SCHEDULING TECHNIQUES IN WIRELESS MESH NETWORKS A thesis
85
SCHEDULING TECHNIQUES IN WIRELESS MESH NETWORKS A thesis Presented to The Faculty of Graduate Studies of The University of Guelph by JASON B. ERNST In partial fulfilment of requirements for the degree of Master of Science April, 2009 © Jason B. Ernst, 2009
Transcript of SCHEDULING TECHNIQUES IN WIRELESS MESH NETWORKS A thesis
SCHEDULING TECHNIQUES IN WIRELESS MESH NETWORKS
A thesis
Presented to
The Faculty of Graduate Studies
of
The University of Guelph
by
JASON B. ERNST
In partial fulfilment of requirements
for the degree of
Master of Science
April, 2009
© Jason B. Ernst, 2009
ABSTRACT
SCHEDULING TECHNIQUES FOR WIRELESS MESH NETWORKS
Jason Bruce Ernst Advisor:
University of Guelph, 2009 Dr. Mieso Denko
Wireless mesh networks (WMN) are a promising technology which provides
wireless broadband connectivity to the Internet. This thesis is an investigation of
scheduling problems in WMNs. First, existing scheduling solutions are discussed and
classified based on technique and implementation framework. Then two novel proposed
schemes are discussed in detail. The first proposed technique is a multiple gateway fair
scheduling scheme. This scheme consists of distributed routing and requirement tables
and a propagation algorithm for scheduling at the gateways. Simulation results confirm
that fair scheduling has better performance than the scheme without fair scheduling and
that multiple gateways are beneficial. The second proposed scheme is a cross-layer
mixed-bias solution. We bias against distance from gateway, size of queue, link quality
and a combined mixed-bias technique. Simulation results confirm that the mixed-bias
approach performs better than IEEE 802.11 DCF for wireless mesh networks with respect
to the metrics used for evaluation.
ACKNOWLEDGEMENTS
First, I would like to thank my advisor, Prof. Mieso Denko and my co-advisor
Prof. Hongbing Fan for all of their insightful guidance and helpful comments and advice
during the development of the ideas presented in my thesis. I am also grateful for the
expertise provided by committee members, Prof. Xining Li and Prof. Obimbo who
provided additional feedback and helpful suggestions.
Additionally, I would like to acknowledge my parents, Bruce and Debbie, my
sister Nicole, Sarah Fewster and many other friends and family who provided support,
inspiration and encouragement to complete my thesis in a timely manner.
I would like to thank my colleagues and friends in the PERWIN research group at
Guelph: Thabo Nkwe, Nikhi Saxena, Dario Guiao, Jayt Jughmohan, Brian Heisler and
Henry Phan for attending my various presentations and providing fresh perspectives and
questions on all aspects of my work as well as additional support and inspiration
throughout my studies and the University of Guelph.
Lastly, I am grateful for the support staff, administration and other faculty
members in the Computing and Information Science department at the University of
Guelph, who were always helpful in submitting forms, booking rooms and other
equipment whenever necessary.
