 14 1002 115 16 8888 127 055 9166 144 12 96763 11 14 7777 12 05 8333 131 1 88144 105 12 6666 1 025 4166 118 08 79515 1 1 55555 09 013 2083 096 06 64426 09 08 4444Cost ec = 114 ec = 057 ec = 076
Table 2 Task execution times and priorities
Task (119879119894) 119901
11199012
1199013
120596119894
Priority (Pr(119879119894))
1 14 16 9 13 1080002 13 19 18 1667 770003 11 13 19 1433 800004 13 8 17 1267 800005 12 13 10 1167 690006 13 16 9 1267 633337 07 15 11 11 426678 5 11 14 10 356679 18 12 20 1667 4433310 21 7 16 1467 14667
119875119886V(119879119894 119901119895) is defined as the earliest time when processor 119901
119895
will be available to begin executing task 119879119894 Hence
119864119878119879 (119879119894 119901119895) =
119900 if 119879119894= 119879entry
max 119875119886V(119879119894 119901119895) 120572
(4)
where 120572 = max119879119895 isin pred (119879119894 )(119864119865119879(119879119895 119901119896) + tr
119895119894
119864119865119879 (119879119894 119901119895) = 119908
119894119895+ 119864119878119879 (119879
119894 119901119895) (5)
Note that the Actual Start Time and Actual Finish Timeof a task 119879
119894on a processor 119901
119895 denoted as AST (119879
119894 119901119895) and
AFT (119879119894 119901119895) can be different from its earliest start EST (119879
119894 119901119895)
and finish EFT (119879119894 119901119895) times if the actual finish time of
another task 119879119896scheduled on the same processor is later than
its EST (119879119896 119901119895) [44]
Figure 1 depicts a workflow application with 10 tasks andthe Table 2 provides its details (given in [17]) The valuespresented in the last column of the table represent the priorityof the tasks The priority of task 119879
119894represented by Pr(119879
119894) is
computed recursively by traversing the DAG upward startingfrom the exit task 119879exit as follows (6)
Pr (119879119894) =
119908119879exit if 119879
119894= 119879exit
119908119894+ 120573 otherwise
(6)
where120573 = max119879119895 isin succ (119879119894)119879119903119894119895 + Pr(119879
119895)
1
2 5 6
7 8
10
3 4
9
18 12 9 1114
19 16 2723 23 1513
1311 17
Figure 1 An example of workflow (given in [17]) with the tasknumbers 119879
119894inside nodes and values of 119889
119894119895function next to the
corresponding edges
33 QoS Parameter Models
331 Energy Model Among the main system-level energy-saving techniques Dynamic Voltage Scaling (DVS) operateson a simple principle decreases the supply voltage (and so theclock frequency) to the CPU so as to consume less power
In this work we use a model of energy derived from thepower consumption model in digital complementary metal-oxide semiconductor (CMOS) logic circuits [44] Under thedynamic power model the processor power is dominated bythe dynamic power which is given by
119875dynamic = 119860119862efV2119891 (7)
where119860 is the number of switches per clock cycle119862ef denotesthe effective charged capacitance V is the supply voltage and119891 denotes the operational frequency Equation (7) shows thatthe supply voltage is the dominant factor hence its reductionwould be most influential to lower power consumption
The energy consumption of the execution of a workflowapplication used in this paper is defined as
119864119888=
119899
sum119894=1
119860119862efV2
119894119891119894119908119903lowast
119894
=
119899
sum119894=1
120574V2119894119891119894119908119903lowast
119894
(8)
The Scientific World Journal 5
where 120574 = 119860119862ef is assumed constant for a given machineV119894 119891119894are the voltage supply and frequency of the processor
on which task 119879119894is executed respectively and 119908119903lowast
119894is the real
completion time of task 119879119894on the scheduled processor In the
idle time the processor turns into sleep mode and thus thevoltage supply and relative frequency are at the lowest levelSo the energy consumption during idle periods of processorsis defined as
119864119894=
119898
sum119895=1
sumidle119895119896isin IDLE119895
120574V2min 119895 119891min 119895 119871119895119896 (9)
where IDLE119895
is the set of idling slots on machine119901119895 Vmin 119895(119891min 119895) is the lowest supply voltage (frequency)
on 119901119895 and 119871
119895119896is the amount of idling time for idle
119895119896
Then the total energy 119864total utilized by the cloud system forcompletion of the workflow application can be defined asfollows
119864total = 119864119888+119864119894 (10)
332 Time Model The completion time 119879total is themakespan from the user submitting a workflow untilreceiving the results It is defined as the actual finish timeof the exit task It is calculated in equation 11 and consistsof the workflow execution time and network transmissiontime The execution time depends on both the workloadand system performance The network transmission timedepends on both the network latency and the input data size
119879total = maxAFT (119879exit) (11)
333 Cost Model As result of the marketization characteris-tic of current services most cloud providers have set a pricefor their services They have fixed the price for transferringbasic data unit (eg per MB) between two services and theprice for processing basic time units (eg per hour)The cost119862total of running a cloud workflow is defined in Formula (12)It consists of processing cost 119862ex and data transfers cost 119862tr
119862total = 119862ex + 119862tr (12)
The processing cost for a given task119879119894depends on the real
completion time of 119879119894on the scheduled processor (119908119903lowast
119894) and
the hourly price of this processor (ec119894) Thus 119862ex is given by
119862ex =119899
sum119894=1
119908119903lowast
119894lowast ec119894 (13)
The data transfer cost (119862tr) is described as follows
119862tr =119899
sum119894=1
119899
sum119895=1
119889119894119895lowast trc119894119895 (14)
where 119889119894119895characterizes the output file size from task119879
119894to task
119879119895 and tr 119888
119894119895represents the cost of communication from the
processor where 119879119894is mapped to another processor where 119879
119895
is mappedThe cost of communication is added to the overallcost only when two tasks have data dependency betweenthem (ie 119889
119894119895gt 0) For two or more tasks running on the
same processor the transfer cost is neglected
334 Scheduling Model Given (1) A cloud provider thatoffers a set 119875 of 119898 heterogeneous processors and (2) auser workflow application composed of a set 119879 of 119899 tasksthat have to be executed on these processors The workflowscheduling problem is to construct a mapping 119872 of tasksto processors (without violating precedence constraints) thatminimizes the following conflicting objectives makespancost and energy consumption as low as possible Thereforethe workflow scheduling problem can be formulated as amathematical optimization problem
MakespanMinimize 119879total (119872)
CostMinimize 119862total (119872)
EnergyMinimize119864total (119872)
(15)
4 Workflow Scheduling Based on DiscreteParticle Swarm Optimization
This section starts with a brief overview on multi-objectivecombinatorial optimization and Particle swarm optimizationalgorithm Afterwards our new Multi-Objective DiscreteSwarmOptimization combined with DVFS technique will bepresented
41 Multi-Objective Optimization A Multi-objective Opti-mization Problem (MOP) with 119898 decision variables and 119899objectives can be formally defined as
Min (119910 = 119891 (119909) = [1198911(119909) 119891
119899(119909)]) (16)
where 119909 = (1199091 119909
119898) isin 119883 is an 119898-dimensional decision
vector 119883 is the search space 119910 = (1199101 119910
119899) isin 119884 is the
objective vector and Y the objective-spaceIn general MOP there is no single optimal solution
with regards to all objectives This is also the case forthe multi-objective optimization problem addressed in thispaper As given in (15) there are three conflicting objectivesminimizing execution time minimizing execution cost andminimizing energy consumption In such problems thedesired solution is considered to be the set of potentialsolutions which are all optimal in some objectives This set isknown as the Pareto optimal setWe provide some definitionsof the Pareto concepts used in MOP as follows (withoutloss of generality we suppose that the objectives are to beminimized)
(i) Pareto dominance For two decision vectors 1199091 and 1199092dominance (denoted by ≺) is defined as follows
1199091≺ 1199092lArrrArr forall119894 119891
119894(1199091) le 119891119894(1199092) and exist119895119891
119895(1199091) lt 119891119895(1199092)
(17)
The decision vector 1199091is said to dominate 1199092 if andonly if 1199091 is as good as 1199092 considering all objectivesand 1199091 is strictly better than 1199092 in at least oneobjective
6 The Scientific World Journal
(ii) Pareto optimally A decision vector 1199091 is said to bePareto optimal if and only if
∄1199092isin 119883 119909
2≺ 1199091 (18)
(iii) Pareto optimal set The Pareto optimal set 119875119878is the set
of all Pareto optimal decision vectors
119875119878= 1199091isin 119883 | ∄119909
2isin 119883 119909
2≺ 1199091 (19)
(iv) Pareto optimal frontThePareto optimal front119875119865is the
image of the Pareto optimal set in the objective space
119875119865= 119891 (119909) = (119891
1(119909) 119891
119899(119909)) | 119909 isin 119875
119904 (20)
1199091 is said to be non-dominated regarding a given set
if 1199091 is not dominated by any decision vectors in theset
Thepareto optimal decision vector cannot be enhanced inany objective without causing degradation in at least anotherobjective A decision vector is said to be Pareto optimal whenit is not dominated in the whole search space
42 Particle Swarm Optimisation
421 The Standard PSO PSO is a population-based stochas-tic optimization technique developed by Kennedy and Eber-hart in 1995 [47] It is inspired by the social behavior of insectcolonies bird flocks fish schools and other animal societies Itis also related to evolutionary computation PSOhas attractedsignificant attention from many researchers due to both itssimplicity of use and optimization via social behavior In factPSO has good performance requires low computational costIt is effective and easy to implement as it uses numericalencoding
A particle in PSO is analogous to a fish or bird movingin the 119863-dimensional search space All particles have fitnessvalues indicating their performances which are problemspecific and velocities which direct the flight of particlesEach particle position at any given time is influenced by bothits best position called 119901119861119890119904119905 and the position of the bestparticle in a problem space referred to as 119892119861119890119904119905 Thereforeparticles tend to fly towards a better search area duringthe search process A particle status on the search spaceis characterized by two elements namely its velocity andposition which are updated in every generation as follows
119881119896+1
119894= 120596119881
119896
119894+ 11988811199031(119901119861119890119904119905
119894minus 119883119896
119894)
+ 11988821199032(119892119861119890119904119905 minus 119883
119896
119894)
119883119896+1
119894= 119883119896
119894+ 119881119896+1
119894
(21)
where 119881119896+1119894
is the velocity of particle 119894 at iteration 119896 + 1119881119896119894is the velocity of particle 119894 at iteration 119896 120596 is the inertia
weight 1198881and 1198882are the acceleration coefficients (cognitive
and social coefficients) 1199031and 1199032are the random numbers
between 0 and 1119883119896119894is the current position of particle 119894 at the
119896th iteration 119901119861119890119904119905119894is the best previous position of the 119894th
particle 119892119861119890119904119905 is the position of best particle in the swarmand119883119896+1
119894is the position of 119894th particle at 119896 + 1 iteration
The procedure for standard PSO is as follows
(1) Initialize a population of particles with random posi-tions and velocities in the search space
(2) Evaluate the objective values of all particles set 119901119861119890119904119905of each particle equal to its current position and set119892119861119890119904119905 equal to the position of the best initial particle
(3) Update the velocity and the position of each particleaccording to (21)
(4) Map the position of each particle in the solution spaceand evaluate its fitness value according to the desiredoptimization fitness function
(5) For each particle compare its current objective valuewith its 119901119861119890119904119905 value If the current value is better thenupdate 119901Best with the current position and objectivevalue
(6) Determine the best particle of the current whole pop-ulation with the best objective value If the objectivevalue is better than that of 119892119861119890119904119905 then update 119892Bestwith the current best particle
(7) If the stopping criterion ismet then output 119892119861119890119904119905 andits objective value otherwise go to Step (2)
The original design of PSO is appropriate for findingsolutions to continuous optimization problems However asthe workflow scheduling discussed in this paper is both adiscrete andmulti objective problem in nature we propose aneffective approach to address this problem by using a discreteversion of the Multi-Objective PSO (MODPSO) combinedwith DVFS techniqueThe key issues of our approach consistof (1) defining the position and velocity of the particlesbased on the features of discrete variables of the workflowscheduling (2) solving the multi-objective aspect of theproblem by modifying PSO so as to generate a set of non-dominated solutions satisfying the different objectives underconsideration instead of one solution
422 Handling Workflow Scheduling Using DVFS-MODPSOIn general a workflow scheduling can be defined by a setof triplets 119872 = [⟨119879
119894 119875119895 119871119896⟩](119894 isin [1 119899] 119895 isin [1119898] 119896 isin
[1 119871(119895)]) 119899 is the number of workflow tasks to be scheduled119898 is the number of processors available in the cloud environ-ment and 119871(119895) is the number of operating points (VSLs) of the119895th processor
For the sake of clarity the variables and rules of DVFS-MODPSO for solving workflow scheduling can be depictedas follows
The position of a particle represents a feasible solutionto the scheduling problem It consists of a set of⟨task(119879
119894) service (119875
119895) VSl (119871
119896)⟩ triplets Each triplet
means that a task (119879119894) is assigned to a processor (119875
119895)
with a voltage scaling level (119871119896) It also indicates
The Scientific World Journal 7
that the position satisfies the precedence constraintbetween tasks
The process of generating a newposition for a selectedparticle in the swarm is depicted in the followingformulas
119881119896+1
119894= 119881119896
119894oplus ((119877
1otimes (119901119861119890119904119905
119894⊖ 119883119896
119894))
oplus (1198772otimes (119892119861119890119904119905 ⊖ 119883
119896
119894)))
(22)
119883119896+1
119894= 119883119896
119894oplus 119881119896+1
119894 (23)
The operator definitions are as follows
(i) The substract operator (⊖) the difference betweentwo particle positions designated as 1199091 and 1199092 isdefined as a set of triplets in which each triplet⟨task serviceVSL⟩ shows whether the contents ofthe corresponding elements in 1199091 are different fromthose of 1199092 or not If so that triplet gets its values(service and VSL) from the position that has thelowest value of theVSL For those triplets that have thesame content in 1199091 and 1199092 their corresponding VSLsare decreased (Note that the scaling of VSL makesfluctuations on the energy makespan and cost)
(ii) The multiply operator (otimes) the multiplication betweennumber and velocity is defined as a set of tripletswhere a threshold 120572 isin [0 1] is defined arandom number 119903 is generated for each triplet⟨task serviceVSL⟩ compare 119903 and 120572 when 119903 ge 120572decrease the triplet VSL otherwise increase it Thisoperator adds the exploration ability to the algorithm
(iii) The add operator (oplus) the addition of two positions isdefined as the reservation of the dominated one
The Pseudocode 1 outlines the general steps of the DVFS-MODPSO algorithm
The algorithm begins by initializing the positions andvelocities of particles To obtain the position of a particle theVSL (voltage and frequency) of each resource is randomlyinitialized firstly then the HEFT algorithm is applied togenerate a feasible and efficient solution minimizing themakespan The process is repeated several times to initializethe positions of all particles of the swarm Initially thevelocity119881 and the best position for each particle 119901119861119890119904119905 areattributed as the particle itself The algorithm maintainsan external archive EA to store non-dominated particlesfound after the evaluation process (based on the paretodominance using the objectives mentioned in 33) After allthese initializations in the main loop of the algorithm thenew velocity and position of each particle are calculatedrespectively after selecting the best overall position in theexternal archive and eventually perform a mutation then theparticle is evaluated and its corresponding 119901119861119890119904119905 is updatedThe external archive is updated after every iteration Oncethe termination condition is reached the external archivecontaining the Pareto front is returned as the result
5 Experimental Evaluations
In this section we describe the overall setup of our experi-mentation effort and the results we have obtained from it tovalidate the new proposed approach
In our experiments we simulate a synthetic workflowapplication described in [17] (see Figure 1) and two commonworkflow structures parallel and hybrid structuresWe chosetwo well known real world applications namely a neuro-science workflow [48] for our parallel application and aprotein annotation workflow [49] developed by London e-Science Centre for our hybrid workflow application Figure 2shows their simplified representationsThe number indicatedin the parentheses next to each node represents the length ofthe task (the number of instructions to execute) in Millionsof Instructions (MI) The input and output files of each taskrange from 10MB to 1024MB
We have performed experiments on 3 4 5 6 and 12resources whose characteristics are shown in Table 1 Wechoose the pricing model associated with Amazon EC2(httpawsamazoncomec2) for the processing costs of thedifferent classes of resources and the pricing model given byAmazon CloudFront (httpawsamazoncomcloudfront)for the costs of transferring data unit between resources
As for DVFS-MODPSO the parameter settings areSwarm size 119878119873119906119898 = 50 particles and the maximum numberof iterations 119866 = 100
We evaluated the performance of our proposed DVFS-MODPSO on all the workflow instances described previouslyDue to the lack of works considering both the heteroge-neous configuration of our cloud and the three QoS metrics(makespan cost energy) at the same time we comparedour results with those of the HEFT heuristic we haveimplemented HEFT is one of themost widely used heuristicsfor DAGs in distributed heterogeneous computing systemswhich optimize the makespan
Figures 3 4 and 5 show sample Pareto fronts obtainedwith the DVFS-MODPSO and the solution computed byHEFT for the synthetic workflow the hybrid workflow andthe parallel workflow respectively
As can be seen from these figures unlike the HEFTalgorithm which gives one solution our proposed approachprovides a set of non dominated solutions These resultsillustrate the basic multi-objective definitions provided inSection 41
Now for comparing these two approaches and in orderto analyze the effectiveness of our proposition in terms ofthe values of makespan cost and energy consumption wehave been inspired by the methodology described in [43]where we compare the solution provided by HEFT to onlyone solution of the Pareto front set computed by our proposedmulti-objective DVFS-MODPSO
The evaluation process follows the next steps For eachworkflow instance we perform a first resolution usingHEFTin order to get one solution 119878 After a second resolution iscomputed using our proposed approach to obtain a set EA ofPareto solutions Next we select one solution 1198781015840 from the setEA of Pareto solutions This solution is the closest one to 119878 inthe sense of Euclidean distance Finally a comparison is donebetween 119878 and 1198781015840
8 The Scientific World Journal
Swarm initialization with HEFT(1) For i = 1 to SNum (SNum is the size of particle swarm)
(a) For j = 1 to m (m the number of processors)(i) Randomly initialize VSL(j) (randomly choose the voltagefrequency of processor from the set of its operating points)
(b) Initialize 119878[119894] with HEFT heuristic (S is the swarm of particles)(c) Initialize the velocity V of each particle
119881 [119894] = 0 (119878 [119894])
(d) Initialize the Personal Best (pBest) of each particle119901119861119890119904119905[119894] = 119878[119894]
(e) Evaluate objectives of each particleEvaluate S[i]
(f) Initialize the Global Best particle (gBest) with the best one among the SNum particlesg119861119890119904119905 = Best particle found in 119878
(2) End For(3) Add the nondominated solutions found in S into EA (EA is the External Archive storing the pareto front)(4) Initialize the iteration number 119905 = 0(5) Repeat until 119905 gt 119866 (119866 is the maximum number of iterations)
(a) For 119894 = 1 to SNum (swarm size)(i) Randomly select the global best particle for S from the External Archive EA and store its position in gBest(ii) Calculate the new velocity 119881[119894] according to (22)(iii) Compute the new position of 119878[119894] according to (23)(iv) If (119905 lt 119866 lowast PMUT ) then (PMUT is the probability of mutation)
Perform mutation on 119878[119894](v) Evaluate S[i]
(b) End For(c) Update the personal best solution of each particle 119878[119894]
if 119878 [119894] ≼ 119901119861119890119904119905119904 [119894] or (119878 [119894] sim 119901119861119890119904119905119904 [119894])Then 119901119861119890119904119905119904 [119894] = 119878 [119894]
(d) Update the External Archive EA(i) Add all new non-dominated solution in S into EAif 119878 [119894] 997410 119909 forall119909 isin 119864119860Then 119864119860 = 119864119860 cup 119878 [119894](ii) Remove all particles dominated by 119878[119894] in EA
119864119860 = 119864119860 minus 119910 isin 119864119860 | 119910 lt 119878[119894]
(iii) If the archive is full then randomly select the article to be replaced from EA(6) Return the pareto front (the set of non-dominated solutions from S and EA)
Pseudocode 1 DVFS-MODPSO based workflow scheduling
The different results are displayed on Figures 6 7 8 and 9and on Tables 3 and 4 They show the improvement achievedby our proposed approach according to the structure of theworkflow applications and the number of processors
Table 3 and Figure 6 show that our proposed approachimproves on average the result obtained by HEFT for thesynthetic workflow instance The makespan is reduced by005 the cost is reduced by 993 and the energy con-sumption is reduced by 2640 Figure 6 shows detailedimprovements of the proposed approach according to thenumber of processors
Table 4 and Figure 7 illustrate the gain obtained by ourapproach according to the kind of real world applicationsTable 4 also shows the detailed improvements when scalingthe number of processors As can be seen the QoS metricsare improved over HEFT in average by 095 for themakespan 108 for cost and 812 for the energy con-sumption when using the hybrid workflow application andaverage improvements of 295 formakespan 2215 for costand 209 for energy consumption when using the parallelworkflow
These evaluations confirm the results obtained from thesynthetic workflow This means that by using our DVFS-MODPSO we are able to improve not only the cost and energybut also themakespan onwhichHEFT algorithm is supposedto provide good results These results remain true for all kindof workflow applications
In our experiments we limited the number of resourcesto 12 as we found that for these kinds of workflowapplications some resources are not used at all Figure 8shows the resource loads when scheduling the syntheticworkflow on 12 resources Furthermore even though thenumber of resources increased the total cost and total energyconsumption could not always decrease as illustrated inFigure 9 From this figure we can see that the QoS metrics(1) firstly decreases as the number of resources increasesuntil achieving the number 6 This can be explained by thefact that when increasing the number of resources thereare fewer tasks executed in a resource therefore tasks canextend their execution times and the resource have more ofchances to scale down their voltages and frequencies whichcan be very effective in reducing total energy consumption
The Scientific World Journal 9
1 2 3 4
5
6 7
8 9 10
11
12
14
13
15
(300000) (600000) (600000) (900000)
(300000)
(150000) (150000)
(300000) (300000) (300000)
(600000)
(300000)
(150000)
(300000)
(600000)
SignalP COILS2 SEG PROSITE
TMHMM
ProsperoHMMer
PSI-BLASTBLAST
IMPALA
PSI-PRED
3D-PSSMSummary
Genomesummary
SCOP
(a)
1 3 5 7
9
11 1210
(400000) (400000) (400000) (400000)
(300000)
(500000) (500000) (500000)
Softmea
2 4 6 8(600000) (600000) (600000) (600000)
Reslice
14 1513
(600000) (600000) (600000)
Convert
Reslice Reslice Reslice
Slicer Slicer
Convert Convert
Slicer
Align wapAlign wapAlign wapAlign wap
(b)
Figure 2 Parallel and Hybrid workflows
10 The Scientific World Journal
75 80 85 90 95 100 8090
100110
120130
140
2060
100140180220
Ener
gy
DVFS-MODPSOHEFT
MakespanCost
Figure 3 Obtained non-dominated solutions for the syntheticworkflow
Ener
gy
DVFS-MODPSOHEFT
MakespanCost
200 400 600 800 100012001400160018002000 400800
12001600
2000
400200
600800
100012001400160018002000
Figure 4 Obtained non-dominated solutions for the hybrid work-flow
(2) After achieving 6 resources the QoS metrics begin torise the reason for this is that time executions are dominatedby interprocessors communications hence decreasing theopportunities for scaling down voltages and frequencies ofresources Consequently the threshold numbers of resourcesthat minimized the QoS metrics could be obtained
6 Conclusion
In this paper we propose a new algorithm called DVFS-Multi-Objective Discrete Particle Swarm Optimization(DVFS-MODPSO) for workflow scheduling in distributedenvironments such as cloud computing infrastructures DV-FS-MODPSO simultaneously optimizes several conflictingobjectives namely the makespan cost and energy in adiscrete space It produces a set of non-dominated solutionsthus providing more flexibility for users to assess theirpreferences and select a schedule that meets their QoS
Ener
gy
DVFS-MODPSOHEFT
Makespan Cost
400200600800
10001200140016001800
240020002200
450 500 550 600 650 700 7501200
14001600
18002000
Figure 5 Obtained non-dominated solutions for the parallel work-flow
0
10
20
30
40
50
3 4 5 6 12
MakespanCostEnergy
Figure 6 Gain over HEFT () according to the number ofprocessors
requirements Our approach exploits the heterogeneityand the marketization of cloud resources in order to findsolutions that optimize makespan and cost Furthermore ituses the Dynamic Voltage and Frequency Scaling (DVFS)technique to reduce energy consumption
We have evaluated our algorithm by simulating it withboth synthetic and real world scientific workflow applicationshaving different structures The results show that the pro-posedDVFS-MODPSO is able to produce a set of good Paretooptimal solutions The results also show that our approachprovides significant improvement not only in terms of the
The Scientific World Journal 11
0
5
10
15
20
25
HybridParallel
Makespan Cost Energy
Figure 7 Improvement according to real world applications
R17
R20
R37
R446
R50
R60
R733
R80
R90
R107
R110
R120
Figure 8 Resource loads for synthetic workflow
Table 3 Improvement for synthetic workflow
Num of proc Gain over Heft ()Makespan Cost Energy
3 0 2126 40084 022 464 15355 0 756 17266 003 652 267512 0 969 3255Average 005 993 2640
cost and the energy consumption but also in term of themakespan for which HEFT algorithm is supposed to givebetter results
Multi-objective optimization in cloud workflow schedul-ing is not a mature domain Most of the existing worksattempt to minimize either the makespan or cost Howeverwe plan to consider other objectives such as reliabilitysecurity in addition to the energy consumption We also
0
50
100
150
200
250
3 4 5 6 12
MakespanCostEnergy
Figure 9 QoS metrics evolutions according to resource numbersfor the synthetic workflow
Table 4 Improvement according to real world applications
Appli Num of proc Gain over Heft ()Makespan Cost Energy
Hybrid
3 405 1155 04024 028 0566 05075 000 1347 11246 000 0117 126512 043 2214 0761
Average 095 1080 0812
Parallel
3 235 5640 25524 1147 2094 32715 000 2142 11666 093 1170 315412 000 031 306
Average 295 2215 209
intend to apply our algorithm in a real world cloud orintegrate it in existing cloud toolkits such as Cloudbus forcomparing with other algorithms
References
[1] D Thain T Tannenbaum and M Livny Condor and the GridGrid Computing Making the Global Infrastructure a RealityJohn Wiley amp Sons Hoboken NJ USA 2003
[2] R Buyya and S Venugopal ldquoThe gridbus toolkit for serviceoriented grid and utility computing an overview and statusreportrdquo in Proceedings of the 1st IEEE International Workshopon Grid Economics and Business Models (GECON rsquo04) pp 19ndash36 IEEE CS Seoul Republic of Korea April 2004
[3] S McGough L Young A Afzal S Newhouse and J Darling-ton ldquoWorkflow enactment in ICENIrdquo in Proceedings of the UK
12 The Scientific World Journal
e-Science All Hands Meeting pp 894ndash900 IOP NottinghamUK September 2004
[4] E Deelman J Blythe Y Gil et al ldquoPegasus mapping scientificworkflows onto the gridrdquo in Grid Computing 2nd EuropeanAcrossGrids Conference AxGrids 2004 Nicosia Cyprus January28ndash30 2004 vol 3165 of Lecture Notes in Computer Science pp11ndash20 Springer New York NY USA 2004
[5] R Buyya S Pandey and C Vecchiola ldquoCloudbus toolkitfor market-oriented cloud computingrdquo in Cloud ComputingProceedings of the 1st International Conference CloudCom 2009Beijing China December 1ndash4 2009 vol 5931 of Lecture Notesin Computer Science pp 24ndash44 Springer New York NY USA2009
[6] Y Yang K Liu J Chen X Liu D Yuan and H Jin ldquoAnalgorithm in SwinDeW-C for scheduling transaction-intensivecost-constrained cloud workflowsrdquo in Proceedings of the 4thIEEE International Conference on eScience (eScience rsquo08) pp374ndash375 Indianapolis Ind USA December 2008
[7] L Ramakrishnan C Koelbel Y Kee et al ldquoVGrADS enablinge-Scienceworkflows on grids and cloudswith fault tolerancerdquo inProceedings of the Conference on High Performance ComputingNetworking Storage and Analysis (SC rsquo09) November 2009
[8] M L Pinedo Scheduling Theory Algorithms and SystemsSpringer Berlin 2008
[9] S Pandey L Wu S M Guru and R Buyya ldquoA particle swarmoptimization-based heuristic for scheduling workflow applica-tions in cloud computing environmentsrdquo in Proceedings of the24th IEEE International Conference on Advanced InformationNetworking and Applications (AINA rsquo10) pp 400ndash407 PerthAustralia April 2010
[10] S Abrishami and M Naghibzadeh ldquoDeadline-constrainedworkflow scheduling in software as a service cloudrdquo ScientiaIranica vol 19 no 3 pp 680ndash689 2012
[11] S Selvarani and G S Sadhasivam ldquoImproved cost-based algo-rithm for task scheduling in cloud computingrdquo in Proceedings ofthe IEEE International Conference on Computational Intelligenceand Computing Research (ICCIC rsquo10) pp 1ndash5 CoimbatoreIndia December 2010
[12] A Verma and S Kaushal ldquoDeadline and budget distributionbased cost-time optimization workflow scheduling algorithmfor cloudrdquo inProceedings of the IJCAon International Conferenceon Recent Advances and Future Trends in Information Technol-ogy (iRAFIT rsquo12) iRAFIT(7) pp 1ndash4 April 2012
[13] R C Correa A Ferreira and P Rebreyend ldquoIntegrating listheuristics into genetic algorithms for multiprocessor schedul-ingrdquo in Proceedings of the 8th IEEE Symposium on Parallel andDistributed Processing (SPDP rsquo96) pp 462ndash469 IEEEComputerSociety Washington DC USA October 1996
[14] E S H Hou N Ansari and H Ren ldquoGenetic algorithm formultiprocessor schedulingrdquo IEEE Transactions on Parallel andDistributed Systems vol 5 no 2 pp 113ndash120 1994
[15] M Grajcar ldquoGenetic list scheduling algorithm for schedulingand allocation on a loosely coupled heterogeneous multi-processor systemrdquo in Proceedings of the 36th Annual DesignAutomation Conference (DAC rsquo99) pp 280ndash285 ACM PressNew York NY USA June 1999
[16] R Sakellariou and H Zhao ldquoA hybrid heuristic for DAGscheduling on heterogeneous systemsrdquo in Proceedings of the13th Heterogeneous Computing Workshop pp 111ndash123 IEEEComputer Society Washington DC USA 2004
[17] H Topcuoglu S Hariri and M Y Wu ldquoPerformance-effectiveand low-complexity task scheduling for heterogeneous comput-ingrdquo IEEE Transactions on Parallel and Distributed Systems vol13 no 3 pp 260ndash274 2002
[18] J Yu and R Buyya ldquoWorkflow scheduling algorithms for gridcomputingrdquo in Metaheuristics for Scheduling in DistributedComputing Environments F Xhafa and A Abraham EdsSpringer Berlin Germany 2008
[19] WN Chen and J Zhang ldquoAn ant colony optimization approachto a grid workflow scheduling problem with various QoSrequirementsrdquo IEEE Transactions on Systems Man and Cyber-netics C vol 39 no 1 pp 29ndash43 2009
[20] X Liu J Chen Z Wu Z Ni D Yuan and Y Yang ldquoHandlingrecoverable temporal violations in scientific workflow systemsa workflow rescheduling based strategyrdquo in Proceedings of the10th IEEEACM International Symposium on Cluster Cloudand Grid Computing (CCGrid rsquo10) pp 534ndash537 MelbourneAustralia May 2010
[21] X Liu Y Yang Y Jiang and J Chen ldquoPreventing temporalviolations in scientific workflows where and howrdquo IEEE Trans-actions on Software Engineering vol 37 no 6 pp 805ndash825 2011
[22] J Yu and R Buyya ldquoScheduling scientific workflow applicationswith deadline and budget constraints using genetic algorithmsrdquoScientific Programming vol 14 no 3-4 pp 217ndash230 2006
[23] L Benini and G Micheli Dynamic Power Management DesignTechniques and CAD Tools Kluwer Academic New York NYUSA 1998
[24] S Irani S Shukla and R Gupta ldquoOnline strategies for dynamicpower management in system with multiple power-savingstatesrdquo ACM Transactions on Embedded Computing Systemsvol 2 no 3 pp 325ndash346 2003
[25] M T Schmitz B M Al-Hashimi and P Eles System-Level Design Techniques for Energy-Efficient Embedded SystemsSpringer New York NY USA 2005
[26] S U Khan and I Ahmad ldquoA cooperative game theoreticaltechnique for joint optimization of energy consumption andresponse time in computational gridsrdquo IEEE Transactions onParallel and Distributed Systems vol 20 no 3 pp 346ndash3602009
[27] Y Li Y Liu and D Qian ldquoA heuristic energy-aware schedulingalgorithm for heterogeneous clustersrdquo in Proceedings of the 15thInternational Conference on Parallel and Distributed Systems(ICPADS rsquo09) pp 407ndash413 Shenzhen China December 2009
[28] L K Goh B Veeravalli and S Viswanathan ldquoDesign of fast andefficient energy-aware gradient-based scheduling algorithmsheterogeneous embeddedmultiprocessor systemsrdquo IEEE Trans-actions on Parallel and Distributed Systems vol 20 no 1 pp 1ndash12 2009
[29] F Yao A Demers and S Shenker ldquoA scheduling model forreduced CPU energyrdquo in Proceedings of the 36th IEEE AnnualSymposium on Foundations of Computer Science pp 374ndash382Milwaukee Wis USA October 1995
[30] Y Shin and K Choi ldquoPower conscious fixed priority schedulingfor hard real-time systemsrdquo in Proceedings of the 36th AnnualDesign Automation Conference (DAC rsquo99) pp 134ndash139 NewOrleans La USA June 1999
[31] I Hong D Kirovski G QuM Potkonjak andM B SrivastavaldquoPower optimization of variable-voltage core-based systemsrdquoIEEE Transactions on Computer-Aided Design of IntegratedCircuits and Systems vol 18 no 12 pp 1702ndash1714 1999
[32] P Pillai and K G Shin ldquoReal-time dynamic voltage scaling forlow-power embedded operating systemsrdquo in Proceedings of the
The Scientific World Journal 13
ACM Symposium on Operating Systems Principles pp 89ndash102October 2001
[33] H Aydin R Melhem D Mosse and P Mejia-Alvarez ldquoPower-aware scheduling for periodic real-time tasksrdquo IEEE Transac-tions on Computers vol 53 no 5 pp 584ndash600 2004
[34] J Zhuo and C Chakrabarti ldquoAn efficient dynamic task schedul-ing algorithm for battery powered DVS systemsrdquo in Proceedingsof the Asian South Pacific Design Automation Conference pp846ndash849 January 2005
[35] R Jejurikar and R Gupta ldquoDynamic slack reclamation withprocrastination scheduling in real-time embedded systemsrdquo inProceedings of the 42nd Design Automation Conference (DACrsquo05) pp 111ndash116 June 2005
[36] J Luo andN K Jha ldquoPower-profile driven variable voltage scal-ing for heterogeneous distributed real-time embedded systemsrdquoin Proceedings of the International Conference on VLSI Designpp 369ndash375 January 2003
[37] M T Schmitz and B M Al-Hashimi ldquoConsidering powervariations of DVS processing elements for energy minimisationin distributed systemsrdquo in Proceedings of the International Sym-posium on System Synthesis (ISSS rsquo01) pp 250ndash255 MontrealCanada October 2001
[38] Y Zhang X Hu and D Z Chen ldquoTask scheduling and voltageselection for energy minimizationrdquo in Proceedings of the 39thAnnual Design Automation Conference (DAC rsquo02) pp 183ndash188June 2002
[39] D Zhu R Melhem and B R Childers ldquoScheduling withdynamic voltagespeed adjustment using slack reclamationin multiprocessor real-time systemsrdquo IEEE Transactions onParallel and Distributed Systems vol 14 no 7 pp 686ndash7002003
[40] D Zhu D Mosse and R Melhem ldquoPower-aware schedulingfor ANDOR graphs in real-time systemsrdquo IEEE Transactionson Parallel and Distributed Systems vol 15 no 9 pp 849ndash8642004
[41] J Kang and S Ranka ldquoDVS based energy minimizationalgorithm for parallel machinesrdquo in Proceedings of the IEEEInternational Parallel and Distributed Processing Symposium(IPDPS rsquo08) pp 1ndash12 Miami Fla USA April 2008
[42] B Rountree D K Lowenthal S Funk V W Freeh B R deSupinski and M Schulz ldquoBounding energy consumption inlarge-scale MPI programsrdquo in Proceedings of the ACMIEEEConference on Supercomputing (SC rsquo07) ACM November 2007
[43] L Wang G von Laszewski J Dayal and F Wang ldquoTowardsenergy aware scheduling for precedence constrained paralleltasks in a cluster with DVFSrdquo in Proceedings of the 10thIEEEACM International Symposium on Cluster Cloud andGrid Computing (CCGrid rsquo10) pp 368ndash377 May 2010
[44] Y C Lee and A Y Zomaya ldquoEnergy conscious schedulingfor distributed computing systems under different operatingconditionsrdquo IEEE Transactions on Parallel and DistributedSystems vol 22 no 8 pp 1374ndash1381 2011
[45] M Mezmaz Y C Lee N Melab E Talbi and A Y ZomayaldquoA bi-objective hybrid genetic algorithm to minimize energyconsumption and makespan for precedence-constrained appli-cations using dynamic voltage scalingrdquo in Proceedings of theIEEE Congress on Evolutionary Computation (CEC rsquo10) pp 1ndash8Barcelona Spain July 2010
[46] M Mezmaz N Melab Y Kessaci et al ldquoA parallel bi-objectivehybrid metaheuristic for energy-aware scheduling for cloudcomputing systemsrdquo Journal of Parallel and Distributed Com-puting vol 71 no 11 pp 1497ndash1508 2011
[47] J Kennedy and R Eberhart ldquoParticle swarm optimizationrdquoin Proceedings of the IEEE International Conference on NeuralNetworks vol 4 pp 1942ndash1948 Perth Australia December1995
[48] Y Zhao M Wilde I Foster et al ldquoGrid middleware servicesfor virtual data discovery composition and integrationrdquo inProceedings of the 2nd Workshop on Middleware for GridComputing (MGC rsquo04) pp 57ndash62 Ontario Canada October2004
[49] A OrsquoBrien S Newhouse and J Darlington ldquoMapping ofscientific workflow within the e-protein project to distributedresourcesrdquo inProceedings of theUK e-Science All HandsMeetingNottingham UK September 2004
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
2 The Scientific World Journal
In order to effectively schedule the tasks and data appli-cations on these cloud environments workflow manage-ment systems require more elaborated scheduling strategiesto meet QoS constraints and the precedence relationshipsbetween workflow tasks The study of workflow schedulingis becoming an important challenge in the area of cloudcomputing
The workflow scheduling in the cloud is a difficultproblem This problem is even more difficult when there areseveral factors to be considered namely (1) the various QoSrequirements of customers like service response time servicecost and so forth (2) the heterogeneity dynamicity andelasticity of cloud services (3) the various ways of combiningthese services to execute workflow tasks (4) the transfer oflarge volumes of data and so forth However the workflowscheduling problem is seen as a combinatorial problemwhere it is impossible to find the globally optimal solutionby using simple algorithms or rules It is well known as anNP-complete problem [8] and depends on the problem size
The workflow scheduling problem has been widely stud-ied in many previous works [9ndash12] Most of these workshave concentrated only on two QoS parameters namelythe deadline and budget In this paper we extend theseworks to handle multiple QoS requirements We address theQoS-based workflow scheduling which aims to minimizethe cost and total time execution of user applications asspecified in the SLA Furthermore the scheduler must alsobe able to schedule workflow tasks so as to maximize theprovider profits by minimizing energy consumption whilepreserving the users QoS preferences We achieve this byusing an iterative method called Multi-objective DiscreteParticle SwarmOptimization (MODPSO) combinedwith theDynamic Voltage and Frequency Scaling (DVFS) techniqueThis last one allows a compromise between system perfor-mance and energy consumption
The proposed approach is assessed by simulation runson a set of synthetic and real-world scientific applicationsSimulation results showed that this new multi-objectivealgorithm significantly improves the performance of relatedapproaches
The remainder of this paper is organized as followsSection 2 reviews several related works Section 3 presentsthe problemmodeling of theQoS basedworkflow schedulingSection 4 describes in detail our scheduling heuristic calledDVFS-MODPSO Section 5 shows an experimental evalu-ation of our heuristic Section 6 concludes the paper anddiscusses some future works
2 Related Work
The workflow scheduling problem in heterogeneous com-puting systems is an NP-hard optimization problem [8]meaning that the amount of computation needed to findoptimum solutions increases exponentially with the problemsize Previous works have proposed many heuristic andmeta-heuristic based approaches [13ndash16] to solve this prob-lem One of the most widely used heuristics for schedulingworkflow application is the Heterogeneous Earliest Finish
Time (HEFT) algorithm developed by Topcuoglu et al [17]HEFT is a static scheduling algorithm that attempts to min-imize execution time (makespan) It preserves the workflowprecedence constraints and produces a good schedule length
Most of these previous works have focused on mini-mizing the workflow execution time without consideringthe usersrsquo budget constraint However with the market-oriented business model in cloud computing environmentswhere users are billed for their consumption of resourcesseveral works that consider usersrsquo budget and deadline havebeen proposed [18ndash21] In [22] a study indicating howto schedule scientific workflow applications with budgetand deadline constraints onto computational grids usinggenetic algorithms is presented Authors in [6] proposedan improved cost-based scheduling algorithm for makingefficient scheduling of tasks to available resources in cloudIn [9] a particle swarm optimization (PSO) based heuristic isused to minimize the execution cost of scheduling workflowapplications to cloud resources