i
TABLE OF CONTENTS
TABLE OF CONTENTS................................................................................................. i
List of Figures ................................................................................................................ iv
List of Acronyms ............................................................................................................ v
Chapter 1 Introduction .................................................................................................... 1
1.1. Introduction to WMNs......................................................................................... 1
1.2. Problem statement................................................................................................ 2
1.3. Contributions........................................................................................................ 3
1.4. Organization of thesis .......................................................................................... 4
Chapter 2 Background and related work ........................................................................ 5
2.1. Definition of fairness ........................................................................................... 5
2.2. Motivation for fair scheduling in WMNs ............................................................ 8
2.3. Classification and comparison of fair scheduling techniques.............................. 8
2.3.1. Hard fairness ................................................................................................. 9
2.3.2. Max-min fairness .......................................................................................... 9
2.3.3. Proportional fairness ................................................................................... 10
2.3.4. Mixed-bias scheduling ................................................................................ 10
2.3.5. Maximum throughput ................................................................................. 11
2.3.6 Comparison of techniques............................................................................ 12
2.3.7. Centralized and decentralized scheduling................................................... 14
2.4. Analysis of limitations and assumptions in current techniques ......................... 15
2.5. Motivation for cross-layer design in WMNs ..................................................... 18
2.6. Overview of cross-layer architectures ............................................................... 19
ii
2.7. Analysis of cross-layer design ........................................................................... 20
2.7.1. Strengths and weaknesses of cross-layer design......................................... 21
2.7.2. Comparison of cross-layer techniques ........................................................ 22
2.7.3. Cross-layer extension of existing scheduling solutions .............................. 24
2.8. Motivation for cross-layer mixed-bias framework ............................................ 25
Chapter 3 Proposed approaches .................................................................................... 27
3.1. System assumptions ........................................................................................... 27
3.2. Overview of fair scheduling approach............................................................... 28
3.3. Overview of cross-layer mixed-bias scheduling................................................ 30
Chapter 4 Detailed descriptions of the proposed approaches ....................................... 34
4.1. Fair scheduling for WMNs with multiple gateways .......................................... 34
4.1.1. Requirement tables...................................................................................... 36
4.1.2. Requirement propagation............................................................................ 38
4.1.3. Clique generation ........................................................................................ 40
4.1.4. Schedule generation .................................................................................... 40
4.2. Cross-layer mixed-bias techniques .................................................................... 41
4.2.1 Biasing against distance from gateways ...................................................... 43
4.2.2. Biasing against low demand (queue size)................................................... 45
4.2.3 Biasing against poor link quality ................................................................. 46
4.2.4 Combined cross-layer mixed-bias scheduling ............................................. 47
Chapter 5 Performance evaluations of the proposed approaches ................................. 49
5.1. Comparison of simulation environments ........................................................... 50
5.1.1. Custom C/C++ simulation environment ..................................................... 50
iii
5.1.2. Network simulation 2.................................................................................. 50
5.1.3. Network simulation 3.................................................................................. 51
5.2 Performance metrics ........................................................................................... 52
5.3. Performance evaluation of fair scheduling with multiple gateways.................. 53
5.3.1. Simulation environment.............................................................................. 54
5.3.2. Performance metrics ................................................................................... 55
5.3.3. Simulation parameters ................................................................................ 55
5.3.4. Effects of number of mesh routers.............................................................. 56
5.3.5. Effects of number of gateways ................................................................... 58
5.3.6. Summary of fair scheduling results ............................................................ 59
5.4. Performance evaluation of cross-layer mixed-bias scheduling ......................... 60
5.4.1. Simulation environment.............................................................................. 61
5.4.2. Performance metrics ................................................................................... 61
5.4.3. Simulation parameters ................................................................................ 62
5.4.4. Effects of number of mesh routers.............................................................. 63
5.4.5. Effects of number of gateways ................................................................... 66
5.4.6. Effects of number of traffic flows............................................................... 68
5.4.7. Summary of cross-layer mixed-bias results................................................ 69
Chapter 6 Conclusions and future work........................................................................ 70
6.1. Conclusions........................................................................................................ 70
6.2. Future work........................................................................................................ 71
REFERENCES ............................................................................................................. 73
iv
List of Figures
Figure 1 Venn diagram showing relationship between throughput and fairness................ 6
Figure 2 Example of a Mobile WMN for various applications ........................................ 16
Figure 3 Wireless Mesh Network (WMN) with multiple gateways ................................ 29
Figure 4 Distributed Requirement Tables Combined at GW ........................................... 38
Figure 5 Average Packet Delivery Ratio with Varying Mesh Routers............................. 56
Figure 6 Average Delay with Varying Mesh Routers ...................................................... 57
Figure 7 Average Packet Delivery Ratio with Varying Gateways ................................... 58
Figure 8 Average Delay with Varying Gateways ............................................................. 59
Figure 9 Packet Delivery Ratio, Two Flows..................................................................... 64
Figure 10 Average End-To-End Delay, Two Flows......................................................... 65
Figure 11 Average End-To-End Delay, Five Flows ......................................................... 66
Figure 12 Average Packet Delivery Ratio, 250 Clients, 50 MR....................................... 67
Figure 13 Average End-To-End Delay, 250 Clients, 50 MR ........................................... 68
Figure 14 – Packet Delivery Ratio, Varying Flows.......................................................... 69
v
List of Acronyms
AODV Ad-hoc On Demand Distance Vector
DCF Distributed Coordination Function
DSDV Destination-Sequenced Distance-Vector
ETX Expected Transmission Count
GPS Global Positioning System
GW Gateway
MAC Medium Access Control
MANET Mobile Ad Hoc Network
MC Mesh Client
MR Mesh Router
NS2 Network Simulation 2
NS3 Network Simulation 3
OLSR Optimized Link State Routing
PDR Packet Delivery Ratio
PER Packet Error Rate
QoS Quality of Service
RTT Round Trip Time
SINR Signal to Noise plus Interference Ratio
STDMA Spatial Time Division Multiple Access
TDMA Time Division Multiple Access
WLAN Wireless Local Area Network
WMN Wireless Mesh Network
1
Chapter 1
Introduction
1.1. Introduction to WMNs
Wireless Mesh Networks (WMNs) have become the focus of much research since
they allow for increased coverage while retaining the attractive features of low cost and
easy deployment. WMNs have been identified as key technology to enhance and
compliment existing network installations as well as provide access where traditional
technology is not available or too costly in install [18]. A WMN is made up of mesh
routers (MRs), which have limited or no mobility, and mesh clients (MCs) which are
often fully mobile. The mesh routers form the backbone of the network allowing the
clients to have access to the network through the backbone. We propose an algorithm for
fair scheduling in WMNs with multiple gateways. We also propose another algorithm for
scheduling which places more emphasis on throughput while retaining a basic level of
throughput called mixed-bias. This technique biases against characteristics of the network
which are detrimental to performance, fairness, or both.
Many protocols currently implemented for WMNs have evolved from traditional
single-hop wireless local area networks (WLAN) and mobile ad-hoc networks
(MANET). However, both of these networks have characteristics which make them very
different from WMNs. While WLANs have relatively static topologies, MANETs on the
other hand are fully mobile. Therefore, using protocols designed solely for either of these
networks alone does not take advantage of some of the most advantageous features of
WMNs. In MANETs all nodes are routers and suffer from limited power and bandwidth.
In a WMN the MRs have greater resources available than the MCs which is a property
2
that may be exploited. Although a lot of research efforts have been made to address these
problems and some new specialized algorithms have been proposed specifically for
WMNs, there are still many challenges in the area. Many of the existing solutions make
many assumptions that can be relaxed to allow for a more general approach to be taken.
There is much motivation for studying fair scheduling in WMNs. Without work
on fair scheduling for wireless mesh networks it would be extremely difficult to design a
commercial network where all users could expect relatively equal service. Many existing
deployments of WMNs do not account for greedy users or flows, which means that it is
possible for extremely unequal service between users. Some existing protocols are
designed without fairness in mind. These protocols focus on other important
characteristics such as high throughput. Greedy nodes may take advantage of the network
by using an unfair share of the network resources such as bandwidth, causing other nodes
to get an unfair share. Many WMN deployments are optimized with respect to
throughput, delay, or some other features that give little regard to fairness. This thesis
provides two approaches to scheduling: One approach that emphasizes fairness while
relaxing the assumption of a single gateway. The second approach is a cross-layer
mixed-bias approach that emphasizes throughput while retaining a basic level of fairness.
1.2. Problem statement
Since a WMN is a multi-hop wireless network, there are unique challenges to deal
with when compared with traditional wired and wireless networks. The wireless channel
is a broadcast medium, meaning that all nodes within a certain range are subjected to
interference and cannot transmit simultaneously. At the same time, it is difficult to sense
whether communication is taking place in other parts of the network because of hidden
3
and exposed node problems [16,24] where an intermediate node may be stuck in between
two nodes which are trying to transmit simultaneously but out of range of each other. The
solution to many of these problems can be scheduling of transmissions. The scheduling
algorithms that exist have problems as well. As mentioned previously, scheduling in
WMNs seems to be a problem of balancing two often disjoint goals; one is ensuring
fairness among client nodes in the network, another is ensuring the network is performing
at as close to capacity as possible. The goal of a good scheduling algorithm is to find a
balance between these two goals. This thesis will provide scheduling techniques for
wireless mesh networks which find a balance between fairness and throughput in WMNs
using both traditional layered approach and emerging cross-layer design based
optimization
1.3. Contributions
The main contributions of this thesis are (i) in-depth comparison and analysis of
existing scheduling techniques [19], cross-layer design approaches [20], and simulation
tools for WMNs (ii) the proposal and implementation of a unique fair scheduling
algorithm for WMNs with multiple gateways (GW)s [19] and (iii) the proposal and
implementation of a unique algorithm for scheduling using a mixed-bias approach. The
main contributions to the fair scheduling approach are distributed routing and
requirement tables at each router which help in generating the scheduling. The main
contributions to the mixed-bias approach are three new mixed-bias techniques: mixed-
bias against queue length, mixed-bias against link quality and lastly a more general
combined mixed-bias approach which combines biasing approaches against many
characteristics. The proposed techniques are compared to similar existing solutions as a
4
baseline to gauge the performance of our approach against the existing techniques. The
performance is evaluated using simulation with respect to two metrics: packet delivery
ratio and average end-to-end delay.