Besides makespan and cost energy consumption isbecoming more and more important in the cloud computingenvironments However cloud providers must adopt mea-sures not only to meet the userrsquo QoS requirements butalso to ensure that their profit margin is not dramaticallyreduced due to high energy consumptions The energyefficiency can conflict with the other QoS requirements(makespan cost) Incorporating the energy consumptioninto the workflow scheduling adds another layer of complex-ity Therefore recent works have concentrated on developingenergy-aware scheduling algorithms They have examinedvarious techniques such as dynamic power managementDynamic Voltage and Frequency Scaling (DVFS) or resourcehibernation [23ndash26] Authors in [27] presented an onlinedynamic power management strategy with many power-saving states They proposed a min-min based energy-awarescheduling algorithm to minimize energy consumption inheterogeneous computing systems In [26] a dynamic slackallocation technique which tries to use idle (slack) timeslots of processors to lower supply voltage (frequencyspeed)is presented These slack time slots occur due to earliercompletion andor dependencies of tasks Several DVS-basedapproaches for slack allocation have been proposed for bothindependent [28ndash35] and precedence-constrained [36ndash42]tasks In [43] an energy-aware scheduling algorithm anddetailed discussion of slack time computation are presentedThis scheduling algorithm reduces voltages during the com-munication phases between parallel tasks on homogeneousprocessors In [44] an Energy Conscious Scheduling heuris-tic (ECS) is proposed The heuristic is devised with relativesuperiority as a novel objective function which takes intoaccount energy andmakespan ECS is used to improve the bi-objective genetic algorithm proposed in [45] This latter hasbeen extended in [46] to a parallel model of their approach
All of these presented works have focused on optimizingeither a single or two objectives but none of them consider therelationships between several objectives namely the relation-ship between energy makespan and cost They do not takeinto account how each one of these criteria can affect othersTo deal with these misses we propose a Multi-Objective
The Scientific World Journal 3
Discrete Particle Swarm Optimization algorithm combinedwith DVFS technique (DVFS-MODPSO) to optimize allthree objectives at the same time Our new approach providesa set of solutions named Pareto solutions (ie non-dominatedsolutions) enabling the user to select the desired tradeoff
To the best of our knowledge none of the previ-ous scheduling approaches deal with the three-dimensionalmakespancostenergy optimization when tackling the prob-lem of scheduling workflow applications on heterogeneouscomputing environments such as the cloud computing oneswhich constitute our key novelties
3 Problem Modeling
In this section we describe our system model in a formalway Our ultimate goal is to distribute workflow tasks amongcloud services so as to optimize both the usersrsquo QoS criteriaand cloud providersrsquo profits by saving energy consumption oftheir servicesTherefore we first present the cloud computingmodel Then we describe our workflow model and the QoSparameters we deal We conclude this section by describingthe scheduling model formalized as a multi-objective opti-mization problem we solve
31 Cloud Computing Model The cloud computing systemused in this work is a set of resources offered by a cloudprovider to run client applications Our cloud model isinspired by the model described in [45] We assume that thecloud is hosted in data centers composed of heterogeneousmachines These data centers deliver a variety of serviceshosted on thousands of IT servers which are made availableas subscription-based services in a pay-as-you-go model Inour model the cloud computing system consists of a set of119875 = 119901
1 1199012 119901
119898 heterogeneous processors which are
fully interconnected The processors have varied processingcapability delivered at different processing prices (see ec ofTable 1) Each processor 119901
119895isin 119875 is DVFS-enabled that is it
can operate with different VSLs (Voltage Scaling Level iedifferent clock frequencies) For each processor 119901
119895isin 119875
a set 119881119895of V VSLs is randomly and uniformly distributed
among three different sets of VSLs (Table 1) We consider thatprocessors consume energy during periods of inactivity thatis when a processor is idling it is assumed that the lowestvoltage is supplied [44] Because clock frequency transitionoverheads usually take a negligible amount of time theseoverheads are not considered in this paper and the inclusionof such an overheadwill have no bearing on the overall modelof the proposed study
Additionally each processor 119901119895isin 119875 has a set of links
119871119901119895= 119897119895 1199011 119897119895 1199012 119897119895 119901119896 1 le 119896 le 119898 where 119897
119895119894isin 119877+ is
the available bandwidthmdashmeasured in Mega bits per second(Mbps)mdashin the link between processors 119901
119895and 119901
119894 with 119897
119895119895=
1 We assume that a message can be transmitted from oneprocessor to another while a task is executed on the recipientprocessor Finally communication between tasks executed onthe same processor is neglected Table 1 shows DVFS levelsRelative speeds (RSpeed) and execution costs (ec) for threeprocessor classes (TURIONMT-34 OMAP PENTIUMM)
32 Workflow Application Model We model a cloud work-flow application as a Directed Acyclic Graph (DAG) denotedas 119866(119881 119864) The set of nodes 119881 = 119879
1 119879
119899 represents
the tasks in the workflow application the set of arcs denotesprecedence constraints and the controldata dependenciesbetween tasks An arc is in the form of 119889
119894119895= (119879119894 119879119895) isin 119864
where119879119894is called the parent task of119879
119895119879119895is the child task of119879
119894
119889119894119895is the data produced by119879
119894and consumed by119879
119895We assume
that a child task cannot be executed until all of its parent taskshave been completed In a given task graph a task with noparent is referred as an entry task and one without any childis called an exit task Since our algorithm involves only oneentry and one exit tasks we add two dummy tasks 119879entry and119879exit which have zero execution time to the beginning andthe end of the workflow respectively These dummy tasks areconnected with zero-weight arcs to the actual entry and exittasks respectively
We assume that each task 119879119894isin 119881 has an associated
basic execution time which is an independent value for eachmachine We denote 119908
119894119895 the basic computation time of a
task 119879119894on a compute resource 119901
119895at maximum speed and
voltage (ie it corresponds to Level 1 in Table 1) The averageexecution time of the task 119879
119894is defined as
119908119894=
119898
sum119895=1
119908119894119895
119898 (1)
Real computation time119908119903119894119895119896
of the task 119879119894on machine 119901
119895
using relative execution speed 119904119896119895is defined as
119908119903119894119895119896=119908119894119895
119904119896119895
(2)
We also assume that every edge (119879119894 119879119895) isin 119864 is associated
with value tr119894119895 representing the time needed to transfer data
from 119879119894to 119879119895 The transfer time can be calculated according
to the bandwidth 119897119909119910
between the resources executing thesetasks (119901
119909and 119901
119910resp) as follows
tr119894119895=119889119894119895
119897119909119910
(3)
However a communication time is only required whentwo tasks are assigned to different processors That is thecommunication time when tasks are assigned to the sameprocessor can be neglected that is 0
In general the execution costs (ec) and transmission costs(trc) are inversely proportional to the execution times andtransmission times respectively
We define pred (119879119894) as the set of all predecessors of 119879
119894and
succ (119879119894) as the set of all successors of 119879
119894 An ancestor of node
119879119894is any node 119879
119896that is contained in pred (119879
119894) or any node
119879119895that is also an ancestor of any node 119879
119896contained in pred
(119879119896)The Earliest Start Time and the Earliest Finish Time of a
task 119879119894on a processor 119901
119895are represented as EST (119879
119894 119901119895) and
EFT (119879119894 119901119895) respectively EST (119879
119894) and EFT (119879
119894) represent the
earliest start and finish times on any processor respectively
4 The Scientific World Journal
Table 1 Voltage Scaling Levels and relative speeds of processors
VSLP1 TURIONMT-34 P2 OMAP P3 PENTIUMM
Volt Freq R Speed Volt Freq R Speed Volt Freq R Speed(V) (Ghz) () (V) (Ghz) () (V) (Ghz) ()
1 12 18 100 135 06 100 148 14 1002 115 16 8888 127 055 9166 144 12 96763 11 14 7777 12 05 8333 131 1 88144 105 12 6666 1 025 4166 118 08 79515 1 1 55555 09 013 2083 096 06 64426 09 08 4444Cost ec = 114 ec = 057 ec = 076
Table 2 Task execution times and priorities
Task (119879119894) 119901
11199012
1199013
120596119894
Priority (Pr(119879119894))
1 14 16 9 13 1080002 13 19 18 1667 770003 11 13 19 1433 800004 13 8 17 1267 800005 12 13 10 1167 690006 13 16 9 1267 633337 07 15 11 11 426678 5 11 14 10 356679 18 12 20 1667 4433310 21 7 16 1467 14667
119875119886V(119879119894 119901119895) is defined as the earliest time when processor 119901
119895
will be available to begin executing task 119879119894 Hence
119864119878119879 (119879119894 119901119895) =
119900 if 119879119894= 119879entry
max 119875119886V(119879119894 119901119895) 120572
(4)
where 120572 = max119879119895 isin pred (119879119894 )(119864119865119879(119879119895 119901119896) + tr
119895119894
119864119865119879 (119879119894 119901119895) = 119908
119894119895+ 119864119878119879 (119879
119894 119901119895) (5)
Note that the Actual Start Time and Actual Finish Timeof a task 119879
119894on a processor 119901
119895 denoted as AST (119879
119894 119901119895) and
AFT (119879119894 119901119895) can be different from its earliest start EST (119879
119894 119901119895)
and finish EFT (119879119894 119901119895) times if the actual finish time of
another task 119879119896scheduled on the same processor is later than
its EST (119879119896 119901119895) [44]
Figure 1 depicts a workflow application with 10 tasks andthe Table 2 provides its details (given in [17]) The valuespresented in the last column of the table represent the priorityof the tasks The priority of task 119879
119894represented by Pr(119879
119894) is
computed recursively by traversing the DAG upward startingfrom the exit task 119879exit as follows (6)
Pr (119879119894) =
119908119879exit if 119879
119894= 119879exit
119908119894+ 120573 otherwise
(6)
where120573 = max119879119895 isin succ (119879119894)119879119903119894119895 + Pr(119879
119895)
1
2 5 6
7 8
10
3 4
9
18 12 9 1114
19 16 2723 23 1513
1311 17
Figure 1 An example of workflow (given in [17]) with the tasknumbers 119879
119894inside nodes and values of 119889
119894119895function next to the
corresponding edges
33 QoS Parameter Models
331 Energy Model Among the main system-level energy-saving techniques Dynamic Voltage Scaling (DVS) operateson a simple principle decreases the supply voltage (and so theclock frequency) to the CPU so as to consume less power
In this work we use a model of energy derived from thepower consumption model in digital complementary metal-oxide semiconductor (CMOS) logic circuits [44] Under thedynamic power model the processor power is dominated bythe dynamic power which is given by
119875dynamic = 119860119862efV2119891 (7)
where119860 is the number of switches per clock cycle119862ef denotesthe effective charged capacitance V is the supply voltage and119891 denotes the operational frequency Equation (7) shows thatthe supply voltage is the dominant factor hence its reductionwould be most influential to lower power consumption
The energy consumption of the execution of a workflowapplication used in this paper is defined as
119864119888=
119899
sum119894=1
119860119862efV2
119894119891119894119908119903lowast
119894
=
119899
sum119894=1
120574V2119894119891119894119908119903lowast
119894
(8)
The Scientific World Journal 5
where 120574 = 119860119862ef is assumed constant for a given machineV119894 119891119894are the voltage supply and frequency of the processor
on which task 119879119894is executed respectively and 119908119903lowast
119894is the real
completion time of task 119879119894on the scheduled processor In the
idle time the processor turns into sleep mode and thus thevoltage supply and relative frequency are at the lowest levelSo the energy consumption during idle periods of processorsis defined as
119864119894=
119898
sum119895=1
sumidle119895119896isin IDLE119895
120574V2min 119895 119891min 119895 119871119895119896 (9)
where IDLE119895
is the set of idling slots on machine119901119895 Vmin 119895(119891min 119895) is the lowest supply voltage (frequency)
on 119901119895 and 119871
119895119896is the amount of idling time for idle
119895119896
Then the total energy 119864total utilized by the cloud system forcompletion of the workflow application can be defined asfollows
119864total = 119864119888+119864119894 (10)
332 Time Model The completion time 119879total is themakespan from the user submitting a workflow untilreceiving the results It is defined as the actual finish timeof the exit task It is calculated in equation 11 and consistsof the workflow execution time and network transmissiontime The execution time depends on both the workloadand system performance The network transmission timedepends on both the network latency and the input data size
119879total = maxAFT (119879exit) (11)
333 Cost Model As result of the marketization characteris-tic of current services most cloud providers have set a pricefor their services They have fixed the price for transferringbasic data unit (eg per MB) between two services and theprice for processing basic time units (eg per hour)The cost119862total of running a cloud workflow is defined in Formula (12)It consists of processing cost 119862ex and data transfers cost 119862tr
119862total = 119862ex + 119862tr (12)
The processing cost for a given task119879119894depends on the real
completion time of 119879119894on the scheduled processor (119908119903lowast
119894) and
the hourly price of this processor (ec119894) Thus 119862ex is given by
119862ex =119899
sum119894=1
119908119903lowast
119894lowast ec119894 (13)
The data transfer cost (119862tr) is described as follows
119862tr =119899
sum119894=1
119899
sum119895=1
119889119894119895lowast trc119894119895 (14)
where 119889119894119895characterizes the output file size from task119879
119894to task
119879119895 and tr 119888
119894119895represents the cost of communication from the
processor where 119879119894is mapped to another processor where 119879
119895
is mappedThe cost of communication is added to the overallcost only when two tasks have data dependency betweenthem (ie 119889
119894119895gt 0) For two or more tasks running on the
same processor the transfer cost is neglected
334 Scheduling Model Given (1) A cloud provider thatoffers a set 119875 of 119898 heterogeneous processors and (2) auser workflow application composed of a set 119879 of 119899 tasksthat have to be executed on these processors The workflowscheduling problem is to construct a mapping 119872 of tasksto processors (without violating precedence constraints) thatminimizes the following conflicting objectives makespancost and energy consumption as low as possible Thereforethe workflow scheduling problem can be formulated as amathematical optimization problem
MakespanMinimize 119879total (119872)
CostMinimize 119862total (119872)
EnergyMinimize119864total (119872)
(15)
4 Workflow Scheduling Based on DiscreteParticle Swarm Optimization
This section starts with a brief overview on multi-objectivecombinatorial optimization and Particle swarm optimizationalgorithm Afterwards our new Multi-Objective DiscreteSwarmOptimization combined with DVFS technique will bepresented
41 Multi-Objective Optimization A Multi-objective Opti-mization Problem (MOP) with 119898 decision variables and 119899objectives can be formally defined as
Min (119910 = 119891 (119909) = [1198911(119909) 119891
119899(119909)]) (16)
where 119909 = (1199091 119909
119898) isin 119883 is an 119898-dimensional decision
vector 119883 is the search space 119910 = (1199101 119910
119899) isin 119884 is the
objective vector and Y the objective-spaceIn general MOP there is no single optimal solution
with regards to all objectives This is also the case forthe multi-objective optimization problem addressed in thispaper As given in (15) there are three conflicting objectivesminimizing execution time minimizing execution cost andminimizing energy consumption In such problems thedesired solution is considered to be the set of potentialsolutions which are all optimal in some objectives This set isknown as the Pareto optimal setWe provide some definitionsof the Pareto concepts used in MOP as follows (withoutloss of generality we suppose that the objectives are to beminimized)
(i) Pareto dominance For two decision vectors 1199091 and 1199092dominance (denoted by ≺) is defined as follows
1199091≺ 1199092lArrrArr forall119894 119891
119894(1199091) le 119891119894(1199092) and exist119895119891
119895(1199091) lt 119891119895(1199092)
(17)
The decision vector 1199091is said to dominate 1199092 if andonly if 1199091 is as good as 1199092 considering all objectivesand 1199091 is strictly better than 1199092 in at least oneobjective
6 The Scientific World Journal
(ii) Pareto optimally A decision vector 1199091 is said to bePareto optimal if and only if
∄1199092isin 119883 119909
2≺ 1199091 (18)
(iii) Pareto optimal set The Pareto optimal set 119875119878is the set
of all Pareto optimal decision vectors
119875119878= 1199091isin 119883 | ∄119909
2isin 119883 119909
2≺ 1199091 (19)
(iv) Pareto optimal frontThePareto optimal front119875119865is the
image of the Pareto optimal set in the objective space
119875119865= 119891 (119909) = (119891
1(119909) 119891
119899(119909)) | 119909 isin 119875
119904 (20)
1199091 is said to be non-dominated regarding a given set
if 1199091 is not dominated by any decision vectors in theset
Thepareto optimal decision vector cannot be enhanced inany objective without causing degradation in at least anotherobjective A decision vector is said to be Pareto optimal whenit is not dominated in the whole search space
42 Particle Swarm Optimisation
421 The Standard PSO PSO is a population-based stochas-tic optimization technique developed by Kennedy and Eber-hart in 1995 [47] It is inspired by the social behavior of insectcolonies bird flocks fish schools and other animal societies Itis also related to evolutionary computation PSOhas attractedsignificant attention from many researchers due to both itssimplicity of use and optimization via social behavior In factPSO has good performance requires low computational costIt is effective and easy to implement as it uses numericalencoding
A particle in PSO is analogous to a fish or bird movingin the 119863-dimensional search space All particles have fitnessvalues indicating their performances which are problemspecific and velocities which direct the flight of particlesEach particle position at any given time is influenced by bothits best position called 119901119861119890119904119905 and the position of the bestparticle in a problem space referred to as 119892119861119890119904119905 Thereforeparticles tend to fly towards a better search area duringthe search process A particle status on the search spaceis characterized by two elements namely its velocity andposition which are updated in every generation as follows
119881119896+1
119894= 120596119881
119896
119894+ 11988811199031(119901119861119890119904119905
119894minus 119883119896
119894)
+ 11988821199032(119892119861119890119904119905 minus 119883
119896
119894)
119883119896+1
119894= 119883119896
119894+ 119881119896+1
119894
(21)
where 119881119896+1119894
is the velocity of particle 119894 at iteration 119896 + 1119881119896119894is the velocity of particle 119894 at iteration 119896 120596 is the inertia
weight 1198881and 1198882are the acceleration coefficients (cognitive
and social coefficients) 1199031and 1199032are the random numbers
between 0 and 1119883119896119894is the current position of particle 119894 at the
119896th iteration 119901119861119890119904119905119894is the best previous position of the 119894th
particle 119892119861119890119904119905 is the position of best particle in the swarmand119883119896+1
119894is the position of 119894th particle at 119896 + 1 iteration
The procedure for standard PSO is as follows
(1) Initialize a population of particles with random posi-tions and velocities in the search space
(2) Evaluate the objective values of all particles set 119901119861119890119904119905of each particle equal to its current position and set119892119861119890119904119905 equal to the position of the best initial particle
(3) Update the velocity and the position of each particleaccording to (21)
(4) Map the position of each particle in the solution spaceand evaluate its fitness value according to the desiredoptimization fitness function
(5) For each particle compare its current objective valuewith its 119901119861119890119904119905 value If the current value is better thenupdate 119901Best with the current position and objectivevalue
(6) Determine the best particle of the current whole pop-ulation with the best objective value If the objectivevalue is better than that of 119892119861119890119904119905 then update 119892Bestwith the current best particle
(7) If the stopping criterion ismet then output 119892119861119890119904119905 andits objective value otherwise go to Step (2)
The original design of PSO is appropriate for findingsolutions to continuous optimization problems However asthe workflow scheduling discussed in this paper is both adiscrete andmulti objective problem in nature we propose aneffective approach to address this problem by using a discreteversion of the Multi-Objective PSO (MODPSO) combinedwith DVFS techniqueThe key issues of our approach consistof (1) defining the position and velocity of the particlesbased on the features of discrete variables of the workflowscheduling (2) solving the multi-objective aspect of theproblem by modifying PSO so as to generate a set of non-dominated solutions satisfying the different objectives underconsideration instead of one solution
422 Handling Workflow Scheduling Using DVFS-MODPSOIn general a workflow scheduling can be defined by a setof triplets 119872 = [⟨119879
119894 119875119895 119871119896⟩](119894 isin [1 119899] 119895 isin [1119898] 119896 isin
[1 119871(119895)]) 119899 is the number of workflow tasks to be scheduled119898 is the number of processors available in the cloud environ-ment and 119871(119895) is the number of operating points (VSLs) of the119895th processor
For the sake of clarity the variables and rules of DVFS-MODPSO for solving workflow scheduling can be depictedas follows
The position of a particle represents a feasible solutionto the scheduling problem It consists of a set of⟨task(119879
119894) service (119875
119895) VSl (119871
119896)⟩ triplets Each triplet
means that a task (119879119894) is assigned to a processor (119875
119895)
with a voltage scaling level (119871119896) It also indicates
The Scientific World Journal 7
that the position satisfies the precedence constraintbetween tasks
The process of generating a newposition for a selectedparticle in the swarm is depicted in the followingformulas
119881119896+1
119894= 119881119896
119894oplus ((119877
1otimes (119901119861119890119904119905
119894⊖ 119883119896
119894))
oplus (1198772otimes (119892119861119890119904119905 ⊖ 119883
119896
119894)))
(22)
119883119896+1
119894= 119883119896
119894oplus 119881119896+1
119894 (23)
The operator definitions are as follows
(i) The substract operator (⊖) the difference betweentwo particle positions designated as 1199091 and 1199092 isdefined as a set of triplets in which each triplet⟨task serviceVSL⟩ shows whether the contents ofthe corresponding elements in 1199091 are different fromthose of 1199092 or not If so that triplet gets its values(service and VSL) from the position that has thelowest value of theVSL For those triplets that have thesame content in 1199091 and 1199092 their corresponding VSLsare decreased (Note that the scaling of VSL makesfluctuations on the energy makespan and cost)
(ii) The multiply operator (otimes) the multiplication betweennumber and velocity is defined as a set of tripletswhere a threshold 120572 isin [0 1] is defined arandom number 119903 is generated for each triplet⟨task serviceVSL⟩ compare 119903 and 120572 when 119903 ge 120572decrease the triplet VSL otherwise increase it Thisoperator adds the exploration ability to the algorithm
(iii) The add operator (oplus) the addition of two positions isdefined as the reservation of the dominated one
The Pseudocode 1 outlines the general steps of the DVFS-MODPSO algorithm
The algorithm begins by initializing the positions andvelocities of particles To obtain the position of a particle theVSL (voltage and frequency) of each resource is randomlyinitialized firstly then the HEFT algorithm is applied togenerate a feasible and efficient solution minimizing themakespan The process is repeated several times to initializethe positions of all particles of the swarm Initially thevelocity119881 and the best position for each particle 119901119861119890119904119905 areattributed as the particle itself The algorithm maintainsan external archive EA to store non-dominated particlesfound after the evaluation process (based on the paretodominance using the objectives mentioned in 33) After allthese initializations in the main loop of the algorithm thenew velocity and position of each particle are calculatedrespectively after selecting the best overall position in theexternal archive and eventually perform a mutation then theparticle is evaluated and its corresponding 119901119861119890119904119905 is updatedThe external archive is updated after every iteration Oncethe termination condition is reached the external archivecontaining the Pareto front is returned as the result
5 Experimental Evaluations
In this section we describe the overall setup of our experi-mentation effort and the results we have obtained from it tovalidate the new proposed approach
In our experiments we simulate a synthetic workflowapplication described in [17] (see Figure 1) and two commonworkflow structures parallel and hybrid structuresWe chosetwo well known real world applications namely a neuro-science workflow [48] for our parallel application and aprotein annotation workflow [49] developed by London e-Science Centre for our hybrid workflow application Figure 2shows their simplified representationsThe number indicatedin the parentheses next to each node represents the length ofthe task (the number of instructions to execute) in Millionsof Instructions (MI) The input and output files of each taskrange from 10MB to 1024MB
We have performed experiments on 3 4 5 6 and 12resources whose characteristics are shown in Table 1 Wechoose the pricing model associated with Amazon EC2(httpawsamazoncomec2) for the processing costs of thedifferent classes of resources and the pricing model given byAmazon CloudFront (httpawsamazoncomcloudfront)for the costs of transferring data unit between resources
As for DVFS-MODPSO the parameter settings areSwarm size 119878119873119906119898 = 50 particles and the maximum numberof iterations 119866 = 100
We evaluated the performance of our proposed DVFS-MODPSO on all the workflow instances described previouslyDue to the lack of works considering both the heteroge-neous configuration of our cloud and the three QoS metrics(makespan cost energy) at the same time we comparedour results with those of the HEFT heuristic we haveimplemented HEFT is one of themost widely used heuristicsfor DAGs in distributed heterogeneous computing systemswhich optimize the makespan
Figures 3 4 and 5 show sample Pareto fronts obtainedwith the DVFS-MODPSO and the solution computed byHEFT for the synthetic workflow the hybrid workflow andthe parallel workflow respectively
As can be seen from these figures unlike the HEFTalgorithm which gives one solution our proposed approachprovides a set of non dominated solutions These resultsillustrate the basic multi-objective definitions provided inSection 41
Now for comparing these two approaches and in orderto analyze the effectiveness of our proposition in terms ofthe values of makespan cost and energy consumption wehave been inspired by the methodology described in [43]where we compare the solution provided by HEFT to onlyone solution of the Pareto front set computed by our proposedmulti-objective DVFS-MODPSO
The evaluation process follows the next steps For eachworkflow instance we perform a first resolution usingHEFTin order to get one solution 119878 After a second resolution iscomputed using our proposed approach to obtain a set EA ofPareto solutions Next we select one solution 1198781015840 from the setEA of Pareto solutions This solution is the closest one to 119878 inthe sense of Euclidean distance Finally a comparison is donebetween 119878 and 1198781015840
8 The Scientific World Journal
Swarm initialization with HEFT(1) For i = 1 to SNum (SNum is the size of particle swarm)
(a) For j = 1 to m (m the number of processors)(i) Randomly initialize VSL(j) (randomly choose the voltagefrequency of processor from the set of its operating points)
(b) Initialize 119878[119894] with HEFT heuristic (S is the swarm of particles)(c) Initialize the velocity V of each particle
119881 [119894] = 0 (119878 [119894])
(d) Initialize the Personal Best (pBest) of each particle119901119861119890119904119905[119894] = 119878[119894]
(e) Evaluate objectives of each particleEvaluate S[i]
(f) Initialize the Global Best particle (gBest) with the best one among the SNum particlesg119861119890119904119905 = Best particle found in 119878
(2) End For(3) Add the nondominated solutions found in S into EA (EA is the External Archive storing the pareto front)(4) Initialize the iteration number 119905 = 0(5) Repeat until 119905 gt 119866 (119866 is the maximum number of iterations)
(a) For 119894 = 1 to SNum (swarm size)(i) Randomly select the global best particle for S from the External Archive EA and store its position in gBest(ii) Calculate the new velocity 119881[119894] according to (22)(iii) Compute the new position of 119878[119894] according to (23)(iv) If (119905 lt 119866 lowast PMUT ) then (PMUT is the probability of mutation)
Perform mutation on 119878[119894](v) Evaluate S[i]
(b) End For(c) Update the personal best solution of each particle 119878[119894]
if 119878 [119894] ≼ 119901119861119890119904119905119904 [119894] or (119878 [119894] sim 119901119861119890119904119905119904 [119894])Then 119901119861119890119904119905119904 [119894] = 119878 [119894]
(d) Update the External Archive EA(i) Add all new non-dominated solution in S into EAif 119878 [119894] 997410 119909 forall119909 isin 119864119860Then 119864119860 = 119864119860 cup 119878 [119894](ii) Remove all particles dominated by 119878[119894] in EA
119864119860 = 119864119860 minus 119910 isin 119864119860 | 119910 lt 119878[119894]
(iii) If the archive is full then randomly select the article to be replaced from EA(6) Return the pareto front (the set of non-dominated solutions from S and EA)
Pseudocode 1 DVFS-MODPSO based workflow scheduling
The different results are displayed on Figures 6 7 8 and 9and on Tables 3 and 4 They show the improvement achievedby our proposed approach according to the structure of theworkflow applications and the number of processors
Table 3 and Figure 6 show that our proposed approachimproves on average the result obtained by HEFT for thesynthetic workflow instance The makespan is reduced by005 the cost is reduced by 993 and the energy con-sumption is reduced by 2640 Figure 6 shows detailedimprovements of the proposed approach according to thenumber of processors
Table 4 and Figure 7 illustrate the gain obtained by ourapproach according to the kind of real world applicationsTable 4 also shows the detailed improvements when scalingthe number of processors As can be seen the QoS metricsare improved over HEFT in average by 095 for themakespan 108 for cost and 812 for the energy con-sumption when using the hybrid workflow application andaverage improvements of 295 formakespan 2215 for costand 209 for energy consumption when using the parallelworkflow
These evaluations confirm the results obtained from thesynthetic workflow This means that by using our DVFS-MODPSO we are able to improve not only the cost and energybut also themakespan onwhichHEFT algorithm is supposedto provide good results These results remain true for all kindof workflow applications
In our experiments we limited the number of resourcesto 12 as we found that for these kinds of workflowapplications some resources are not used at all Figure 8shows the resource loads when scheduling the syntheticworkflow on 12 resources Furthermore even though thenumber of resources increased the total cost and total energyconsumption could not always decrease as illustrated inFigure 9 From this figure we can see that the QoS metrics(1) firstly decreases as the number of resources increasesuntil achieving the number 6 This can be explained by thefact that when increasing the number of resources thereare fewer tasks executed in a resource therefore tasks canextend their execution times and the resource have more ofchances to scale down their voltages and frequencies whichcan be very effective in reducing total energy consumption
The Scientific World Journal 9
1 2 3 4
5
6 7
8 9 10
11
12
14
13
15
(300000) (600000) (600000) (900000)
(300000)
(150000) (150000)
(300000) (300000) (300000)
(600000)
(300000)
(150000)
(300000)
(600000)
SignalP COILS2 SEG PROSITE
TMHMM
ProsperoHMMer
PSI-BLASTBLAST
IMPALA
PSI-PRED
3D-PSSMSummary
Genomesummary
SCOP
(a)
1 3 5 7
9
11 1210
(400000) (400000) (400000) (400000)
(300000)
(500000) (500000) (500000)
Softmea
2 4 6 8(600000) (600000) (600000) (600000)
Reslice
14 1513
(600000) (600000) (600000)
Convert
Reslice Reslice Reslice
Slicer Slicer
Convert Convert
Slicer
Align wapAlign wapAlign wapAlign wap
(b)
Figure 2 Parallel and Hybrid workflows
10 The Scientific World Journal
75 80 85 90 95 100 8090
100110
120130
140
2060
100140180220
Ener
gy
DVFS-MODPSOHEFT
MakespanCost
Figure 3 Obtained non-dominated solutions for the syntheticworkflow
Ener
gy
DVFS-MODPSOHEFT
MakespanCost
200 400 600 800 100012001400160018002000 400800
12001600
2000
400200
600800
100012001400160018002000
Figure 4 Obtained non-dominated solutions for the hybrid work-flow
(2) After achieving 6 resources the QoS metrics begin torise the reason for this is that time executions are dominatedby interprocessors communications hence decreasing theopportunities for scaling down voltages and frequencies ofresources Consequently the threshold numbers of resourcesthat minimized the QoS metrics could be obtained
6 Conclusion
In this paper we propose a new algorithm called DVFS-Multi-Objective Discrete Particle Swarm Optimization(DVFS-MODPSO) for workflow scheduling in distributedenvironments such as cloud computing infrastructures DV-FS-MODPSO simultaneously optimizes several conflictingobjectives namely the makespan cost and energy in adiscrete space It produces a set of non-dominated solutionsthus providing more flexibility for users to assess theirpreferences and select a schedule that meets their QoS
Ener
gy
DVFS-MODPSOHEFT
Makespan Cost
400200600800
10001200140016001800
240020002200
450 500 550 600 650 700 7501200
14001600
18002000
Figure 5 Obtained non-dominated solutions for the parallel work-flow
0
10
20
30
40
50
3 4 5 6 12
MakespanCostEnergy
Figure 6 Gain over HEFT () according to the number ofprocessors
requirements Our approach exploits the heterogeneityand the marketization of cloud resources in order to findsolutions that optimize makespan and cost Furthermore ituses the Dynamic Voltage and Frequency Scaling (DVFS)technique to reduce energy consumption
We have evaluated our algorithm by simulating it withboth synthetic and real world scientific workflow applicationshaving different structures The results show that the pro-posedDVFS-MODPSO is able to produce a set of good Paretooptimal solutions The results also show that our approachprovides significant improvement not only in terms of the
The Scientific World Journal 11
0
5
10
15
20
25
HybridParallel
Makespan Cost Energy
Figure 7 Improvement according to real world applications
R17
R20
R37
R446
R50
R60
R733
R80
R90
R107
R110
R120
Figure 8 Resource loads for synthetic workflow
Table 3 Improvement for synthetic workflow
Num of proc Gain over Heft ()Makespan Cost Energy
3 0 2126 40084 022 464 15355 0 756 17266 003 652 267512 0 969 3255Average 005 993 2640
cost and the energy consumption but also in term of themakespan for which HEFT algorithm is supposed to givebetter results
Multi-objective optimization in cloud workflow schedul-ing is not a mature domain Most of the existing worksattempt to minimize either the makespan or cost Howeverwe plan to consider other objectives such as reliabilitysecurity in addition to the energy consumption We also
0
50
100
150
200
250
3 4 5 6 12
MakespanCostEnergy
Figure 9 QoS metrics evolutions according to resource numbersfor the synthetic workflow
Table 4 Improvement according to real world applications
Appli Num of proc Gain over Heft ()Makespan Cost Energy
Hybrid
3 405 1155 04024 028 0566 05075 000 1347 11246 000 0117 126512 043 2214 0761
Average 095 1080 0812
Parallel
3 235 5640 25524 1147 2094 32715 000 2142 11666 093 1170 315412 000 031 306
Average 295 2215 209
intend to apply our algorithm in a real world cloud orintegrate it in existing cloud toolkits such as Cloudbus forcomparing with other algorithms
References
[1] D Thain T Tannenbaum and M Livny Condor and the GridGrid Computing Making the Global Infrastructure a RealityJohn Wiley amp Sons Hoboken NJ USA 2003
[2] R Buyya and S Venugopal ldquoThe gridbus toolkit for serviceoriented grid and utility computing an overview and statusreportrdquo in Proceedings of the 1st IEEE International Workshopon Grid Economics and Business Models (GECON rsquo04) pp 19ndash36 IEEE CS Seoul Republic of Korea April 2004
[3] S McGough L Young A Afzal S Newhouse and J Darling-ton ldquoWorkflow enactment in ICENIrdquo in Proceedings of the UK
12 The Scientific World Journal
e-Science All Hands Meeting pp 894ndash900 IOP NottinghamUK September 2004
[4] E Deelman J Blythe Y Gil et al ldquoPegasus mapping scientificworkflows onto the gridrdquo in Grid Computing 2nd EuropeanAcrossGrids Conference AxGrids 2004 Nicosia Cyprus January28ndash30 2004 vol 3165 of Lecture Notes in Computer Science pp11ndash20 Springer New York NY USA 2004
[5] R Buyya S Pandey and C Vecchiola ldquoCloudbus toolkitfor market-oriented cloud computingrdquo in Cloud ComputingProceedings of the 1st International Conference CloudCom 2009Beijing China December 1ndash4 2009 vol 5931 of Lecture Notesin Computer Science pp 24ndash44 Springer New York NY USA2009
[6] Y Yang K Liu J Chen X Liu D Yuan and H Jin ldquoAnalgorithm in SwinDeW-C for scheduling transaction-intensivecost-constrained cloud workflowsrdquo in Proceedings of the 4thIEEE International Conference on eScience (eScience rsquo08) pp374ndash375 Indianapolis Ind USA December 2008
[7] L Ramakrishnan C Koelbel Y Kee et al ldquoVGrADS enablinge-Scienceworkflows on grids and cloudswith fault tolerancerdquo inProceedings of the Conference on High Performance ComputingNetworking Storage and Analysis (SC rsquo09) November 2009
[8] M L Pinedo Scheduling Theory Algorithms and SystemsSpringer Berlin 2008
[9] S Pandey L Wu S M Guru and R Buyya ldquoA particle swarmoptimization-based heuristic for scheduling workflow applica-tions in cloud computing environmentsrdquo in Proceedings of the24th IEEE International Conference on Advanced InformationNetworking and Applications (AINA rsquo10) pp 400ndash407 PerthAustralia April 2010
[10] S Abrishami and M Naghibzadeh ldquoDeadline-constrainedworkflow scheduling in software as a service cloudrdquo ScientiaIranica vol 19 no 3 pp 680ndash689 2012
[11] S Selvarani and G S Sadhasivam ldquoImproved cost-based algo-rithm for task scheduling in cloud computingrdquo in Proceedings ofthe IEEE International Conference on Computational Intelligenceand Computing Research (ICCIC rsquo10) pp 1ndash5 CoimbatoreIndia December 2010
[12] A Verma and S Kaushal ldquoDeadline and budget distributionbased cost-time optimization workflow scheduling algorithmfor cloudrdquo inProceedings of the IJCAon International Conferenceon Recent Advances and Future Trends in Information Technol-ogy (iRAFIT rsquo12) iRAFIT(7) pp 1ndash4 April 2012
[13] R C Correa A Ferreira and P Rebreyend ldquoIntegrating listheuristics into genetic algorithms for multiprocessor schedul-ingrdquo in Proceedings of the 8th IEEE Symposium on Parallel andDistributed Processing (SPDP rsquo96) pp 462ndash469 IEEEComputerSociety Washington DC USA October 1996
[14] E S H Hou N Ansari and H Ren ldquoGenetic algorithm formultiprocessor schedulingrdquo IEEE Transactions on Parallel andDistributed Systems vol 5 no 2 pp 113ndash120 1994
[15] M Grajcar ldquoGenetic list scheduling algorithm for schedulingand allocation on a loosely coupled heterogeneous multi-processor systemrdquo in Proceedings of the 36th Annual DesignAutomation Conference (DAC rsquo99) pp 280ndash285 ACM PressNew York NY USA June 1999
[16] R Sakellariou and H Zhao ldquoA hybrid heuristic for DAGscheduling on heterogeneous systemsrdquo in Proceedings of the13th Heterogeneous Computing Workshop pp 111ndash123 IEEEComputer Society Washington DC USA 2004
[17] H Topcuoglu S Hariri and M Y Wu ldquoPerformance-effectiveand low-complexity task scheduling for heterogeneous comput-ingrdquo IEEE Transactions on Parallel and Distributed Systems vol13 no 3 pp 260ndash274 2002
[18] J Yu and R Buyya ldquoWorkflow scheduling algorithms for gridcomputingrdquo in Metaheuristics for Scheduling in DistributedComputing Environments F Xhafa and A Abraham EdsSpringer Berlin Germany 2008
[19] WN Chen and J Zhang ldquoAn ant colony optimization approachto a grid workflow scheduling problem with various QoSrequirementsrdquo IEEE Transactions on Systems Man and Cyber-netics C vol 39 no 1 pp 29ndash43 2009
[20] X Liu J Chen Z Wu Z Ni D Yuan and Y Yang ldquoHandlingrecoverable temporal violations in scientific workflow systemsa workflow rescheduling based strategyrdquo in Proceedings of the10th IEEEACM International Symposium on Cluster Cloudand Grid Computing (CCGrid rsquo10) pp 534ndash537 MelbourneAustralia May 2010
[21] X Liu Y Yang Y Jiang and J Chen ldquoPreventing temporalviolations in scientific workflows where and howrdquo IEEE Trans-actions on Software Engineering vol 37 no 6 pp 805ndash825 2011
[22] J Yu and R Buyya ldquoScheduling scientific workflow applicationswith deadline and budget constraints using genetic algorithmsrdquoScientific Programming vol 14 no 3-4 pp 217ndash230 2006
[23] L Benini and G Micheli Dynamic Power Management DesignTechniques and CAD Tools Kluwer Academic New York NYUSA 1998
[24] S Irani S Shukla and R Gupta ldquoOnline strategies for dynamicpower management in system with multiple power-savingstatesrdquo ACM Transactions on Embedded Computing Systemsvol 2 no 3 pp 325ndash346 2003
[25] M T Schmitz B M Al-Hashimi and P Eles System-Level Design Techniques for Energy-Efficient Embedded SystemsSpringer New York NY USA 2005
[26] S U Khan and I Ahmad ldquoA cooperative game theoreticaltechnique for joint optimization of energy consumption andresponse time in computational gridsrdquo IEEE Transactions onParallel and Distributed Systems vol 20 no 3 pp 346ndash3602009