1.4. Organization of thesis
The remainder of this thesis is organized as follows. Chapter 2 will give
background and related work on scheduling in WMNs. We will define fairness, give
motivations for studying scheduling, and compare and contrast existing work. This will
be followed by an overview on cross-layering. Cross-layering is an important technique
for enhancing performance in wireless networks and will be used in this thesis. We will
also compare and analyse scheduling and cross-layer based techniques to provide a
thorough review of the literature in the area. Chapter 3 will include a brief overview of
both the fair scheduling and mixed-bias techniques which are proposed in this thesis.
Chapter 4 presents detailed description of both of the approaches along with in-depth
descriptions of the algorithms and components of each solution individually. This
includes discussions on distributed routing tables, the requirement propagation in the
WMN, and the scheduling for the fair approach. For the mixed-bias approach, four
mixed-bias techniques are presented formally in detail. In Chapter 5 experimental results
are presented. We also provide a discussion of the performance of our proposed
approaches in terms of packet delivery ratio and average end-to-end delay. Finally,
Chapter 6 concludes the thesis and discusses future research directions.
5
Chapter 2
Background and related work
In this chapter we first define fairness with respect to wireless mesh networks. We
then give a brief introduction to fair scheduling techniques. This is followed by a
classification, comparison, and analysis of current scheduling solutions. The literature
review establishes where our proposals stand in comparison to the existing work. We
identify areas where more research could be accomplished in the future. Moreover we
identify work that is most similar to our own. Lastly, we describe how cross-layer design
can be used to further improve scheduling in wireless mesh networks and why a mixed-
biased cross-layer approach is a promising technique for cross-layer scheduling.
2.1. Definition of fairness
A number of scheduling and resource allocation techniques have been proposed
for WMN in literature [14,27,28,30,31,37,39,40,41,55]. The trend is a tradeoff between
the throughput and fairness using a constant weighting system or a dynamic weighting
system that changes the weights over time to achieve a long-term fairness. It is important
to note that fairness could occur at different points in a wireless mesh network. Some
researchers have proposed per-mesh-router fairness or per-link fairness [42]. There is also
a notion of “uplink-downlink fairness” [31,35,46] because the mechanisms in some
current solutions, such as IEEE distributed coordination function (DCF) [46] allows for
inequality between the directions of flow in WMNs. In other words, an improvement in
downlink throughput may severely affect performance of the uplink or vice-versa.
However, more recently [42,54] has focused on per-client fairness. The motivation
behind this is that in commercial applications each user is paying an equal amount of
6
money for services from the network so each user should get equal Quality of Service
(QoS). It is also important to consider which metrics fairness is being defined with
respect to. For example, a scheduling algorithm could provide fairness in terms of the
possible throughput available but the delay may not be equal. Certain nodes in the
network may remain starved for traffic while other nodes are free to communicate for
various reasons. It is also important to consider that fairness and scheduling is affected by
intruders in the system. Many of the existing solutions for scheduling in WMNs rely on
the assumption of co-operation between nodes, and this is not always the case in real
world networks.
The Venn-diagram shown in Figure 1 helps to illustrate the tradeoff between
throughput and fairness in existing solutions. In one circle is throughput and in the other
is fairness. On the right side of the diagram in Figure 1, the algorithms which favour
throughput exist where fairness is a very low priority.
Figure 1 Venn diagram showing relationship between throughput and fairness as well as some fair
scheduling techniques [19]
These algorithms usually give preference to flows which are least expensive by
some criteria. These criteria may be distance from the gateway, delay, small flows and
Throughput
Fairness
Max Throughput
Hard-Fairness,
Round Robin
Max-Min
Proportional,
Weighted
Mixed Bias
7
other similar metrics. However, this approach allows for starvation or reduced QoS for
flows which do not meet the criteria. Preference is given to greedy flows. On the right
side of the diagram in Figure 1 is absolute or hard-fairness. This side gives little priority
to throughput and ensures that each client gets a fair share of the network resources. This
may be achieved by using a time division mechanism or other similar approaches.
The problem with this approach is that not all flows require the same amount of
resources at all times so the resources may remain unused at times resulting in poor
throughput. One approach which aims for a balance between the competing goals of
fairness and throughput, denoted as max-min fairness [62] works by maximizing the
minimum data rates for each flow. It results in higher throughput than hard-fairness,
however, the overall throughput is still much less than maximum throughput and leaves
much to be desired. The most interesting definition of fairness then is a compromise
between hard-fairness and maximum throughput. In [9,31,37,47,54,59] this approach has
been denoted as proportional fairness. Proportional fairness assigns priority to certain
flows based on criteria such as the number of hops or amount of resources requested.
Similarly, the max-min approach has also been modified with a proportional factor as
well yielding improved results. A new approach called mixed-bias is a hybrid approach
which emphasizes throughput while still providing a basic level of fairness. In the scheme
proposed in [54], a portion of the resources are assigned to a strong biasing against nodes
which are far away from each other. In order to prevent starvation, however, another
portion is assigned to a proportional or max-min scheme as well. This is one of the first
approaches that is able to offer a minimum level of fairness while retaining throughput
which is often even greater than of proportional fairness or max-min.
8
2.2. Motivation for fair scheduling in WMNs
The first motivation for studying fairness in WMNs is in networks where users
are paying equal amounts of money for service and expect a similar quality of service
(QoS). Often existing solutions focus on either the problem of throughput or the problem
of fairness. It is often difficult to create a solution which addresses both of these problems
since they are divergent goals. Recently, however, with works like those of [54], it is
possible to have high throughput solutions that avoid node starvation.
Mesh Clients (MCs) which are far away from the gateways (in terms of hops)
often receive much lower QoS than those which are very close. This is because while the
farther users’ packets are traversing all of the hops along the path, there is a transmission
and queuing delay at each hop. The nodes which are close to the gateways do not
experience this and can often transmit many packets while the farther nodes are still
waiting for one packet to arrive. However, if we give each node enough time to transmit
regardless of distance the throughput of the network decreases dramatically. This is
because the delay increases greatly by giving each node enough time to transmit
regardless of distance to the GW. Some nodes may end up waiting almost indefinitely
while other nodes are transmitting.
2.3. Classification and comparison of fair scheduling techniques
Fair scheduling protocols for wireless mesh networks can be classified into five
categories. These categories in order of fairness from the most fair to the least fair are:
Hard-fairness [37,9,42,46,55], max-min [62,54], proportional fairness [9,31,37,47,54,59],
mixed-bias [54] and maximum throughput [7]. This classification and the relationship
among them are shown visually in Figure 1.
9
2.3.1. Hard fairness
Hard fairness [37,9,42,46,55] is also known as round-robin scheduling. It has
been used in some of the earliest wireless networks and in simplistic network models
since it is the least complex. It is the fairest scheme since each node is guaranteed exactly
equal amount of time in order. In networks where the nodes only require a small
proportion of resources hard fairness causes problems. Since each node is given time to
transmit at regular intervals, if the node does not have any data to send, the time is
wasted. This leads to very low overall throughput. At the same time, however, the
problem of node starvation does not exist. Resources are assigned to each node as shown
in Eq. 1.
(Eq.1)
Where: R is the resources allocated to the node
2.3.2. Max-min fairness
Max-min fairness [54,62] allocates resources in order of increasing demand. The
minimum amount of resources assigned to each node is maximized. So if there are more
than enough resources for each node, every node gets what it needs. If there is not, the
resources are split evenly. This means that the nodes which require fewer resources get a
higher proportion of their need satisfied. The nodes which require more resources end up
dropping many packets and thus the network ends up with still quite low packet delivery
ratio. This type of scheme works best in situations where there is not large differences in
resources requested at each node. This can be a problem in a mesh network because
intuitively, the nodes closer to the gateways will experience much higher traffic than
nodethethroughflowsR
#
1=
10
those on the outside of the network, yet may end up dropping many of the packets
anyway. This may be partially solved by increasing the resource capacity of nodes closest
to the gateways.