[27] Y Li Y Liu and D Qian ldquoA heuristic energy-aware schedulingalgorithm for heterogeneous clustersrdquo in Proceedings of the 15thInternational Conference on Parallel and Distributed Systems(ICPADS rsquo09) pp 407ndash413 Shenzhen China December 2009
[28] L K Goh B Veeravalli and S Viswanathan ldquoDesign of fast andefficient energy-aware gradient-based scheduling algorithmsheterogeneous embeddedmultiprocessor systemsrdquo IEEE Trans-actions on Parallel and Distributed Systems vol 20 no 1 pp 1ndash12 2009
[29] F Yao A Demers and S Shenker ldquoA scheduling model forreduced CPU energyrdquo in Proceedings of the 36th IEEE AnnualSymposium on Foundations of Computer Science pp 374ndash382Milwaukee Wis USA October 1995
[30] Y Shin and K Choi ldquoPower conscious fixed priority schedulingfor hard real-time systemsrdquo in Proceedings of the 36th AnnualDesign Automation Conference (DAC rsquo99) pp 134ndash139 NewOrleans La USA June 1999
[31] I Hong D Kirovski G QuM Potkonjak andM B SrivastavaldquoPower optimization of variable-voltage core-based systemsrdquoIEEE Transactions on Computer-Aided Design of IntegratedCircuits and Systems vol 18 no 12 pp 1702ndash1714 1999
[32] P Pillai and K G Shin ldquoReal-time dynamic voltage scaling forlow-power embedded operating systemsrdquo in Proceedings of the
The Scientific World Journal 13
ACM Symposium on Operating Systems Principles pp 89ndash102October 2001
[33] H Aydin R Melhem D Mosse and P Mejia-Alvarez ldquoPower-aware scheduling for periodic real-time tasksrdquo IEEE Transac-tions on Computers vol 53 no 5 pp 584ndash600 2004
[34] J Zhuo and C Chakrabarti ldquoAn efficient dynamic task schedul-ing algorithm for battery powered DVS systemsrdquo in Proceedingsof the Asian South Pacific Design Automation Conference pp846ndash849 January 2005
[35] R Jejurikar and R Gupta ldquoDynamic slack reclamation withprocrastination scheduling in real-time embedded systemsrdquo inProceedings of the 42nd Design Automation Conference (DACrsquo05) pp 111ndash116 June 2005
[36] J Luo andN K Jha ldquoPower-profile driven variable voltage scal-ing for heterogeneous distributed real-time embedded systemsrdquoin Proceedings of the International Conference on VLSI Designpp 369ndash375 January 2003
[37] M T Schmitz and B M Al-Hashimi ldquoConsidering powervariations of DVS processing elements for energy minimisationin distributed systemsrdquo in Proceedings of the International Sym-posium on System Synthesis (ISSS rsquo01) pp 250ndash255 MontrealCanada October 2001
[38] Y Zhang X Hu and D Z Chen ldquoTask scheduling and voltageselection for energy minimizationrdquo in Proceedings of the 39thAnnual Design Automation Conference (DAC rsquo02) pp 183ndash188June 2002
[39] D Zhu R Melhem and B R Childers ldquoScheduling withdynamic voltagespeed adjustment using slack reclamationin multiprocessor real-time systemsrdquo IEEE Transactions onParallel and Distributed Systems vol 14 no 7 pp 686ndash7002003
[40] D Zhu D Mosse and R Melhem ldquoPower-aware schedulingfor ANDOR graphs in real-time systemsrdquo IEEE Transactionson Parallel and Distributed Systems vol 15 no 9 pp 849ndash8642004
[41] J Kang and S Ranka ldquoDVS based energy minimizationalgorithm for parallel machinesrdquo in Proceedings of the IEEEInternational Parallel and Distributed Processing Symposium(IPDPS rsquo08) pp 1ndash12 Miami Fla USA April 2008
[42] B Rountree D K Lowenthal S Funk V W Freeh B R deSupinski and M Schulz ldquoBounding energy consumption inlarge-scale MPI programsrdquo in Proceedings of the ACMIEEEConference on Supercomputing (SC rsquo07) ACM November 2007
[43] L Wang G von Laszewski J Dayal and F Wang ldquoTowardsenergy aware scheduling for precedence constrained paralleltasks in a cluster with DVFSrdquo in Proceedings of the 10thIEEEACM International Symposium on Cluster Cloud andGrid Computing (CCGrid rsquo10) pp 368ndash377 May 2010
[44] Y C Lee and A Y Zomaya ldquoEnergy conscious schedulingfor distributed computing systems under different operatingconditionsrdquo IEEE Transactions on Parallel and DistributedSystems vol 22 no 8 pp 1374ndash1381 2011
[45] M Mezmaz Y C Lee N Melab E Talbi and A Y ZomayaldquoA bi-objective hybrid genetic algorithm to minimize energyconsumption and makespan for precedence-constrained appli-cations using dynamic voltage scalingrdquo in Proceedings of theIEEE Congress on Evolutionary Computation (CEC rsquo10) pp 1ndash8Barcelona Spain July 2010
[46] M Mezmaz N Melab Y Kessaci et al ldquoA parallel bi-objectivehybrid metaheuristic for energy-aware scheduling for cloudcomputing systemsrdquo Journal of Parallel and Distributed Com-puting vol 71 no 11 pp 1497ndash1508 2011
[47] J Kennedy and R Eberhart ldquoParticle swarm optimizationrdquoin Proceedings of the IEEE International Conference on NeuralNetworks vol 4 pp 1942ndash1948 Perth Australia December1995
[48] Y Zhao M Wilde I Foster et al ldquoGrid middleware servicesfor virtual data discovery composition and integrationrdquo inProceedings of the 2nd Workshop on Middleware for GridComputing (MGC rsquo04) pp 57ndash62 Ontario Canada October2004
[49] A OrsquoBrien S Newhouse and J Darlington ldquoMapping ofscientific workflow within the e-protein project to distributedresourcesrdquo inProceedings of theUK e-Science All HandsMeetingNottingham UK September 2004
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World Journal 3
Discrete Particle Swarm Optimization algorithm combinedwith DVFS technique (DVFS-MODPSO) to optimize allthree objectives at the same time Our new approach providesa set of solutions named Pareto solutions (ie non-dominatedsolutions) enabling the user to select the desired tradeoff
To the best of our knowledge none of the previ-ous scheduling approaches deal with the three-dimensionalmakespancostenergy optimization when tackling the prob-lem of scheduling workflow applications on heterogeneouscomputing environments such as the cloud computing oneswhich constitute our key novelties
3 Problem Modeling
In this section we describe our system model in a formalway Our ultimate goal is to distribute workflow tasks amongcloud services so as to optimize both the usersrsquo QoS criteriaand cloud providersrsquo profits by saving energy consumption oftheir servicesTherefore we first present the cloud computingmodel Then we describe our workflow model and the QoSparameters we deal We conclude this section by describingthe scheduling model formalized as a multi-objective opti-mization problem we solve
31 Cloud Computing Model The cloud computing systemused in this work is a set of resources offered by a cloudprovider to run client applications Our cloud model isinspired by the model described in [45] We assume that thecloud is hosted in data centers composed of heterogeneousmachines These data centers deliver a variety of serviceshosted on thousands of IT servers which are made availableas subscription-based services in a pay-as-you-go model Inour model the cloud computing system consists of a set of119875 = 119901
1 1199012 119901
119898 heterogeneous processors which are
fully interconnected The processors have varied processingcapability delivered at different processing prices (see ec ofTable 1) Each processor 119901
119895isin 119875 is DVFS-enabled that is it
can operate with different VSLs (Voltage Scaling Level iedifferent clock frequencies) For each processor 119901
119895isin 119875
a set 119881119895of V VSLs is randomly and uniformly distributed
among three different sets of VSLs (Table 1) We consider thatprocessors consume energy during periods of inactivity thatis when a processor is idling it is assumed that the lowestvoltage is supplied [44] Because clock frequency transitionoverheads usually take a negligible amount of time theseoverheads are not considered in this paper and the inclusionof such an overheadwill have no bearing on the overall modelof the proposed study
Additionally each processor 119901119895isin 119875 has a set of links
119871119901119895= 119897119895 1199011 119897119895 1199012 119897119895 119901119896 1 le 119896 le 119898 where 119897
119895119894isin 119877+ is
the available bandwidthmdashmeasured in Mega bits per second(Mbps)mdashin the link between processors 119901
119895and 119901
119894 with 119897
119895119895=
1 We assume that a message can be transmitted from oneprocessor to another while a task is executed on the recipientprocessor Finally communication between tasks executed onthe same processor is neglected Table 1 shows DVFS levelsRelative speeds (RSpeed) and execution costs (ec) for threeprocessor classes (TURIONMT-34 OMAP PENTIUMM)
32 Workflow Application Model We model a cloud work-flow application as a Directed Acyclic Graph (DAG) denotedas 119866(119881 119864) The set of nodes 119881 = 119879
1 119879
119899 represents
the tasks in the workflow application the set of arcs denotesprecedence constraints and the controldata dependenciesbetween tasks An arc is in the form of 119889
119894119895= (119879119894 119879119895) isin 119864
where119879119894is called the parent task of119879
119895119879119895is the child task of119879
119894
119889119894119895is the data produced by119879
119894and consumed by119879
119895We assume
that a child task cannot be executed until all of its parent taskshave been completed In a given task graph a task with noparent is referred as an entry task and one without any childis called an exit task Since our algorithm involves only oneentry and one exit tasks we add two dummy tasks 119879entry and119879exit which have zero execution time to the beginning andthe end of the workflow respectively These dummy tasks areconnected with zero-weight arcs to the actual entry and exittasks respectively
We assume that each task 119879119894isin 119881 has an associated
basic execution time which is an independent value for eachmachine We denote 119908
119894119895 the basic computation time of a
task 119879119894on a compute resource 119901
119895at maximum speed and
voltage (ie it corresponds to Level 1 in Table 1) The averageexecution time of the task 119879
119894is defined as
119908119894=
119898
sum119895=1
119908119894119895
119898 (1)
Real computation time119908119903119894119895119896
of the task 119879119894on machine 119901
119895
using relative execution speed 119904119896119895is defined as
119908119903119894119895119896=119908119894119895
119904119896119895
(2)
We also assume that every edge (119879119894 119879119895) isin 119864 is associated
with value tr119894119895 representing the time needed to transfer data
from 119879119894to 119879119895 The transfer time can be calculated according
to the bandwidth 119897119909119910
between the resources executing thesetasks (119901
119909and 119901
119910resp) as follows
tr119894119895=119889119894119895
119897119909119910
(3)
However a communication time is only required whentwo tasks are assigned to different processors That is thecommunication time when tasks are assigned to the sameprocessor can be neglected that is 0
In general the execution costs (ec) and transmission costs(trc) are inversely proportional to the execution times andtransmission times respectively
We define pred (119879119894) as the set of all predecessors of 119879
119894and
succ (119879119894) as the set of all successors of 119879
119894 An ancestor of node
119879119894is any node 119879
119896that is contained in pred (119879
119894) or any node
119879119895that is also an ancestor of any node 119879
119896contained in pred
(119879119896)The Earliest Start Time and the Earliest Finish Time of a
task 119879119894on a processor 119901
119895are represented as EST (119879
119894 119901119895) and
EFT (119879119894 119901119895) respectively EST (119879
119894) and EFT (119879
119894) represent the
earliest start and finish times on any processor respectively
4 The Scientific World Journal
Table 1 Voltage Scaling Levels and relative speeds of processors
VSLP1 TURIONMT-34 P2 OMAP P3 PENTIUMM
Volt Freq R Speed Volt Freq R Speed Volt Freq R Speed(V) (Ghz) () (V) (Ghz) () (V) (Ghz) ()
1 12 18 100 135 06 100 148 14 1002 115 16 8888 127 055 9166 144 12 96763 11 14 7777 12 05 8333 131 1 88144 105 12 6666 1 025 4166 118 08 79515 1 1 55555 09 013 2083 096 06 64426 09 08 4444Cost ec = 114 ec = 057 ec = 076
Table 2 Task execution times and priorities
Task (119879119894) 119901
11199012
1199013
120596119894
Priority (Pr(119879119894))
1 14 16 9 13 1080002 13 19 18 1667 770003 11 13 19 1433 800004 13 8 17 1267 800005 12 13 10 1167 690006 13 16 9 1267 633337 07 15 11 11 426678 5 11 14 10 356679 18 12 20 1667 4433310 21 7 16 1467 14667
119875119886V(119879119894 119901119895) is defined as the earliest time when processor 119901
119895
will be available to begin executing task 119879119894 Hence
119864119878119879 (119879119894 119901119895) =
119900 if 119879119894= 119879entry
max 119875119886V(119879119894 119901119895) 120572
(4)
where 120572 = max119879119895 isin pred (119879119894 )(119864119865119879(119879119895 119901119896) + tr
119895119894
119864119865119879 (119879119894 119901119895) = 119908
119894119895+ 119864119878119879 (119879
119894 119901119895) (5)
Note that the Actual Start Time and Actual Finish Timeof a task 119879
119894on a processor 119901
119895 denoted as AST (119879
119894 119901119895) and
AFT (119879119894 119901119895) can be different from its earliest start EST (119879
119894 119901119895)
and finish EFT (119879119894 119901119895) times if the actual finish time of
another task 119879119896scheduled on the same processor is later than
its EST (119879119896 119901119895) [44]
Figure 1 depicts a workflow application with 10 tasks andthe Table 2 provides its details (given in [17]) The valuespresented in the last column of the table represent the priorityof the tasks The priority of task 119879
119894represented by Pr(119879
119894) is
computed recursively by traversing the DAG upward startingfrom the exit task 119879exit as follows (6)
Pr (119879119894) =
119908119879exit if 119879
119894= 119879exit
119908119894+ 120573 otherwise
(6)
where120573 = max119879119895 isin succ (119879119894)119879119903119894119895 + Pr(119879
119895)
1
2 5 6
7 8
10
3 4
9
18 12 9 1114
19 16 2723 23 1513
1311 17
Figure 1 An example of workflow (given in [17]) with the tasknumbers 119879
119894inside nodes and values of 119889
119894119895function next to the
corresponding edges
33 QoS Parameter Models
331 Energy Model Among the main system-level energy-saving techniques Dynamic Voltage Scaling (DVS) operateson a simple principle decreases the supply voltage (and so theclock frequency) to the CPU so as to consume less power
In this work we use a model of energy derived from thepower consumption model in digital complementary metal-oxide semiconductor (CMOS) logic circuits [44] Under thedynamic power model the processor power is dominated bythe dynamic power which is given by
119875dynamic = 119860119862efV2119891 (7)
where119860 is the number of switches per clock cycle119862ef denotesthe effective charged capacitance V is the supply voltage and119891 denotes the operational frequency Equation (7) shows thatthe supply voltage is the dominant factor hence its reductionwould be most influential to lower power consumption
The energy consumption of the execution of a workflowapplication used in this paper is defined as
119864119888=
119899
sum119894=1
119860119862efV2
119894119891119894119908119903lowast
119894
=
119899
sum119894=1
120574V2119894119891119894119908119903lowast
119894
(8)
The Scientific World Journal 5
where 120574 = 119860119862ef is assumed constant for a given machineV119894 119891119894are the voltage supply and frequency of the processor
on which task 119879119894is executed respectively and 119908119903lowast
119894is the real
completion time of task 119879119894on the scheduled processor In the
idle time the processor turns into sleep mode and thus thevoltage supply and relative frequency are at the lowest levelSo the energy consumption during idle periods of processorsis defined as
119864119894=
119898
sum119895=1
sumidle119895119896isin IDLE119895
120574V2min 119895 119891min 119895 119871119895119896 (9)
where IDLE119895
is the set of idling slots on machine119901119895 Vmin 119895(119891min 119895) is the lowest supply voltage (frequency)
on 119901119895 and 119871
119895119896is the amount of idling time for idle
119895119896
Then the total energy 119864total utilized by the cloud system forcompletion of the workflow application can be defined asfollows
119864total = 119864119888+119864119894 (10)
332 Time Model The completion time 119879total is themakespan from the user submitting a workflow untilreceiving the results It is defined as the actual finish timeof the exit task It is calculated in equation 11 and consistsof the workflow execution time and network transmissiontime The execution time depends on both the workloadand system performance The network transmission timedepends on both the network latency and the input data size
119879total = maxAFT (119879exit) (11)
333 Cost Model As result of the marketization characteris-tic of current services most cloud providers have set a pricefor their services They have fixed the price for transferringbasic data unit (eg per MB) between two services and theprice for processing basic time units (eg per hour)The cost119862total of running a cloud workflow is defined in Formula (12)It consists of processing cost 119862ex and data transfers cost 119862tr
119862total = 119862ex + 119862tr (12)
The processing cost for a given task119879119894depends on the real
completion time of 119879119894on the scheduled processor (119908119903lowast
119894) and
the hourly price of this processor (ec119894) Thus 119862ex is given by
119862ex =119899
sum119894=1
119908119903lowast
119894lowast ec119894 (13)
The data transfer cost (119862tr) is described as follows
119862tr =119899
sum119894=1
119899
sum119895=1
119889119894119895lowast trc119894119895 (14)
where 119889119894119895characterizes the output file size from task119879
119894to task
119879119895 and tr 119888
119894119895represents the cost of communication from the
processor where 119879119894is mapped to another processor where 119879
119895
is mappedThe cost of communication is added to the overallcost only when two tasks have data dependency betweenthem (ie 119889
119894119895gt 0) For two or more tasks running on the
same processor the transfer cost is neglected
334 Scheduling Model Given (1) A cloud provider thatoffers a set 119875 of 119898 heterogeneous processors and (2) auser workflow application composed of a set 119879 of 119899 tasksthat have to be executed on these processors The workflowscheduling problem is to construct a mapping 119872 of tasksto processors (without violating precedence constraints) thatminimizes the following conflicting objectives makespancost and energy consumption as low as possible Thereforethe workflow scheduling problem can be formulated as amathematical optimization problem
MakespanMinimize 119879total (119872)
CostMinimize 119862total (119872)
EnergyMinimize119864total (119872)
(15)
4 Workflow Scheduling Based on DiscreteParticle Swarm Optimization
This section starts with a brief overview on multi-objectivecombinatorial optimization and Particle swarm optimizationalgorithm Afterwards our new Multi-Objective DiscreteSwarmOptimization combined with DVFS technique will bepresented
41 Multi-Objective Optimization A Multi-objective Opti-mization Problem (MOP) with 119898 decision variables and 119899objectives can be formally defined as
Min (119910 = 119891 (119909) = [1198911(119909) 119891
119899(119909)]) (16)
where 119909 = (1199091 119909
119898) isin 119883 is an 119898-dimensional decision
vector 119883 is the search space 119910 = (1199101 119910
119899) isin 119884 is the
objective vector and Y the objective-spaceIn general MOP there is no single optimal solution
with regards to all objectives This is also the case forthe multi-objective optimization problem addressed in thispaper As given in (15) there are three conflicting objectivesminimizing execution time minimizing execution cost andminimizing energy consumption In such problems thedesired solution is considered to be the set of potentialsolutions which are all optimal in some objectives This set isknown as the Pareto optimal setWe provide some definitionsof the Pareto concepts used in MOP as follows (withoutloss of generality we suppose that the objectives are to beminimized)
(i) Pareto dominance For two decision vectors 1199091 and 1199092dominance (denoted by ≺) is defined as follows
1199091≺ 1199092lArrrArr forall119894 119891
119894(1199091) le 119891119894(1199092) and exist119895119891
119895(1199091) lt 119891119895(1199092)
(17)
The decision vector 1199091is said to dominate 1199092 if andonly if 1199091 is as good as 1199092 considering all objectivesand 1199091 is strictly better than 1199092 in at least oneobjective
6 The Scientific World Journal
(ii) Pareto optimally A decision vector 1199091 is said to bePareto optimal if and only if
∄1199092isin 119883 119909
2≺ 1199091 (18)
(iii) Pareto optimal set The Pareto optimal set 119875119878is the set
of all Pareto optimal decision vectors
119875119878= 1199091isin 119883 | ∄119909
2isin 119883 119909
2≺ 1199091 (19)
(iv) Pareto optimal frontThePareto optimal front119875119865is the
image of the Pareto optimal set in the objective space
119875119865= 119891 (119909) = (119891
1(119909) 119891
119899(119909)) | 119909 isin 119875
119904 (20)
1199091 is said to be non-dominated regarding a given set
if 1199091 is not dominated by any decision vectors in theset
Thepareto optimal decision vector cannot be enhanced inany objective without causing degradation in at least anotherobjective A decision vector is said to be Pareto optimal whenit is not dominated in the whole search space
42 Particle Swarm Optimisation
421 The Standard PSO PSO is a population-based stochas-tic optimization technique developed by Kennedy and Eber-hart in 1995 [47] It is inspired by the social behavior of insectcolonies bird flocks fish schools and other animal societies Itis also related to evolutionary computation PSOhas attractedsignificant attention from many researchers due to both itssimplicity of use and optimization via social behavior In factPSO has good performance requires low computational costIt is effective and easy to implement as it uses numericalencoding
A particle in PSO is analogous to a fish or bird movingin the 119863-dimensional search space All particles have fitnessvalues indicating their performances which are problemspecific and velocities which direct the flight of particlesEach particle position at any given time is influenced by bothits best position called 119901119861119890119904119905 and the position of the bestparticle in a problem space referred to as 119892119861119890119904119905 Thereforeparticles tend to fly towards a better search area duringthe search process A particle status on the search spaceis characterized by two elements namely its velocity andposition which are updated in every generation as follows
119881119896+1
119894= 120596119881
119896
119894+ 11988811199031(119901119861119890119904119905
119894minus 119883119896
119894)
+ 11988821199032(119892119861119890119904119905 minus 119883
119896
119894)
119883119896+1
119894= 119883119896
119894+ 119881119896+1
119894
(21)
where 119881119896+1119894
is the velocity of particle 119894 at iteration 119896 + 1119881119896119894is the velocity of particle 119894 at iteration 119896 120596 is the inertia
weight 1198881and 1198882are the acceleration coefficients (cognitive
and social coefficients) 1199031and 1199032are the random numbers
between 0 and 1119883119896119894is the current position of particle 119894 at the
119896th iteration 119901119861119890119904119905119894is the best previous position of the 119894th
particle 119892119861119890119904119905 is the position of best particle in the swarmand119883119896+1
119894is the position of 119894th particle at 119896 + 1 iteration
The procedure for standard PSO is as follows
(1) Initialize a population of particles with random posi-tions and velocities in the search space
(2) Evaluate the objective values of all particles set 119901119861119890119904119905of each particle equal to its current position and set119892119861119890119904119905 equal to the position of the best initial particle
(3) Update the velocity and the position of each particleaccording to (21)
(4) Map the position of each particle in the solution spaceand evaluate its fitness value according to the desiredoptimization fitness function
(5) For each particle compare its current objective valuewith its 119901119861119890119904119905 value If the current value is better thenupdate 119901Best with the current position and objectivevalue
(6) Determine the best particle of the current whole pop-ulation with the best objective value If the objectivevalue is better than that of 119892119861119890119904119905 then update 119892Bestwith the current best particle
(7) If the stopping criterion ismet then output 119892119861119890119904119905 andits objective value otherwise go to Step (2)
The original design of PSO is appropriate for findingsolutions to continuous optimization problems However asthe workflow scheduling discussed in this paper is both adiscrete andmulti objective problem in nature we propose aneffective approach to address this problem by using a discreteversion of the Multi-Objective PSO (MODPSO) combinedwith DVFS techniqueThe key issues of our approach consistof (1) defining the position and velocity of the particlesbased on the features of discrete variables of the workflowscheduling (2) solving the multi-objective aspect of theproblem by modifying PSO so as to generate a set of non-dominated solutions satisfying the different objectives underconsideration instead of one solution
422 Handling Workflow Scheduling Using DVFS-MODPSOIn general a workflow scheduling can be defined by a setof triplets 119872 = [⟨119879
119894 119875119895 119871119896⟩](119894 isin [1 119899] 119895 isin [1119898] 119896 isin
[1 119871(119895)]) 119899 is the number of workflow tasks to be scheduled119898 is the number of processors available in the cloud environ-ment and 119871(119895) is the number of operating points (VSLs) of the119895th processor
For the sake of clarity the variables and rules of DVFS-MODPSO for solving workflow scheduling can be depictedas follows
The position of a particle represents a feasible solutionto the scheduling problem It consists of a set of⟨task(119879
119894) service (119875
119895) VSl (119871
119896)⟩ triplets Each triplet
means that a task (119879119894) is assigned to a processor (119875
119895)
with a voltage scaling level (119871119896) It also indicates
The Scientific World Journal 7
that the position satisfies the precedence constraintbetween tasks
The process of generating a newposition for a selectedparticle in the swarm is depicted in the followingformulas
119881119896+1
119894= 119881119896
119894oplus ((119877
1otimes (119901119861119890119904119905
119894⊖ 119883119896
119894))
oplus (1198772otimes (119892119861119890119904119905 ⊖ 119883
119896
119894)))
(22)
119883119896+1
119894= 119883119896
119894oplus 119881119896+1
119894 (23)
The operator definitions are as follows
(i) The substract operator (⊖) the difference betweentwo particle positions designated as 1199091 and 1199092 isdefined as a set of triplets in which each triplet⟨task serviceVSL⟩ shows whether the contents ofthe corresponding elements in 1199091 are different fromthose of 1199092 or not If so that triplet gets its values(service and VSL) from the position that has thelowest value of theVSL For those triplets that have thesame content in 1199091 and 1199092 their corresponding VSLsare decreased (Note that the scaling of VSL makesfluctuations on the energy makespan and cost)
(ii) The multiply operator (otimes) the multiplication betweennumber and velocity is defined as a set of tripletswhere a threshold 120572 isin [0 1] is defined arandom number 119903 is generated for each triplet⟨task serviceVSL⟩ compare 119903 and 120572 when 119903 ge 120572decrease the triplet VSL otherwise increase it Thisoperator adds the exploration ability to the algorithm
(iii) The add operator (oplus) the addition of two positions isdefined as the reservation of the dominated one
The Pseudocode 1 outlines the general steps of the DVFS-MODPSO algorithm
The algorithm begins by initializing the positions andvelocities of particles To obtain the position of a particle theVSL (voltage and frequency) of each resource is randomlyinitialized firstly then the HEFT algorithm is applied togenerate a feasible and efficient solution minimizing themakespan The process is repeated several times to initializethe positions of all particles of the swarm Initially thevelocity119881 and the best position for each particle 119901119861119890119904119905 areattributed as the particle itself The algorithm maintainsan external archive EA to store non-dominated particlesfound after the evaluation process (based on the paretodominance using the objectives mentioned in 33) After allthese initializations in the main loop of the algorithm thenew velocity and position of each particle are calculatedrespectively after selecting the best overall position in theexternal archive and eventually perform a mutation then theparticle is evaluated and its corresponding 119901119861119890119904119905 is updatedThe external archive is updated after every iteration Oncethe termination condition is reached the external archivecontaining the Pareto front is returned as the result
5 Experimental Evaluations
In this section we describe the overall setup of our experi-mentation effort and the results we have obtained from it tovalidate the new proposed approach
In our experiments we simulate a synthetic workflowapplication described in [17] (see Figure 1) and two commonworkflow structures parallel and hybrid structuresWe chosetwo well known real world applications namely a neuro-science workflow [48] for our parallel application and aprotein annotation workflow [49] developed by London e-Science Centre for our hybrid workflow application Figure 2shows their simplified representationsThe number indicatedin the parentheses next to each node represents the length ofthe task (the number of instructions to execute) in Millionsof Instructions (MI) The input and output files of each taskrange from 10MB to 1024MB
We have performed experiments on 3 4 5 6 and 12resources whose characteristics are shown in Table 1 Wechoose the pricing model associated with Amazon EC2(httpawsamazoncomec2) for the processing costs of thedifferent classes of resources and the pricing model given byAmazon CloudFront (httpawsamazoncomcloudfront)for the costs of transferring data unit between resources
As for DVFS-MODPSO the parameter settings areSwarm size 119878119873119906119898 = 50 particles and the maximum numberof iterations 119866 = 100
We evaluated the performance of our proposed DVFS-MODPSO on all the workflow instances described previouslyDue to the lack of works considering both the heteroge-neous configuration of our cloud and the three QoS metrics(makespan cost energy) at the same time we comparedour results with those of the HEFT heuristic we haveimplemented HEFT is one of themost widely used heuristicsfor DAGs in distributed heterogeneous computing systemswhich optimize the makespan
Figures 3 4 and 5 show sample Pareto fronts obtainedwith the DVFS-MODPSO and the solution computed byHEFT for the synthetic workflow the hybrid workflow andthe parallel workflow respectively
As can be seen from these figures unlike the HEFTalgorithm which gives one solution our proposed approachprovides a set of non dominated solutions These resultsillustrate the basic multi-objective definitions provided inSection 41
Now for comparing these two approaches and in orderto analyze the effectiveness of our proposition in terms ofthe values of makespan cost and energy consumption wehave been inspired by the methodology described in [43]where we compare the solution provided by HEFT to onlyone solution of the Pareto front set computed by our proposedmulti-objective DVFS-MODPSO
The evaluation process follows the next steps For eachworkflow instance we perform a first resolution usingHEFTin order to get one solution 119878 After a second resolution iscomputed using our proposed approach to obtain a set EA ofPareto solutions Next we select one solution 1198781015840 from the setEA of Pareto solutions This solution is the closest one to 119878 inthe sense of Euclidean distance Finally a comparison is donebetween 119878 and 1198781015840
8 The Scientific World Journal
Swarm initialization with HEFT(1) For i = 1 to SNum (SNum is the size of particle swarm)
(a) For j = 1 to m (m the number of processors)(i) Randomly initialize VSL(j) (randomly choose the voltagefrequency of processor from the set of its operating points)
(b) Initialize 119878[119894] with HEFT heuristic (S is the swarm of particles)(c) Initialize the velocity V of each particle
119881 [119894] = 0 (119878 [119894])
(d) Initialize the Personal Best (pBest) of each particle119901119861119890119904119905[119894] = 119878[119894]
(e) Evaluate objectives of each particleEvaluate S[i]
(f) Initialize the Global Best particle (gBest) with the best one among the SNum particlesg119861119890119904119905 = Best particle found in 119878
(2) End For(3) Add the nondominated solutions found in S into EA (EA is the External Archive storing the pareto front)(4) Initialize the iteration number 119905 = 0(5) Repeat until 119905 gt 119866 (119866 is the maximum number of iterations)
(a) For 119894 = 1 to SNum (swarm size)(i) Randomly select the global best particle for S from the External Archive EA and store its position in gBest(ii) Calculate the new velocity 119881[119894] according to (22)(iii) Compute the new position of 119878[119894] according to (23)(iv) If (119905 lt 119866 lowast PMUT ) then (PMUT is the probability of mutation)
Perform mutation on 119878[119894](v) Evaluate S[i]
(b) End For(c) Update the personal best solution of each particle 119878[119894]
if 119878 [119894] ≼ 119901119861119890119904119905119904 [119894] or (119878 [119894] sim 119901119861119890119904119905119904 [119894])Then 119901119861119890119904119905119904 [119894] = 119878 [119894]
(d) Update the External Archive EA(i) Add all new non-dominated solution in S into EAif 119878 [119894] 997410 119909 forall119909 isin 119864119860Then 119864119860 = 119864119860 cup 119878 [119894](ii) Remove all particles dominated by 119878[119894] in EA
119864119860 = 119864119860 minus 119910 isin 119864119860 | 119910 lt 119878[119894]
(iii) If the archive is full then randomly select the article to be replaced from EA(6) Return the pareto front (the set of non-dominated solutions from S and EA)
Pseudocode 1 DVFS-MODPSO based workflow scheduling
The different results are displayed on Figures 6 7 8 and 9and on Tables 3 and 4 They show the improvement achievedby our proposed approach according to the structure of theworkflow applications and the number of processors
Table 3 and Figure 6 show that our proposed approachimproves on average the result obtained by HEFT for thesynthetic workflow instance The makespan is reduced by005 the cost is reduced by 993 and the energy con-sumption is reduced by 2640 Figure 6 shows detailedimprovements of the proposed approach according to thenumber of processors
Table 4 and Figure 7 illustrate the gain obtained by ourapproach according to the kind of real world applicationsTable 4 also shows the detailed improvements when scalingthe number of processors As can be seen the QoS metricsare improved over HEFT in average by 095 for themakespan 108 for cost and 812 for the energy con-sumption when using the hybrid workflow application andaverage improvements of 295 formakespan 2215 for costand 209 for energy consumption when using the parallelworkflow
These evaluations confirm the results obtained from thesynthetic workflow This means that by using our DVFS-MODPSO we are able to improve not only the cost and energybut also themakespan onwhichHEFT algorithm is supposedto provide good results These results remain true for all kindof workflow applications
In our experiments we limited the number of resourcesto 12 as we found that for these kinds of workflowapplications some resources are not used at all Figure 8shows the resource loads when scheduling the syntheticworkflow on 12 resources Furthermore even though thenumber of resources increased the total cost and total energyconsumption could not always decrease as illustrated inFigure 9 From this figure we can see that the QoS metrics(1) firstly decreases as the number of resources increasesuntil achieving the number 6 This can be explained by thefact that when increasing the number of resources thereare fewer tasks executed in a resource therefore tasks canextend their execution times and the resource have more ofchances to scale down their voltages and frequencies whichcan be very effective in reducing total energy consumption
The Scientific World Journal 9
1 2 3 4
5
6 7
8 9 10
11
12
14
13
15
(300000) (600000) (600000) (900000)
(300000)
(150000) (150000)
(300000) (300000) (300000)
(600000)
(300000)
(150000)
(300000)
(600000)
SignalP COILS2 SEG PROSITE
TMHMM
ProsperoHMMer
PSI-BLASTBLAST
IMPALA
PSI-PRED
3D-PSSMSummary
Genomesummary
SCOP
(a)
1 3 5 7
9
11 1210
(400000) (400000) (400000) (400000)
(300000)
(500000) (500000) (500000)
Softmea
2 4 6 8(600000) (600000) (600000) (600000)
Reslice
14 1513
(600000) (600000) (600000)
Convert
Reslice Reslice Reslice
Slicer Slicer
Convert Convert
Slicer
Align wapAlign wapAlign wapAlign wap
(b)
Figure 2 Parallel and Hybrid workflows
10 The Scientific World Journal
75 80 85 90 95 100 8090
100110
120130
140
2060
100140180220
Ener
gy
DVFS-MODPSOHEFT
MakespanCost
Figure 3 Obtained non-dominated solutions for the syntheticworkflow
Ener
gy
DVFS-MODPSOHEFT
MakespanCost
200 400 600 800 100012001400160018002000 400800
12001600
2000
400200
600800
100012001400160018002000
Figure 4 Obtained non-dominated solutions for the hybrid work-flow
(2) After achieving 6 resources the QoS metrics begin torise the reason for this is that time executions are dominatedby interprocessors communications hence decreasing theopportunities for scaling down voltages and frequencies ofresources Consequently the threshold numbers of resourcesthat minimized the QoS metrics could be obtained
6 Conclusion
In this paper we propose a new algorithm called DVFS-Multi-Objective Discrete Particle Swarm Optimization(DVFS-MODPSO) for workflow scheduling in distributedenvironments such as cloud computing infrastructures DV-FS-MODPSO simultaneously optimizes several conflictingobjectives namely the makespan cost and energy in adiscrete space It produces a set of non-dominated solutionsthus providing more flexibility for users to assess theirpreferences and select a schedule that meets their QoS
Ener
gy
DVFS-MODPSOHEFT
Makespan Cost
400200600800
10001200140016001800
240020002200
450 500 550 600 650 700 7501200
14001600
18002000
Figure 5 Obtained non-dominated solutions for the parallel work-flow
0
10
20
30
40
50
3 4 5 6 12
MakespanCostEnergy
Figure 6 Gain over HEFT () according to the number ofprocessors
requirements Our approach exploits the heterogeneityand the marketization of cloud resources in order to findsolutions that optimize makespan and cost Furthermore ituses the Dynamic Voltage and Frequency Scaling (DVFS)technique to reduce energy consumption
We have evaluated our algorithm by simulating it withboth synthetic and real world scientific workflow applicationshaving different structures The results show that the pro-posedDVFS-MODPSO is able to produce a set of good Paretooptimal solutions The results also show that our approachprovides significant improvement not only in terms of the
The Scientific World Journal 11
0
5
10
15
20
25
HybridParallel
Makespan Cost Energy
Figure 7 Improvement according to real world applications
R17
R20
R37
R446
R50
R60
R733
R80
R90
R107
R110
R120
Figure 8 Resource loads for synthetic workflow
Table 3 Improvement for synthetic workflow
Num of proc Gain over Heft ()Makespan Cost Energy
3 0 2126 40084 022 464 15355 0 756 17266 003 652 267512 0 969 3255Average 005 993 2640
cost and the energy consumption but also in term of themakespan for which HEFT algorithm is supposed to givebetter results
Multi-objective optimization in cloud workflow schedul-ing is not a mature domain Most of the existing worksattempt to minimize either the makespan or cost Howeverwe plan to consider other objectives such as reliabilitysecurity in addition to the energy consumption We also
0
50
100
150
200
250
3 4 5 6 12
MakespanCostEnergy
Figure 9 QoS metrics evolutions according to resource numbersfor the synthetic workflow
Table 4 Improvement according to real world applications
Appli Num of proc Gain over Heft ()Makespan Cost Energy
Hybrid
3 405 1155 04024 028 0566 05075 000 1347 11246 000 0117 126512 043 2214 0761
Average 095 1080 0812
Parallel
3 235 5640 25524 1147 2094 32715 000 2142 11666 093 1170 315412 000 031 306
Average 295 2215 209
intend to apply our algorithm in a real world cloud orintegrate it in existing cloud toolkits such as Cloudbus forcomparing with other algorithms
References
[1] D Thain T Tannenbaum and M Livny Condor and the GridGrid Computing Making the Global Infrastructure a RealityJohn Wiley amp Sons Hoboken NJ USA 2003
[2] R Buyya and S Venugopal ldquoThe gridbus toolkit for serviceoriented grid and utility computing an overview and statusreportrdquo in Proceedings of the 1st IEEE International Workshopon Grid Economics and Business Models (GECON rsquo04) pp 19ndash36 IEEE CS Seoul Republic of Korea April 2004
[3] S McGough L Young A Afzal S Newhouse and J Darling-ton ldquoWorkflow enactment in ICENIrdquo in Proceedings of the UK
12 The Scientific World Journal
e-Science All Hands Meeting pp 894ndash900 IOP NottinghamUK September 2004
[4] E Deelman J Blythe Y Gil et al ldquoPegasus mapping scientificworkflows onto the gridrdquo in Grid Computing 2nd EuropeanAcrossGrids Conference AxGrids 2004 Nicosia Cyprus January28ndash30 2004 vol 3165 of Lecture Notes in Computer Science pp11ndash20 Springer New York NY USA 2004
[5] R Buyya S Pandey and C Vecchiola ldquoCloudbus toolkitfor market-oriented cloud computingrdquo in Cloud ComputingProceedings of the 1st International Conference CloudCom 2009Beijing China December 1ndash4 2009 vol 5931 of Lecture Notesin Computer Science pp 24ndash44 Springer New York NY USA2009
[6] Y Yang K Liu J Chen X Liu D Yuan and H Jin ldquoAnalgorithm in SwinDeW-C for scheduling transaction-intensivecost-constrained cloud workflowsrdquo in Proceedings of the 4thIEEE International Conference on eScience (eScience rsquo08) pp374ndash375 Indianapolis Ind USA December 2008
[7] L Ramakrishnan C Koelbel Y Kee et al ldquoVGrADS enablinge-Scienceworkflows on grids and cloudswith fault tolerancerdquo inProceedings of the Conference on High Performance ComputingNetworking Storage and Analysis (SC rsquo09) November 2009
[8] M L Pinedo Scheduling Theory Algorithms and SystemsSpringer Berlin 2008
[9] S Pandey L Wu S M Guru and R Buyya ldquoA particle swarmoptimization-based heuristic for scheduling workflow applica-tions in cloud computing environmentsrdquo in Proceedings of the24th IEEE International Conference on Advanced InformationNetworking and Applications (AINA rsquo10) pp 400ndash407 PerthAustralia April 2010
[10] S Abrishami and M Naghibzadeh ldquoDeadline-constrainedworkflow scheduling in software as a service cloudrdquo ScientiaIranica vol 19 no 3 pp 680ndash689 2012
[11] S Selvarani and G S Sadhasivam ldquoImproved cost-based algo-rithm for task scheduling in cloud computingrdquo in Proceedings ofthe IEEE International Conference on Computational Intelligenceand Computing Research (ICCIC rsquo10) pp 1ndash5 CoimbatoreIndia December 2010
[12] A Verma and S Kaushal ldquoDeadline and budget distributionbased cost-time optimization workflow scheduling algorithmfor cloudrdquo inProceedings of the IJCAon International Conferenceon Recent Advances and Future Trends in Information Technol-ogy (iRAFIT rsquo12) iRAFIT(7) pp 1ndash4 April 2012
[13] R C Correa A Ferreira and P Rebreyend ldquoIntegrating listheuristics into genetic algorithms for multiprocessor schedul-ingrdquo in Proceedings of the 8th IEEE Symposium on Parallel andDistributed Processing (SPDP rsquo96) pp 462ndash469 IEEEComputerSociety Washington DC USA October 1996
[14] E S H Hou N Ansari and H Ren ldquoGenetic algorithm formultiprocessor schedulingrdquo IEEE Transactions on Parallel andDistributed Systems vol 5 no 2 pp 113ndash120 1994
[15] M Grajcar ldquoGenetic list scheduling algorithm for schedulingand allocation on a loosely coupled heterogeneous multi-processor systemrdquo in Proceedings of the 36th Annual DesignAutomation Conference (DAC rsquo99) pp 280ndash285 ACM PressNew York NY USA June 1999
[16] R Sakellariou and H Zhao ldquoA hybrid heuristic for DAGscheduling on heterogeneous systemsrdquo in Proceedings of the13th Heterogeneous Computing Workshop pp 111ndash123 IEEEComputer Society Washington DC USA 2004
[17] H Topcuoglu S Hariri and M Y Wu ldquoPerformance-effectiveand low-complexity task scheduling for heterogeneous comput-ingrdquo IEEE Transactions on Parallel and Distributed Systems vol13 no 3 pp 260ndash274 2002
[18] J Yu and R Buyya ldquoWorkflow scheduling algorithms for gridcomputingrdquo in Metaheuristics for Scheduling in DistributedComputing Environments F Xhafa and A Abraham EdsSpringer Berlin Germany 2008
[19] WN Chen and J Zhang ldquoAn ant colony optimization approachto a grid workflow scheduling problem with various QoSrequirementsrdquo IEEE Transactions on Systems Man and Cyber-netics C vol 39 no 1 pp 29ndash43 2009
[20] X Liu J Chen Z Wu Z Ni D Yuan and Y Yang ldquoHandlingrecoverable temporal violations in scientific workflow systemsa workflow rescheduling based strategyrdquo in Proceedings of the10th IEEEACM International Symposium on Cluster Cloudand Grid Computing (CCGrid rsquo10) pp 534ndash537 MelbourneAustralia May 2010
[21] X Liu Y Yang Y Jiang and J Chen ldquoPreventing temporalviolations in scientific workflows where and howrdquo IEEE Trans-actions on Software Engineering vol 37 no 6 pp 805ndash825 2011
[22] J Yu and R Buyya ldquoScheduling scientific workflow applicationswith deadline and budget constraints using genetic algorithmsrdquoScientific Programming vol 14 no 3-4 pp 217ndash230 2006
[23] L Benini and G Micheli Dynamic Power Management DesignTechniques and CAD Tools Kluwer Academic New York NYUSA 1998
[24] S Irani S Shukla and R Gupta ldquoOnline strategies for dynamicpower management in system with multiple power-savingstatesrdquo ACM Transactions on Embedded Computing Systemsvol 2 no 3 pp 325ndash346 2003
[25] M T Schmitz B M Al-Hashimi and P Eles System-Level Design Techniques for Energy-Efficient Embedded SystemsSpringer New York NY USA 2005
[26] S U Khan and I Ahmad ldquoA cooperative game theoreticaltechnique for joint optimization of energy consumption andresponse time in computational gridsrdquo IEEE Transactions onParallel and Distributed Systems vol 20 no 3 pp 346ndash3602009
[27] Y Li Y Liu and D Qian ldquoA heuristic energy-aware schedulingalgorithm for heterogeneous clustersrdquo in Proceedings of the 15thInternational Conference on Parallel and Distributed Systems(ICPADS rsquo09) pp 407ndash413 Shenzhen China December 2009
[28] L K Goh B Veeravalli and S Viswanathan ldquoDesign of fast andefficient energy-aware gradient-based scheduling algorithmsheterogeneous embeddedmultiprocessor systemsrdquo IEEE Trans-actions on Parallel and Distributed Systems vol 20 no 1 pp 1ndash12 2009
[29] F Yao A Demers and S Shenker ldquoA scheduling model forreduced CPU energyrdquo in Proceedings of the 36th IEEE AnnualSymposium on Foundations of Computer Science pp 374ndash382Milwaukee Wis USA October 1995
[30] Y Shin and K Choi ldquoPower conscious fixed priority schedulingfor hard real-time systemsrdquo in Proceedings of the 36th AnnualDesign Automation Conference (DAC rsquo99) pp 134ndash139 NewOrleans La USA June 1999
[31] I Hong D Kirovski G QuM Potkonjak andM B SrivastavaldquoPower optimization of variable-voltage core-based systemsrdquoIEEE Transactions on Computer-Aided Design of IntegratedCircuits and Systems vol 18 no 12 pp 1702ndash1714 1999
[32] P Pillai and K G Shin ldquoReal-time dynamic voltage scaling forlow-power embedded operating systemsrdquo in Proceedings of the
The Scientific World Journal 13
ACM Symposium on Operating Systems Principles pp 89ndash102October 2001
[33] H Aydin R Melhem D Mosse and P Mejia-Alvarez ldquoPower-aware scheduling for periodic real-time tasksrdquo IEEE Transac-tions on Computers vol 53 no 5 pp 584ndash600 2004
[34] J Zhuo and C Chakrabarti ldquoAn efficient dynamic task schedul-ing algorithm for battery powered DVS systemsrdquo in Proceedingsof the Asian South Pacific Design Automation Conference pp846ndash849 January 2005
[35] R Jejurikar and R Gupta ldquoDynamic slack reclamation withprocrastination scheduling in real-time embedded systemsrdquo inProceedings of the 42nd Design Automation Conference (DACrsquo05) pp 111ndash116 June 2005
[36] J Luo andN K Jha ldquoPower-profile driven variable voltage scal-ing for heterogeneous distributed real-time embedded systemsrdquoin Proceedings of the International Conference on VLSI Designpp 369ndash375 January 2003
[37] M T Schmitz and B M Al-Hashimi ldquoConsidering powervariations of DVS processing elements for energy minimisationin distributed systemsrdquo in Proceedings of the International Sym-posium on System Synthesis (ISSS rsquo01) pp 250ndash255 MontrealCanada October 2001
[38] Y Zhang X Hu and D Z Chen ldquoTask scheduling and voltageselection for energy minimizationrdquo in Proceedings of the 39thAnnual Design Automation Conference (DAC rsquo02) pp 183ndash188June 2002
[39] D Zhu R Melhem and B R Childers ldquoScheduling withdynamic voltagespeed adjustment using slack reclamationin multiprocessor real-time systemsrdquo IEEE Transactions onParallel and Distributed Systems vol 14 no 7 pp 686ndash7002003
[40] D Zhu D Mosse and R Melhem ldquoPower-aware schedulingfor ANDOR graphs in real-time systemsrdquo IEEE Transactionson Parallel and Distributed Systems vol 15 no 9 pp 849ndash8642004
[41] J Kang and S Ranka ldquoDVS based energy minimizationalgorithm for parallel machinesrdquo in Proceedings of the IEEEInternational Parallel and Distributed Processing Symposium(IPDPS rsquo08) pp 1ndash12 Miami Fla USA April 2008
[42] B Rountree D K Lowenthal S Funk V W Freeh B R deSupinski and M Schulz ldquoBounding energy consumption inlarge-scale MPI programsrdquo in Proceedings of the ACMIEEEConference on Supercomputing (SC rsquo07) ACM November 2007
[43] L Wang G von Laszewski J Dayal and F Wang ldquoTowardsenergy aware scheduling for precedence constrained paralleltasks in a cluster with DVFSrdquo in Proceedings of the 10thIEEEACM International Symposium on Cluster Cloud andGrid Computing (CCGrid rsquo10) pp 368ndash377 May 2010
[44] Y C Lee and A Y Zomaya ldquoEnergy conscious schedulingfor distributed computing systems under different operatingconditionsrdquo IEEE Transactions on Parallel and DistributedSystems vol 22 no 8 pp 1374ndash1381 2011
[45] M Mezmaz Y C Lee N Melab E Talbi and A Y ZomayaldquoA bi-objective hybrid genetic algorithm to minimize energyconsumption and makespan for precedence-constrained appli-cations using dynamic voltage scalingrdquo in Proceedings of theIEEE Congress on Evolutionary Computation (CEC rsquo10) pp 1ndash8Barcelona Spain July 2010
[46] M Mezmaz N Melab Y Kessaci et al ldquoA parallel bi-objectivehybrid metaheuristic for energy-aware scheduling for cloudcomputing systemsrdquo Journal of Parallel and Distributed Com-puting vol 71 no 11 pp 1497ndash1508 2011
[47] J Kennedy and R Eberhart ldquoParticle swarm optimizationrdquoin Proceedings of the IEEE International Conference on NeuralNetworks vol 4 pp 1942ndash1948 Perth Australia December1995
[48] Y Zhao M Wilde I Foster et al ldquoGrid middleware servicesfor virtual data discovery composition and integrationrdquo inProceedings of the 2nd Workshop on Middleware for GridComputing (MGC rsquo04) pp 57ndash62 Ontario Canada October2004
[49] A OrsquoBrien S Newhouse and J Darlington ldquoMapping ofscientific workflow within the e-protein project to distributedresourcesrdquo inProceedings of theUK e-Science All HandsMeetingNottingham UK September 2004
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
4 The Scientific World Journal
Table 1 Voltage Scaling Levels and relative speeds of processors
VSLP1 TURIONMT-34 P2 OMAP P3 PENTIUMM
Volt Freq R Speed Volt Freq R Speed Volt Freq R Speed(V) (Ghz) () (V) (Ghz) () (V) (Ghz) ()
1 12 18 100 135 06 100 148 14 1002 115 16 8888 127 055 9166 144 12 96763 11 14 7777 12 05 8333 131 1 88144 105 12 6666 1 025 4166 118 08 79515 1 1 55555 09 013 2083 096 06 64426 09 08 4444Cost ec = 114 ec = 057 ec = 076
Table 2 Task execution times and priorities
Task (119879119894) 119901
11199012
1199013
120596119894
Priority (Pr(119879119894))
1 14 16 9 13 1080002 13 19 18 1667 770003 11 13 19 1433 800004 13 8 17 1267 800005 12 13 10 1167 690006 13 16 9 1267 633337 07 15 11 11 426678 5 11 14 10 356679 18 12 20 1667 4433310 21 7 16 1467 14667
119875119886V(119879119894 119901119895) is defined as the earliest time when processor 119901
119895
will be available to begin executing task 119879119894 Hence
119864119878119879 (119879119894 119901119895) =
119900 if 119879119894= 119879entry
max 119875119886V(119879119894 119901119895) 120572
(4)
where 120572 = max119879119895 isin pred (119879119894 )(119864119865119879(119879119895 119901119896) + tr
119895119894
119864119865119879 (119879119894 119901119895) = 119908
119894119895+ 119864119878119879 (119879
119894 119901119895) (5)
Note that the Actual Start Time and Actual Finish Timeof a task 119879
119894on a processor 119901
119895 denoted as AST (119879
119894 119901119895) and
AFT (119879119894 119901119895) can be different from its earliest start EST (119879
119894 119901119895)
and finish EFT (119879119894 119901119895) times if the actual finish time of
another task 119879119896scheduled on the same processor is later than
its EST (119879119896 119901119895) [44]
Figure 1 depicts a workflow application with 10 tasks andthe Table 2 provides its details (given in [17]) The valuespresented in the last column of the table represent the priorityof the tasks The priority of task 119879
119894represented by Pr(119879
119894) is
computed recursively by traversing the DAG upward startingfrom the exit task 119879exit as follows (6)
Pr (119879119894) =
119908119879exit if 119879
119894= 119879exit
119908119894+ 120573 otherwise
(6)
where120573 = max119879119895 isin succ (119879119894)119879119903119894119895 + Pr(119879
119895)
1
2 5 6
7 8
10
3 4
9
18 12 9 1114
19 16 2723 23 1513
1311 17
Figure 1 An example of workflow (given in [17]) with the tasknumbers 119879
119894inside nodes and values of 119889
119894119895function next to the
corresponding edges
33 QoS Parameter Models
331 Energy Model Among the main system-level energy-saving techniques Dynamic Voltage Scaling (DVS) operateson a simple principle decreases the supply voltage (and so theclock frequency) to the CPU so as to consume less power
In this work we use a model of energy derived from thepower consumption model in digital complementary metal-oxide semiconductor (CMOS) logic circuits [44] Under thedynamic power model the processor power is dominated bythe dynamic power which is given by
119875dynamic = 119860119862efV2119891 (7)
where119860 is the number of switches per clock cycle119862ef denotesthe effective charged capacitance V is the supply voltage and119891 denotes the operational frequency Equation (7) shows thatthe supply voltage is the dominant factor hence its reductionwould be most influential to lower power consumption
The energy consumption of the execution of a workflowapplication used in this paper is defined as
119864119888=
119899
sum119894=1
119860119862efV2
119894119891119894119908119903lowast
119894
=
119899
sum119894=1
120574V2119894119891119894119908119903lowast
119894
(8)
The Scientific World Journal 5
where 120574 = 119860119862ef is assumed constant for a given machineV119894 119891119894are the voltage supply and frequency of the processor
on which task 119879119894is executed respectively and 119908119903lowast
119894is the real
completion time of task 119879119894on the scheduled processor In the
idle time the processor turns into sleep mode and thus thevoltage supply and relative frequency are at the lowest levelSo the energy consumption during idle periods of processorsis defined as
119864119894=
119898
sum119895=1
sumidle119895119896isin IDLE119895
120574V2min 119895 119891min 119895 119871119895119896 (9)
where IDLE119895
is the set of idling slots on machine119901119895 Vmin 119895(119891min 119895) is the lowest supply voltage (frequency)
on 119901119895 and 119871
119895119896is the amount of idling time for idle
119895119896
Then the total energy 119864total utilized by the cloud system forcompletion of the workflow application can be defined asfollows
119864total = 119864119888+119864119894 (10)
332 Time Model The completion time 119879total is themakespan from the user submitting a workflow untilreceiving the results It is defined as the actual finish timeof the exit task It is calculated in equation 11 and consistsof the workflow execution time and network transmissiontime The execution time depends on both the workloadand system performance The network transmission timedepends on both the network latency and the input data size
119879total = maxAFT (119879exit) (11)
333 Cost Model As result of the marketization characteris-tic of current services most cloud providers have set a pricefor their services They have fixed the price for transferringbasic data unit (eg per MB) between two services and theprice for processing basic time units (eg per hour)The cost119862total of running a cloud workflow is defined in Formula (12)It consists of processing cost 119862ex and data transfers cost 119862tr
119862total = 119862ex + 119862tr (12)
The processing cost for a given task119879119894depends on the real
completion time of 119879119894on the scheduled processor (119908119903lowast
119894) and
the hourly price of this processor (ec119894) Thus 119862ex is given by
119862ex =119899
sum119894=1
119908119903lowast
119894lowast ec119894 (13)
The data transfer cost (119862tr) is described as follows
119862tr =119899
sum119894=1
119899
sum119895=1
119889119894119895lowast trc119894119895 (14)
where 119889119894119895characterizes the output file size from task119879
119894to task
119879119895 and tr 119888
119894119895represents the cost of communication from the
processor where 119879119894is mapped to another processor where 119879
119895
is mappedThe cost of communication is added to the overallcost only when two tasks have data dependency betweenthem (ie 119889
119894119895gt 0) For two or more tasks running on the
same processor the transfer cost is neglected
334 Scheduling Model Given (1) A cloud provider thatoffers a set 119875 of 119898 heterogeneous processors and (2) auser workflow application composed of a set 119879 of 119899 tasksthat have to be executed on these processors The workflowscheduling problem is to construct a mapping 119872 of tasksto processors (without violating precedence constraints) thatminimizes the following conflicting objectives makespancost and energy consumption as low as possible Thereforethe workflow scheduling problem can be formulated as amathematical optimization problem
MakespanMinimize 119879total (119872)
CostMinimize 119862total (119872)
EnergyMinimize119864total (119872)
(15)
4 Workflow Scheduling Based on DiscreteParticle Swarm Optimization
This section starts with a brief overview on multi-objectivecombinatorial optimization and Particle swarm optimizationalgorithm Afterwards our new Multi-Objective DiscreteSwarmOptimization combined with DVFS technique will bepresented
41 Multi-Objective Optimization A Multi-objective Opti-mization Problem (MOP) with 119898 decision variables and 119899objectives can be formally defined as
Min (119910 = 119891 (119909) = [1198911(119909) 119891
119899(119909)]) (16)
where 119909 = (1199091 119909
119898) isin 119883 is an 119898-dimensional decision
vector 119883 is the search space 119910 = (1199101 119910
119899) isin 119884 is the
objective vector and Y the objective-spaceIn general MOP there is no single optimal solution
with regards to all objectives This is also the case forthe multi-objective optimization problem addressed in thispaper As given in (15) there are three conflicting objectivesminimizing execution time minimizing execution cost andminimizing energy consumption In such problems thedesired solution is considered to be the set of potentialsolutions which are all optimal in some objectives This set isknown as the Pareto optimal setWe provide some definitionsof the Pareto concepts used in MOP as follows (withoutloss of generality we suppose that the objectives are to beminimized)
(i) Pareto dominance For two decision vectors 1199091 and 1199092dominance (denoted by ≺) is defined as follows
1199091≺ 1199092lArrrArr forall119894 119891
119894(1199091) le 119891119894(1199092) and exist119895119891
119895(1199091) lt 119891119895(1199092)
(17)
The decision vector 1199091is said to dominate 1199092 if andonly if 1199091 is as good as 1199092 considering all objectivesand 1199091 is strictly better than 1199092 in at least oneobjective
6 The Scientific World Journal
(ii) Pareto optimally A decision vector 1199091 is said to bePareto optimal if and only if
∄1199092isin 119883 119909
2≺ 1199091 (18)
(iii) Pareto optimal set The Pareto optimal set 119875119878is the set
of all Pareto optimal decision vectors
119875119878= 1199091isin 119883 | ∄119909
2isin 119883 119909
2≺ 1199091 (19)
(iv) Pareto optimal frontThePareto optimal front119875119865is the
image of the Pareto optimal set in the objective space
119875119865= 119891 (119909) = (119891
1(119909) 119891
119899(119909)) | 119909 isin 119875
119904 (20)
1199091 is said to be non-dominated regarding a given set
if 1199091 is not dominated by any decision vectors in theset
Thepareto optimal decision vector cannot be enhanced inany objective without causing degradation in at least anotherobjective A decision vector is said to be Pareto optimal whenit is not dominated in the whole search space
42 Particle Swarm Optimisation
421 The Standard PSO PSO is a population-based stochas-tic optimization technique developed by Kennedy and Eber-hart in 1995 [47] It is inspired by the social behavior of insectcolonies bird flocks fish schools and other animal societies Itis also related to evolutionary computation PSOhas attractedsignificant attention from many researchers due to both itssimplicity of use and optimization via social behavior In factPSO has good performance requires low computational costIt is effective and easy to implement as it uses numericalencoding
A particle in PSO is analogous to a fish or bird movingin the 119863-dimensional search space All particles have fitnessvalues indicating their performances which are problemspecific and velocities which direct the flight of particlesEach particle position at any given time is influenced by bothits best position called 119901119861119890119904119905 and the position of the bestparticle in a problem space referred to as 119892119861119890119904119905 Thereforeparticles tend to fly towards a better search area duringthe search process A particle status on the search spaceis characterized by two elements namely its velocity andposition which are updated in every generation as follows
119881119896+1
119894= 120596119881
119896
119894+ 11988811199031(119901119861119890119904119905
119894minus 119883119896
119894)
+ 11988821199032(119892119861119890119904119905 minus 119883
119896
119894)
119883119896+1
119894= 119883119896
119894+ 119881119896+1
119894
(21)
where 119881119896+1119894
is the velocity of particle 119894 at iteration 119896 + 1119881119896119894is the velocity of particle 119894 at iteration 119896 120596 is the inertia
weight 1198881and 1198882are the acceleration coefficients (cognitive
and social coefficients) 1199031and 1199032are the random numbers
between 0 and 1119883119896119894is the current position of particle 119894 at the
119896th iteration 119901119861119890119904119905119894is the best previous position of the 119894th
particle 119892119861119890119904119905 is the position of best particle in the swarmand119883119896+1
119894is the position of 119894th particle at 119896 + 1 iteration
The procedure for standard PSO is as follows
(1) Initialize a population of particles with random posi-tions and velocities in the search space
(2) Evaluate the objective values of all particles set 119901119861119890119904119905of each particle equal to its current position and set119892119861119890119904119905 equal to the position of the best initial particle
(3) Update the velocity and the position of each particleaccording to (21)
(4) Map the position of each particle in the solution spaceand evaluate its fitness value according to the desiredoptimization fitness function
(5) For each particle compare its current objective valuewith its 119901119861119890119904119905 value If the current value is better thenupdate 119901Best with the current position and objectivevalue
(6) Determine the best particle of the current whole pop-ulation with the best objective value If the objectivevalue is better than that of 119892119861119890119904119905 then update 119892Bestwith the current best particle
(7) If the stopping criterion ismet then output 119892119861119890119904119905 andits objective value otherwise go to Step (2)
The original design of PSO is appropriate for findingsolutions to continuous optimization problems However asthe workflow scheduling discussed in this paper is both adiscrete andmulti objective problem in nature we propose aneffective approach to address this problem by using a discreteversion of the Multi-Objective PSO (MODPSO) combinedwith DVFS techniqueThe key issues of our approach consistof (1) defining the position and velocity of the particlesbased on the features of discrete variables of the workflowscheduling (2) solving the multi-objective aspect of theproblem by modifying PSO so as to generate a set of non-dominated solutions satisfying the different objectives underconsideration instead of one solution
422 Handling Workflow Scheduling Using DVFS-MODPSOIn general a workflow scheduling can be defined by a setof triplets 119872 = [⟨119879
119894 119875119895 119871119896⟩](119894 isin [1 119899] 119895 isin [1119898] 119896 isin
[1 119871(119895)]) 119899 is the number of workflow tasks to be scheduled119898 is the number of processors available in the cloud environ-ment and 119871(119895) is the number of operating points (VSLs) of the119895th processor
For the sake of clarity the variables and rules of DVFS-MODPSO for solving workflow scheduling can be depictedas follows
The position of a particle represents a feasible solutionto the scheduling problem It consists of a set of⟨task(119879
119894) service (119875
119895) VSl (119871
119896)⟩ triplets Each triplet
means that a task (119879119894) is assigned to a processor (119875
119895)
with a voltage scaling level (119871119896) It also indicates
The Scientific World Journal 7
that the position satisfies the precedence constraintbetween tasks
The process of generating a newposition for a selectedparticle in the swarm is depicted in the followingformulas
119881119896+1
119894= 119881119896
119894oplus ((119877
1otimes (119901119861119890119904119905
119894⊖ 119883119896
119894))
oplus (1198772otimes (119892119861119890119904119905 ⊖ 119883
119896
119894)))
(22)
119883119896+1
119894= 119883119896
119894oplus 119881119896+1
119894 (23)
The operator definitions are as follows
(i) The substract operator (⊖) the difference betweentwo particle positions designated as 1199091 and 1199092 isdefined as a set of triplets in which each triplet⟨task serviceVSL⟩ shows whether the contents ofthe corresponding elements in 1199091 are different fromthose of 1199092 or not If so that triplet gets its values(service and VSL) from the position that has thelowest value of theVSL For those triplets that have thesame content in 1199091 and 1199092 their corresponding VSLsare decreased (Note that the scaling of VSL makesfluctuations on the energy makespan and cost)
(ii) The multiply operator (otimes) the multiplication betweennumber and velocity is defined as a set of tripletswhere a threshold 120572 isin [0 1] is defined arandom number 119903 is generated for each triplet⟨task serviceVSL⟩ compare 119903 and 120572 when 119903 ge 120572decrease the triplet VSL otherwise increase it Thisoperator adds the exploration ability to the algorithm
(iii) The add operator (oplus) the addition of two positions isdefined as the reservation of the dominated one
The Pseudocode 1 outlines the general steps of the DVFS-MODPSO algorithm
The algorithm begins by initializing the positions andvelocities of particles To obtain the position of a particle theVSL (voltage and frequency) of each resource is randomlyinitialized firstly then the HEFT algorithm is applied togenerate a feasible and efficient solution minimizing themakespan The process is repeated several times to initializethe positions of all particles of the swarm Initially thevelocity119881 and the best position for each particle 119901119861119890119904119905 areattributed as the particle itself The algorithm maintainsan external archive EA to store non-dominated particlesfound after the evaluation process (based on the paretodominance using the objectives mentioned in 33) After allthese initializations in the main loop of the algorithm thenew velocity and position of each particle are calculatedrespectively after selecting the best overall position in theexternal archive and eventually perform a mutation then theparticle is evaluated and its corresponding 119901119861119890119904119905 is updatedThe external archive is updated after every iteration Oncethe termination condition is reached the external archivecontaining the Pareto front is returned as the result
5 Experimental Evaluations
In this section we describe the overall setup of our experi-mentation effort and the results we have obtained from it tovalidate the new proposed approach
In our experiments we simulate a synthetic workflowapplication described in [17] (see Figure 1) and two commonworkflow structures parallel and hybrid structuresWe chosetwo well known real world applications namely a neuro-science workflow [48] for our parallel application and aprotein annotation workflow [49] developed by London e-Science Centre for our hybrid workflow application Figure 2shows their simplified representationsThe number indicatedin the parentheses next to each node represents the length ofthe task (the number of instructions to execute) in Millionsof Instructions (MI) The input and output files of each taskrange from 10MB to 1024MB
We have performed experiments on 3 4 5 6 and 12resources whose characteristics are shown in Table 1 Wechoose the pricing model associated with Amazon EC2(httpawsamazoncomec2) for the processing costs of thedifferent classes of resources and the pricing model given byAmazon CloudFront (httpawsamazoncomcloudfront)for the costs of transferring data unit between resources
As for DVFS-MODPSO the parameter settings areSwarm size 119878119873119906119898 = 50 particles and the maximum numberof iterations 119866 = 100
We evaluated the performance of our proposed DVFS-MODPSO on all the workflow instances described previouslyDue to the lack of works considering both the heteroge-neous configuration of our cloud and the three QoS metrics(makespan cost energy) at the same time we comparedour results with those of the HEFT heuristic we haveimplemented HEFT is one of themost widely used heuristicsfor DAGs in distributed heterogeneous computing systemswhich optimize the makespan
Figures 3 4 and 5 show sample Pareto fronts obtainedwith the DVFS-MODPSO and the solution computed byHEFT for the synthetic workflow the hybrid workflow andthe parallel workflow respectively
As can be seen from these figures unlike the HEFTalgorithm which gives one solution our proposed approachprovides a set of non dominated solutions These resultsillustrate the basic multi-objective definitions provided inSection 41
Now for comparing these two approaches and in orderto analyze the effectiveness of our proposition in terms ofthe values of makespan cost and energy consumption wehave been inspired by the methodology described in [43]where we compare the solution provided by HEFT to onlyone solution of the Pareto front set computed by our proposedmulti-objective DVFS-MODPSO
The evaluation process follows the next steps For eachworkflow instance we perform a first resolution usingHEFTin order to get one solution 119878 After a second resolution iscomputed using our proposed approach to obtain a set EA ofPareto solutions Next we select one solution 1198781015840 from the setEA of Pareto solutions This solution is the closest one to 119878 inthe sense of Euclidean distance Finally a comparison is donebetween 119878 and 1198781015840
8 The Scientific World Journal
Swarm initialization with HEFT(1) For i = 1 to SNum (SNum is the size of particle swarm)
(a) For j = 1 to m (m the number of processors)(i) Randomly initialize VSL(j) (randomly choose the voltagefrequency of processor from the set of its operating points)
(b) Initialize 119878[119894] with HEFT heuristic (S is the swarm of particles)(c) Initialize the velocity V of each particle
119881 [119894] = 0 (119878 [119894])
(d) Initialize the Personal Best (pBest) of each particle119901119861119890119904119905[119894] = 119878[119894]
(e) Evaluate objectives of each particleEvaluate S[i]
(f) Initialize the Global Best particle (gBest) with the best one among the SNum particlesg119861119890119904119905 = Best particle found in 119878
(2) End For(3) Add the nondominated solutions found in S into EA (EA is the External Archive storing the pareto front)(4) Initialize the iteration number 119905 = 0(5) Repeat until 119905 gt 119866 (119866 is the maximum number of iterations)
(a) For 119894 = 1 to SNum (swarm size)(i) Randomly select the global best particle for S from the External Archive EA and store its position in gBest(ii) Calculate the new velocity 119881[119894] according to (22)(iii) Compute the new position of 119878[119894] according to (23)(iv) If (119905 lt 119866 lowast PMUT ) then (PMUT is the probability of mutation)
Perform mutation on 119878[119894](v) Evaluate S[i]
(b) End For(c) Update the personal best solution of each particle 119878[119894]
if 119878 [119894] ≼ 119901119861119890119904119905119904 [119894] or (119878 [119894] sim 119901119861119890119904119905119904 [119894])Then 119901119861119890119904119905119904 [119894] = 119878 [119894]
(d) Update the External Archive EA(i) Add all new non-dominated solution in S into EAif 119878 [119894] 997410 119909 forall119909 isin 119864119860Then 119864119860 = 119864119860 cup 119878 [119894](ii) Remove all particles dominated by 119878[119894] in EA
119864119860 = 119864119860 minus 119910 isin 119864119860 | 119910 lt 119878[119894]
(iii) If the archive is full then randomly select the article to be replaced from EA(6) Return the pareto front (the set of non-dominated solutions from S and EA)
Pseudocode 1 DVFS-MODPSO based workflow scheduling
The different results are displayed on Figures 6 7 8 and 9and on Tables 3 and 4 They show the improvement achievedby our proposed approach according to the structure of theworkflow applications and the number of processors
Table 3 and Figure 6 show that our proposed approachimproves on average the result obtained by HEFT for thesynthetic workflow instance The makespan is reduced by005 the cost is reduced by 993 and the energy con-sumption is reduced by 2640 Figure 6 shows detailedimprovements of the proposed approach according to thenumber of processors
Table 4 and Figure 7 illustrate the gain obtained by ourapproach according to the kind of real world applicationsTable 4 also shows the detailed improvements when scalingthe number of processors As can be seen the QoS metricsare improved over HEFT in average by 095 for themakespan 108 for cost and 812 for the energy con-sumption when using the hybrid workflow application andaverage improvements of 295 formakespan 2215 for costand 209 for energy consumption when using the parallelworkflow
These evaluations confirm the results obtained from thesynthetic workflow This means that by using our DVFS-MODPSO we are able to improve not only the cost and energybut also themakespan onwhichHEFT algorithm is supposedto provide good results These results remain true for all kindof workflow applications
In our experiments we limited the number of resourcesto 12 as we found that for these kinds of workflowapplications some resources are not used at all Figure 8shows the resource loads when scheduling the syntheticworkflow on 12 resources Furthermore even though thenumber of resources increased the total cost and total energyconsumption could not always decrease as illustrated inFigure 9 From this figure we can see that the QoS metrics(1) firstly decreases as the number of resources increasesuntil achieving the number 6 This can be explained by thefact that when increasing the number of resources thereare fewer tasks executed in a resource therefore tasks canextend their execution times and the resource have more ofchances to scale down their voltages and frequencies whichcan be very effective in reducing total energy consumption
The Scientific World Journal 9
1 2 3 4
5
6 7
8 9 10
11
12
14
13
15
(300000) (600000) (600000) (900000)
(300000)
(150000) (150000)
(300000) (300000) (300000)
(600000)
(300000)
(150000)
(300000)
(600000)
SignalP COILS2 SEG PROSITE
TMHMM
ProsperoHMMer
PSI-BLASTBLAST
IMPALA
PSI-PRED
3D-PSSMSummary
Genomesummary
SCOP
(a)
1 3 5 7
9
11 1210
(400000) (400000) (400000) (400000)
(300000)
(500000) (500000) (500000)
Softmea
2 4 6 8(600000) (600000) (600000) (600000)
Reslice
14 1513
(600000) (600000) (600000)
Convert
Reslice Reslice Reslice
Slicer Slicer
Convert Convert
Slicer
Align wapAlign wapAlign wapAlign wap
(b)
Figure 2 Parallel and Hybrid workflows
10 The Scientific World Journal
75 80 85 90 95 100 8090
100110
120130
140
2060
100140180220
Ener
gy
DVFS-MODPSOHEFT
MakespanCost
Figure 3 Obtained non-dominated solutions for the syntheticworkflow
Ener
gy
DVFS-MODPSOHEFT
MakespanCost
200 400 600 800 100012001400160018002000 400800
12001600
2000
400200
600800
100012001400160018002000
Figure 4 Obtained non-dominated solutions for the hybrid work-flow
(2) After achieving 6 resources the QoS metrics begin torise the reason for this is that time executions are dominatedby interprocessors communications hence decreasing theopportunities for scaling down voltages and frequencies ofresources Consequently the threshold numbers of resourcesthat minimized the QoS metrics could be obtained
6 Conclusion
In this paper we propose a new algorithm called DVFS-Multi-Objective Discrete Particle Swarm Optimization(DVFS-MODPSO) for workflow scheduling in distributedenvironments such as cloud computing infrastructures DV-FS-MODPSO simultaneously optimizes several conflictingobjectives namely the makespan cost and energy in adiscrete space It produces a set of non-dominated solutionsthus providing more flexibility for users to assess theirpreferences and select a schedule that meets their QoS
Ener
gy
DVFS-MODPSOHEFT
Makespan Cost
400200600800
10001200140016001800
240020002200
450 500 550 600 650 700 7501200
14001600
18002000
Figure 5 Obtained non-dominated solutions for the parallel work-flow
0
10
20
30
40
50
3 4 5 6 12
MakespanCostEnergy
Figure 6 Gain over HEFT () according to the number ofprocessors
requirements Our approach exploits the heterogeneityand the marketization of cloud resources in order to findsolutions that optimize makespan and cost Furthermore ituses the Dynamic Voltage and Frequency Scaling (DVFS)technique to reduce energy consumption
We have evaluated our algorithm by simulating it withboth synthetic and real world scientific workflow applicationshaving different structures The results show that the pro-posedDVFS-MODPSO is able to produce a set of good Paretooptimal solutions The results also show that our approachprovides significant improvement not only in terms of the
The Scientific World Journal 11
0
5
10
15
20
25
HybridParallel
Makespan Cost Energy
Figure 7 Improvement according to real world applications
R17
R20
R37
R446
R50
R60
R733
R80
R90
R107
R110
R120
Figure 8 Resource loads for synthetic workflow
Table 3 Improvement for synthetic workflow
Num of proc Gain over Heft ()Makespan Cost Energy
3 0 2126 40084 022 464 15355 0 756 17266 003 652 267512 0 969 3255Average 005 993 2640
cost and the energy consumption but also in term of themakespan for which HEFT algorithm is supposed to givebetter results
Multi-objective optimization in cloud workflow schedul-ing is not a mature domain Most of the existing worksattempt to minimize either the makespan or cost Howeverwe plan to consider other objectives such as reliabilitysecurity in addition to the energy consumption We also
0
50
100
150
200
250
3 4 5 6 12
MakespanCostEnergy
Figure 9 QoS metrics evolutions according to resource numbersfor the synthetic workflow
Table 4 Improvement according to real world applications
Appli Num of proc Gain over Heft ()Makespan Cost Energy
Hybrid
3 405 1155 04024 028 0566 05075 000 1347 11246 000 0117 126512 043 2214 0761
Average 095 1080 0812
Parallel
3 235 5640 25524 1147 2094 32715 000 2142 11666 093 1170 315412 000 031 306
Average 295 2215 209
intend to apply our algorithm in a real world cloud orintegrate it in existing cloud toolkits such as Cloudbus forcomparing with other algorithms
References
[1] D Thain T Tannenbaum and M Livny Condor and the GridGrid Computing Making the Global Infrastructure a RealityJohn Wiley amp Sons Hoboken NJ USA 2003
[2] R Buyya and S Venugopal ldquoThe gridbus toolkit for serviceoriented grid and utility computing an overview and statusreportrdquo in Proceedings of the 1st IEEE International Workshopon Grid Economics and Business Models (GECON rsquo04) pp 19ndash36 IEEE CS Seoul Republic of Korea April 2004
[3] S McGough L Young A Afzal S Newhouse and J Darling-ton ldquoWorkflow enactment in ICENIrdquo in Proceedings of the UK
12 The Scientific World Journal
e-Science All Hands Meeting pp 894ndash900 IOP NottinghamUK September 2004
[4] E Deelman J Blythe Y Gil et al ldquoPegasus mapping scientificworkflows onto the gridrdquo in Grid Computing 2nd EuropeanAcrossGrids Conference AxGrids 2004 Nicosia Cyprus January28ndash30 2004 vol 3165 of Lecture Notes in Computer Science pp11ndash20 Springer New York NY USA 2004
[5] R Buyya S Pandey and C Vecchiola ldquoCloudbus toolkitfor market-oriented cloud computingrdquo in Cloud ComputingProceedings of the 1st International Conference CloudCom 2009Beijing China December 1ndash4 2009 vol 5931 of Lecture Notesin Computer Science pp 24ndash44 Springer New York NY USA2009
[6] Y Yang K Liu J Chen X Liu D Yuan and H Jin ldquoAnalgorithm in SwinDeW-C for scheduling transaction-intensivecost-constrained cloud workflowsrdquo in Proceedings of the 4thIEEE International Conference on eScience (eScience rsquo08) pp374ndash375 Indianapolis Ind USA December 2008
[7] L Ramakrishnan C Koelbel Y Kee et al ldquoVGrADS enablinge-Scienceworkflows on grids and cloudswith fault tolerancerdquo inProceedings of the Conference on High Performance ComputingNetworking Storage and Analysis (SC rsquo09) November 2009
[8] M L Pinedo Scheduling Theory Algorithms and SystemsSpringer Berlin 2008
[9] S Pandey L Wu S M Guru and R Buyya ldquoA particle swarmoptimization-based heuristic for scheduling workflow applica-tions in cloud computing environmentsrdquo in Proceedings of the24th IEEE International Conference on Advanced InformationNetworking and Applications (AINA rsquo10) pp 400ndash407 PerthAustralia April 2010
[10] S Abrishami and M Naghibzadeh ldquoDeadline-constrainedworkflow scheduling in software as a service cloudrdquo ScientiaIranica vol 19 no 3 pp 680ndash689 2012
[11] S Selvarani and G S Sadhasivam ldquoImproved cost-based algo-rithm for task scheduling in cloud computingrdquo in Proceedings ofthe IEEE International Conference on Computational Intelligenceand Computing Research (ICCIC rsquo10) pp 1ndash5 CoimbatoreIndia December 2010
[12] A Verma and S Kaushal ldquoDeadline and budget distributionbased cost-time optimization workflow scheduling algorithmfor cloudrdquo inProceedings of the IJCAon International Conferenceon Recent Advances and Future Trends in Information Technol-ogy (iRAFIT rsquo12) iRAFIT(7) pp 1ndash4 April 2012
[13] R C Correa A Ferreira and P Rebreyend ldquoIntegrating listheuristics into genetic algorithms for multiprocessor schedul-ingrdquo in Proceedings of the 8th IEEE Symposium on Parallel andDistributed Processing (SPDP rsquo96) pp 462ndash469 IEEEComputerSociety Washington DC USA October 1996
[14] E S H Hou N Ansari and H Ren ldquoGenetic algorithm formultiprocessor schedulingrdquo IEEE Transactions on Parallel andDistributed Systems vol 5 no 2 pp 113ndash120 1994
[15] M Grajcar ldquoGenetic list scheduling algorithm for schedulingand allocation on a loosely coupled heterogeneous multi-processor systemrdquo in Proceedings of the 36th Annual DesignAutomation Conference (DAC rsquo99) pp 280ndash285 ACM PressNew York NY USA June 1999
[16] R Sakellariou and H Zhao ldquoA hybrid heuristic for DAGscheduling on heterogeneous systemsrdquo in Proceedings of the13th Heterogeneous Computing Workshop pp 111ndash123 IEEEComputer Society Washington DC USA 2004
[17] H Topcuoglu S Hariri and M Y Wu ldquoPerformance-effectiveand low-complexity task scheduling for heterogeneous comput-ingrdquo IEEE Transactions on Parallel and Distributed Systems vol13 no 3 pp 260ndash274 2002
[18] J Yu and R Buyya ldquoWorkflow scheduling algorithms for gridcomputingrdquo in Metaheuristics for Scheduling in DistributedComputing Environments F Xhafa and A Abraham EdsSpringer Berlin Germany 2008
[19] WN Chen and J Zhang ldquoAn ant colony optimization approachto a grid workflow scheduling problem with various QoSrequirementsrdquo IEEE Transactions on Systems Man and Cyber-netics C vol 39 no 1 pp 29ndash43 2009
[20] X Liu J Chen Z Wu Z Ni D Yuan and Y Yang ldquoHandlingrecoverable temporal violations in scientific workflow systemsa workflow rescheduling based strategyrdquo in Proceedings of the10th IEEEACM International Symposium on Cluster Cloudand Grid Computing (CCGrid rsquo10) pp 534ndash537 MelbourneAustralia May 2010
[21] X Liu Y Yang Y Jiang and J Chen ldquoPreventing temporalviolations in scientific workflows where and howrdquo IEEE Trans-actions on Software Engineering vol 37 no 6 pp 805ndash825 2011
[22] J Yu and R Buyya ldquoScheduling scientific workflow applicationswith deadline and budget constraints using genetic algorithmsrdquoScientific Programming vol 14 no 3-4 pp 217ndash230 2006
[23] L Benini and G Micheli Dynamic Power Management DesignTechniques and CAD Tools Kluwer Academic New York NYUSA 1998
[24] S Irani S Shukla and R Gupta ldquoOnline strategies for dynamicpower management in system with multiple power-savingstatesrdquo ACM Transactions on Embedded Computing Systemsvol 2 no 3 pp 325ndash346 2003
[25] M T Schmitz B M Al-Hashimi and P Eles System-Level Design Techniques for Energy-Efficient Embedded SystemsSpringer New York NY USA 2005
[26] S U Khan and I Ahmad ldquoA cooperative game theoreticaltechnique for joint optimization of energy consumption andresponse time in computational gridsrdquo IEEE Transactions onParallel and Distributed Systems vol 20 no 3 pp 346ndash3602009
[27] Y Li Y Liu and D Qian ldquoA heuristic energy-aware schedulingalgorithm for heterogeneous clustersrdquo in Proceedings of the 15thInternational Conference on Parallel and Distributed Systems(ICPADS rsquo09) pp 407ndash413 Shenzhen China December 2009
[28] L K Goh B Veeravalli and S Viswanathan ldquoDesign of fast andefficient energy-aware gradient-based scheduling algorithmsheterogeneous embeddedmultiprocessor systemsrdquo IEEE Trans-actions on Parallel and Distributed Systems vol 20 no 1 pp 1ndash12 2009
[29] F Yao A Demers and S Shenker ldquoA scheduling model forreduced CPU energyrdquo in Proceedings of the 36th IEEE AnnualSymposium on Foundations of Computer Science pp 374ndash382Milwaukee Wis USA October 1995
[30] Y Shin and K Choi ldquoPower conscious fixed priority schedulingfor hard real-time systemsrdquo in Proceedings of the 36th AnnualDesign Automation Conference (DAC rsquo99) pp 134ndash139 NewOrleans La USA June 1999
[31] I Hong D Kirovski G QuM Potkonjak andM B SrivastavaldquoPower optimization of variable-voltage core-based systemsrdquoIEEE Transactions on Computer-Aided Design of IntegratedCircuits and Systems vol 18 no 12 pp 1702ndash1714 1999
[32] P Pillai and K G Shin ldquoReal-time dynamic voltage scaling forlow-power embedded operating systemsrdquo in Proceedings of the
The Scientific World Journal 13
ACM Symposium on Operating Systems Principles pp 89ndash102October 2001
[33] H Aydin R Melhem D Mosse and P Mejia-Alvarez ldquoPower-aware scheduling for periodic real-time tasksrdquo IEEE Transac-tions on Computers vol 53 no 5 pp 584ndash600 2004
[34] J Zhuo and C Chakrabarti ldquoAn efficient dynamic task schedul-ing algorithm for battery powered DVS systemsrdquo in Proceedingsof the Asian South Pacific Design Automation Conference pp846ndash849 January 2005
[35] R Jejurikar and R Gupta ldquoDynamic slack reclamation withprocrastination scheduling in real-time embedded systemsrdquo inProceedings of the 42nd Design Automation Conference (DACrsquo05) pp 111ndash116 June 2005
[36] J Luo andN K Jha ldquoPower-profile driven variable voltage scal-ing for heterogeneous distributed real-time embedded systemsrdquoin Proceedings of the International Conference on VLSI Designpp 369ndash375 January 2003
[37] M T Schmitz and B M Al-Hashimi ldquoConsidering powervariations of DVS processing elements for energy minimisationin distributed systemsrdquo in Proceedings of the International Sym-posium on System Synthesis (ISSS rsquo01) pp 250ndash255 MontrealCanada October 2001
[38] Y Zhang X Hu and D Z Chen ldquoTask scheduling and voltageselection for energy minimizationrdquo in Proceedings of the 39thAnnual Design Automation Conference (DAC rsquo02) pp 183ndash188June 2002
[39] D Zhu R Melhem and B R Childers ldquoScheduling withdynamic voltagespeed adjustment using slack reclamationin multiprocessor real-time systemsrdquo IEEE Transactions onParallel and Distributed Systems vol 14 no 7 pp 686ndash7002003
[40] D Zhu D Mosse and R Melhem ldquoPower-aware schedulingfor ANDOR graphs in real-time systemsrdquo IEEE Transactionson Parallel and Distributed Systems vol 15 no 9 pp 849ndash8642004
[41] J Kang and S Ranka ldquoDVS based energy minimizationalgorithm for parallel machinesrdquo in Proceedings of the IEEEInternational Parallel and Distributed Processing Symposium(IPDPS rsquo08) pp 1ndash12 Miami Fla USA April 2008
[42] B Rountree D K Lowenthal S Funk V W Freeh B R deSupinski and M Schulz ldquoBounding energy consumption inlarge-scale MPI programsrdquo in Proceedings of the ACMIEEEConference on Supercomputing (SC rsquo07) ACM November 2007
[43] L Wang G von Laszewski J Dayal and F Wang ldquoTowardsenergy aware scheduling for precedence constrained paralleltasks in a cluster with DVFSrdquo in Proceedings of the 10thIEEEACM International Symposium on Cluster Cloud andGrid Computing (CCGrid rsquo10) pp 368ndash377 May 2010
[44] Y C Lee and A Y Zomaya ldquoEnergy conscious schedulingfor distributed computing systems under different operatingconditionsrdquo IEEE Transactions on Parallel and DistributedSystems vol 22 no 8 pp 1374ndash1381 2011
[45] M Mezmaz Y C Lee N Melab E Talbi and A Y ZomayaldquoA bi-objective hybrid genetic algorithm to minimize energyconsumption and makespan for precedence-constrained appli-cations using dynamic voltage scalingrdquo in Proceedings of theIEEE Congress on Evolutionary Computation (CEC rsquo10) pp 1ndash8Barcelona Spain July 2010
[46] M Mezmaz N Melab Y Kessaci et al ldquoA parallel bi-objectivehybrid metaheuristic for energy-aware scheduling for cloudcomputing systemsrdquo Journal of Parallel and Distributed Com-puting vol 71 no 11 pp 1497ndash1508 2011
[47] J Kennedy and R Eberhart ldquoParticle swarm optimizationrdquoin Proceedings of the IEEE International Conference on NeuralNetworks vol 4 pp 1942ndash1948 Perth Australia December1995
[48] Y Zhao M Wilde I Foster et al ldquoGrid middleware servicesfor virtual data discovery composition and integrationrdquo inProceedings of the 2nd Workshop on Middleware for GridComputing (MGC rsquo04) pp 57ndash62 Ontario Canada October2004
[49] A OrsquoBrien S Newhouse and J Darlington ldquoMapping ofscientific workflow within the e-protein project to distributedresourcesrdquo inProceedings of theUK e-Science All HandsMeetingNottingham UK September 2004
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World Journal 5
where 120574 = 119860119862ef is assumed constant for a given machineV119894 119891119894are the voltage supply and frequency of the processor
on which task 119879119894is executed respectively and 119908119903lowast
119894is the real
completion time of task 119879119894on the scheduled processor In the
idle time the processor turns into sleep mode and thus thevoltage supply and relative frequency are at the lowest levelSo the energy consumption during idle periods of processorsis defined as
119864119894=
119898
sum119895=1
sumidle119895119896isin IDLE119895
120574V2min 119895 119891min 119895 119871119895119896 (9)
where IDLE119895
is the set of idling slots on machine119901119895 Vmin 119895(119891min 119895) is the lowest supply voltage (frequency)
on 119901119895 and 119871
119895119896is the amount of idling time for idle
119895119896
Then the total energy 119864total utilized by the cloud system forcompletion of the workflow application can be defined asfollows
119864total = 119864119888+119864119894 (10)
332 Time Model The completion time 119879total is themakespan from the user submitting a workflow untilreceiving the results It is defined as the actual finish timeof the exit task It is calculated in equation 11 and consistsof the workflow execution time and network transmissiontime The execution time depends on both the workloadand system performance The network transmission timedepends on both the network latency and the input data size
119879total = maxAFT (119879exit) (11)
333 Cost Model As result of the marketization characteris-tic of current services most cloud providers have set a pricefor their services They have fixed the price for transferringbasic data unit (eg per MB) between two services and theprice for processing basic time units (eg per hour)The cost119862total of running a cloud workflow is defined in Formula (12)It consists of processing cost 119862ex and data transfers cost 119862tr
119862total = 119862ex + 119862tr (12)
The processing cost for a given task119879119894depends on the real
completion time of 119879119894on the scheduled processor (119908119903lowast
119894) and
the hourly price of this processor (ec119894) Thus 119862ex is given by
119862ex =119899
sum119894=1
119908119903lowast
119894lowast ec119894 (13)
The data transfer cost (119862tr) is described as follows
119862tr =119899
sum119894=1
119899
sum119895=1
119889119894119895lowast trc119894119895 (14)
where 119889119894119895characterizes the output file size from task119879
119894to task
119879119895 and tr 119888
119894119895represents the cost of communication from the
processor where 119879119894is mapped to another processor where 119879
119895
is mappedThe cost of communication is added to the overallcost only when two tasks have data dependency betweenthem (ie 119889
119894119895gt 0) For two or more tasks running on the
same processor the transfer cost is neglected
334 Scheduling Model Given (1) A cloud provider thatoffers a set 119875 of 119898 heterogeneous processors and (2) auser workflow application composed of a set 119879 of 119899 tasksthat have to be executed on these processors The workflowscheduling problem is to construct a mapping 119872 of tasksto processors (without violating precedence constraints) thatminimizes the following conflicting objectives makespancost and energy consumption as low as possible Thereforethe workflow scheduling problem can be formulated as amathematical optimization problem
MakespanMinimize 119879total (119872)
CostMinimize 119862total (119872)
EnergyMinimize119864total (119872)
(15)
4 Workflow Scheduling Based on DiscreteParticle Swarm Optimization
This section starts with a brief overview on multi-objectivecombinatorial optimization and Particle swarm optimizationalgorithm Afterwards our new Multi-Objective DiscreteSwarmOptimization combined with DVFS technique will bepresented
41 Multi-Objective Optimization A Multi-objective Opti-mization Problem (MOP) with 119898 decision variables and 119899objectives can be formally defined as
Min (119910 = 119891 (119909) = [1198911(119909) 119891
119899(119909)]) (16)
where 119909 = (1199091 119909
119898) isin 119883 is an 119898-dimensional decision
vector 119883 is the search space 119910 = (1199101 119910
119899) isin 119884 is the
objective vector and Y the objective-spaceIn general MOP there is no single optimal solution
with regards to all objectives This is also the case forthe multi-objective optimization problem addressed in thispaper As given in (15) there are three conflicting objectivesminimizing execution time minimizing execution cost andminimizing energy consumption In such problems thedesired solution is considered to be the set of potentialsolutions which are all optimal in some objectives This set isknown as the Pareto optimal setWe provide some definitionsof the Pareto concepts used in MOP as follows (withoutloss of generality we suppose that the objectives are to beminimized)
(i) Pareto dominance For two decision vectors 1199091 and 1199092dominance (denoted by ≺) is defined as follows
1199091≺ 1199092lArrrArr forall119894 119891
119894(1199091) le 119891119894(1199092) and exist119895119891
119895(1199091) lt 119891119895(1199092)
(17)
The decision vector 1199091is said to dominate 1199092 if andonly if 1199091 is as good as 1199092 considering all objectivesand 1199091 is strictly better than 1199092 in at least oneobjective
6 The Scientific World Journal
(ii) Pareto optimally A decision vector 1199091 is said to bePareto optimal if and only if
∄1199092isin 119883 119909
2≺ 1199091 (18)
(iii) Pareto optimal set The Pareto optimal set 119875119878is the set
of all Pareto optimal decision vectors
119875119878= 1199091isin 119883 | ∄119909
2isin 119883 119909
2≺ 1199091 (19)
(iv) Pareto optimal frontThePareto optimal front119875119865is the
image of the Pareto optimal set in the objective space
119875119865= 119891 (119909) = (119891
1(119909) 119891
119899(119909)) | 119909 isin 119875
119904 (20)
1199091 is said to be non-dominated regarding a given set
if 1199091 is not dominated by any decision vectors in theset
Thepareto optimal decision vector cannot be enhanced inany objective without causing degradation in at least anotherobjective A decision vector is said to be Pareto optimal whenit is not dominated in the whole search space
42 Particle Swarm Optimisation
421 The Standard PSO PSO is a population-based stochas-tic optimization technique developed by Kennedy and Eber-hart in 1995 [47] It is inspired by the social behavior of insectcolonies bird flocks fish schools and other animal societies Itis also related to evolutionary computation PSOhas attractedsignificant attention from many researchers due to both itssimplicity of use and optimization via social behavior In factPSO has good performance requires low computational costIt is effective and easy to implement as it uses numericalencoding
A particle in PSO is analogous to a fish or bird movingin the 119863-dimensional search space All particles have fitnessvalues indicating their performances which are problemspecific and velocities which direct the flight of particlesEach particle position at any given time is influenced by bothits best position called 119901119861119890119904119905 and the position of the bestparticle in a problem space referred to as 119892119861119890119904119905 Thereforeparticles tend to fly towards a better search area duringthe search process A particle status on the search spaceis characterized by two elements namely its velocity andposition which are updated in every generation as follows
119881119896+1
119894= 120596119881
119896
119894+ 11988811199031(119901119861119890119904119905
119894minus 119883119896
119894)
+ 11988821199032(119892119861119890119904119905 minus 119883
119896
119894)
119883119896+1
119894= 119883119896
119894+ 119881119896+1
119894
(21)
where 119881119896+1119894
is the velocity of particle 119894 at iteration 119896 + 1119881119896119894is the velocity of particle 119894 at iteration 119896 120596 is the inertia
weight 1198881and 1198882are the acceleration coefficients (cognitive
and social coefficients) 1199031and 1199032are the random numbers
between 0 and 1119883119896119894is the current position of particle 119894 at the
119896th iteration 119901119861119890119904119905119894is the best previous position of the 119894th
particle 119892119861119890119904119905 is the position of best particle in the swarmand119883119896+1
119894is the position of 119894th particle at 119896 + 1 iteration
The procedure for standard PSO is as follows
(1) Initialize a population of particles with random posi-tions and velocities in the search space
(2) Evaluate the objective values of all particles set 119901119861119890119904119905of each particle equal to its current position and set119892119861119890119904119905 equal to the position of the best initial particle
(3) Update the velocity and the position of each particleaccording to (21)
(4) Map the position of each particle in the solution spaceand evaluate its fitness value according to the desiredoptimization fitness function
(5) For each particle compare its current objective valuewith its 119901119861119890119904119905 value If the current value is better thenupdate 119901Best with the current position and objectivevalue
(6) Determine the best particle of the current whole pop-ulation with the best objective value If the objectivevalue is better than that of 119892119861119890119904119905 then update 119892Bestwith the current best particle
(7) If the stopping criterion ismet then output 119892119861119890119904119905 andits objective value otherwise go to Step (2)
The original design of PSO is appropriate for findingsolutions to continuous optimization problems However asthe workflow scheduling discussed in this paper is both adiscrete andmulti objective problem in nature we propose aneffective approach to address this problem by using a discreteversion of the Multi-Objective PSO (MODPSO) combinedwith DVFS techniqueThe key issues of our approach consistof (1) defining the position and velocity of the particlesbased on the features of discrete variables of the workflowscheduling (2) solving the multi-objective aspect of theproblem by modifying PSO so as to generate a set of non-dominated solutions satisfying the different objectives underconsideration instead of one solution
422 Handling Workflow Scheduling Using DVFS-MODPSOIn general a workflow scheduling can be defined by a setof triplets 119872 = [⟨119879
119894 119875119895 119871119896⟩](119894 isin [1 119899] 119895 isin [1119898] 119896 isin
[1 119871(119895)]) 119899 is the number of workflow tasks to be scheduled119898 is the number of processors available in the cloud environ-ment and 119871(119895) is the number of operating points (VSLs) of the119895th processor
For the sake of clarity the variables and rules of DVFS-MODPSO for solving workflow scheduling can be depictedas follows
The position of a particle represents a feasible solutionto the scheduling problem It consists of a set of⟨task(119879
119894) service (119875
119895) VSl (119871
119896)⟩ triplets Each triplet
means that a task (119879119894) is assigned to a processor (119875
119895)
with a voltage scaling level (119871119896) It also indicates
The Scientific World Journal 7
that the position satisfies the precedence constraintbetween tasks
The process of generating a newposition for a selectedparticle in the swarm is depicted in the followingformulas
119881119896+1
119894= 119881119896
119894oplus ((119877
1otimes (119901119861119890119904119905
119894⊖ 119883119896
119894))
oplus (1198772otimes (119892119861119890119904119905 ⊖ 119883
119896
119894)))
(22)
119883119896+1
119894= 119883119896
119894oplus 119881119896+1
119894 (23)
The operator definitions are as follows
(i) The substract operator (⊖) the difference betweentwo particle positions designated as 1199091 and 1199092 isdefined as a set of triplets in which each triplet⟨task serviceVSL⟩ shows whether the contents ofthe corresponding elements in 1199091 are different fromthose of 1199092 or not If so that triplet gets its values(service and VSL) from the position that has thelowest value of theVSL For those triplets that have thesame content in 1199091 and 1199092 their corresponding VSLsare decreased (Note that the scaling of VSL makesfluctuations on the energy makespan and cost)
(ii) The multiply operator (otimes) the multiplication betweennumber and velocity is defined as a set of tripletswhere a threshold 120572 isin [0 1] is defined arandom number 119903 is generated for each triplet⟨task serviceVSL⟩ compare 119903 and 120572 when 119903 ge 120572decrease the triplet VSL otherwise increase it Thisoperator adds the exploration ability to the algorithm
(iii) The add operator (oplus) the addition of two positions isdefined as the reservation of the dominated one
The Pseudocode 1 outlines the general steps of the DVFS-MODPSO algorithm
The algorithm begins by initializing the positions andvelocities of particles To obtain the position of a particle theVSL (voltage and frequency) of each resource is randomlyinitialized firstly then the HEFT algorithm is applied togenerate a feasible and efficient solution minimizing themakespan The process is repeated several times to initializethe positions of all particles of the swarm Initially thevelocity119881 and the best position for each particle 119901119861119890119904119905 areattributed as the particle itself The algorithm maintainsan external archive EA to store non-dominated particlesfound after the evaluation process (based on the paretodominance using the objectives mentioned in 33) After allthese initializations in the main loop of the algorithm thenew velocity and position of each particle are calculatedrespectively after selecting the best overall position in theexternal archive and eventually perform a mutation then theparticle is evaluated and its corresponding 119901119861119890119904119905 is updatedThe external archive is updated after every iteration Oncethe termination condition is reached the external archivecontaining the Pareto front is returned as the result
5 Experimental Evaluations
In this section we describe the overall setup of our experi-mentation effort and the results we have obtained from it tovalidate the new proposed approach
In our experiments we simulate a synthetic workflowapplication described in [17] (see Figure 1) and two commonworkflow structures parallel and hybrid structuresWe chosetwo well known real world applications namely a neuro-science workflow [48] for our parallel application and aprotein annotation workflow [49] developed by London e-Science Centre for our hybrid workflow application Figure 2shows their simplified representationsThe number indicatedin the parentheses next to each node represents the length ofthe task (the number of instructions to execute) in Millionsof Instructions (MI) The input and output files of each taskrange from 10MB to 1024MB
We have performed experiments on 3 4 5 6 and 12resources whose characteristics are shown in Table 1 Wechoose the pricing model associated with Amazon EC2(httpawsamazoncomec2) for the processing costs of thedifferent classes of resources and the pricing model given byAmazon CloudFront (httpawsamazoncomcloudfront)for the costs of transferring data unit between resources
As for DVFS-MODPSO the parameter settings areSwarm size 119878119873119906119898 = 50 particles and the maximum numberof iterations 119866 = 100
We evaluated the performance of our proposed DVFS-MODPSO on all the workflow instances described previouslyDue to the lack of works considering both the heteroge-neous configuration of our cloud and the three QoS metrics(makespan cost energy) at the same time we comparedour results with those of the HEFT heuristic we haveimplemented HEFT is one of themost widely used heuristicsfor DAGs in distributed heterogeneous computing systemswhich optimize the makespan
Figures 3 4 and 5 show sample Pareto fronts obtainedwith the DVFS-MODPSO and the solution computed byHEFT for the synthetic workflow the hybrid workflow andthe parallel workflow respectively
As can be seen from these figures unlike the HEFTalgorithm which gives one solution our proposed approachprovides a set of non dominated solutions These resultsillustrate the basic multi-objective definitions provided inSection 41
Now for comparing these two approaches and in orderto analyze the effectiveness of our proposition in terms ofthe values of makespan cost and energy consumption wehave been inspired by the methodology described in [43]where we compare the solution provided by HEFT to onlyone solution of the Pareto front set computed by our proposedmulti-objective DVFS-MODPSO
The evaluation process follows the next steps For eachworkflow instance we perform a first resolution usingHEFTin order to get one solution 119878 After a second resolution iscomputed using our proposed approach to obtain a set EA ofPareto solutions Next we select one solution 1198781015840 from the setEA of Pareto solutions This solution is the closest one to 119878 inthe sense of Euclidean distance Finally a comparison is donebetween 119878 and 1198781015840
8 The Scientific World Journal
Swarm initialization with HEFT(1) For i = 1 to SNum (SNum is the size of particle swarm)
(a) For j = 1 to m (m the number of processors)(i) Randomly initialize VSL(j) (randomly choose the voltagefrequency of processor from the set of its operating points)
(b) Initialize 119878[119894] with HEFT heuristic (S is the swarm of particles)(c) Initialize the velocity V of each particle
119881 [119894] = 0 (119878 [119894])
(d) Initialize the Personal Best (pBest) of each particle119901119861119890119904119905[119894] = 119878[119894]
(e) Evaluate objectives of each particleEvaluate S[i]
(f) Initialize the Global Best particle (gBest) with the best one among the SNum particlesg119861119890119904119905 = Best particle found in 119878
(2) End For(3) Add the nondominated solutions found in S into EA (EA is the External Archive storing the pareto front)(4) Initialize the iteration number 119905 = 0(5) Repeat until 119905 gt 119866 (119866 is the maximum number of iterations)
(a) For 119894 = 1 to SNum (swarm size)(i) Randomly select the global best particle for S from the External Archive EA and store its position in gBest(ii) Calculate the new velocity 119881[119894] according to (22)(iii) Compute the new position of 119878[119894] according to (23)(iv) If (119905 lt 119866 lowast PMUT ) then (PMUT is the probability of mutation)
Perform mutation on 119878[119894](v) Evaluate S[i]
(b) End For(c) Update the personal best solution of each particle 119878[119894]
if 119878 [119894] ≼ 119901119861119890119904119905119904 [119894] or (119878 [119894] sim 119901119861119890119904119905119904 [119894])Then 119901119861119890119904119905119904 [119894] = 119878 [119894]
(d) Update the External Archive EA(i) Add all new non-dominated solution in S into EAif 119878 [119894] 997410 119909 forall119909 isin 119864119860Then 119864119860 = 119864119860 cup 119878 [119894](ii) Remove all particles dominated by 119878[119894] in EA
119864119860 = 119864119860 minus 119910 isin 119864119860 | 119910 lt 119878[119894]
(iii) If the archive is full then randomly select the article to be replaced from EA(6) Return the pareto front (the set of non-dominated solutions from S and EA)
Pseudocode 1 DVFS-MODPSO based workflow scheduling
The different results are displayed on Figures 6 7 8 and 9and on Tables 3 and 4 They show the improvement achievedby our proposed approach according to the structure of theworkflow applications and the number of processors
Table 3 and Figure 6 show that our proposed approachimproves on average the result obtained by HEFT for thesynthetic workflow instance The makespan is reduced by005 the cost is reduced by 993 and the energy con-sumption is reduced by 2640 Figure 6 shows detailedimprovements of the proposed approach according to thenumber of processors
Table 4 and Figure 7 illustrate the gain obtained by ourapproach according to the kind of real world applicationsTable 4 also shows the detailed improvements when scalingthe number of processors As can be seen the QoS metricsare improved over HEFT in average by 095 for themakespan 108 for cost and 812 for the energy con-sumption when using the hybrid workflow application andaverage improvements of 295 formakespan 2215 for costand 209 for energy consumption when using the parallelworkflow
These evaluations confirm the results obtained from thesynthetic workflow This means that by using our DVFS-MODPSO we are able to improve not only the cost and energybut also themakespan onwhichHEFT algorithm is supposedto provide good results These results remain true for all kindof workflow applications
In our experiments we limited the number of resourcesto 12 as we found that for these kinds of workflowapplications some resources are not used at all Figure 8shows the resource loads when scheduling the syntheticworkflow on 12 resources Furthermore even though thenumber of resources increased the total cost and total energyconsumption could not always decrease as illustrated inFigure 9 From this figure we can see that the QoS metrics(1) firstly decreases as the number of resources increasesuntil achieving the number 6 This can be explained by thefact that when increasing the number of resources thereare fewer tasks executed in a resource therefore tasks canextend their execution times and the resource have more ofchances to scale down their voltages and frequencies whichcan be very effective in reducing total energy consumption
The Scientific World Journal 9
1 2 3 4
5
6 7
8 9 10
11
12
14
13
15
(300000) (600000) (600000) (900000)
(300000)
(150000) (150000)
(300000) (300000) (300000)
(600000)
(300000)
(150000)
(300000)
(600000)
SignalP COILS2 SEG PROSITE
TMHMM
ProsperoHMMer
PSI-BLASTBLAST
IMPALA
PSI-PRED
3D-PSSMSummary
Genomesummary
SCOP
(a)
1 3 5 7
9
11 1210
(400000) (400000) (400000) (400000)
(300000)
(500000) (500000) (500000)
Softmea
2 4 6 8(600000) (600000) (600000) (600000)
Reslice
14 1513
(600000) (600000) (600000)
Convert
Reslice Reslice Reslice
Slicer Slicer
Convert Convert
Slicer
Align wapAlign wapAlign wapAlign wap
(b)
Figure 2 Parallel and Hybrid workflows
10 The Scientific World Journal
75 80 85 90 95 100 8090
100110
120130
140
2060
100140180220
Ener
gy
DVFS-MODPSOHEFT
MakespanCost
Figure 3 Obtained non-dominated solutions for the syntheticworkflow
Ener
gy
DVFS-MODPSOHEFT
MakespanCost
200 400 600 800 100012001400160018002000 400800
12001600
2000
400200
600800
100012001400160018002000
Figure 4 Obtained non-dominated solutions for the hybrid work-flow
(2) After achieving 6 resources the QoS metrics begin torise the reason for this is that time executions are dominatedby interprocessors communications hence decreasing theopportunities for scaling down voltages and frequencies ofresources Consequently the threshold numbers of resourcesthat minimized the QoS metrics could be obtained
6 Conclusion
In this paper we propose a new algorithm called DVFS-Multi-Objective Discrete Particle Swarm Optimization(DVFS-MODPSO) for workflow scheduling in distributedenvironments such as cloud computing infrastructures DV-FS-MODPSO simultaneously optimizes several conflictingobjectives namely the makespan cost and energy in adiscrete space It produces a set of non-dominated solutionsthus providing more flexibility for users to assess theirpreferences and select a schedule that meets their QoS
Ener
gy
DVFS-MODPSOHEFT
Makespan Cost
400200600800
10001200140016001800
240020002200
450 500 550 600 650 700 7501200
14001600
18002000
Figure 5 Obtained non-dominated solutions for the parallel work-flow
0
10
20
30
40
50
3 4 5 6 12
MakespanCostEnergy
Figure 6 Gain over HEFT () according to the number ofprocessors
requirements Our approach exploits the heterogeneityand the marketization of cloud resources in order to findsolutions that optimize makespan and cost Furthermore ituses the Dynamic Voltage and Frequency Scaling (DVFS)technique to reduce energy consumption
We have evaluated our algorithm by simulating it withboth synthetic and real world scientific workflow applicationshaving different structures The results show that the pro-posedDVFS-MODPSO is able to produce a set of good Paretooptimal solutions The results also show that our approachprovides significant improvement not only in terms of the
The Scientific World Journal 11
0
5
10
15
20
25
HybridParallel
Makespan Cost Energy
Figure 7 Improvement according to real world applications
R17
R20
R37
R446
R50
R60
R733
R80
R90
R107
R110
R120
Figure 8 Resource loads for synthetic workflow
Table 3 Improvement for synthetic workflow
Num of proc Gain over Heft ()Makespan Cost Energy
3 0 2126 40084 022 464 15355 0 756 17266 003 652 267512 0 969 3255Average 005 993 2640
cost and the energy consumption but also in term of themakespan for which HEFT algorithm is supposed to givebetter results
Multi-objective optimization in cloud workflow schedul-ing is not a mature domain Most of the existing worksattempt to minimize either the makespan or cost Howeverwe plan to consider other objectives such as reliabilitysecurity in addition to the energy consumption We also
0
50
100
150
200
250
3 4 5 6 12
MakespanCostEnergy
Figure 9 QoS metrics evolutions according to resource numbersfor the synthetic workflow
Table 4 Improvement according to real world applications
Appli Num of proc Gain over Heft ()Makespan Cost Energy
Hybrid
3 405 1155 04024 028 0566 05075 000 1347 11246 000 0117 126512 043 2214 0761
Average 095 1080 0812
Parallel
3 235 5640 25524 1147 2094 32715 000 2142 11666 093 1170 315412 000 031 306
Average 295 2215 209
intend to apply our algorithm in a real world cloud orintegrate it in existing cloud toolkits such as Cloudbus forcomparing with other algorithms
References
[1] D Thain T Tannenbaum and M Livny Condor and the GridGrid Computing Making the Global Infrastructure a RealityJohn Wiley amp Sons Hoboken NJ USA 2003
[2] R Buyya and S Venugopal ldquoThe gridbus toolkit for serviceoriented grid and utility computing an overview and statusreportrdquo in Proceedings of the 1st IEEE International Workshopon Grid Economics and Business Models (GECON rsquo04) pp 19ndash36 IEEE CS Seoul Republic of Korea April 2004
[3] S McGough L Young A Afzal S Newhouse and J Darling-ton ldquoWorkflow enactment in ICENIrdquo in Proceedings of the UK
12 The Scientific World Journal
e-Science All Hands Meeting pp 894ndash900 IOP NottinghamUK September 2004
[4] E Deelman J Blythe Y Gil et al ldquoPegasus mapping scientificworkflows onto the gridrdquo in Grid Computing 2nd EuropeanAcrossGrids Conference AxGrids 2004 Nicosia Cyprus January28ndash30 2004 vol 3165 of Lecture Notes in Computer Science pp11ndash20 Springer New York NY USA 2004
[5] R Buyya S Pandey and C Vecchiola ldquoCloudbus toolkitfor market-oriented cloud computingrdquo in Cloud ComputingProceedings of the 1st International Conference CloudCom 2009Beijing China December 1ndash4 2009 vol 5931 of Lecture Notesin Computer Science pp 24ndash44 Springer New York NY USA2009
[6] Y Yang K Liu J Chen X Liu D Yuan and H Jin ldquoAnalgorithm in SwinDeW-C for scheduling transaction-intensivecost-constrained cloud workflowsrdquo in Proceedings of the 4thIEEE International Conference on eScience (eScience rsquo08) pp374ndash375 Indianapolis Ind USA December 2008
[7] L Ramakrishnan C Koelbel Y Kee et al ldquoVGrADS enablinge-Scienceworkflows on grids and cloudswith fault tolerancerdquo inProceedings of the Conference on High Performance ComputingNetworking Storage and Analysis (SC rsquo09) November 2009
[8] M L Pinedo Scheduling Theory Algorithms and SystemsSpringer Berlin 2008
[9] S Pandey L Wu S M Guru and R Buyya ldquoA particle swarmoptimization-based heuristic for scheduling workflow applica-tions in cloud computing environmentsrdquo in Proceedings of the24th IEEE International Conference on Advanced InformationNetworking and Applications (AINA rsquo10) pp 400ndash407 PerthAustralia April 2010
[10] S Abrishami and M Naghibzadeh ldquoDeadline-constrainedworkflow scheduling in software as a service cloudrdquo ScientiaIranica vol 19 no 3 pp 680ndash689 2012
[11] S Selvarani and G S Sadhasivam ldquoImproved cost-based algo-rithm for task scheduling in cloud computingrdquo in Proceedings ofthe IEEE International Conference on Computational Intelligenceand Computing Research (ICCIC rsquo10) pp 1ndash5 CoimbatoreIndia December 2010
[12] A Verma and S Kaushal ldquoDeadline and budget distributionbased cost-time optimization workflow scheduling algorithmfor cloudrdquo inProceedings of the IJCAon International Conferenceon Recent Advances and Future Trends in Information Technol-ogy (iRAFIT rsquo12) iRAFIT(7) pp 1ndash4 April 2012
[13] R C Correa A Ferreira and P Rebreyend ldquoIntegrating listheuristics into genetic algorithms for multiprocessor schedul-ingrdquo in Proceedings of the 8th IEEE Symposium on Parallel andDistributed Processing (SPDP rsquo96) pp 462ndash469 IEEEComputerSociety Washington DC USA October 1996
[14] E S H Hou N Ansari and H Ren ldquoGenetic algorithm formultiprocessor schedulingrdquo IEEE Transactions on Parallel andDistributed Systems vol 5 no 2 pp 113ndash120 1994
[15] M Grajcar ldquoGenetic list scheduling algorithm for schedulingand allocation on a loosely coupled heterogeneous multi-processor systemrdquo in Proceedings of the 36th Annual DesignAutomation Conference (DAC rsquo99) pp 280ndash285 ACM PressNew York NY USA June 1999
[16] R Sakellariou and H Zhao ldquoA hybrid heuristic for DAGscheduling on heterogeneous systemsrdquo in Proceedings of the13th Heterogeneous Computing Workshop pp 111ndash123 IEEEComputer Society Washington DC USA 2004
[17] H Topcuoglu S Hariri and M Y Wu ldquoPerformance-effectiveand low-complexity task scheduling for heterogeneous comput-ingrdquo IEEE Transactions on Parallel and Distributed Systems vol13 no 3 pp 260ndash274 2002
[18] J Yu and R Buyya ldquoWorkflow scheduling algorithms for gridcomputingrdquo in Metaheuristics for Scheduling in DistributedComputing Environments F Xhafa and A Abraham EdsSpringer Berlin Germany 2008
[19] WN Chen and J Zhang ldquoAn ant colony optimization approachto a grid workflow scheduling problem with various QoSrequirementsrdquo IEEE Transactions on Systems Man and Cyber-netics C vol 39 no 1 pp 29ndash43 2009
[20] X Liu J Chen Z Wu Z Ni D Yuan and Y Yang ldquoHandlingrecoverable temporal violations in scientific workflow systemsa workflow rescheduling based strategyrdquo in Proceedings of the10th IEEEACM International Symposium on Cluster Cloudand Grid Computing (CCGrid rsquo10) pp 534ndash537 MelbourneAustralia May 2010
[21] X Liu Y Yang Y Jiang and J Chen ldquoPreventing temporalviolations in scientific workflows where and howrdquo IEEE Trans-actions on Software Engineering vol 37 no 6 pp 805ndash825 2011
[22] J Yu and R Buyya ldquoScheduling scientific workflow applicationswith deadline and budget constraints using genetic algorithmsrdquoScientific Programming vol 14 no 3-4 pp 217ndash230 2006
[23] L Benini and G Micheli Dynamic Power Management DesignTechniques and CAD Tools Kluwer Academic New York NYUSA 1998
[24] S Irani S Shukla and R Gupta ldquoOnline strategies for dynamicpower management in system with multiple power-savingstatesrdquo ACM Transactions on Embedded Computing Systemsvol 2 no 3 pp 325ndash346 2003
[25] M T Schmitz B M Al-Hashimi and P Eles System-Level Design Techniques for Energy-Efficient Embedded SystemsSpringer New York NY USA 2005
[26] S U Khan and I Ahmad ldquoA cooperative game theoreticaltechnique for joint optimization of energy consumption andresponse time in computational gridsrdquo IEEE Transactions onParallel and Distributed Systems vol 20 no 3 pp 346ndash3602009
[27] Y Li Y Liu and D Qian ldquoA heuristic energy-aware schedulingalgorithm for heterogeneous clustersrdquo in Proceedings of the 15thInternational Conference on Parallel and Distributed Systems(ICPADS rsquo09) pp 407ndash413 Shenzhen China December 2009
[28] L K Goh B Veeravalli and S Viswanathan ldquoDesign of fast andefficient energy-aware gradient-based scheduling algorithmsheterogeneous embeddedmultiprocessor systemsrdquo IEEE Trans-actions on Parallel and Distributed Systems vol 20 no 1 pp 1ndash12 2009
[29] F Yao A Demers and S Shenker ldquoA scheduling model forreduced CPU energyrdquo in Proceedings of the 36th IEEE AnnualSymposium on Foundations of Computer Science pp 374ndash382Milwaukee Wis USA October 1995
[30] Y Shin and K Choi ldquoPower conscious fixed priority schedulingfor hard real-time systemsrdquo in Proceedings of the 36th AnnualDesign Automation Conference (DAC rsquo99) pp 134ndash139 NewOrleans La USA June 1999
[31] I Hong D Kirovski G QuM Potkonjak andM B SrivastavaldquoPower optimization of variable-voltage core-based systemsrdquoIEEE Transactions on Computer-Aided Design of IntegratedCircuits and Systems vol 18 no 12 pp 1702ndash1714 1999
[32] P Pillai and K G Shin ldquoReal-time dynamic voltage scaling forlow-power embedded operating systemsrdquo in Proceedings of the
The Scientific World Journal 13
ACM Symposium on Operating Systems Principles pp 89ndash102October 2001
[33] H Aydin R Melhem D Mosse and P Mejia-Alvarez ldquoPower-aware scheduling for periodic real-time tasksrdquo IEEE Transac-tions on Computers vol 53 no 5 pp 584ndash600 2004
[34] J Zhuo and C Chakrabarti ldquoAn efficient dynamic task schedul-ing algorithm for battery powered DVS systemsrdquo in Proceedingsof the Asian South Pacific Design Automation Conference pp846ndash849 January 2005
[35] R Jejurikar and R Gupta ldquoDynamic slack reclamation withprocrastination scheduling in real-time embedded systemsrdquo inProceedings of the 42nd Design Automation Conference (DACrsquo05) pp 111ndash116 June 2005
[36] J Luo andN K Jha ldquoPower-profile driven variable voltage scal-ing for heterogeneous distributed real-time embedded systemsrdquoin Proceedings of the International Conference on VLSI Designpp 369ndash375 January 2003
[37] M T Schmitz and B M Al-Hashimi ldquoConsidering powervariations of DVS processing elements for energy minimisationin distributed systemsrdquo in Proceedings of the International Sym-posium on System Synthesis (ISSS rsquo01) pp 250ndash255 MontrealCanada October 2001
[38] Y Zhang X Hu and D Z Chen ldquoTask scheduling and voltageselection for energy minimizationrdquo in Proceedings of the 39thAnnual Design Automation Conference (DAC rsquo02) pp 183ndash188June 2002
[39] D Zhu R Melhem and B R Childers ldquoScheduling withdynamic voltagespeed adjustment using slack reclamationin multiprocessor real-time systemsrdquo IEEE Transactions onParallel and Distributed Systems vol 14 no 7 pp 686ndash7002003
[40] D Zhu D Mosse and R Melhem ldquoPower-aware schedulingfor ANDOR graphs in real-time systemsrdquo IEEE Transactionson Parallel and Distributed Systems vol 15 no 9 pp 849ndash8642004
[41] J Kang and S Ranka ldquoDVS based energy minimizationalgorithm for parallel machinesrdquo in Proceedings of the IEEEInternational Parallel and Distributed Processing Symposium(IPDPS rsquo08) pp 1ndash12 Miami Fla USA April 2008
[42] B Rountree D K Lowenthal S Funk V W Freeh B R deSupinski and M Schulz ldquoBounding energy consumption inlarge-scale MPI programsrdquo in Proceedings of the ACMIEEEConference on Supercomputing (SC rsquo07) ACM November 2007
[43] L Wang G von Laszewski J Dayal and F Wang ldquoTowardsenergy aware scheduling for precedence constrained paralleltasks in a cluster with DVFSrdquo in Proceedings of the 10thIEEEACM International Symposium on Cluster Cloud andGrid Computing (CCGrid rsquo10) pp 368ndash377 May 2010
[44] Y C Lee and A Y Zomaya ldquoEnergy conscious schedulingfor distributed computing systems under different operatingconditionsrdquo IEEE Transactions on Parallel and DistributedSystems vol 22 no 8 pp 1374ndash1381 2011
[45] M Mezmaz Y C Lee N Melab E Talbi and A Y ZomayaldquoA bi-objective hybrid genetic algorithm to minimize energyconsumption and makespan for precedence-constrained appli-cations using dynamic voltage scalingrdquo in Proceedings of theIEEE Congress on Evolutionary Computation (CEC rsquo10) pp 1ndash8Barcelona Spain July 2010
[46] M Mezmaz N Melab Y Kessaci et al ldquoA parallel bi-objectivehybrid metaheuristic for energy-aware scheduling for cloudcomputing systemsrdquo Journal of Parallel and Distributed Com-puting vol 71 no 11 pp 1497ndash1508 2011
[47] J Kennedy and R Eberhart ldquoParticle swarm optimizationrdquoin Proceedings of the IEEE International Conference on NeuralNetworks vol 4 pp 1942ndash1948 Perth Australia December1995
[48] Y Zhao M Wilde I Foster et al ldquoGrid middleware servicesfor virtual data discovery composition and integrationrdquo inProceedings of the 2nd Workshop on Middleware for GridComputing (MGC rsquo04) pp 57ndash62 Ontario Canada October2004
[49] A OrsquoBrien S Newhouse and J Darlington ldquoMapping ofscientific workflow within the e-protein project to distributedresourcesrdquo inProceedings of theUK e-Science All HandsMeetingNottingham UK September 2004
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
6 The Scientific World Journal
(ii) Pareto optimally A decision vector 1199091 is said to bePareto optimal if and only if
∄1199092isin 119883 119909
2≺ 1199091 (18)
(iii) Pareto optimal set The Pareto optimal set 119875119878is the set
of all Pareto optimal decision vectors
119875119878= 1199091isin 119883 | ∄119909
2isin 119883 119909
2≺ 1199091 (19)
(iv) Pareto optimal frontThePareto optimal front119875119865is the
image of the Pareto optimal set in the objective space
119875119865= 119891 (119909) = (119891
1(119909) 119891
119899(119909)) | 119909 isin 119875
119904 (20)
1199091 is said to be non-dominated regarding a given set
if 1199091 is not dominated by any decision vectors in theset
Thepareto optimal decision vector cannot be enhanced inany objective without causing degradation in at least anotherobjective A decision vector is said to be Pareto optimal whenit is not dominated in the whole search space
42 Particle Swarm Optimisation
421 The Standard PSO PSO is a population-based stochas-tic optimization technique developed by Kennedy and Eber-hart in 1995 [47] It is inspired by the social behavior of insectcolonies bird flocks fish schools and other animal societies Itis also related to evolutionary computation PSOhas attractedsignificant attention from many researchers due to both itssimplicity of use and optimization via social behavior In factPSO has good performance requires low computational costIt is effective and easy to implement as it uses numericalencoding
A particle in PSO is analogous to a fish or bird movingin the 119863-dimensional search space All particles have fitnessvalues indicating their performances which are problemspecific and velocities which direct the flight of particlesEach particle position at any given time is influenced by bothits best position called 119901119861119890119904119905 and the position of the bestparticle in a problem space referred to as 119892119861119890119904119905 Thereforeparticles tend to fly towards a better search area duringthe search process A particle status on the search spaceis characterized by two elements namely its velocity andposition which are updated in every generation as follows
119881119896+1
119894= 120596119881
119896
119894+ 11988811199031(119901119861119890119904119905
119894minus 119883119896
119894)
+ 11988821199032(119892119861119890119904119905 minus 119883
119896
119894)
119883119896+1
119894= 119883119896
119894+ 119881119896+1
119894
(21)
where 119881119896+1119894
is the velocity of particle 119894 at iteration 119896 + 1119881119896119894is the velocity of particle 119894 at iteration 119896 120596 is the inertia
weight 1198881and 1198882are the acceleration coefficients (cognitive
and social coefficients) 1199031and 1199032are the random numbers
between 0 and 1119883119896119894is the current position of particle 119894 at the
119896th iteration 119901119861119890119904119905119894is the best previous position of the 119894th
particle 119892119861119890119904119905 is the position of best particle in the swarmand119883119896+1
119894is the position of 119894th particle at 119896 + 1 iteration
The procedure for standard PSO is as follows
(1) Initialize a population of particles with random posi-tions and velocities in the search space
(2) Evaluate the objective values of all particles set 119901119861119890119904119905of each particle equal to its current position and set119892119861119890119904119905 equal to the position of the best initial particle
(3) Update the velocity and the position of each particleaccording to (21)
(4) Map the position of each particle in the solution spaceand evaluate its fitness value according to the desiredoptimization fitness function
(5) For each particle compare its current objective valuewith its 119901119861119890119904119905 value If the current value is better thenupdate 119901Best with the current position and objectivevalue
(6) Determine the best particle of the current whole pop-ulation with the best objective value If the objectivevalue is better than that of 119892119861119890119904119905 then update 119892Bestwith the current best particle
(7) If the stopping criterion ismet then output 119892119861119890119904119905 andits objective value otherwise go to Step (2)
The original design of PSO is appropriate for findingsolutions to continuous optimization problems However asthe workflow scheduling discussed in this paper is both adiscrete andmulti objective problem in nature we propose aneffective approach to address this problem by using a discreteversion of the Multi-Objective PSO (MODPSO) combinedwith DVFS techniqueThe key issues of our approach consistof (1) defining the position and velocity of the particlesbased on the features of discrete variables of the workflowscheduling (2) solving the multi-objective aspect of theproblem by modifying PSO so as to generate a set of non-dominated solutions satisfying the different objectives underconsideration instead of one solution
422 Handling Workflow Scheduling Using DVFS-MODPSOIn general a workflow scheduling can be defined by a setof triplets 119872 = [⟨119879
119894 119875119895 119871119896⟩](119894 isin [1 119899] 119895 isin [1119898] 119896 isin
[1 119871(119895)]) 119899 is the number of workflow tasks to be scheduled119898 is the number of processors available in the cloud environ-ment and 119871(119895) is the number of operating points (VSLs) of the119895th processor
For the sake of clarity the variables and rules of DVFS-MODPSO for solving workflow scheduling can be depictedas follows
The position of a particle represents a feasible solutionto the scheduling problem It consists of a set of⟨task(119879
119894) service (119875
119895) VSl (119871
119896)⟩ triplets Each triplet
means that a task (119879119894) is assigned to a processor (119875
119895)
with a voltage scaling level (119871119896) It also indicates
The Scientific World Journal 7
that the position satisfies the precedence constraintbetween tasks
The process of generating a newposition for a selectedparticle in the swarm is depicted in the followingformulas
119881119896+1
119894= 119881119896
119894oplus ((119877
1otimes (119901119861119890119904119905
119894⊖ 119883119896
119894))
oplus (1198772otimes (119892119861119890119904119905 ⊖ 119883
119896
119894)))
(22)
119883119896+1
119894= 119883119896
119894oplus 119881119896+1
119894 (23)
The operator definitions are as follows
(i) The substract operator (⊖) the difference betweentwo particle positions designated as 1199091 and 1199092 isdefined as a set of triplets in which each triplet⟨task serviceVSL⟩ shows whether the contents ofthe corresponding elements in 1199091 are different fromthose of 1199092 or not If so that triplet gets its values(service and VSL) from the position that has thelowest value of theVSL For those triplets that have thesame content in 1199091 and 1199092 their corresponding VSLsare decreased (Note that the scaling of VSL makesfluctuations on the energy makespan and cost)
(ii) The multiply operator (otimes) the multiplication betweennumber and velocity is defined as a set of tripletswhere a threshold 120572 isin [0 1] is defined arandom number 119903 is generated for each triplet⟨task serviceVSL⟩ compare 119903 and 120572 when 119903 ge 120572decrease the triplet VSL otherwise increase it Thisoperator adds the exploration ability to the algorithm
(iii) The add operator (oplus) the addition of two positions isdefined as the reservation of the dominated one
The Pseudocode 1 outlines the general steps of the DVFS-MODPSO algorithm
The algorithm begins by initializing the positions andvelocities of particles To obtain the position of a particle theVSL (voltage and frequency) of each resource is randomlyinitialized firstly then the HEFT algorithm is applied togenerate a feasible and efficient solution minimizing themakespan The process is repeated several times to initializethe positions of all particles of the swarm Initially thevelocity119881 and the best position for each particle 119901119861119890119904119905 areattributed as the particle itself The algorithm maintainsan external archive EA to store non-dominated particlesfound after the evaluation process (based on the paretodominance using the objectives mentioned in 33) After allthese initializations in the main loop of the algorithm thenew velocity and position of each particle are calculatedrespectively after selecting the best overall position in theexternal archive and eventually perform a mutation then theparticle is evaluated and its corresponding 119901119861119890119904119905 is updatedThe external archive is updated after every iteration Oncethe termination condition is reached the external archivecontaining the Pareto front is returned as the result
5 Experimental Evaluations
In this section we describe the overall setup of our experi-mentation effort and the results we have obtained from it tovalidate the new proposed approach
In our experiments we simulate a synthetic workflowapplication described in [17] (see Figure 1) and two commonworkflow structures parallel and hybrid structuresWe chosetwo well known real world applications namely a neuro-science workflow [48] for our parallel application and aprotein annotation workflow [49] developed by London e-Science Centre for our hybrid workflow application Figure 2shows their simplified representationsThe number indicatedin the parentheses next to each node represents the length ofthe task (the number of instructions to execute) in Millionsof Instructions (MI) The input and output files of each taskrange from 10MB to 1024MB
We have performed experiments on 3 4 5 6 and 12resources whose characteristics are shown in Table 1 Wechoose the pricing model associated with Amazon EC2(httpawsamazoncomec2) for the processing costs of thedifferent classes of resources and the pricing model given byAmazon CloudFront (httpawsamazoncomcloudfront)for the costs of transferring data unit between resources
As for DVFS-MODPSO the parameter settings areSwarm size 119878119873119906119898 = 50 particles and the maximum numberof iterations 119866 = 100
We evaluated the performance of our proposed DVFS-MODPSO on all the workflow instances described previouslyDue to the lack of works considering both the heteroge-neous configuration of our cloud and the three QoS metrics(makespan cost energy) at the same time we comparedour results with those of the HEFT heuristic we haveimplemented HEFT is one of themost widely used heuristicsfor DAGs in distributed heterogeneous computing systemswhich optimize the makespan
Figures 3 4 and 5 show sample Pareto fronts obtainedwith the DVFS-MODPSO and the solution computed byHEFT for the synthetic workflow the hybrid workflow andthe parallel workflow respectively
As can be seen from these figures unlike the HEFTalgorithm which gives one solution our proposed approachprovides a set of non dominated solutions These resultsillustrate the basic multi-objective definitions provided inSection 41
Now for comparing these two approaches and in orderto analyze the effectiveness of our proposition in terms ofthe values of makespan cost and energy consumption wehave been inspired by the methodology described in [43]where we compare the solution provided by HEFT to onlyone solution of the Pareto front set computed by our proposedmulti-objective DVFS-MODPSO
The evaluation process follows the next steps For eachworkflow instance we perform a first resolution usingHEFTin order to get one solution 119878 After a second resolution iscomputed using our proposed approach to obtain a set EA ofPareto solutions Next we select one solution 1198781015840 from the setEA of Pareto solutions This solution is the closest one to 119878 inthe sense of Euclidean distance Finally a comparison is donebetween 119878 and 1198781015840
8 The Scientific World Journal
Swarm initialization with HEFT(1) For i = 1 to SNum (SNum is the size of particle swarm)
(a) For j = 1 to m (m the number of processors)(i) Randomly initialize VSL(j) (randomly choose the voltagefrequency of processor from the set of its operating points)
(b) Initialize 119878[119894] with HEFT heuristic (S is the swarm of particles)(c) Initialize the velocity V of each particle
119881 [119894] = 0 (119878 [119894])
(d) Initialize the Personal Best (pBest) of each particle119901119861119890119904119905[119894] = 119878[119894]
(e) Evaluate objectives of each particleEvaluate S[i]
(f) Initialize the Global Best particle (gBest) with the best one among the SNum particlesg119861119890119904119905 = Best particle found in 119878
(2) End For(3) Add the nondominated solutions found in S into EA (EA is the External Archive storing the pareto front)(4) Initialize the iteration number 119905 = 0(5) Repeat until 119905 gt 119866 (119866 is the maximum number of iterations)
(a) For 119894 = 1 to SNum (swarm size)(i) Randomly select the global best particle for S from the External Archive EA and store its position in gBest(ii) Calculate the new velocity 119881[119894] according to (22)(iii) Compute the new position of 119878[119894] according to (23)(iv) If (119905 lt 119866 lowast PMUT ) then (PMUT is the probability of mutation)
Perform mutation on 119878[119894](v) Evaluate S[i]
(b) End For(c) Update the personal best solution of each particle 119878[119894]
if 119878 [119894] ≼ 119901119861119890119904119905119904 [119894] or (119878 [119894] sim 119901119861119890119904119905119904 [119894])Then 119901119861119890119904119905119904 [119894] = 119878 [119894]
(d) Update the External Archive EA(i) Add all new non-dominated solution in S into EAif 119878 [119894] 997410 119909 forall119909 isin 119864119860Then 119864119860 = 119864119860 cup 119878 [119894](ii) Remove all particles dominated by 119878[119894] in EA
119864119860 = 119864119860 minus 119910 isin 119864119860 | 119910 lt 119878[119894]
(iii) If the archive is full then randomly select the article to be replaced from EA(6) Return the pareto front (the set of non-dominated solutions from S and EA)
Pseudocode 1 DVFS-MODPSO based workflow scheduling
The different results are displayed on Figures 6 7 8 and 9and on Tables 3 and 4 They show the improvement achievedby our proposed approach according to the structure of theworkflow applications and the number of processors
Table 3 and Figure 6 show that our proposed approachimproves on average the result obtained by HEFT for thesynthetic workflow instance The makespan is reduced by005 the cost is reduced by 993 and the energy con-sumption is reduced by 2640 Figure 6 shows detailedimprovements of the proposed approach according to thenumber of processors
Table 4 and Figure 7 illustrate the gain obtained by ourapproach according to the kind of real world applicationsTable 4 also shows the detailed improvements when scalingthe number of processors As can be seen the QoS metricsare improved over HEFT in average by 095 for themakespan 108 for cost and 812 for the energy con-sumption when using the hybrid workflow application andaverage improvements of 295 formakespan 2215 for costand 209 for energy consumption when using the parallelworkflow
These evaluations confirm the results obtained from thesynthetic workflow This means that by using our DVFS-MODPSO we are able to improve not only the cost and energybut also themakespan onwhichHEFT algorithm is supposedto provide good results These results remain true for all kindof workflow applications
In our experiments we limited the number of resourcesto 12 as we found that for these kinds of workflowapplications some resources are not used at all Figure 8shows the resource loads when scheduling the syntheticworkflow on 12 resources Furthermore even though thenumber of resources increased the total cost and total energyconsumption could not always decrease as illustrated inFigure 9 From this figure we can see that the QoS metrics(1) firstly decreases as the number of resources increasesuntil achieving the number 6 This can be explained by thefact that when increasing the number of resources thereare fewer tasks executed in a resource therefore tasks canextend their execution times and the resource have more ofchances to scale down their voltages and frequencies whichcan be very effective in reducing total energy consumption
The Scientific World Journal 9
1 2 3 4
5
6 7
8 9 10
11
12
14
13
15
(300000) (600000) (600000) (900000)
(300000)
(150000) (150000)
(300000) (300000) (300000)
(600000)
(300000)
(150000)
(300000)
(600000)
SignalP COILS2 SEG PROSITE
TMHMM
ProsperoHMMer
PSI-BLASTBLAST
IMPALA
PSI-PRED
3D-PSSMSummary
Genomesummary
SCOP
(a)
1 3 5 7
9
11 1210
(400000) (400000) (400000) (400000)
(300000)
(500000) (500000) (500000)
Softmea
2 4 6 8(600000) (600000) (600000) (600000)
Reslice
14 1513
(600000) (600000) (600000)
Convert
Reslice Reslice Reslice
Slicer Slicer
Convert Convert
Slicer
Align wapAlign wapAlign wapAlign wap
(b)
Figure 2 Parallel and Hybrid workflows
10 The Scientific World Journal
75 80 85 90 95 100 8090
100110
120130
140
2060
100140180220
Ener
gy
DVFS-MODPSOHEFT
MakespanCost
Figure 3 Obtained non-dominated solutions for the syntheticworkflow
Ener
gy
DVFS-MODPSOHEFT
MakespanCost
200 400 600 800 100012001400160018002000 400800
12001600
2000
400200
600800
100012001400160018002000
Figure 4 Obtained non-dominated solutions for the hybrid work-flow
(2) After achieving 6 resources the QoS metrics begin torise the reason for this is that time executions are dominatedby interprocessors communications hence decreasing theopportunities for scaling down voltages and frequencies ofresources Consequently the threshold numbers of resourcesthat minimized the QoS metrics could be obtained
6 Conclusion
In this paper we propose a new algorithm called DVFS-Multi-Objective Discrete Particle Swarm Optimization(DVFS-MODPSO) for workflow scheduling in distributedenvironments such as cloud computing infrastructures DV-FS-MODPSO simultaneously optimizes several conflictingobjectives namely the makespan cost and energy in adiscrete space It produces a set of non-dominated solutionsthus providing more flexibility for users to assess theirpreferences and select a schedule that meets their QoS
Ener
gy
DVFS-MODPSOHEFT
Makespan Cost
400200600800
10001200140016001800
240020002200
450 500 550 600 650 700 7501200
14001600
18002000
Figure 5 Obtained non-dominated solutions for the parallel work-flow
0
10
20
30
40
50
3 4 5 6 12
MakespanCostEnergy
Figure 6 Gain over HEFT () according to the number ofprocessors
requirements Our approach exploits the heterogeneityand the marketization of cloud resources in order to findsolutions that optimize makespan and cost Furthermore ituses the Dynamic Voltage and Frequency Scaling (DVFS)technique to reduce energy consumption
We have evaluated our algorithm by simulating it withboth synthetic and real world scientific workflow applicationshaving different structures The results show that the pro-posedDVFS-MODPSO is able to produce a set of good Paretooptimal solutions The results also show that our approachprovides significant improvement not only in terms of the
The Scientific World Journal 11
0
5
10
15
20
25
HybridParallel
Makespan Cost Energy
Figure 7 Improvement according to real world applications
R17
R20
R37
R446
R50
R60
R733
R80
R90
R107
R110
R120
Figure 8 Resource loads for synthetic workflow
Table 3 Improvement for synthetic workflow
Num of proc Gain over Heft ()Makespan Cost Energy
3 0 2126 40084 022 464 15355 0 756 17266 003 652 267512 0 969 3255Average 005 993 2640
cost and the energy consumption but also in term of themakespan for which HEFT algorithm is supposed to givebetter results
Multi-objective optimization in cloud workflow schedul-ing is not a mature domain Most of the existing worksattempt to minimize either the makespan or cost Howeverwe plan to consider other objectives such as reliabilitysecurity in addition to the energy consumption We also
0
50
100
150
200
250
3 4 5 6 12
MakespanCostEnergy
Figure 9 QoS metrics evolutions according to resource numbersfor the synthetic workflow
Table 4 Improvement according to real world applications
Appli Num of proc Gain over Heft ()Makespan Cost Energy
Hybrid
3 405 1155 04024 028 0566 05075 000 1347 11246 000 0117 126512 043 2214 0761
Average 095 1080 0812
Parallel
3 235 5640 25524 1147 2094 32715 000 2142 11666 093 1170 315412 000 031 306
Average 295 2215 209
intend to apply our algorithm in a real world cloud orintegrate it in existing cloud toolkits such as Cloudbus forcomparing with other algorithms
References
[1] D Thain T Tannenbaum and M Livny Condor and the GridGrid Computing Making the Global Infrastructure a RealityJohn Wiley amp Sons Hoboken NJ USA 2003
[2] R Buyya and S Venugopal ldquoThe gridbus toolkit for serviceoriented grid and utility computing an overview and statusreportrdquo in Proceedings of the 1st IEEE International Workshopon Grid Economics and Business Models (GECON rsquo04) pp 19ndash36 IEEE CS Seoul Republic of Korea April 2004
[3] S McGough L Young A Afzal S Newhouse and J Darling-ton ldquoWorkflow enactment in ICENIrdquo in Proceedings of the UK
12 The Scientific World Journal
e-Science All Hands Meeting pp 894ndash900 IOP NottinghamUK September 2004
[4] E Deelman J Blythe Y Gil et al ldquoPegasus mapping scientificworkflows onto the gridrdquo in Grid Computing 2nd EuropeanAcrossGrids Conference AxGrids 2004 Nicosia Cyprus January28ndash30 2004 vol 3165 of Lecture Notes in Computer Science pp11ndash20 Springer New York NY USA 2004
[5] R Buyya S Pandey and C Vecchiola ldquoCloudbus toolkitfor market-oriented cloud computingrdquo in Cloud ComputingProceedings of the 1st International Conference CloudCom 2009Beijing China December 1ndash4 2009 vol 5931 of Lecture Notesin Computer Science pp 24ndash44 Springer New York NY USA2009
[6] Y Yang K Liu J Chen X Liu D Yuan and H Jin ldquoAnalgorithm in SwinDeW-C for scheduling transaction-intensivecost-constrained cloud workflowsrdquo in Proceedings of the 4thIEEE International Conference on eScience (eScience rsquo08) pp374ndash375 Indianapolis Ind USA December 2008
[7] L Ramakrishnan C Koelbel Y Kee et al ldquoVGrADS enablinge-Scienceworkflows on grids and cloudswith fault tolerancerdquo inProceedings of the Conference on High Performance ComputingNetworking Storage and Analysis (SC rsquo09) November 2009
[8] M L Pinedo Scheduling Theory Algorithms and SystemsSpringer Berlin 2008
[9] S Pandey L Wu S M Guru and R Buyya ldquoA particle swarmoptimization-based heuristic for scheduling workflow applica-tions in cloud computing environmentsrdquo in Proceedings of the24th IEEE International Conference on Advanced InformationNetworking and Applications (AINA rsquo10) pp 400ndash407 PerthAustralia April 2010
[10] S Abrishami and M Naghibzadeh ldquoDeadline-constrainedworkflow scheduling in software as a service cloudrdquo ScientiaIranica vol 19 no 3 pp 680ndash689 2012
[11] S Selvarani and G S Sadhasivam ldquoImproved cost-based algo-rithm for task scheduling in cloud computingrdquo in Proceedings ofthe IEEE International Conference on Computational Intelligenceand Computing Research (ICCIC rsquo10) pp 1ndash5 CoimbatoreIndia December 2010
[12] A Verma and S Kaushal ldquoDeadline and budget distributionbased cost-time optimization workflow scheduling algorithmfor cloudrdquo inProceedings of the IJCAon International Conferenceon Recent Advances and Future Trends in Information Technol-ogy (iRAFIT rsquo12) iRAFIT(7) pp 1ndash4 April 2012
[13] R C Correa A Ferreira and P Rebreyend ldquoIntegrating listheuristics into genetic algorithms for multiprocessor schedul-ingrdquo in Proceedings of the 8th IEEE Symposium on Parallel andDistributed Processing (SPDP rsquo96) pp 462ndash469 IEEEComputerSociety Washington DC USA October 1996
[14] E S H Hou N Ansari and H Ren ldquoGenetic algorithm formultiprocessor schedulingrdquo IEEE Transactions on Parallel andDistributed Systems vol 5 no 2 pp 113ndash120 1994
[15] M Grajcar ldquoGenetic list scheduling algorithm for schedulingand allocation on a loosely coupled heterogeneous multi-processor systemrdquo in Proceedings of the 36th Annual DesignAutomation Conference (DAC rsquo99) pp 280ndash285 ACM PressNew York NY USA June 1999
[16] R Sakellariou and H Zhao ldquoA hybrid heuristic for DAGscheduling on heterogeneous systemsrdquo in Proceedings of the13th Heterogeneous Computing Workshop pp 111ndash123 IEEEComputer Society Washington DC USA 2004
[17] H Topcuoglu S Hariri and M Y Wu ldquoPerformance-effectiveand low-complexity task scheduling for heterogeneous comput-ingrdquo IEEE Transactions on Parallel and Distributed Systems vol13 no 3 pp 260ndash274 2002
[18] J Yu and R Buyya ldquoWorkflow scheduling algorithms for gridcomputingrdquo in Metaheuristics for Scheduling in DistributedComputing Environments F Xhafa and A Abraham EdsSpringer Berlin Germany 2008
[19] WN Chen and J Zhang ldquoAn ant colony optimization approachto a grid workflow scheduling problem with various QoSrequirementsrdquo IEEE Transactions on Systems Man and Cyber-netics C vol 39 no 1 pp 29ndash43 2009
[20] X Liu J Chen Z Wu Z Ni D Yuan and Y Yang ldquoHandlingrecoverable temporal violations in scientific workflow systemsa workflow rescheduling based strategyrdquo in Proceedings of the10th IEEEACM International Symposium on Cluster Cloudand Grid Computing (CCGrid rsquo10) pp 534ndash537 MelbourneAustralia May 2010
[21] X Liu Y Yang Y Jiang and J Chen ldquoPreventing temporalviolations in scientific workflows where and howrdquo IEEE Trans-actions on Software Engineering vol 37 no 6 pp 805ndash825 2011
[22] J Yu and R Buyya ldquoScheduling scientific workflow applicationswith deadline and budget constraints using genetic algorithmsrdquoScientific Programming vol 14 no 3-4 pp 217ndash230 2006
[23] L Benini and G Micheli Dynamic Power Management DesignTechniques and CAD Tools Kluwer Academic New York NYUSA 1998
[24] S Irani S Shukla and R Gupta ldquoOnline strategies for dynamicpower management in system with multiple power-savingstatesrdquo ACM Transactions on Embedded Computing Systemsvol 2 no 3 pp 325ndash346 2003
[25] M T Schmitz B M Al-Hashimi and P Eles System-Level Design Techniques for Energy-Efficient Embedded SystemsSpringer New York NY USA 2005
[26] S U Khan and I Ahmad ldquoA cooperative game theoreticaltechnique for joint optimization of energy consumption andresponse time in computational gridsrdquo IEEE Transactions onParallel and Distributed Systems vol 20 no 3 pp 346ndash3602009
[27] Y Li Y Liu and D Qian ldquoA heuristic energy-aware schedulingalgorithm for heterogeneous clustersrdquo in Proceedings of the 15thInternational Conference on Parallel and Distributed Systems(ICPADS rsquo09) pp 407ndash413 Shenzhen China December 2009
[28] L K Goh B Veeravalli and S Viswanathan ldquoDesign of fast andefficient energy-aware gradient-based scheduling algorithmsheterogeneous embeddedmultiprocessor systemsrdquo IEEE Trans-actions on Parallel and Distributed Systems vol 20 no 1 pp 1ndash12 2009
[29] F Yao A Demers and S Shenker ldquoA scheduling model forreduced CPU energyrdquo in Proceedings of the 36th IEEE AnnualSymposium on Foundations of Computer Science pp 374ndash382Milwaukee Wis USA October 1995
[30] Y Shin and K Choi ldquoPower conscious fixed priority schedulingfor hard real-time systemsrdquo in Proceedings of the 36th AnnualDesign Automation Conference (DAC rsquo99) pp 134ndash139 NewOrleans La USA June 1999
[31] I Hong D Kirovski G QuM Potkonjak andM B SrivastavaldquoPower optimization of variable-voltage core-based systemsrdquoIEEE Transactions on Computer-Aided Design of IntegratedCircuits and Systems vol 18 no 12 pp 1702ndash1714 1999
[32] P Pillai and K G Shin ldquoReal-time dynamic voltage scaling forlow-power embedded operating systemsrdquo in Proceedings of the
The Scientific World Journal 13
ACM Symposium on Operating Systems Principles pp 89ndash102October 2001
[33] H Aydin R Melhem D Mosse and P Mejia-Alvarez ldquoPower-aware scheduling for periodic real-time tasksrdquo IEEE Transac-tions on Computers vol 53 no 5 pp 584ndash600 2004
[34] J Zhuo and C Chakrabarti ldquoAn efficient dynamic task schedul-ing algorithm for battery powered DVS systemsrdquo in Proceedingsof the Asian South Pacific Design Automation Conference pp846ndash849 January 2005
[35] R Jejurikar and R Gupta ldquoDynamic slack reclamation withprocrastination scheduling in real-time embedded systemsrdquo inProceedings of the 42nd Design Automation Conference (DACrsquo05) pp 111ndash116 June 2005
[36] J Luo andN K Jha ldquoPower-profile driven variable voltage scal-ing for heterogeneous distributed real-time embedded systemsrdquoin Proceedings of the International Conference on VLSI Designpp 369ndash375 January 2003
[37] M T Schmitz and B M Al-Hashimi ldquoConsidering powervariations of DVS processing elements for energy minimisationin distributed systemsrdquo in Proceedings of the International Sym-posium on System Synthesis (ISSS rsquo01) pp 250ndash255 MontrealCanada October 2001
[38] Y Zhang X Hu and D Z Chen ldquoTask scheduling and voltageselection for energy minimizationrdquo in Proceedings of the 39thAnnual Design Automation Conference (DAC rsquo02) pp 183ndash188June 2002
[39] D Zhu R Melhem and B R Childers ldquoScheduling withdynamic voltagespeed adjustment using slack reclamationin multiprocessor real-time systemsrdquo IEEE Transactions onParallel and Distributed Systems vol 14 no 7 pp 686ndash7002003
[40] D Zhu D Mosse and R Melhem ldquoPower-aware schedulingfor ANDOR graphs in real-time systemsrdquo IEEE Transactionson Parallel and Distributed Systems vol 15 no 9 pp 849ndash8642004
[41] J Kang and S Ranka ldquoDVS based energy minimizationalgorithm for parallel machinesrdquo in Proceedings of the IEEEInternational Parallel and Distributed Processing Symposium(IPDPS rsquo08) pp 1ndash12 Miami Fla USA April 2008
[42] B Rountree D K Lowenthal S Funk V W Freeh B R deSupinski and M Schulz ldquoBounding energy consumption inlarge-scale MPI programsrdquo in Proceedings of the ACMIEEEConference on Supercomputing (SC rsquo07) ACM November 2007
[43] L Wang G von Laszewski J Dayal and F Wang ldquoTowardsenergy aware scheduling for precedence constrained paralleltasks in a cluster with DVFSrdquo in Proceedings of the 10thIEEEACM International Symposium on Cluster Cloud andGrid Computing (CCGrid rsquo10) pp 368ndash377 May 2010
[44] Y C Lee and A Y Zomaya ldquoEnergy conscious schedulingfor distributed computing systems under different operatingconditionsrdquo IEEE Transactions on Parallel and DistributedSystems vol 22 no 8 pp 1374ndash1381 2011
[45] M Mezmaz Y C Lee N Melab E Talbi and A Y ZomayaldquoA bi-objective hybrid genetic algorithm to minimize energyconsumption and makespan for precedence-constrained appli-cations using dynamic voltage scalingrdquo in Proceedings of theIEEE Congress on Evolutionary Computation (CEC rsquo10) pp 1ndash8Barcelona Spain July 2010
[46] M Mezmaz N Melab Y Kessaci et al ldquoA parallel bi-objectivehybrid metaheuristic for energy-aware scheduling for cloudcomputing systemsrdquo Journal of Parallel and Distributed Com-puting vol 71 no 11 pp 1497ndash1508 2011
[47] J Kennedy and R Eberhart ldquoParticle swarm optimizationrdquoin Proceedings of the IEEE International Conference on NeuralNetworks vol 4 pp 1942ndash1948 Perth Australia December1995
[48] Y Zhao M Wilde I Foster et al ldquoGrid middleware servicesfor virtual data discovery composition and integrationrdquo inProceedings of the 2nd Workshop on Middleware for GridComputing (MGC rsquo04) pp 57ndash62 Ontario Canada October2004
[49] A OrsquoBrien S Newhouse and J Darlington ldquoMapping ofscientific workflow within the e-protein project to distributedresourcesrdquo inProceedings of theUK e-Science All HandsMeetingNottingham UK September 2004
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World Journal 7
that the position satisfies the precedence constraintbetween tasks
The process of generating a newposition for a selectedparticle in the swarm is depicted in the followingformulas
119881119896+1
119894= 119881119896
119894oplus ((119877
1otimes (119901119861119890119904119905
119894⊖ 119883119896
119894))
oplus (1198772otimes (119892119861119890119904119905 ⊖ 119883
119896
119894)))
(22)
119883119896+1
119894= 119883119896
119894oplus 119881119896+1
119894 (23)
The operator definitions are as follows
(i) The substract operator (⊖) the difference betweentwo particle positions designated as 1199091 and 1199092 isdefined as a set of triplets in which each triplet⟨task serviceVSL⟩ shows whether the contents ofthe corresponding elements in 1199091 are different fromthose of 1199092 or not If so that triplet gets its values(service and VSL) from the position that has thelowest value of theVSL For those triplets that have thesame content in 1199091 and 1199092 their corresponding VSLsare decreased (Note that the scaling of VSL makesfluctuations on the energy makespan and cost)
(ii) The multiply operator (otimes) the multiplication betweennumber and velocity is defined as a set of tripletswhere a threshold 120572 isin [0 1] is defined arandom number 119903 is generated for each triplet⟨task serviceVSL⟩ compare 119903 and 120572 when 119903 ge 120572decrease the triplet VSL otherwise increase it Thisoperator adds the exploration ability to the algorithm
(iii) The add operator (oplus) the addition of two positions isdefined as the reservation of the dominated one
The Pseudocode 1 outlines the general steps of the DVFS-MODPSO algorithm
The algorithm begins by initializing the positions andvelocities of particles To obtain the position of a particle theVSL (voltage and frequency) of each resource is randomlyinitialized firstly then the HEFT algorithm is applied togenerate a feasible and efficient solution minimizing themakespan The process is repeated several times to initializethe positions of all particles of the swarm Initially thevelocity119881 and the best position for each particle 119901119861119890119904119905 areattributed as the particle itself The algorithm maintainsan external archive EA to store non-dominated particlesfound after the evaluation process (based on the paretodominance using the objectives mentioned in 33) After allthese initializations in the main loop of the algorithm thenew velocity and position of each particle are calculatedrespectively after selecting the best overall position in theexternal archive and eventually perform a mutation then theparticle is evaluated and its corresponding 119901119861119890119904119905 is updatedThe external archive is updated after every iteration Oncethe termination condition is reached the external archivecontaining the Pareto front is returned as the result
5 Experimental Evaluations
In this section we describe the overall setup of our experi-mentation effort and the results we have obtained from it tovalidate the new proposed approach
In our experiments we simulate a synthetic workflowapplication described in [17] (see Figure 1) and two commonworkflow structures parallel and hybrid structuresWe chosetwo well known real world applications namely a neuro-science workflow [48] for our parallel application and aprotein annotation workflow [49] developed by London e-Science Centre for our hybrid workflow application Figure 2shows their simplified representationsThe number indicatedin the parentheses next to each node represents the length ofthe task (the number of instructions to execute) in Millionsof Instructions (MI) The input and output files of each taskrange from 10MB to 1024MB
We have performed experiments on 3 4 5 6 and 12resources whose characteristics are shown in Table 1 Wechoose the pricing model associated with Amazon EC2(httpawsamazoncomec2) for the processing costs of thedifferent classes of resources and the pricing model given byAmazon CloudFront (httpawsamazoncomcloudfront)for the costs of transferring data unit between resources
As for DVFS-MODPSO the parameter settings areSwarm size 119878119873119906119898 = 50 particles and the maximum numberof iterations 119866 = 100
We evaluated the performance of our proposed DVFS-MODPSO on all the workflow instances described previouslyDue to the lack of works considering both the heteroge-neous configuration of our cloud and the three QoS metrics(makespan cost energy) at the same time we comparedour results with those of the HEFT heuristic we haveimplemented HEFT is one of themost widely used heuristicsfor DAGs in distributed heterogeneous computing systemswhich optimize the makespan
Figures 3 4 and 5 show sample Pareto fronts obtainedwith the DVFS-MODPSO and the solution computed byHEFT for the synthetic workflow the hybrid workflow andthe parallel workflow respectively
As can be seen from these figures unlike the HEFTalgorithm which gives one solution our proposed approachprovides a set of non dominated solutions These resultsillustrate the basic multi-objective definitions provided inSection 41
Now for comparing these two approaches and in orderto analyze the effectiveness of our proposition in terms ofthe values of makespan cost and energy consumption wehave been inspired by the methodology described in [43]where we compare the solution provided by HEFT to onlyone solution of the Pareto front set computed by our proposedmulti-objective DVFS-MODPSO
The evaluation process follows the next steps For eachworkflow instance we perform a first resolution usingHEFTin order to get one solution 119878 After a second resolution iscomputed using our proposed approach to obtain a set EA ofPareto solutions Next we select one solution 1198781015840 from the setEA of Pareto solutions This solution is the closest one to 119878 inthe sense of Euclidean distance Finally a comparison is donebetween 119878 and 1198781015840
8 The Scientific World Journal
Swarm initialization with HEFT(1) For i = 1 to SNum (SNum is the size of particle swarm)
(a) For j = 1 to m (m the number of processors)(i) Randomly initialize VSL(j) (randomly choose the voltagefrequency of processor from the set of its operating points)
(b) Initialize 119878[119894] with HEFT heuristic (S is the swarm of particles)(c) Initialize the velocity V of each particle
119881 [119894] = 0 (119878 [119894])
(d) Initialize the Personal Best (pBest) of each particle119901119861119890119904119905[119894] = 119878[119894]
(e) Evaluate objectives of each particleEvaluate S[i]
(f) Initialize the Global Best particle (gBest) with the best one among the SNum particlesg119861119890119904119905 = Best particle found in 119878
(2) End For(3) Add the nondominated solutions found in S into EA (EA is the External Archive storing the pareto front)(4) Initialize the iteration number 119905 = 0(5) Repeat until 119905 gt 119866 (119866 is the maximum number of iterations)
(a) For 119894 = 1 to SNum (swarm size)(i) Randomly select the global best particle for S from the External Archive EA and store its position in gBest(ii) Calculate the new velocity 119881[119894] according to (22)(iii) Compute the new position of 119878[119894] according to (23)(iv) If (119905 lt 119866 lowast PMUT ) then (PMUT is the probability of mutation)
Perform mutation on 119878[119894](v) Evaluate S[i]
(b) End For(c) Update the personal best solution of each particle 119878[119894]
if 119878 [119894] ≼ 119901119861119890119904119905119904 [119894] or (119878 [119894] sim 119901119861119890119904119905119904 [119894])Then 119901119861119890119904119905119904 [119894] = 119878 [119894]
(d) Update the External Archive EA(i) Add all new non-dominated solution in S into EAif 119878 [119894] 997410 119909 forall119909 isin 119864119860Then 119864119860 = 119864119860 cup 119878 [119894](ii) Remove all particles dominated by 119878[119894] in EA
119864119860 = 119864119860 minus 119910 isin 119864119860 | 119910 lt 119878[119894]
(iii) If the archive is full then randomly select the article to be replaced from EA(6) Return the pareto front (the set of non-dominated solutions from S and EA)
Pseudocode 1 DVFS-MODPSO based workflow scheduling
The different results are displayed on Figures 6 7 8 and 9and on Tables 3 and 4 They show the improvement achievedby our proposed approach according to the structure of theworkflow applications and the number of processors
Table 3 and Figure 6 show that our proposed approachimproves on average the result obtained by HEFT for thesynthetic workflow instance The makespan is reduced by005 the cost is reduced by 993 and the energy con-sumption is reduced by 2640 Figure 6 shows detailedimprovements of the proposed approach according to thenumber of processors
Table 4 and Figure 7 illustrate the gain obtained by ourapproach according to the kind of real world applicationsTable 4 also shows the detailed improvements when scalingthe number of processors As can be seen the QoS metricsare improved over HEFT in average by 095 for themakespan 108 for cost and 812 for the energy con-sumption when using the hybrid workflow application andaverage improvements of 295 formakespan 2215 for costand 209 for energy consumption when using the parallelworkflow
These evaluations confirm the results obtained from thesynthetic workflow This means that by using our DVFS-MODPSO we are able to improve not only the cost and energybut also themakespan onwhichHEFT algorithm is supposedto provide good results These results remain true for all kindof workflow applications
In our experiments we limited the number of resourcesto 12 as we found that for these kinds of workflowapplications some resources are not used at all Figure 8shows the resource loads when scheduling the syntheticworkflow on 12 resources Furthermore even though thenumber of resources increased the total cost and total energyconsumption could not always decrease as illustrated inFigure 9 From this figure we can see that the QoS metrics(1) firstly decreases as the number of resources increasesuntil achieving the number 6 This can be explained by thefact that when increasing the number of resources thereare fewer tasks executed in a resource therefore tasks canextend their execution times and the resource have more ofchances to scale down their voltages and frequencies whichcan be very effective in reducing total energy consumption
The Scientific World Journal 9
1 2 3 4
5
6 7
8 9 10
11
12
14
13
15
(300000) (600000) (600000) (900000)
(300000)
(150000) (150000)
(300000) (300000) (300000)
(600000)
(300000)
(150000)
(300000)
(600000)
SignalP COILS2 SEG PROSITE
TMHMM
ProsperoHMMer
PSI-BLASTBLAST
IMPALA
PSI-PRED
3D-PSSMSummary
Genomesummary
SCOP
(a)
1 3 5 7
9
11 1210
(400000) (400000) (400000) (400000)
(300000)
(500000) (500000) (500000)
Softmea
2 4 6 8(600000) (600000) (600000) (600000)
Reslice
14 1513
(600000) (600000) (600000)
Convert
Reslice Reslice Reslice
Slicer Slicer
Convert Convert
Slicer
Align wapAlign wapAlign wapAlign wap
(b)
Figure 2 Parallel and Hybrid workflows
10 The Scientific World Journal
75 80 85 90 95 100 8090
100110
120130
140
2060
100140180220
Ener
gy
DVFS-MODPSOHEFT
MakespanCost
Figure 3 Obtained non-dominated solutions for the syntheticworkflow
Ener
gy
DVFS-MODPSOHEFT
MakespanCost
200 400 600 800 100012001400160018002000 400800
12001600
2000
400200
600800
100012001400160018002000
Figure 4 Obtained non-dominated solutions for the hybrid work-flow
(2) After achieving 6 resources the QoS metrics begin torise the reason for this is that time executions are dominatedby interprocessors communications hence decreasing theopportunities for scaling down voltages and frequencies ofresources Consequently the threshold numbers of resourcesthat minimized the QoS metrics could be obtained
6 Conclusion
In this paper we propose a new algorithm called DVFS-Multi-Objective Discrete Particle Swarm Optimization(DVFS-MODPSO) for workflow scheduling in distributedenvironments such as cloud computing infrastructures DV-FS-MODPSO simultaneously optimizes several conflictingobjectives namely the makespan cost and energy in adiscrete space It produces a set of non-dominated solutionsthus providing more flexibility for users to assess theirpreferences and select a schedule that meets their QoS
Ener
gy
DVFS-MODPSOHEFT
Makespan Cost
400200600800
10001200140016001800
240020002200
450 500 550 600 650 700 7501200
14001600
18002000
Figure 5 Obtained non-dominated solutions for the parallel work-flow
0
10
20
30
40
50
3 4 5 6 12
MakespanCostEnergy
Figure 6 Gain over HEFT () according to the number ofprocessors
requirements Our approach exploits the heterogeneityand the marketization of cloud resources in order to findsolutions that optimize makespan and cost Furthermore ituses the Dynamic Voltage and Frequency Scaling (DVFS)technique to reduce energy consumption
We have evaluated our algorithm by simulating it withboth synthetic and real world scientific workflow applicationshaving different structures The results show that the pro-posedDVFS-MODPSO is able to produce a set of good Paretooptimal solutions The results also show that our approachprovides significant improvement not only in terms of the
The Scientific World Journal 11
0
5
10
15
20
25
HybridParallel
Makespan Cost Energy
Figure 7 Improvement according to real world applications
R17
R20
R37
R446
R50
R60
R733
R80
R90
R107
R110
R120
Figure 8 Resource loads for synthetic workflow
Table 3 Improvement for synthetic workflow
Num of proc Gain over Heft ()Makespan Cost Energy
3 0 2126 40084 022 464 15355 0 756 17266 003 652 267512 0 969 3255Average 005 993 2640
cost and the energy consumption but also in term of themakespan for which HEFT algorithm is supposed to givebetter results
Multi-objective optimization in cloud workflow schedul-ing is not a mature domain Most of the existing worksattempt to minimize either the makespan or cost Howeverwe plan to consider other objectives such as reliabilitysecurity in addition to the energy consumption We also
0
50
100
150
200
250
3 4 5 6 12
MakespanCostEnergy
Figure 9 QoS metrics evolutions according to resource numbersfor the synthetic workflow
Table 4 Improvement according to real world applications
Appli Num of proc Gain over Heft ()Makespan Cost Energy
Hybrid
3 405 1155 04024 028 0566 05075 000 1347 11246 000 0117 126512 043 2214 0761
Average 095 1080 0812
Parallel
3 235 5640 25524 1147 2094 32715 000 2142 11666 093 1170 315412 000 031 306
Average 295 2215 209
intend to apply our algorithm in a real world cloud orintegrate it in existing cloud toolkits such as Cloudbus forcomparing with other algorithms
References
[1] D Thain T Tannenbaum and M Livny Condor and the GridGrid Computing Making the Global Infrastructure a RealityJohn Wiley amp Sons Hoboken NJ USA 2003
[2] R Buyya and S Venugopal ldquoThe gridbus toolkit for serviceoriented grid and utility computing an overview and statusreportrdquo in Proceedings of the 1st IEEE International Workshopon Grid Economics and Business Models (GECON rsquo04) pp 19ndash36 IEEE CS Seoul Republic of Korea April 2004
[3] S McGough L Young A Afzal S Newhouse and J Darling-ton ldquoWorkflow enactment in ICENIrdquo in Proceedings of the UK
12 The Scientific World Journal
e-Science All Hands Meeting pp 894ndash900 IOP NottinghamUK September 2004
[4] E Deelman J Blythe Y Gil et al ldquoPegasus mapping scientificworkflows onto the gridrdquo in Grid Computing 2nd EuropeanAcrossGrids Conference AxGrids 2004 Nicosia Cyprus January28ndash30 2004 vol 3165 of Lecture Notes in Computer Science pp11ndash20 Springer New York NY USA 2004
[5] R Buyya S Pandey and C Vecchiola ldquoCloudbus toolkitfor market-oriented cloud computingrdquo in Cloud ComputingProceedings of the 1st International Conference CloudCom 2009Beijing China December 1ndash4 2009 vol 5931 of Lecture Notesin Computer Science pp 24ndash44 Springer New York NY USA2009
[6] Y Yang K Liu J Chen X Liu D Yuan and H Jin ldquoAnalgorithm in SwinDeW-C for scheduling transaction-intensivecost-constrained cloud workflowsrdquo in Proceedings of the 4thIEEE International Conference on eScience (eScience rsquo08) pp374ndash375 Indianapolis Ind USA December 2008
[7] L Ramakrishnan C Koelbel Y Kee et al ldquoVGrADS enablinge-Scienceworkflows on grids and cloudswith fault tolerancerdquo inProceedings of the Conference on High Performance ComputingNetworking Storage and Analysis (SC rsquo09) November 2009
[8] M L Pinedo Scheduling Theory Algorithms and SystemsSpringer Berlin 2008
[9] S Pandey L Wu S M Guru and R Buyya ldquoA particle swarmoptimization-based heuristic for scheduling workflow applica-tions in cloud computing environmentsrdquo in Proceedings of the24th IEEE International Conference on Advanced InformationNetworking and Applications (AINA rsquo10) pp 400ndash407 PerthAustralia April 2010
[10] S Abrishami and M Naghibzadeh ldquoDeadline-constrainedworkflow scheduling in software as a service cloudrdquo ScientiaIranica vol 19 no 3 pp 680ndash689 2012
[11] S Selvarani and G S Sadhasivam ldquoImproved cost-based algo-rithm for task scheduling in cloud computingrdquo in Proceedings ofthe IEEE International Conference on Computational Intelligenceand Computing Research (ICCIC rsquo10) pp 1ndash5 CoimbatoreIndia December 2010
[12] A Verma and S Kaushal ldquoDeadline and budget distributionbased cost-time optimization workflow scheduling algorithmfor cloudrdquo inProceedings of the IJCAon International Conferenceon Recent Advances and Future Trends in Information Technol-ogy (iRAFIT rsquo12) iRAFIT(7) pp 1ndash4 April 2012
[13] R C Correa A Ferreira and P Rebreyend ldquoIntegrating listheuristics into genetic algorithms for multiprocessor schedul-ingrdquo in Proceedings of the 8th IEEE Symposium on Parallel andDistributed Processing (SPDP rsquo96) pp 462ndash469 IEEEComputerSociety Washington DC USA October 1996
[14] E S H Hou N Ansari and H Ren ldquoGenetic algorithm formultiprocessor schedulingrdquo IEEE Transactions on Parallel andDistributed Systems vol 5 no 2 pp 113ndash120 1994
[15] M Grajcar ldquoGenetic list scheduling algorithm for schedulingand allocation on a loosely coupled heterogeneous multi-processor systemrdquo in Proceedings of the 36th Annual DesignAutomation Conference (DAC rsquo99) pp 280ndash285 ACM PressNew York NY USA June 1999
[16] R Sakellariou and H Zhao ldquoA hybrid heuristic for DAGscheduling on heterogeneous systemsrdquo in Proceedings of the13th Heterogeneous Computing Workshop pp 111ndash123 IEEEComputer Society Washington DC USA 2004
[17] H Topcuoglu S Hariri and M Y Wu ldquoPerformance-effectiveand low-complexity task scheduling for heterogeneous comput-ingrdquo IEEE Transactions on Parallel and Distributed Systems vol13 no 3 pp 260ndash274 2002
[18] J Yu and R Buyya ldquoWorkflow scheduling algorithms for gridcomputingrdquo in Metaheuristics for Scheduling in DistributedComputing Environments F Xhafa and A Abraham EdsSpringer Berlin Germany 2008
[19] WN Chen and J Zhang ldquoAn ant colony optimization approachto a grid workflow scheduling problem with various QoSrequirementsrdquo IEEE Transactions on Systems Man and Cyber-netics C vol 39 no 1 pp 29ndash43 2009
[20] X Liu J Chen Z Wu Z Ni D Yuan and Y Yang ldquoHandlingrecoverable temporal violations in scientific workflow systemsa workflow rescheduling based strategyrdquo in Proceedings of the10th IEEEACM International Symposium on Cluster Cloudand Grid Computing (CCGrid rsquo10) pp 534ndash537 MelbourneAustralia May 2010
[21] X Liu Y Yang Y Jiang and J Chen ldquoPreventing temporalviolations in scientific workflows where and howrdquo IEEE Trans-actions on Software Engineering vol 37 no 6 pp 805ndash825 2011
[22] J Yu and R Buyya ldquoScheduling scientific workflow applicationswith deadline and budget constraints using genetic algorithmsrdquoScientific Programming vol 14 no 3-4 pp 217ndash230 2006
[23] L Benini and G Micheli Dynamic Power Management DesignTechniques and CAD Tools Kluwer Academic New York NYUSA 1998
[24] S Irani S Shukla and R Gupta ldquoOnline strategies for dynamicpower management in system with multiple power-savingstatesrdquo ACM Transactions on Embedded Computing Systemsvol 2 no 3 pp 325ndash346 2003
[25] M T Schmitz B M Al-Hashimi and P Eles System-Level Design Techniques for Energy-Efficient Embedded SystemsSpringer New York NY USA 2005
[26] S U Khan and I Ahmad ldquoA cooperative game theoreticaltechnique for joint optimization of energy consumption andresponse time in computational gridsrdquo IEEE Transactions onParallel and Distributed Systems vol 20 no 3 pp 346ndash3602009
[27] Y Li Y Liu and D Qian ldquoA heuristic energy-aware schedulingalgorithm for heterogeneous clustersrdquo in Proceedings of the 15thInternational Conference on Parallel and Distributed Systems(ICPADS rsquo09) pp 407ndash413 Shenzhen China December 2009
[28] L K Goh B Veeravalli and S Viswanathan ldquoDesign of fast andefficient energy-aware gradient-based scheduling algorithmsheterogeneous embeddedmultiprocessor systemsrdquo IEEE Trans-actions on Parallel and Distributed Systems vol 20 no 1 pp 1ndash12 2009
[29] F Yao A Demers and S Shenker ldquoA scheduling model forreduced CPU energyrdquo in Proceedings of the 36th IEEE AnnualSymposium on Foundations of Computer Science pp 374ndash382Milwaukee Wis USA October 1995
[30] Y Shin and K Choi ldquoPower conscious fixed priority schedulingfor hard real-time systemsrdquo in Proceedings of the 36th AnnualDesign Automation Conference (DAC rsquo99) pp 134ndash139 NewOrleans La USA June 1999
[31] I Hong D Kirovski G QuM Potkonjak andM B SrivastavaldquoPower optimization of variable-voltage core-based systemsrdquoIEEE Transactions on Computer-Aided Design of IntegratedCircuits and Systems vol 18 no 12 pp 1702ndash1714 1999
[32] P Pillai and K G Shin ldquoReal-time dynamic voltage scaling forlow-power embedded operating systemsrdquo in Proceedings of the
The Scientific World Journal 13
ACM Symposium on Operating Systems Principles pp 89ndash102October 2001
[33] H Aydin R Melhem D Mosse and P Mejia-Alvarez ldquoPower-aware scheduling for periodic real-time tasksrdquo IEEE Transac-tions on Computers vol 53 no 5 pp 584ndash600 2004
[34] J Zhuo and C Chakrabarti ldquoAn efficient dynamic task schedul-ing algorithm for battery powered DVS systemsrdquo in Proceedingsof the Asian South Pacific Design Automation Conference pp846ndash849 January 2005
[35] R Jejurikar and R Gupta ldquoDynamic slack reclamation withprocrastination scheduling in real-time embedded systemsrdquo inProceedings of the 42nd Design Automation Conference (DACrsquo05) pp 111ndash116 June 2005
[36] J Luo andN K Jha ldquoPower-profile driven variable voltage scal-ing for heterogeneous distributed real-time embedded systemsrdquoin Proceedings of the International Conference on VLSI Designpp 369ndash375 January 2003
[37] M T Schmitz and B M Al-Hashimi ldquoConsidering powervariations of DVS processing elements for energy minimisationin distributed systemsrdquo in Proceedings of the International Sym-posium on System Synthesis (ISSS rsquo01) pp 250ndash255 MontrealCanada October 2001
[38] Y Zhang X Hu and D Z Chen ldquoTask scheduling and voltageselection for energy minimizationrdquo in Proceedings of the 39thAnnual Design Automation Conference (DAC rsquo02) pp 183ndash188June 2002
[39] D Zhu R Melhem and B R Childers ldquoScheduling withdynamic voltagespeed adjustment using slack reclamationin multiprocessor real-time systemsrdquo IEEE Transactions onParallel and Distributed Systems vol 14 no 7 pp 686ndash7002003
[40] D Zhu D Mosse and R Melhem ldquoPower-aware schedulingfor ANDOR graphs in real-time systemsrdquo IEEE Transactionson Parallel and Distributed Systems vol 15 no 9 pp 849ndash8642004
[41] J Kang and S Ranka ldquoDVS based energy minimizationalgorithm for parallel machinesrdquo in Proceedings of the IEEEInternational Parallel and Distributed Processing Symposium(IPDPS rsquo08) pp 1ndash12 Miami Fla USA April 2008
[42] B Rountree D K Lowenthal S Funk V W Freeh B R deSupinski and M Schulz ldquoBounding energy consumption inlarge-scale MPI programsrdquo in Proceedings of the ACMIEEEConference on Supercomputing (SC rsquo07) ACM November 2007
[43] L Wang G von Laszewski J Dayal and F Wang ldquoTowardsenergy aware scheduling for precedence constrained paralleltasks in a cluster with DVFSrdquo in Proceedings of the 10thIEEEACM International Symposium on Cluster Cloud andGrid Computing (CCGrid rsquo10) pp 368ndash377 May 2010
[44] Y C Lee and A Y Zomaya ldquoEnergy conscious schedulingfor distributed computing systems under different operatingconditionsrdquo IEEE Transactions on Parallel and DistributedSystems vol 22 no 8 pp 1374ndash1381 2011
[45] M Mezmaz Y C Lee N Melab E Talbi and A Y ZomayaldquoA bi-objective hybrid genetic algorithm to minimize energyconsumption and makespan for precedence-constrained appli-cations using dynamic voltage scalingrdquo in Proceedings of theIEEE Congress on Evolutionary Computation (CEC rsquo10) pp 1ndash8Barcelona Spain July 2010
[46] M Mezmaz N Melab Y Kessaci et al ldquoA parallel bi-objectivehybrid metaheuristic for energy-aware scheduling for cloudcomputing systemsrdquo Journal of Parallel and Distributed Com-puting vol 71 no 11 pp 1497ndash1508 2011
[47] J Kennedy and R Eberhart ldquoParticle swarm optimizationrdquoin Proceedings of the IEEE International Conference on NeuralNetworks vol 4 pp 1942ndash1948 Perth Australia December1995
[48] Y Zhao M Wilde I Foster et al ldquoGrid middleware servicesfor virtual data discovery composition and integrationrdquo inProceedings of the 2nd Workshop on Middleware for GridComputing (MGC rsquo04) pp 57ndash62 Ontario Canada October2004
[49] A OrsquoBrien S Newhouse and J Darlington ldquoMapping ofscientific workflow within the e-protein project to distributedresourcesrdquo inProceedings of theUK e-Science All HandsMeetingNottingham UK September 2004
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
8 The Scientific World Journal
Swarm initialization with HEFT(1) For i = 1 to SNum (SNum is the size of particle swarm)
(a) For j = 1 to m (m the number of processors)(i) Randomly initialize VSL(j) (randomly choose the voltagefrequency of processor from the set of its operating points)
(b) Initialize 119878[119894] with HEFT heuristic (S is the swarm of particles)(c) Initialize the velocity V of each particle
119881 [119894] = 0 (119878 [119894])
(d) Initialize the Personal Best (pBest) of each particle119901119861119890119904119905[119894] = 119878[119894]
(e) Evaluate objectives of each particleEvaluate S[i]
(f) Initialize the Global Best particle (gBest) with the best one among the SNum particlesg119861119890119904119905 = Best particle found in 119878
(2) End For(3) Add the nondominated solutions found in S into EA (EA is the External Archive storing the pareto front)(4) Initialize the iteration number 119905 = 0(5) Repeat until 119905 gt 119866 (119866 is the maximum number of iterations)
(a) For 119894 = 1 to SNum (swarm size)(i) Randomly select the global best particle for S from the External Archive EA and store its position in gBest(ii) Calculate the new velocity 119881[119894] according to (22)(iii) Compute the new position of 119878[119894] according to (23)(iv) If (119905 lt 119866 lowast PMUT ) then (PMUT is the probability of mutation)
Perform mutation on 119878[119894](v) Evaluate S[i]
(b) End For(c) Update the personal best solution of each particle 119878[119894]
if 119878 [119894] ≼ 119901119861119890119904119905119904 [119894] or (119878 [119894] sim 119901119861119890119904119905119904 [119894])Then 119901119861119890119904119905119904 [119894] = 119878 [119894]
(d) Update the External Archive EA(i) Add all new non-dominated solution in S into EAif 119878 [119894] 997410 119909 forall119909 isin 119864119860Then 119864119860 = 119864119860 cup 119878 [119894](ii) Remove all particles dominated by 119878[119894] in EA
119864119860 = 119864119860 minus 119910 isin 119864119860 | 119910 lt 119878[119894]
(iii) If the archive is full then randomly select the article to be replaced from EA(6) Return the pareto front (the set of non-dominated solutions from S and EA)
Pseudocode 1 DVFS-MODPSO based workflow scheduling
The different results are displayed on Figures 6 7 8 and 9and on Tables 3 and 4 They show the improvement achievedby our proposed approach according to the structure of theworkflow applications and the number of processors
Table 3 and Figure 6 show that our proposed approachimproves on average the result obtained by HEFT for thesynthetic workflow instance The makespan is reduced by005 the cost is reduced by 993 and the energy con-sumption is reduced by 2640 Figure 6 shows detailedimprovements of the proposed approach according to thenumber of processors
Table 4 and Figure 7 illustrate the gain obtained by ourapproach according to the kind of real world applicationsTable 4 also shows the detailed improvements when scalingthe number of processors As can be seen the QoS metricsare improved over HEFT in average by 095 for themakespan 108 for cost and 812 for the energy con-sumption when using the hybrid workflow application andaverage improvements of 295 formakespan 2215 for costand 209 for energy consumption when using the parallelworkflow
These evaluations confirm the results obtained from thesynthetic workflow This means that by using our DVFS-MODPSO we are able to improve not only the cost and energybut also themakespan onwhichHEFT algorithm is supposedto provide good results These results remain true for all kindof workflow applications
In our experiments we limited the number of resourcesto 12 as we found that for these kinds of workflowapplications some resources are not used at all Figure 8shows the resource loads when scheduling the syntheticworkflow on 12 resources Furthermore even though thenumber of resources increased the total cost and total energyconsumption could not always decrease as illustrated inFigure 9 From this figure we can see that the QoS metrics(1) firstly decreases as the number of resources increasesuntil achieving the number 6 This can be explained by thefact that when increasing the number of resources thereare fewer tasks executed in a resource therefore tasks canextend their execution times and the resource have more ofchances to scale down their voltages and frequencies whichcan be very effective in reducing total energy consumption
The Scientific World Journal 9
1 2 3 4
5
6 7
8 9 10
11
12
14
13
15
(300000) (600000) (600000) (900000)
(300000)
(150000) (150000)
(300000) (300000) (300000)
(600000)
(300000)
(150000)
(300000)
(600000)
SignalP COILS2 SEG PROSITE
TMHMM
ProsperoHMMer
PSI-BLASTBLAST
IMPALA
PSI-PRED
3D-PSSMSummary
Genomesummary
SCOP
(a)
1 3 5 7
9
11 1210
(400000) (400000) (400000) (400000)
(300000)
(500000) (500000) (500000)
Softmea
2 4 6 8(600000) (600000) (600000) (600000)
Reslice
14 1513
(600000) (600000) (600000)
Convert
Reslice Reslice Reslice
Slicer Slicer
Convert Convert
Slicer
Align wapAlign wapAlign wapAlign wap
(b)
Figure 2 Parallel and Hybrid workflows
10 The Scientific World Journal
75 80 85 90 95 100 8090
100110
120130
140
2060
100140180220
Ener
gy
DVFS-MODPSOHEFT
MakespanCost
Figure 3 Obtained non-dominated solutions for the syntheticworkflow
Ener
gy
DVFS-MODPSOHEFT
MakespanCost
200 400 600 800 100012001400160018002000 400800
12001600
2000
400200
600800
100012001400160018002000
Figure 4 Obtained non-dominated solutions for the hybrid work-flow
(2) After achieving 6 resources the QoS metrics begin torise the reason for this is that time executions are dominatedby interprocessors communications hence decreasing theopportunities for scaling down voltages and frequencies ofresources Consequently the threshold numbers of resourcesthat minimized the QoS metrics could be obtained
6 Conclusion
In this paper we propose a new algorithm called DVFS-Multi-Objective Discrete Particle Swarm Optimization(DVFS-MODPSO) for workflow scheduling in distributedenvironments such as cloud computing infrastructures DV-FS-MODPSO simultaneously optimizes several conflictingobjectives namely the makespan cost and energy in adiscrete space It produces a set of non-dominated solutionsthus providing more flexibility for users to assess theirpreferences and select a schedule that meets their QoS
Ener
gy
DVFS-MODPSOHEFT
Makespan Cost
400200600800
10001200140016001800
240020002200
450 500 550 600 650 700 7501200
14001600
18002000
Figure 5 Obtained non-dominated solutions for the parallel work-flow
0
10
20
30
40
50
3 4 5 6 12
MakespanCostEnergy
Figure 6 Gain over HEFT () according to the number ofprocessors
requirements Our approach exploits the heterogeneityand the marketization of cloud resources in order to findsolutions that optimize makespan and cost Furthermore ituses the Dynamic Voltage and Frequency Scaling (DVFS)technique to reduce energy consumption
We have evaluated our algorithm by simulating it withboth synthetic and real world scientific workflow applicationshaving different structures The results show that the pro-posedDVFS-MODPSO is able to produce a set of good Paretooptimal solutions The results also show that our approachprovides significant improvement not only in terms of the
The Scientific World Journal 11
0
5
10
15
20
25
HybridParallel
Makespan Cost Energy
Figure 7 Improvement according to real world applications
R17
R20
R37
R446
R50
R60
R733
R80
R90
R107
R110
R120
Figure 8 Resource loads for synthetic workflow
Table 3 Improvement for synthetic workflow
Num of proc Gain over Heft ()Makespan Cost Energy
3 0 2126 40084 022 464 15355 0 756 17266 003 652 267512 0 969 3255Average 005 993 2640
cost and the energy consumption but also in term of themakespan for which HEFT algorithm is supposed to givebetter results
Multi-objective optimization in cloud workflow schedul-ing is not a mature domain Most of the existing worksattempt to minimize either the makespan or cost Howeverwe plan to consider other objectives such as reliabilitysecurity in addition to the energy consumption We also
0
50
100
150
200
250
3 4 5 6 12
MakespanCostEnergy
Figure 9 QoS metrics evolutions according to resource numbersfor the synthetic workflow
Table 4 Improvement according to real world applications
Appli Num of proc Gain over Heft ()Makespan Cost Energy
Hybrid
3 405 1155 04024 028 0566 05075 000 1347 11246 000 0117 126512 043 2214 0761
Average 095 1080 0812
Parallel
3 235 5640 25524 1147 2094 32715 000 2142 11666 093 1170 315412 000 031 306
Average 295 2215 209
intend to apply our algorithm in a real world cloud orintegrate it in existing cloud toolkits such as Cloudbus forcomparing with other algorithms
References
[1] D Thain T Tannenbaum and M Livny Condor and the GridGrid Computing Making the Global Infrastructure a RealityJohn Wiley amp Sons Hoboken NJ USA 2003
[2] R Buyya and S Venugopal ldquoThe gridbus toolkit for serviceoriented grid and utility computing an overview and statusreportrdquo in Proceedings of the 1st IEEE International Workshopon Grid Economics and Business Models (GECON rsquo04) pp 19ndash36 IEEE CS Seoul Republic of Korea April 2004
[3] S McGough L Young A Afzal S Newhouse and J Darling-ton ldquoWorkflow enactment in ICENIrdquo in Proceedings of the UK
12 The Scientific World Journal
e-Science All Hands Meeting pp 894ndash900 IOP NottinghamUK September 2004
[4] E Deelman J Blythe Y Gil et al ldquoPegasus mapping scientificworkflows onto the gridrdquo in Grid Computing 2nd EuropeanAcrossGrids Conference AxGrids 2004 Nicosia Cyprus January28ndash30 2004 vol 3165 of Lecture Notes in Computer Science pp11ndash20 Springer New York NY USA 2004
[5] R Buyya S Pandey and C Vecchiola ldquoCloudbus toolkitfor market-oriented cloud computingrdquo in Cloud ComputingProceedings of the 1st International Conference CloudCom 2009Beijing China December 1ndash4 2009 vol 5931 of Lecture Notesin Computer Science pp 24ndash44 Springer New York NY USA2009
[6] Y Yang K Liu J Chen X Liu D Yuan and H Jin ldquoAnalgorithm in SwinDeW-C for scheduling transaction-intensivecost-constrained cloud workflowsrdquo in Proceedings of the 4thIEEE International Conference on eScience (eScience rsquo08) pp374ndash375 Indianapolis Ind USA December 2008
[7] L Ramakrishnan C Koelbel Y Kee et al ldquoVGrADS enablinge-Scienceworkflows on grids and cloudswith fault tolerancerdquo inProceedings of the Conference on High Performance ComputingNetworking Storage and Analysis (SC rsquo09) November 2009
[8] M L Pinedo Scheduling Theory Algorithms and SystemsSpringer Berlin 2008
[9] S Pandey L Wu S M Guru and R Buyya ldquoA particle swarmoptimization-based heuristic for scheduling workflow applica-tions in cloud computing environmentsrdquo in Proceedings of the24th IEEE International Conference on Advanced InformationNetworking and Applications (AINA rsquo10) pp 400ndash407 PerthAustralia April 2010
[10] S Abrishami and M Naghibzadeh ldquoDeadline-constrainedworkflow scheduling in software as a service cloudrdquo ScientiaIranica vol 19 no 3 pp 680ndash689 2012
[11] S Selvarani and G S Sadhasivam ldquoImproved cost-based algo-rithm for task scheduling in cloud computingrdquo in Proceedings ofthe IEEE International Conference on Computational Intelligenceand Computing Research (ICCIC rsquo10) pp 1ndash5 CoimbatoreIndia December 2010
[12] A Verma and S Kaushal ldquoDeadline and budget distributionbased cost-time optimization workflow scheduling algorithmfor cloudrdquo inProceedings of the IJCAon International Conferenceon Recent Advances and Future Trends in Information Technol-ogy (iRAFIT rsquo12) iRAFIT(7) pp 1ndash4 April 2012
[13] R C Correa A Ferreira and P Rebreyend ldquoIntegrating listheuristics into genetic algorithms for multiprocessor schedul-ingrdquo in Proceedings of the 8th IEEE Symposium on Parallel andDistributed Processing (SPDP rsquo96) pp 462ndash469 IEEEComputerSociety Washington DC USA October 1996
[14] E S H Hou N Ansari and H Ren ldquoGenetic algorithm formultiprocessor schedulingrdquo IEEE Transactions on Parallel andDistributed Systems vol 5 no 2 pp 113ndash120 1994
[15] M Grajcar ldquoGenetic list scheduling algorithm for schedulingand allocation on a loosely coupled heterogeneous multi-processor systemrdquo in Proceedings of the 36th Annual DesignAutomation Conference (DAC rsquo99) pp 280ndash285 ACM PressNew York NY USA June 1999
[16] R Sakellariou and H Zhao ldquoA hybrid heuristic for DAGscheduling on heterogeneous systemsrdquo in Proceedings of the13th Heterogeneous Computing Workshop pp 111ndash123 IEEEComputer Society Washington DC USA 2004
[17] H Topcuoglu S Hariri and M Y Wu ldquoPerformance-effectiveand low-complexity task scheduling for heterogeneous comput-ingrdquo IEEE Transactions on Parallel and Distributed Systems vol13 no 3 pp 260ndash274 2002
[18] J Yu and R Buyya ldquoWorkflow scheduling algorithms for gridcomputingrdquo in Metaheuristics for Scheduling in DistributedComputing Environments F Xhafa and A Abraham EdsSpringer Berlin Germany 2008
[19] WN Chen and J Zhang ldquoAn ant colony optimization approachto a grid workflow scheduling problem with various QoSrequirementsrdquo IEEE Transactions on Systems Man and Cyber-netics C vol 39 no 1 pp 29ndash43 2009
[20] X Liu J Chen Z Wu Z Ni D Yuan and Y Yang ldquoHandlingrecoverable temporal violations in scientific workflow systemsa workflow rescheduling based strategyrdquo in Proceedings of the10th IEEEACM International Symposium on Cluster Cloudand Grid Computing (CCGrid rsquo10) pp 534ndash537 MelbourneAustralia May 2010
[21] X Liu Y Yang Y Jiang and J Chen ldquoPreventing temporalviolations in scientific workflows where and howrdquo IEEE Trans-actions on Software Engineering vol 37 no 6 pp 805ndash825 2011
[22] J Yu and R Buyya ldquoScheduling scientific workflow applicationswith deadline and budget constraints using genetic algorithmsrdquoScientific Programming vol 14 no 3-4 pp 217ndash230 2006
[23] L Benini and G Micheli Dynamic Power Management DesignTechniques and CAD Tools Kluwer Academic New York NYUSA 1998
[24] S Irani S Shukla and R Gupta ldquoOnline strategies for dynamicpower management in system with multiple power-savingstatesrdquo ACM Transactions on Embedded Computing Systemsvol 2 no 3 pp 325ndash346 2003
[25] M T Schmitz B M Al-Hashimi and P Eles System-Level Design Techniques for Energy-Efficient Embedded SystemsSpringer New York NY USA 2005
[26] S U Khan and I Ahmad ldquoA cooperative game theoreticaltechnique for joint optimization of energy consumption andresponse time in computational gridsrdquo IEEE Transactions onParallel and Distributed Systems vol 20 no 3 pp 346ndash3602009
[27] Y Li Y Liu and D Qian ldquoA heuristic energy-aware schedulingalgorithm for heterogeneous clustersrdquo in Proceedings of the 15thInternational Conference on Parallel and Distributed Systems(ICPADS rsquo09) pp 407ndash413 Shenzhen China December 2009
[28] L K Goh B Veeravalli and S Viswanathan ldquoDesign of fast andefficient energy-aware gradient-based scheduling algorithmsheterogeneous embeddedmultiprocessor systemsrdquo IEEE Trans-actions on Parallel and Distributed Systems vol 20 no 1 pp 1ndash12 2009
[29] F Yao A Demers and S Shenker ldquoA scheduling model forreduced CPU energyrdquo in Proceedings of the 36th IEEE AnnualSymposium on Foundations of Computer Science pp 374ndash382Milwaukee Wis USA October 1995
[30] Y Shin and K Choi ldquoPower conscious fixed priority schedulingfor hard real-time systemsrdquo in Proceedings of the 36th AnnualDesign Automation Conference (DAC rsquo99) pp 134ndash139 NewOrleans La USA June 1999
[31] I Hong D Kirovski G QuM Potkonjak andM B SrivastavaldquoPower optimization of variable-voltage core-based systemsrdquoIEEE Transactions on Computer-Aided Design of IntegratedCircuits and Systems vol 18 no 12 pp 1702ndash1714 1999
[32] P Pillai and K G Shin ldquoReal-time dynamic voltage scaling forlow-power embedded operating systemsrdquo in Proceedings of the
The Scientific World Journal 13
ACM Symposium on Operating Systems Principles pp 89ndash102October 2001
[33] H Aydin R Melhem D Mosse and P Mejia-Alvarez ldquoPower-aware scheduling for periodic real-time tasksrdquo IEEE Transac-tions on Computers vol 53 no 5 pp 584ndash600 2004
[34] J Zhuo and C Chakrabarti ldquoAn efficient dynamic task schedul-ing algorithm for battery powered DVS systemsrdquo in Proceedingsof the Asian South Pacific Design Automation Conference pp846ndash849 January 2005
[35] R Jejurikar and R Gupta ldquoDynamic slack reclamation withprocrastination scheduling in real-time embedded systemsrdquo inProceedings of the 42nd Design Automation Conference (DACrsquo05) pp 111ndash116 June 2005
[36] J Luo andN K Jha ldquoPower-profile driven variable voltage scal-ing for heterogeneous distributed real-time embedded systemsrdquoin Proceedings of the International Conference on VLSI Designpp 369ndash375 January 2003
[37] M T Schmitz and B M Al-Hashimi ldquoConsidering powervariations of DVS processing elements for energy minimisationin distributed systemsrdquo in Proceedings of the International Sym-posium on System Synthesis (ISSS rsquo01) pp 250ndash255 MontrealCanada October 2001
[38] Y Zhang X Hu and D Z Chen ldquoTask scheduling and voltageselection for energy minimizationrdquo in Proceedings of the 39thAnnual Design Automation Conference (DAC rsquo02) pp 183ndash188June 2002
[39] D Zhu R Melhem and B R Childers ldquoScheduling withdynamic voltagespeed adjustment using slack reclamationin multiprocessor real-time systemsrdquo IEEE Transactions onParallel and Distributed Systems vol 14 no 7 pp 686ndash7002003
[40] D Zhu D Mosse and R Melhem ldquoPower-aware schedulingfor ANDOR graphs in real-time systemsrdquo IEEE Transactionson Parallel and Distributed Systems vol 15 no 9 pp 849ndash8642004
[41] J Kang and S Ranka ldquoDVS based energy minimizationalgorithm for parallel machinesrdquo in Proceedings of the IEEEInternational Parallel and Distributed Processing Symposium(IPDPS rsquo08) pp 1ndash12 Miami Fla USA April 2008
[42] B Rountree D K Lowenthal S Funk V W Freeh B R deSupinski and M Schulz ldquoBounding energy consumption inlarge-scale MPI programsrdquo in Proceedings of the ACMIEEEConference on Supercomputing (SC rsquo07) ACM November 2007
[43] L Wang G von Laszewski J Dayal and F Wang ldquoTowardsenergy aware scheduling for precedence constrained paralleltasks in a cluster with DVFSrdquo in Proceedings of the 10thIEEEACM International Symposium on Cluster Cloud andGrid Computing (CCGrid rsquo10) pp 368ndash377 May 2010
[44] Y C Lee and A Y Zomaya ldquoEnergy conscious schedulingfor distributed computing systems under different operatingconditionsrdquo IEEE Transactions on Parallel and DistributedSystems vol 22 no 8 pp 1374ndash1381 2011
[45] M Mezmaz Y C Lee N Melab E Talbi and A Y ZomayaldquoA bi-objective hybrid genetic algorithm to minimize energyconsumption and makespan for precedence-constrained appli-cations using dynamic voltage scalingrdquo in Proceedings of theIEEE Congress on Evolutionary Computation (CEC rsquo10) pp 1ndash8Barcelona Spain July 2010
[46] M Mezmaz N Melab Y Kessaci et al ldquoA parallel bi-objectivehybrid metaheuristic for energy-aware scheduling for cloudcomputing systemsrdquo Journal of Parallel and Distributed Com-puting vol 71 no 11 pp 1497ndash1508 2011
[47] J Kennedy and R Eberhart ldquoParticle swarm optimizationrdquoin Proceedings of the IEEE International Conference on NeuralNetworks vol 4 pp 1942ndash1948 Perth Australia December1995
[48] Y Zhao M Wilde I Foster et al ldquoGrid middleware servicesfor virtual data discovery composition and integrationrdquo inProceedings of the 2nd Workshop on Middleware for GridComputing (MGC rsquo04) pp 57ndash62 Ontario Canada October2004
[49] A OrsquoBrien S Newhouse and J Darlington ldquoMapping ofscientific workflow within the e-protein project to distributedresourcesrdquo inProceedings of theUK e-Science All HandsMeetingNottingham UK September 2004
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World Journal 9
1 2 3 4
5
6 7
8 9 10
11
12
14
13
15
(300000) (600000) (600000) (900000)
(300000)
(150000) (150000)
(300000) (300000) (300000)
(600000)
(300000)
(150000)
(300000)
(600000)
SignalP COILS2 SEG PROSITE
TMHMM
ProsperoHMMer
PSI-BLASTBLAST
IMPALA
PSI-PRED
3D-PSSMSummary
Genomesummary
SCOP
(a)
1 3 5 7
9
11 1210
(400000) (400000) (400000) (400000)
(300000)
(500000) (500000) (500000)
Softmea
2 4 6 8(600000) (600000) (600000) (600000)
Reslice
14 1513
(600000) (600000) (600000)
Convert
Reslice Reslice Reslice
Slicer Slicer
Convert Convert
Slicer
Align wapAlign wapAlign wapAlign wap
(b)
Figure 2 Parallel and Hybrid workflows
10 The Scientific World Journal
75 80 85 90 95 100 8090
100110
120130
140
2060
100140180220
Ener
gy
DVFS-MODPSOHEFT
MakespanCost
Figure 3 Obtained non-dominated solutions for the syntheticworkflow
Ener
gy
DVFS-MODPSOHEFT
MakespanCost
200 400 600 800 100012001400160018002000 400800
12001600
2000
400200
600800
100012001400160018002000
Figure 4 Obtained non-dominated solutions for the hybrid work-flow
(2) After achieving 6 resources the QoS metrics begin torise the reason for this is that time executions are dominatedby interprocessors communications hence decreasing theopportunities for scaling down voltages and frequencies ofresources Consequently the threshold numbers of resourcesthat minimized the QoS metrics could be obtained
6 Conclusion
In this paper we propose a new algorithm called DVFS-Multi-Objective Discrete Particle Swarm Optimization(DVFS-MODPSO) for workflow scheduling in distributedenvironments such as cloud computing infrastructures DV-FS-MODPSO simultaneously optimizes several conflictingobjectives namely the makespan cost and energy in adiscrete space It produces a set of non-dominated solutionsthus providing more flexibility for users to assess theirpreferences and select a schedule that meets their QoS
Ener
gy
DVFS-MODPSOHEFT
Makespan Cost
400200600800
10001200140016001800
240020002200
450 500 550 600 650 700 7501200
14001600
18002000
Figure 5 Obtained non-dominated solutions for the parallel work-flow
0
10
20
30
40
50
3 4 5 6 12
MakespanCostEnergy
Figure 6 Gain over HEFT () according to the number ofprocessors
requirements Our approach exploits the heterogeneityand the marketization of cloud resources in order to findsolutions that optimize makespan and cost Furthermore ituses the Dynamic Voltage and Frequency Scaling (DVFS)technique to reduce energy consumption
We have evaluated our algorithm by simulating it withboth synthetic and real world scientific workflow applicationshaving different structures The results show that the pro-posedDVFS-MODPSO is able to produce a set of good Paretooptimal solutions The results also show that our approachprovides significant improvement not only in terms of the
The Scientific World Journal 11
0
5
10
15
20
25
HybridParallel
Makespan Cost Energy
Figure 7 Improvement according to real world applications
R17
R20
R37
R446
R50
R60
R733
R80
R90
R107
R110
R120
Figure 8 Resource loads for synthetic workflow
Table 3 Improvement for synthetic workflow
Num of proc Gain over Heft ()Makespan Cost Energy
3 0 2126 40084 022 464 15355 0 756 17266 003 652 267512 0 969 3255Average 005 993 2640
cost and the energy consumption but also in term of themakespan for which HEFT algorithm is supposed to givebetter results
Multi-objective optimization in cloud workflow schedul-ing is not a mature domain Most of the existing worksattempt to minimize either the makespan or cost Howeverwe plan to consider other objectives such as reliabilitysecurity in addition to the energy consumption We also
0
50
100
150
200
250
3 4 5 6 12
MakespanCostEnergy
Figure 9 QoS metrics evolutions according to resource numbersfor the synthetic workflow
Table 4 Improvement according to real world applications
Appli Num of proc Gain over Heft ()Makespan Cost Energy
Hybrid
3 405 1155 04024 028 0566 05075 000 1347 11246 000 0117 126512 043 2214 0761
Average 095 1080 0812
Parallel
3 235 5640 25524 1147 2094 32715 000 2142 11666 093 1170 315412 000 031 306
Average 295 2215 209
intend to apply our algorithm in a real world cloud orintegrate it in existing cloud toolkits such as Cloudbus forcomparing with other algorithms
References
[1] D Thain T Tannenbaum and M Livny Condor and the GridGrid Computing Making the Global Infrastructure a RealityJohn Wiley amp Sons Hoboken NJ USA 2003
[2] R Buyya and S Venugopal ldquoThe gridbus toolkit for serviceoriented grid and utility computing an overview and statusreportrdquo in Proceedings of the 1st IEEE International Workshopon Grid Economics and Business Models (GECON rsquo04) pp 19ndash36 IEEE CS Seoul Republic of Korea April 2004
[3] S McGough L Young A Afzal S Newhouse and J Darling-ton ldquoWorkflow enactment in ICENIrdquo in Proceedings of the UK
12 The Scientific World Journal
e-Science All Hands Meeting pp 894ndash900 IOP NottinghamUK September 2004
[4] E Deelman J Blythe Y Gil et al ldquoPegasus mapping scientificworkflows onto the gridrdquo in Grid Computing 2nd EuropeanAcrossGrids Conference AxGrids 2004 Nicosia Cyprus January28ndash30 2004 vol 3165 of Lecture Notes in Computer Science pp11ndash20 Springer New York NY USA 2004
[5] R Buyya S Pandey and C Vecchiola ldquoCloudbus toolkitfor market-oriented cloud computingrdquo in Cloud ComputingProceedings of the 1st International Conference CloudCom 2009Beijing China December 1ndash4 2009 vol 5931 of Lecture Notesin Computer Science pp 24ndash44 Springer New York NY USA2009
[6] Y Yang K Liu J Chen X Liu D Yuan and H Jin ldquoAnalgorithm in SwinDeW-C for scheduling transaction-intensivecost-constrained cloud workflowsrdquo in Proceedings of the 4thIEEE International Conference on eScience (eScience rsquo08) pp374ndash375 Indianapolis Ind USA December 2008
[7] L Ramakrishnan C Koelbel Y Kee et al ldquoVGrADS enablinge-Scienceworkflows on grids and cloudswith fault tolerancerdquo inProceedings of the Conference on High Performance ComputingNetworking Storage and Analysis (SC rsquo09) November 2009
[8] M L Pinedo Scheduling Theory Algorithms and SystemsSpringer Berlin 2008
[9] S Pandey L Wu S M Guru and R Buyya ldquoA particle swarmoptimization-based heuristic for scheduling workflow applica-tions in cloud computing environmentsrdquo in Proceedings of the24th IEEE International Conference on Advanced InformationNetworking and Applications (AINA rsquo10) pp 400ndash407 PerthAustralia April 2010
[10] S Abrishami and M Naghibzadeh ldquoDeadline-constrainedworkflow scheduling in software as a service cloudrdquo ScientiaIranica vol 19 no 3 pp 680ndash689 2012
[11] S Selvarani and G S Sadhasivam ldquoImproved cost-based algo-rithm for task scheduling in cloud computingrdquo in Proceedings ofthe IEEE International Conference on Computational Intelligenceand Computing Research (ICCIC rsquo10) pp 1ndash5 CoimbatoreIndia December 2010
[12] A Verma and S Kaushal ldquoDeadline and budget distributionbased cost-time optimization workflow scheduling algorithmfor cloudrdquo inProceedings of the IJCAon International Conferenceon Recent Advances and Future Trends in Information Technol-ogy (iRAFIT rsquo12) iRAFIT(7) pp 1ndash4 April 2012
[13] R C Correa A Ferreira and P Rebreyend ldquoIntegrating listheuristics into genetic algorithms for multiprocessor schedul-ingrdquo in Proceedings of the 8th IEEE Symposium on Parallel andDistributed Processing (SPDP rsquo96) pp 462ndash469 IEEEComputerSociety Washington DC USA October 1996
[14] E S H Hou N Ansari and H Ren ldquoGenetic algorithm formultiprocessor schedulingrdquo IEEE Transactions on Parallel andDistributed Systems vol 5 no 2 pp 113ndash120 1994
[15] M Grajcar ldquoGenetic list scheduling algorithm for schedulingand allocation on a loosely coupled heterogeneous multi-processor systemrdquo in Proceedings of the 36th Annual DesignAutomation Conference (DAC rsquo99) pp 280ndash285 ACM PressNew York NY USA June 1999
[16] R Sakellariou and H Zhao ldquoA hybrid heuristic for DAGscheduling on heterogeneous systemsrdquo in Proceedings of the13th Heterogeneous Computing Workshop pp 111ndash123 IEEEComputer Society Washington DC USA 2004
[17] H Topcuoglu S Hariri and M Y Wu ldquoPerformance-effectiveand low-complexity task scheduling for heterogeneous comput-ingrdquo IEEE Transactions on Parallel and Distributed Systems vol13 no 3 pp 260ndash274 2002
[18] J Yu and R Buyya ldquoWorkflow scheduling algorithms for gridcomputingrdquo in Metaheuristics for Scheduling in DistributedComputing Environments F Xhafa and A Abraham EdsSpringer Berlin Germany 2008
[19] WN Chen and J Zhang ldquoAn ant colony optimization approachto a grid workflow scheduling problem with various QoSrequirementsrdquo IEEE Transactions on Systems Man and Cyber-netics C vol 39 no 1 pp 29ndash43 2009
[20] X Liu J Chen Z Wu Z Ni D Yuan and Y Yang ldquoHandlingrecoverable temporal violations in scientific workflow systemsa workflow rescheduling based strategyrdquo in Proceedings of the10th IEEEACM International Symposium on Cluster Cloudand Grid Computing (CCGrid rsquo10) pp 534ndash537 MelbourneAustralia May 2010
[21] X Liu Y Yang Y Jiang and J Chen ldquoPreventing temporalviolations in scientific workflows where and howrdquo IEEE Trans-actions on Software Engineering vol 37 no 6 pp 805ndash825 2011
[22] J Yu and R Buyya ldquoScheduling scientific workflow applicationswith deadline and budget constraints using genetic algorithmsrdquoScientific Programming vol 14 no 3-4 pp 217ndash230 2006
[23] L Benini and G Micheli Dynamic Power Management DesignTechniques and CAD Tools Kluwer Academic New York NYUSA 1998
[24] S Irani S Shukla and R Gupta ldquoOnline strategies for dynamicpower management in system with multiple power-savingstatesrdquo ACM Transactions on Embedded Computing Systemsvol 2 no 3 pp 325ndash346 2003
[25] M T Schmitz B M Al-Hashimi and P Eles System-Level Design Techniques for Energy-Efficient Embedded SystemsSpringer New York NY USA 2005
[26] S U Khan and I Ahmad ldquoA cooperative game theoreticaltechnique for joint optimization of energy consumption andresponse time in computational gridsrdquo IEEE Transactions onParallel and Distributed Systems vol 20 no 3 pp 346ndash3602009
[27] Y Li Y Liu and D Qian ldquoA heuristic energy-aware schedulingalgorithm for heterogeneous clustersrdquo in Proceedings of the 15thInternational Conference on Parallel and Distributed Systems(ICPADS rsquo09) pp 407ndash413 Shenzhen China December 2009
[28] L K Goh B Veeravalli and S Viswanathan ldquoDesign of fast andefficient energy-aware gradient-based scheduling algorithmsheterogeneous embeddedmultiprocessor systemsrdquo IEEE Trans-actions on Parallel and Distributed Systems vol 20 no 1 pp 1ndash12 2009
[29] F Yao A Demers and S Shenker ldquoA scheduling model forreduced CPU energyrdquo in Proceedings of the 36th IEEE AnnualSymposium on Foundations of Computer Science pp 374ndash382Milwaukee Wis USA October 1995
[30] Y Shin and K Choi ldquoPower conscious fixed priority schedulingfor hard real-time systemsrdquo in Proceedings of the 36th AnnualDesign Automation Conference (DAC rsquo99) pp 134ndash139 NewOrleans La USA June 1999
[31] I Hong D Kirovski G QuM Potkonjak andM B SrivastavaldquoPower optimization of variable-voltage core-based systemsrdquoIEEE Transactions on Computer-Aided Design of IntegratedCircuits and Systems vol 18 no 12 pp 1702ndash1714 1999
[32] P Pillai and K G Shin ldquoReal-time dynamic voltage scaling forlow-power embedded operating systemsrdquo in Proceedings of the
The Scientific World Journal 13
ACM Symposium on Operating Systems Principles pp 89ndash102October 2001
[33] H Aydin R Melhem D Mosse and P Mejia-Alvarez ldquoPower-aware scheduling for periodic real-time tasksrdquo IEEE Transac-tions on Computers vol 53 no 5 pp 584ndash600 2004
[34] J Zhuo and C Chakrabarti ldquoAn efficient dynamic task schedul-ing algorithm for battery powered DVS systemsrdquo in Proceedingsof the Asian South Pacific Design Automation Conference pp846ndash849 January 2005
[35] R Jejurikar and R Gupta ldquoDynamic slack reclamation withprocrastination scheduling in real-time embedded systemsrdquo inProceedings of the 42nd Design Automation Conference (DACrsquo05) pp 111ndash116 June 2005
[36] J Luo andN K Jha ldquoPower-profile driven variable voltage scal-ing for heterogeneous distributed real-time embedded systemsrdquoin Proceedings of the International Conference on VLSI Designpp 369ndash375 January 2003
[37] M T Schmitz and B M Al-Hashimi ldquoConsidering powervariations of DVS processing elements for energy minimisationin distributed systemsrdquo in Proceedings of the International Sym-posium on System Synthesis (ISSS rsquo01) pp 250ndash255 MontrealCanada October 2001
[38] Y Zhang X Hu and D Z Chen ldquoTask scheduling and voltageselection for energy minimizationrdquo in Proceedings of the 39thAnnual Design Automation Conference (DAC rsquo02) pp 183ndash188June 2002
[39] D Zhu R Melhem and B R Childers ldquoScheduling withdynamic voltagespeed adjustment using slack reclamationin multiprocessor real-time systemsrdquo IEEE Transactions onParallel and Distributed Systems vol 14 no 7 pp 686ndash7002003
[40] D Zhu D Mosse and R Melhem ldquoPower-aware schedulingfor ANDOR graphs in real-time systemsrdquo IEEE Transactionson Parallel and Distributed Systems vol 15 no 9 pp 849ndash8642004
[41] J Kang and S Ranka ldquoDVS based energy minimizationalgorithm for parallel machinesrdquo in Proceedings of the IEEEInternational Parallel and Distributed Processing Symposium(IPDPS rsquo08) pp 1ndash12 Miami Fla USA April 2008
[42] B Rountree D K Lowenthal S Funk V W Freeh B R deSupinski and M Schulz ldquoBounding energy consumption inlarge-scale MPI programsrdquo in Proceedings of the ACMIEEEConference on Supercomputing (SC rsquo07) ACM November 2007
[43] L Wang G von Laszewski J Dayal and F Wang ldquoTowardsenergy aware scheduling for precedence constrained paralleltasks in a cluster with DVFSrdquo in Proceedings of the 10thIEEEACM International Symposium on Cluster Cloud andGrid Computing (CCGrid rsquo10) pp 368ndash377 May 2010
[44] Y C Lee and A Y Zomaya ldquoEnergy conscious schedulingfor distributed computing systems under different operatingconditionsrdquo IEEE Transactions on Parallel and DistributedSystems vol 22 no 8 pp 1374ndash1381 2011
[45] M Mezmaz Y C Lee N Melab E Talbi and A Y ZomayaldquoA bi-objective hybrid genetic algorithm to minimize energyconsumption and makespan for precedence-constrained appli-cations using dynamic voltage scalingrdquo in Proceedings of theIEEE Congress on Evolutionary Computation (CEC rsquo10) pp 1ndash8Barcelona Spain July 2010
[46] M Mezmaz N Melab Y Kessaci et al ldquoA parallel bi-objectivehybrid metaheuristic for energy-aware scheduling for cloudcomputing systemsrdquo Journal of Parallel and Distributed Com-puting vol 71 no 11 pp 1497ndash1508 2011
[47] J Kennedy and R Eberhart ldquoParticle swarm optimizationrdquoin Proceedings of the IEEE International Conference on NeuralNetworks vol 4 pp 1942ndash1948 Perth Australia December1995
[48] Y Zhao M Wilde I Foster et al ldquoGrid middleware servicesfor virtual data discovery composition and integrationrdquo inProceedings of the 2nd Workshop on Middleware for GridComputing (MGC rsquo04) pp 57ndash62 Ontario Canada October2004
[49] A OrsquoBrien S Newhouse and J Darlington ldquoMapping ofscientific workflow within the e-protein project to distributedresourcesrdquo inProceedings of theUK e-Science All HandsMeetingNottingham UK September 2004
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
10 The Scientific World Journal
75 80 85 90 95 100 8090
100110
120130
140
2060
100140180220
Ener
gy
DVFS-MODPSOHEFT
MakespanCost
Figure 3 Obtained non-dominated solutions for the syntheticworkflow
Ener
gy
DVFS-MODPSOHEFT
MakespanCost
200 400 600 800 100012001400160018002000 400800
12001600
2000
400200
600800
100012001400160018002000
Figure 4 Obtained non-dominated solutions for the hybrid work-flow
(2) After achieving 6 resources the QoS metrics begin torise the reason for this is that time executions are dominatedby interprocessors communications hence decreasing theopportunities for scaling down voltages and frequencies ofresources Consequently the threshold numbers of resourcesthat minimized the QoS metrics could be obtained
6 Conclusion
In this paper we propose a new algorithm called DVFS-Multi-Objective Discrete Particle Swarm Optimization(DVFS-MODPSO) for workflow scheduling in distributedenvironments such as cloud computing infrastructures DV-FS-MODPSO simultaneously optimizes several conflictingobjectives namely the makespan cost and energy in adiscrete space It produces a set of non-dominated solutionsthus providing more flexibility for users to assess theirpreferences and select a schedule that meets their QoS
Ener
gy
DVFS-MODPSOHEFT
Makespan Cost
400200600800
10001200140016001800
240020002200
450 500 550 600 650 700 7501200
14001600
18002000
Figure 5 Obtained non-dominated solutions for the parallel work-flow
0
10
20
30
40
50
3 4 5 6 12
MakespanCostEnergy
Figure 6 Gain over HEFT () according to the number ofprocessors
requirements Our approach exploits the heterogeneityand the marketization of cloud resources in order to findsolutions that optimize makespan and cost Furthermore ituses the Dynamic Voltage and Frequency Scaling (DVFS)technique to reduce energy consumption
We have evaluated our algorithm by simulating it withboth synthetic and real world scientific workflow applicationshaving different structures The results show that the pro-posedDVFS-MODPSO is able to produce a set of good Paretooptimal solutions The results also show that our approachprovides significant improvement not only in terms of the
The Scientific World Journal 11
0
5
10
15
20
25
HybridParallel
Makespan Cost Energy
Figure 7 Improvement according to real world applications
R17
R20
R37
R446
R50
R60
R733
R80
R90
R107
R110
R120
Figure 8 Resource loads for synthetic workflow
Table 3 Improvement for synthetic workflow
Num of proc Gain over Heft ()Makespan Cost Energy
3 0 2126 40084 022 464 15355 0 756 17266 003 652 267512 0 969 3255Average 005 993 2640
cost and the energy consumption but also in term of themakespan for which HEFT algorithm is supposed to givebetter results
Multi-objective optimization in cloud workflow schedul-ing is not a mature domain Most of the existing worksattempt to minimize either the makespan or cost Howeverwe plan to consider other objectives such as reliabilitysecurity in addition to the energy consumption We also
0
50
100
150
200
250
3 4 5 6 12
MakespanCostEnergy
Figure 9 QoS metrics evolutions according to resource numbersfor the synthetic workflow
Table 4 Improvement according to real world applications
Appli Num of proc Gain over Heft ()Makespan Cost Energy
Hybrid
3 405 1155 04024 028 0566 05075 000 1347 11246 000 0117 126512 043 2214 0761
Average 095 1080 0812
Parallel
3 235 5640 25524 1147 2094 32715 000 2142 11666 093 1170 315412 000 031 306
Average 295 2215 209
intend to apply our algorithm in a real world cloud orintegrate it in existing cloud toolkits such as Cloudbus forcomparing with other algorithms
References
[1] D Thain T Tannenbaum and M Livny Condor and the GridGrid Computing Making the Global Infrastructure a RealityJohn Wiley amp Sons Hoboken NJ USA 2003
[2] R Buyya and S Venugopal ldquoThe gridbus toolkit for serviceoriented grid and utility computing an overview and statusreportrdquo in Proceedings of the 1st IEEE International Workshopon Grid Economics and Business Models (GECON rsquo04) pp 19ndash36 IEEE CS Seoul Republic of Korea April 2004
[3] S McGough L Young A Afzal S Newhouse and J Darling-ton ldquoWorkflow enactment in ICENIrdquo in Proceedings of the UK
12 The Scientific World Journal
e-Science All Hands Meeting pp 894ndash900 IOP NottinghamUK September 2004
[4] E Deelman J Blythe Y Gil et al ldquoPegasus mapping scientificworkflows onto the gridrdquo in Grid Computing 2nd EuropeanAcrossGrids Conference AxGrids 2004 Nicosia Cyprus January28ndash30 2004 vol 3165 of Lecture Notes in Computer Science pp11ndash20 Springer New York NY USA 2004
[5] R Buyya S Pandey and C Vecchiola ldquoCloudbus toolkitfor market-oriented cloud computingrdquo in Cloud ComputingProceedings of the 1st International Conference CloudCom 2009Beijing China December 1ndash4 2009 vol 5931 of Lecture Notesin Computer Science pp 24ndash44 Springer New York NY USA2009
[6] Y Yang K Liu J Chen X Liu D Yuan and H Jin ldquoAnalgorithm in SwinDeW-C for scheduling transaction-intensivecost-constrained cloud workflowsrdquo in Proceedings of the 4thIEEE International Conference on eScience (eScience rsquo08) pp374ndash375 Indianapolis Ind USA December 2008
[7] L Ramakrishnan C Koelbel Y Kee et al ldquoVGrADS enablinge-Scienceworkflows on grids and cloudswith fault tolerancerdquo inProceedings of the Conference on High Performance ComputingNetworking Storage and Analysis (SC rsquo09) November 2009
[8] M L Pinedo Scheduling Theory Algorithms and SystemsSpringer Berlin 2008
[9] S Pandey L Wu S M Guru and R Buyya ldquoA particle swarmoptimization-based heuristic for scheduling workflow applica-tions in cloud computing environmentsrdquo in Proceedings of the24th IEEE International Conference on Advanced InformationNetworking and Applications (AINA rsquo10) pp 400ndash407 PerthAustralia April 2010
[10] S Abrishami and M Naghibzadeh ldquoDeadline-constrainedworkflow scheduling in software as a service cloudrdquo ScientiaIranica vol 19 no 3 pp 680ndash689 2012
[11] S Selvarani and G S Sadhasivam ldquoImproved cost-based algo-rithm for task scheduling in cloud computingrdquo in Proceedings ofthe IEEE International Conference on Computational Intelligenceand Computing Research (ICCIC rsquo10) pp 1ndash5 CoimbatoreIndia December 2010
[12] A Verma and S Kaushal ldquoDeadline and budget distributionbased cost-time optimization workflow scheduling algorithmfor cloudrdquo inProceedings of the IJCAon International Conferenceon Recent Advances and Future Trends in Information Technol-ogy (iRAFIT rsquo12) iRAFIT(7) pp 1ndash4 April 2012
[13] R C Correa A Ferreira and P Rebreyend ldquoIntegrating listheuristics into genetic algorithms for multiprocessor schedul-ingrdquo in Proceedings of the 8th IEEE Symposium on Parallel andDistributed Processing (SPDP rsquo96) pp 462ndash469 IEEEComputerSociety Washington DC USA October 1996
[14] E S H Hou N Ansari and H Ren ldquoGenetic algorithm formultiprocessor schedulingrdquo IEEE Transactions on Parallel andDistributed Systems vol 5 no 2 pp 113ndash120 1994
[15] M Grajcar ldquoGenetic list scheduling algorithm for schedulingand allocation on a loosely coupled heterogeneous multi-processor systemrdquo in Proceedings of the 36th Annual DesignAutomation Conference (DAC rsquo99) pp 280ndash285 ACM PressNew York NY USA June 1999
[16] R Sakellariou and H Zhao ldquoA hybrid heuristic for DAGscheduling on heterogeneous systemsrdquo in Proceedings of the13th Heterogeneous Computing Workshop pp 111ndash123 IEEEComputer Society Washington DC USA 2004
[17] H Topcuoglu S Hariri and M Y Wu ldquoPerformance-effectiveand low-complexity task scheduling for heterogeneous comput-ingrdquo IEEE Transactions on Parallel and Distributed Systems vol13 no 3 pp 260ndash274 2002
[18] J Yu and R Buyya ldquoWorkflow scheduling algorithms for gridcomputingrdquo in Metaheuristics for Scheduling in DistributedComputing Environments F Xhafa and A Abraham EdsSpringer Berlin Germany 2008
[19] WN Chen and J Zhang ldquoAn ant colony optimization approachto a grid workflow scheduling problem with various QoSrequirementsrdquo IEEE Transactions on Systems Man and Cyber-netics C vol 39 no 1 pp 29ndash43 2009
[20] X Liu J Chen Z Wu Z Ni D Yuan and Y Yang ldquoHandlingrecoverable temporal violations in scientific workflow systemsa workflow rescheduling based strategyrdquo in Proceedings of the10th IEEEACM International Symposium on Cluster Cloudand Grid Computing (CCGrid rsquo10) pp 534ndash537 MelbourneAustralia May 2010
[21] X Liu Y Yang Y Jiang and J Chen ldquoPreventing temporalviolations in scientific workflows where and howrdquo IEEE Trans-actions on Software Engineering vol 37 no 6 pp 805ndash825 2011
[22] J Yu and R Buyya ldquoScheduling scientific workflow applicationswith deadline and budget constraints using genetic algorithmsrdquoScientific Programming vol 14 no 3-4 pp 217ndash230 2006
[23] L Benini and G Micheli Dynamic Power Management DesignTechniques and CAD Tools Kluwer Academic New York NYUSA 1998
[24] S Irani S Shukla and R Gupta ldquoOnline strategies for dynamicpower management in system with multiple power-savingstatesrdquo ACM Transactions on Embedded Computing Systemsvol 2 no 3 pp 325ndash346 2003
[25] M T Schmitz B M Al-Hashimi and P Eles System-Level Design Techniques for Energy-Efficient Embedded SystemsSpringer New York NY USA 2005
[26] S U Khan and I Ahmad ldquoA cooperative game theoreticaltechnique for joint optimization of energy consumption andresponse time in computational gridsrdquo IEEE Transactions onParallel and Distributed Systems vol 20 no 3 pp 346ndash3602009
[27] Y Li Y Liu and D Qian ldquoA heuristic energy-aware schedulingalgorithm for heterogeneous clustersrdquo in Proceedings of the 15thInternational Conference on Parallel and Distributed Systems(ICPADS rsquo09) pp 407ndash413 Shenzhen China December 2009
[28] L K Goh B Veeravalli and S Viswanathan ldquoDesign of fast andefficient energy-aware gradient-based scheduling algorithmsheterogeneous embeddedmultiprocessor systemsrdquo IEEE Trans-actions on Parallel and Distributed Systems vol 20 no 1 pp 1ndash12 2009
[29] F Yao A Demers and S Shenker ldquoA scheduling model forreduced CPU energyrdquo in Proceedings of the 36th IEEE AnnualSymposium on Foundations of Computer Science pp 374ndash382Milwaukee Wis USA October 1995
[30] Y Shin and K Choi ldquoPower conscious fixed priority schedulingfor hard real-time systemsrdquo in Proceedings of the 36th AnnualDesign Automation Conference (DAC rsquo99) pp 134ndash139 NewOrleans La USA June 1999
[31] I Hong D Kirovski G QuM Potkonjak andM B SrivastavaldquoPower optimization of variable-voltage core-based systemsrdquoIEEE Transactions on Computer-Aided Design of IntegratedCircuits and Systems vol 18 no 12 pp 1702ndash1714 1999
[32] P Pillai and K G Shin ldquoReal-time dynamic voltage scaling forlow-power embedded operating systemsrdquo in Proceedings of the
The Scientific World Journal 13
ACM Symposium on Operating Systems Principles pp 89ndash102October 2001
[33] H Aydin R Melhem D Mosse and P Mejia-Alvarez ldquoPower-aware scheduling for periodic real-time tasksrdquo IEEE Transac-tions on Computers vol 53 no 5 pp 584ndash600 2004
[34] J Zhuo and C Chakrabarti ldquoAn efficient dynamic task schedul-ing algorithm for battery powered DVS systemsrdquo in Proceedingsof the Asian South Pacific Design Automation Conference pp846ndash849 January 2005
[35] R Jejurikar and R Gupta ldquoDynamic slack reclamation withprocrastination scheduling in real-time embedded systemsrdquo inProceedings of the 42nd Design Automation Conference (DACrsquo05) pp 111ndash116 June 2005
[36] J Luo andN K Jha ldquoPower-profile driven variable voltage scal-ing for heterogeneous distributed real-time embedded systemsrdquoin Proceedings of the International Conference on VLSI Designpp 369ndash375 January 2003
[37] M T Schmitz and B M Al-Hashimi ldquoConsidering powervariations of DVS processing elements for energy minimisationin distributed systemsrdquo in Proceedings of the International Sym-posium on System Synthesis (ISSS rsquo01) pp 250ndash255 MontrealCanada October 2001
[38] Y Zhang X Hu and D Z Chen ldquoTask scheduling and voltageselection for energy minimizationrdquo in Proceedings of the 39thAnnual Design Automation Conference (DAC rsquo02) pp 183ndash188June 2002
[39] D Zhu R Melhem and B R Childers ldquoScheduling withdynamic voltagespeed adjustment using slack reclamationin multiprocessor real-time systemsrdquo IEEE Transactions onParallel and Distributed Systems vol 14 no 7 pp 686ndash7002003
[40] D Zhu D Mosse and R Melhem ldquoPower-aware schedulingfor ANDOR graphs in real-time systemsrdquo IEEE Transactionson Parallel and Distributed Systems vol 15 no 9 pp 849ndash8642004
[41] J Kang and S Ranka ldquoDVS based energy minimizationalgorithm for parallel machinesrdquo in Proceedings of the IEEEInternational Parallel and Distributed Processing Symposium(IPDPS rsquo08) pp 1ndash12 Miami Fla USA April 2008
[42] B Rountree D K Lowenthal S Funk V W Freeh B R deSupinski and M Schulz ldquoBounding energy consumption inlarge-scale MPI programsrdquo in Proceedings of the ACMIEEEConference on Supercomputing (SC rsquo07) ACM November 2007
[43] L Wang G von Laszewski J Dayal and F Wang ldquoTowardsenergy aware scheduling for precedence constrained paralleltasks in a cluster with DVFSrdquo in Proceedings of the 10thIEEEACM International Symposium on Cluster Cloud andGrid Computing (CCGrid rsquo10) pp 368ndash377 May 2010
[44] Y C Lee and A Y Zomaya ldquoEnergy conscious schedulingfor distributed computing systems under different operatingconditionsrdquo IEEE Transactions on Parallel and DistributedSystems vol 22 no 8 pp 1374ndash1381 2011
[45] M Mezmaz Y C Lee N Melab E Talbi and A Y ZomayaldquoA bi-objective hybrid genetic algorithm to minimize energyconsumption and makespan for precedence-constrained appli-cations using dynamic voltage scalingrdquo in Proceedings of theIEEE Congress on Evolutionary Computation (CEC rsquo10) pp 1ndash8Barcelona Spain July 2010
[46] M Mezmaz N Melab Y Kessaci et al ldquoA parallel bi-objectivehybrid metaheuristic for energy-aware scheduling for cloudcomputing systemsrdquo Journal of Parallel and Distributed Com-puting vol 71 no 11 pp 1497ndash1508 2011
[47] J Kennedy and R Eberhart ldquoParticle swarm optimizationrdquoin Proceedings of the IEEE International Conference on NeuralNetworks vol 4 pp 1942ndash1948 Perth Australia December1995
[48] Y Zhao M Wilde I Foster et al ldquoGrid middleware servicesfor virtual data discovery composition and integrationrdquo inProceedings of the 2nd Workshop on Middleware for GridComputing (MGC rsquo04) pp 57ndash62 Ontario Canada October2004
[49] A OrsquoBrien S Newhouse and J Darlington ldquoMapping ofscientific workflow within the e-protein project to distributedresourcesrdquo inProceedings of theUK e-Science All HandsMeetingNottingham UK September 2004
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World Journal 11
0
5
10
15
20
25
HybridParallel
Makespan Cost Energy
Figure 7 Improvement according to real world applications
R17
R20
R37
R446
R50
R60
R733
R80
R90
R107
R110
R120
Figure 8 Resource loads for synthetic workflow
Table 3 Improvement for synthetic workflow
Num of proc Gain over Heft ()Makespan Cost Energy
3 0 2126 40084 022 464 15355 0 756 17266 003 652 267512 0 969 3255Average 005 993 2640
cost and the energy consumption but also in term of themakespan for which HEFT algorithm is supposed to givebetter results
Multi-objective optimization in cloud workflow schedul-ing is not a mature domain Most of the existing worksattempt to minimize either the makespan or cost Howeverwe plan to consider other objectives such as reliabilitysecurity in addition to the energy consumption We also
0
50
100
150
200
250
3 4 5 6 12
MakespanCostEnergy
Figure 9 QoS metrics evolutions according to resource numbersfor the synthetic workflow
Table 4 Improvement according to real world applications
Appli Num of proc Gain over Heft ()Makespan Cost Energy
Hybrid
3 405 1155 04024 028 0566 05075 000 1347 11246 000 0117 126512 043 2214 0761
Average 095 1080 0812
Parallel
3 235 5640 25524 1147 2094 32715 000 2142 11666 093 1170 315412 000 031 306
Average 295 2215 209
intend to apply our algorithm in a real world cloud orintegrate it in existing cloud toolkits such as Cloudbus forcomparing with other algorithms
References
[1] D Thain T Tannenbaum and M Livny Condor and the GridGrid Computing Making the Global Infrastructure a RealityJohn Wiley amp Sons Hoboken NJ USA 2003
[2] R Buyya and S Venugopal ldquoThe gridbus toolkit for serviceoriented grid and utility computing an overview and statusreportrdquo in Proceedings of the 1st IEEE International Workshopon Grid Economics and Business Models (GECON rsquo04) pp 19ndash36 IEEE CS Seoul Republic of Korea April 2004
[3] S McGough L Young A Afzal S Newhouse and J Darling-ton ldquoWorkflow enactment in ICENIrdquo in Proceedings of the UK
12 The Scientific World Journal
e-Science All Hands Meeting pp 894ndash900 IOP NottinghamUK September 2004
[4] E Deelman J Blythe Y Gil et al ldquoPegasus mapping scientificworkflows onto the gridrdquo in Grid Computing 2nd EuropeanAcrossGrids Conference AxGrids 2004 Nicosia Cyprus January28ndash30 2004 vol 3165 of Lecture Notes in Computer Science pp11ndash20 Springer New York NY USA 2004
[5] R Buyya S Pandey and C Vecchiola ldquoCloudbus toolkitfor market-oriented cloud computingrdquo in Cloud ComputingProceedings of the 1st International Conference CloudCom 2009Beijing China December 1ndash4 2009 vol 5931 of Lecture Notesin Computer Science pp 24ndash44 Springer New York NY USA2009
[6] Y Yang K Liu J Chen X Liu D Yuan and H Jin ldquoAnalgorithm in SwinDeW-C for scheduling transaction-intensivecost-constrained cloud workflowsrdquo in Proceedings of the 4thIEEE International Conference on eScience (eScience rsquo08) pp374ndash375 Indianapolis Ind USA December 2008
[7] L Ramakrishnan C Koelbel Y Kee et al ldquoVGrADS enablinge-Scienceworkflows on grids and cloudswith fault tolerancerdquo inProceedings of the Conference on High Performance ComputingNetworking Storage and Analysis (SC rsquo09) November 2009
[8] M L Pinedo Scheduling Theory Algorithms and SystemsSpringer Berlin 2008
[9] S Pandey L Wu S M Guru and R Buyya ldquoA particle swarmoptimization-based heuristic for scheduling workflow applica-tions in cloud computing environmentsrdquo in Proceedings of the24th IEEE International Conference on Advanced InformationNetworking and Applications (AINA rsquo10) pp 400ndash407 PerthAustralia April 2010
[10] S Abrishami and M Naghibzadeh ldquoDeadline-constrainedworkflow scheduling in software as a service cloudrdquo ScientiaIranica vol 19 no 3 pp 680ndash689 2012
[11] S Selvarani and G S Sadhasivam ldquoImproved cost-based algo-rithm for task scheduling in cloud computingrdquo in Proceedings ofthe IEEE International Conference on Computational Intelligenceand Computing Research (ICCIC rsquo10) pp 1ndash5 CoimbatoreIndia December 2010
[12] A Verma and S Kaushal ldquoDeadline and budget distributionbased cost-time optimization workflow scheduling algorithmfor cloudrdquo inProceedings of the IJCAon International Conferenceon Recent Advances and Future Trends in Information Technol-ogy (iRAFIT rsquo12) iRAFIT(7) pp 1ndash4 April 2012
[13] R C Correa A Ferreira and P Rebreyend ldquoIntegrating listheuristics into genetic algorithms for multiprocessor schedul-ingrdquo in Proceedings of the 8th IEEE Symposium on Parallel andDistributed Processing (SPDP rsquo96) pp 462ndash469 IEEEComputerSociety Washington DC USA October 1996
[14] E S H Hou N Ansari and H Ren ldquoGenetic algorithm formultiprocessor schedulingrdquo IEEE Transactions on Parallel andDistributed Systems vol 5 no 2 pp 113ndash120 1994
[15] M Grajcar ldquoGenetic list scheduling algorithm for schedulingand allocation on a loosely coupled heterogeneous multi-processor systemrdquo in Proceedings of the 36th Annual DesignAutomation Conference (DAC rsquo99) pp 280ndash285 ACM PressNew York NY USA June 1999
[16] R Sakellariou and H Zhao ldquoA hybrid heuristic for DAGscheduling on heterogeneous systemsrdquo in Proceedings of the13th Heterogeneous Computing Workshop pp 111ndash123 IEEEComputer Society Washington DC USA 2004
[17] H Topcuoglu S Hariri and M Y Wu ldquoPerformance-effectiveand low-complexity task scheduling for heterogeneous comput-ingrdquo IEEE Transactions on Parallel and Distributed Systems vol13 no 3 pp 260ndash274 2002
[18] J Yu and R Buyya ldquoWorkflow scheduling algorithms for gridcomputingrdquo in Metaheuristics for Scheduling in DistributedComputing Environments F Xhafa and A Abraham EdsSpringer Berlin Germany 2008
[19] WN Chen and J Zhang ldquoAn ant colony optimization approachto a grid workflow scheduling problem with various QoSrequirementsrdquo IEEE Transactions on Systems Man and Cyber-netics C vol 39 no 1 pp 29ndash43 2009
[20] X Liu J Chen Z Wu Z Ni D Yuan and Y Yang ldquoHandlingrecoverable temporal violations in scientific workflow systemsa workflow rescheduling based strategyrdquo in Proceedings of the10th IEEEACM International Symposium on Cluster Cloudand Grid Computing (CCGrid rsquo10) pp 534ndash537 MelbourneAustralia May 2010
[21] X Liu Y Yang Y Jiang and J Chen ldquoPreventing temporalviolations in scientific workflows where and howrdquo IEEE Trans-actions on Software Engineering vol 37 no 6 pp 805ndash825 2011
[22] J Yu and R Buyya ldquoScheduling scientific workflow applicationswith deadline and budget constraints using genetic algorithmsrdquoScientific Programming vol 14 no 3-4 pp 217ndash230 2006
[23] L Benini and G Micheli Dynamic Power Management DesignTechniques and CAD Tools Kluwer Academic New York NYUSA 1998
[24] S Irani S Shukla and R Gupta ldquoOnline strategies for dynamicpower management in system with multiple power-savingstatesrdquo ACM Transactions on Embedded Computing Systemsvol 2 no 3 pp 325ndash346 2003
[25] M T Schmitz B M Al-Hashimi and P Eles System-Level Design Techniques for Energy-Efficient Embedded SystemsSpringer New York NY USA 2005
[26] S U Khan and I Ahmad ldquoA cooperative game theoreticaltechnique for joint optimization of energy consumption andresponse time in computational gridsrdquo IEEE Transactions onParallel and Distributed Systems vol 20 no 3 pp 346ndash3602009
[27] Y Li Y Liu and D Qian ldquoA heuristic energy-aware schedulingalgorithm for heterogeneous clustersrdquo in Proceedings of the 15thInternational Conference on Parallel and Distributed Systems(ICPADS rsquo09) pp 407ndash413 Shenzhen China December 2009
[28] L K Goh B Veeravalli and S Viswanathan ldquoDesign of fast andefficient energy-aware gradient-based scheduling algorithmsheterogeneous embeddedmultiprocessor systemsrdquo IEEE Trans-actions on Parallel and Distributed Systems vol 20 no 1 pp 1ndash12 2009
[29] F Yao A Demers and S Shenker ldquoA scheduling model forreduced CPU energyrdquo in Proceedings of the 36th IEEE AnnualSymposium on Foundations of Computer Science pp 374ndash382Milwaukee Wis USA October 1995
[30] Y Shin and K Choi ldquoPower conscious fixed priority schedulingfor hard real-time systemsrdquo in Proceedings of the 36th AnnualDesign Automation Conference (DAC rsquo99) pp 134ndash139 NewOrleans La USA June 1999
[31] I Hong D Kirovski G QuM Potkonjak andM B SrivastavaldquoPower optimization of variable-voltage core-based systemsrdquoIEEE Transactions on Computer-Aided Design of IntegratedCircuits and Systems vol 18 no 12 pp 1702ndash1714 1999
[32] P Pillai and K G Shin ldquoReal-time dynamic voltage scaling forlow-power embedded operating systemsrdquo in Proceedings of the
The Scientific World Journal 13
ACM Symposium on Operating Systems Principles pp 89ndash102October 2001
[33] H Aydin R Melhem D Mosse and P Mejia-Alvarez ldquoPower-aware scheduling for periodic real-time tasksrdquo IEEE Transac-tions on Computers vol 53 no 5 pp 584ndash600 2004
[34] J Zhuo and C Chakrabarti ldquoAn efficient dynamic task schedul-ing algorithm for battery powered DVS systemsrdquo in Proceedingsof the Asian South Pacific Design Automation Conference pp846ndash849 January 2005
[35] R Jejurikar and R Gupta ldquoDynamic slack reclamation withprocrastination scheduling in real-time embedded systemsrdquo inProceedings of the 42nd Design Automation Conference (DACrsquo05) pp 111ndash116 June 2005
[36] J Luo andN K Jha ldquoPower-profile driven variable voltage scal-ing for heterogeneous distributed real-time embedded systemsrdquoin Proceedings of the International Conference on VLSI Designpp 369ndash375 January 2003
[37] M T Schmitz and B M Al-Hashimi ldquoConsidering powervariations of DVS processing elements for energy minimisationin distributed systemsrdquo in Proceedings of the International Sym-posium on System Synthesis (ISSS rsquo01) pp 250ndash255 MontrealCanada October 2001
[38] Y Zhang X Hu and D Z Chen ldquoTask scheduling and voltageselection for energy minimizationrdquo in Proceedings of the 39thAnnual Design Automation Conference (DAC rsquo02) pp 183ndash188June 2002
[39] D Zhu R Melhem and B R Childers ldquoScheduling withdynamic voltagespeed adjustment using slack reclamationin multiprocessor real-time systemsrdquo IEEE Transactions onParallel and Distributed Systems vol 14 no 7 pp 686ndash7002003
[40] D Zhu D Mosse and R Melhem ldquoPower-aware schedulingfor ANDOR graphs in real-time systemsrdquo IEEE Transactionson Parallel and Distributed Systems vol 15 no 9 pp 849ndash8642004
[41] J Kang and S Ranka ldquoDVS based energy minimizationalgorithm for parallel machinesrdquo in Proceedings of the IEEEInternational Parallel and Distributed Processing Symposium(IPDPS rsquo08) pp 1ndash12 Miami Fla USA April 2008
[42] B Rountree D K Lowenthal S Funk V W Freeh B R deSupinski and M Schulz ldquoBounding energy consumption inlarge-scale MPI programsrdquo in Proceedings of the ACMIEEEConference on Supercomputing (SC rsquo07) ACM November 2007
[43] L Wang G von Laszewski J Dayal and F Wang ldquoTowardsenergy aware scheduling for precedence constrained paralleltasks in a cluster with DVFSrdquo in Proceedings of the 10thIEEEACM International Symposium on Cluster Cloud andGrid Computing (CCGrid rsquo10) pp 368ndash377 May 2010
[44] Y C Lee and A Y Zomaya ldquoEnergy conscious schedulingfor distributed computing systems under different operatingconditionsrdquo IEEE Transactions on Parallel and DistributedSystems vol 22 no 8 pp 1374ndash1381 2011
[45] M Mezmaz Y C Lee N Melab E Talbi and A Y ZomayaldquoA bi-objective hybrid genetic algorithm to minimize energyconsumption and makespan for precedence-constrained appli-cations using dynamic voltage scalingrdquo in Proceedings of theIEEE Congress on Evolutionary Computation (CEC rsquo10) pp 1ndash8Barcelona Spain July 2010
[46] M Mezmaz N Melab Y Kessaci et al ldquoA parallel bi-objectivehybrid metaheuristic for energy-aware scheduling for cloudcomputing systemsrdquo Journal of Parallel and Distributed Com-puting vol 71 no 11 pp 1497ndash1508 2011
[47] J Kennedy and R Eberhart ldquoParticle swarm optimizationrdquoin Proceedings of the IEEE International Conference on NeuralNetworks vol 4 pp 1942ndash1948 Perth Australia December1995
[48] Y Zhao M Wilde I Foster et al ldquoGrid middleware servicesfor virtual data discovery composition and integrationrdquo inProceedings of the 2nd Workshop on Middleware for GridComputing (MGC rsquo04) pp 57ndash62 Ontario Canada October2004
[49] A OrsquoBrien S Newhouse and J Darlington ldquoMapping ofscientific workflow within the e-protein project to distributedresourcesrdquo inProceedings of theUK e-Science All HandsMeetingNottingham UK September 2004
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
12 The Scientific World Journal
e-Science All Hands Meeting pp 894ndash900 IOP NottinghamUK September 2004
[4] E Deelman J Blythe Y Gil et al ldquoPegasus mapping scientificworkflows onto the gridrdquo in Grid Computing 2nd EuropeanAcrossGrids Conference AxGrids 2004 Nicosia Cyprus January28ndash30 2004 vol 3165 of Lecture Notes in Computer Science pp11ndash20 Springer New York NY USA 2004
[5] R Buyya S Pandey and C Vecchiola ldquoCloudbus toolkitfor market-oriented cloud computingrdquo in Cloud ComputingProceedings of the 1st International Conference CloudCom 2009Beijing China December 1ndash4 2009 vol 5931 of Lecture Notesin Computer Science pp 24ndash44 Springer New York NY USA2009
[6] Y Yang K Liu J Chen X Liu D Yuan and H Jin ldquoAnalgorithm in SwinDeW-C for scheduling transaction-intensivecost-constrained cloud workflowsrdquo in Proceedings of the 4thIEEE International Conference on eScience (eScience rsquo08) pp374ndash375 Indianapolis Ind USA December 2008
[7] L Ramakrishnan C Koelbel Y Kee et al ldquoVGrADS enablinge-Scienceworkflows on grids and cloudswith fault tolerancerdquo inProceedings of the Conference on High Performance ComputingNetworking Storage and Analysis (SC rsquo09) November 2009
[8] M L Pinedo Scheduling Theory Algorithms and SystemsSpringer Berlin 2008
[9] S Pandey L Wu S M Guru and R Buyya ldquoA particle swarmoptimization-based heuristic for scheduling workflow applica-tions in cloud computing environmentsrdquo in Proceedings of the24th IEEE International Conference on Advanced InformationNetworking and Applications (AINA rsquo10) pp 400ndash407 PerthAustralia April 2010
[10] S Abrishami and M Naghibzadeh ldquoDeadline-constrainedworkflow scheduling in software as a service cloudrdquo ScientiaIranica vol 19 no 3 pp 680ndash689 2012
[11] S Selvarani and G S Sadhasivam ldquoImproved cost-based algo-rithm for task scheduling in cloud computingrdquo in Proceedings ofthe IEEE International Conference on Computational Intelligenceand Computing Research (ICCIC rsquo10) pp 1ndash5 CoimbatoreIndia December 2010
[12] A Verma and S Kaushal ldquoDeadline and budget distributionbased cost-time optimization workflow scheduling algorithmfor cloudrdquo inProceedings of the IJCAon International Conferenceon Recent Advances and Future Trends in Information Technol-ogy (iRAFIT rsquo12) iRAFIT(7) pp 1ndash4 April 2012
[13] R C Correa A Ferreira and P Rebreyend ldquoIntegrating listheuristics into genetic algorithms for multiprocessor schedul-ingrdquo in Proceedings of the 8th IEEE Symposium on Parallel andDistributed Processing (SPDP rsquo96) pp 462ndash469 IEEEComputerSociety Washington DC USA October 1996
[14] E S H Hou N Ansari and H Ren ldquoGenetic algorithm formultiprocessor schedulingrdquo IEEE Transactions on Parallel andDistributed Systems vol 5 no 2 pp 113ndash120 1994
[15] M Grajcar ldquoGenetic list scheduling algorithm for schedulingand allocation on a loosely coupled heterogeneous multi-processor systemrdquo in Proceedings of the 36th Annual DesignAutomation Conference (DAC rsquo99) pp 280ndash285 ACM PressNew York NY USA June 1999
[16] R Sakellariou and H Zhao ldquoA hybrid heuristic for DAGscheduling on heterogeneous systemsrdquo in Proceedings of the13th Heterogeneous Computing Workshop pp 111ndash123 IEEEComputer Society Washington DC USA 2004
[17] H Topcuoglu S Hariri and M Y Wu ldquoPerformance-effectiveand low-complexity task scheduling for heterogeneous comput-ingrdquo IEEE Transactions on Parallel and Distributed Systems vol13 no 3 pp 260ndash274 2002
[18] J Yu and R Buyya ldquoWorkflow scheduling algorithms for gridcomputingrdquo in Metaheuristics for Scheduling in DistributedComputing Environments F Xhafa and A Abraham EdsSpringer Berlin Germany 2008
[19] WN Chen and J Zhang ldquoAn ant colony optimization approachto a grid workflow scheduling problem with various QoSrequirementsrdquo IEEE Transactions on Systems Man and Cyber-netics C vol 39 no 1 pp 29ndash43 2009
[20] X Liu J Chen Z Wu Z Ni D Yuan and Y Yang ldquoHandlingrecoverable temporal violations in scientific workflow systemsa workflow rescheduling based strategyrdquo in Proceedings of the10th IEEEACM International Symposium on Cluster Cloudand Grid Computing (CCGrid rsquo10) pp 534ndash537 MelbourneAustralia May 2010
[21] X Liu Y Yang Y Jiang and J Chen ldquoPreventing temporalviolations in scientific workflows where and howrdquo IEEE Trans-actions on Software Engineering vol 37 no 6 pp 805ndash825 2011
[22] J Yu and R Buyya ldquoScheduling scientific workflow applicationswith deadline and budget constraints using genetic algorithmsrdquoScientific Programming vol 14 no 3-4 pp 217ndash230 2006
[23] L Benini and G Micheli Dynamic Power Management DesignTechniques and CAD Tools Kluwer Academic New York NYUSA 1998
[24] S Irani S Shukla and R Gupta ldquoOnline strategies for dynamicpower management in system with multiple power-savingstatesrdquo ACM Transactions on Embedded Computing Systemsvol 2 no 3 pp 325ndash346 2003
[25] M T Schmitz B M Al-Hashimi and P Eles System-Level Design Techniques for Energy-Efficient Embedded SystemsSpringer New York NY USA 2005
[26] S U Khan and I Ahmad ldquoA cooperative game theoreticaltechnique for joint optimization of energy consumption andresponse time in computational gridsrdquo IEEE Transactions onParallel and Distributed Systems vol 20 no 3 pp 346ndash3602009
[27] Y Li Y Liu and D Qian ldquoA heuristic energy-aware schedulingalgorithm for heterogeneous clustersrdquo in Proceedings of the 15thInternational Conference on Parallel and Distributed Systems(ICPADS rsquo09) pp 407ndash413 Shenzhen China December 2009
[28] L K Goh B Veeravalli and S Viswanathan ldquoDesign of fast andefficient energy-aware gradient-based scheduling algorithmsheterogeneous embeddedmultiprocessor systemsrdquo IEEE Trans-actions on Parallel and Distributed Systems vol 20 no 1 pp 1ndash12 2009
[29] F Yao A Demers and S Shenker ldquoA scheduling model forreduced CPU energyrdquo in Proceedings of the 36th IEEE AnnualSymposium on Foundations of Computer Science pp 374ndash382Milwaukee Wis USA October 1995
[30] Y Shin and K Choi ldquoPower conscious fixed priority schedulingfor hard real-time systemsrdquo in Proceedings of the 36th AnnualDesign Automation Conference (DAC rsquo99) pp 134ndash139 NewOrleans La USA June 1999
[31] I Hong D Kirovski G QuM Potkonjak andM B SrivastavaldquoPower optimization of variable-voltage core-based systemsrdquoIEEE Transactions on Computer-Aided Design of IntegratedCircuits and Systems vol 18 no 12 pp 1702ndash1714 1999
[32] P Pillai and K G Shin ldquoReal-time dynamic voltage scaling forlow-power embedded operating systemsrdquo in Proceedings of the
The Scientific World Journal 13
ACM Symposium on Operating Systems Principles pp 89ndash102October 2001
[33] H Aydin R Melhem D Mosse and P Mejia-Alvarez ldquoPower-aware scheduling for periodic real-time tasksrdquo IEEE Transac-tions on Computers vol 53 no 5 pp 584ndash600 2004
[34] J Zhuo and C Chakrabarti ldquoAn efficient dynamic task schedul-ing algorithm for battery powered DVS systemsrdquo in Proceedingsof the Asian South Pacific Design Automation Conference pp846ndash849 January 2005
[35] R Jejurikar and R Gupta ldquoDynamic slack reclamation withprocrastination scheduling in real-time embedded systemsrdquo inProceedings of the 42nd Design Automation Conference (DACrsquo05) pp 111ndash116 June 2005
[36] J Luo andN K Jha ldquoPower-profile driven variable voltage scal-ing for heterogeneous distributed real-time embedded systemsrdquoin Proceedings of the International Conference on VLSI Designpp 369ndash375 January 2003
[37] M T Schmitz and B M Al-Hashimi ldquoConsidering powervariations of DVS processing elements for energy minimisationin distributed systemsrdquo in Proceedings of the International Sym-posium on System Synthesis (ISSS rsquo01) pp 250ndash255 MontrealCanada October 2001
[38] Y Zhang X Hu and D Z Chen ldquoTask scheduling and voltageselection for energy minimizationrdquo in Proceedings of the 39thAnnual Design Automation Conference (DAC rsquo02) pp 183ndash188June 2002
[39] D Zhu R Melhem and B R Childers ldquoScheduling withdynamic voltagespeed adjustment using slack reclamationin multiprocessor real-time systemsrdquo IEEE Transactions onParallel and Distributed Systems vol 14 no 7 pp 686ndash7002003
[40] D Zhu D Mosse and R Melhem ldquoPower-aware schedulingfor ANDOR graphs in real-time systemsrdquo IEEE Transactionson Parallel and Distributed Systems vol 15 no 9 pp 849ndash8642004
[41] J Kang and S Ranka ldquoDVS based energy minimizationalgorithm for parallel machinesrdquo in Proceedings of the IEEEInternational Parallel and Distributed Processing Symposium(IPDPS rsquo08) pp 1ndash12 Miami Fla USA April 2008
[42] B Rountree D K Lowenthal S Funk V W Freeh B R deSupinski and M Schulz ldquoBounding energy consumption inlarge-scale MPI programsrdquo in Proceedings of the ACMIEEEConference on Supercomputing (SC rsquo07) ACM November 2007
[43] L Wang G von Laszewski J Dayal and F Wang ldquoTowardsenergy aware scheduling for precedence constrained paralleltasks in a cluster with DVFSrdquo in Proceedings of the 10thIEEEACM International Symposium on Cluster Cloud andGrid Computing (CCGrid rsquo10) pp 368ndash377 May 2010
[44] Y C Lee and A Y Zomaya ldquoEnergy conscious schedulingfor distributed computing systems under different operatingconditionsrdquo IEEE Transactions on Parallel and DistributedSystems vol 22 no 8 pp 1374ndash1381 2011
[45] M Mezmaz Y C Lee N Melab E Talbi and A Y ZomayaldquoA bi-objective hybrid genetic algorithm to minimize energyconsumption and makespan for precedence-constrained appli-cations using dynamic voltage scalingrdquo in Proceedings of theIEEE Congress on Evolutionary Computation (CEC rsquo10) pp 1ndash8Barcelona Spain July 2010
[46] M Mezmaz N Melab Y Kessaci et al ldquoA parallel bi-objectivehybrid metaheuristic for energy-aware scheduling for cloudcomputing systemsrdquo Journal of Parallel and Distributed Com-puting vol 71 no 11 pp 1497ndash1508 2011
[47] J Kennedy and R Eberhart ldquoParticle swarm optimizationrdquoin Proceedings of the IEEE International Conference on NeuralNetworks vol 4 pp 1942ndash1948 Perth Australia December1995
[48] Y Zhao M Wilde I Foster et al ldquoGrid middleware servicesfor virtual data discovery composition and integrationrdquo inProceedings of the 2nd Workshop on Middleware for GridComputing (MGC rsquo04) pp 57ndash62 Ontario Canada October2004
[49] A OrsquoBrien S Newhouse and J Darlington ldquoMapping ofscientific workflow within the e-protein project to distributedresourcesrdquo inProceedings of theUK e-Science All HandsMeetingNottingham UK September 2004
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World Journal 13
ACM Symposium on Operating Systems Principles pp 89ndash102October 2001
[33] H Aydin R Melhem D Mosse and P Mejia-Alvarez ldquoPower-aware scheduling for periodic real-time tasksrdquo IEEE Transac-tions on Computers vol 53 no 5 pp 584ndash600 2004
[34] J Zhuo and C Chakrabarti ldquoAn efficient dynamic task schedul-ing algorithm for battery powered DVS systemsrdquo in Proceedingsof the Asian South Pacific Design Automation Conference pp846ndash849 January 2005
[35] R Jejurikar and R Gupta ldquoDynamic slack reclamation withprocrastination scheduling in real-time embedded systemsrdquo inProceedings of the 42nd Design Automation Conference (DACrsquo05) pp 111ndash116 June 2005
[36] J Luo andN K Jha ldquoPower-profile driven variable voltage scal-ing for heterogeneous distributed real-time embedded systemsrdquoin Proceedings of the International Conference on VLSI Designpp 369ndash375 January 2003
[37] M T Schmitz and B M Al-Hashimi ldquoConsidering powervariations of DVS processing elements for energy minimisationin distributed systemsrdquo in Proceedings of the International Sym-posium on System Synthesis (ISSS rsquo01) pp 250ndash255 MontrealCanada October 2001
[38] Y Zhang X Hu and D Z Chen ldquoTask scheduling and voltageselection for energy minimizationrdquo in Proceedings of the 39thAnnual Design Automation Conference (DAC rsquo02) pp 183ndash188June 2002
[39] D Zhu R Melhem and B R Childers ldquoScheduling withdynamic voltagespeed adjustment using slack reclamationin multiprocessor real-time systemsrdquo IEEE Transactions onParallel and Distributed Systems vol 14 no 7 pp 686ndash7002003
[40] D Zhu D Mosse and R Melhem ldquoPower-aware schedulingfor ANDOR graphs in real-time systemsrdquo IEEE Transactionson Parallel and Distributed Systems vol 15 no 9 pp 849ndash8642004
[41] J Kang and S Ranka ldquoDVS based energy minimizationalgorithm for parallel machinesrdquo in Proceedings of the IEEEInternational Parallel and Distributed Processing Symposium(IPDPS rsquo08) pp 1ndash12 Miami Fla USA April 2008
[42] B Rountree D K Lowenthal S Funk V W Freeh B R deSupinski and M Schulz ldquoBounding energy consumption inlarge-scale MPI programsrdquo in Proceedings of the ACMIEEEConference on Supercomputing (SC rsquo07) ACM November 2007
[43] L Wang G von Laszewski J Dayal and F Wang ldquoTowardsenergy aware scheduling for precedence constrained paralleltasks in a cluster with DVFSrdquo in Proceedings of the 10thIEEEACM International Symposium on Cluster Cloud andGrid Computing (CCGrid rsquo10) pp 368ndash377 May 2010
[44] Y C Lee and A Y Zomaya ldquoEnergy conscious schedulingfor distributed computing systems under different operatingconditionsrdquo IEEE Transactions on Parallel and DistributedSystems vol 22 no 8 pp 1374ndash1381 2011
[45] M Mezmaz Y C Lee N Melab E Talbi and A Y ZomayaldquoA bi-objective hybrid genetic algorithm to minimize energyconsumption and makespan for precedence-constrained appli-cations using dynamic voltage scalingrdquo in Proceedings of theIEEE Congress on Evolutionary Computation (CEC rsquo10) pp 1ndash8Barcelona Spain July 2010
[46] M Mezmaz N Melab Y Kessaci et al ldquoA parallel bi-objectivehybrid metaheuristic for energy-aware scheduling for cloudcomputing systemsrdquo Journal of Parallel and Distributed Com-puting vol 71 no 11 pp 1497ndash1508 2011
[47] J Kennedy and R Eberhart ldquoParticle swarm optimizationrdquoin Proceedings of the IEEE International Conference on NeuralNetworks vol 4 pp 1942ndash1948 Perth Australia December1995
[48] Y Zhao M Wilde I Foster et al ldquoGrid middleware servicesfor virtual data discovery composition and integrationrdquo inProceedings of the 2nd Workshop on Middleware for GridComputing (MGC rsquo04) pp 57ndash62 Ontario Canada October2004
[49] A OrsquoBrien S Newhouse and J Darlington ldquoMapping ofscientific workflow within the e-protein project to distributedresourcesrdquo inProceedings of theUK e-Science All HandsMeetingNottingham UK September 2004
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014