2.3.3. Proportional fairness
Proportional fairness [9,31,37,47,54,59] allocates resources proportional to some
characteristic in the network. For example, one may choose to give priority to nodes
which are close to the gateways in a wireless mesh network. The amount of resources
allocated then would be proportional to how close the node is to the gateway. The
strength of the proportionality can be controlled depending on the proportionality factor
as can be seen in Eq. 2.
(Eq. 2)
Where: R is the resources allocated to the node
c is the characteristic which priority is given to, c > 0
β is the proportionality factor, β > 0
2.3.4. Mixed-bias scheduling
Mixed-bias [54] scheduling allows for different levels of control over resources.
Rather than just allowing for one bias, this scheme mixes two different biasing levels
together. A certain proportion of the resources are assigned to one factor and the rest to
another factor as shown in Eq. 3. This allows the scheduling algorithm to provide two
different biasing levels or “mixed-biasing” against a certain characteristic. Rather than
just strongly biasing against that characteristic which may result in certain nodes to be
βcR
1=
11
starved, the mixed-biasing allows for a combination of weak and strong biasing meaning
that a portion of the resources are reserved to provide a minimum service level, even for
the nodes which are undesirable in terms of certain characteristics. This is shown
mathematically in Eq. 3.
(Eq. 3)
Where: R is the resources allocated to the node
c is the characteristic which priority is given to, c > 0
β1, β2 are the proportionality factors, β1, β2 > 0
α is the fraction of resources assigned to each bias α >= 0
2.3.5. Maximum throughput
Maximum throughput [7] scheduling has only one goal. As the name suggests,
this goal is to maximize throughput. This is the only concern of this particular type of
scheduling. Whichever node requires the most resources, or can transmit the fastest or
most data gets access to the resources first. This ensures a very high throughput,
however, there is a limitation with this approach. Nodes which have less priority, such
as; those far away from gateways, those with fewer users, fewer flows, or less demanding
traffic are essentially ignored. If enough time passes, all of the packets waiting in the
queues at these MRs are dropped causing some nodes to be starved for traffic. This
causes performance problem and should be avoided.
21
1ββ
αα
ccR
−+=
12
2.3.6 Comparison of techniques
In general, as fairness increases throughput decreases, although in [59] a study
compared round-robin to proportional fairness and found that in certain situations (such
as indoors) both perform similarly due to fading and interference effects. If this is the
case it is better to use the fair scheduling approach because fairness is an important
consideration in networks, especially if ensuring fairness produces little or no noticeable
effect in the overall throughput. In some cases, ensuring fairness may actually increase
the overall throughput. For example, consider the case where a node close to the GW is
using far more resources than other nodes in the network. A farther node may request
resources very infrequently, however, because of the closer node, its requests may be
ignored or rarely serviced because the closer node will almost always get its requests to
the GW before the farther node.
Having a network which offers high throughput to only some users is not ideal. It
is better to have lower overall throughput, but similar throughput available to all users.
This does not necessarily mean that instantaneous throughput must be fair. Long term
fairness is also desirable. At a discrete point in time, certain nodes may not require their
full share of the resources. So it may be better to distribute some of the free resources to
other nodes. There are two causes to unfairness in a wireless mesh network: (i) A small
portion of users are consistently requesting more than their share of resources and; (ii)
All or most users in the network are consistently requesting more than their share. The
first situation results in nodes which are starved due to the greedy nodes. The second
situation results in the complete breakdown of the network if sufficient scheduling is not
implemented. In an extremely congested network, there are no nodes left to borrow
13
resources from and so a good scheduling algorithm is required to balance the share of
resources. The scheduling scheme closest to achieving this goal is mixed-bias scheduling.
Proportional fairness and max-min scheduling are actually subsets of this type since they
both can be incorporated into mixed-biasing techniques as a biasing strategy.
The most common scheduling technique in literature is proportional fairness
[9,31,37,47,54,59]. While using a weighting or proportionality factor is a reasonable idea
there are other more recent techniques that have emerged. In [37] proportional fairness is
partially used. When the resources in the network are plentiful, the system defaults to
maximum throughput and when it becomes congested or busy it enforces fairness in the
network.
Another approach, as mentioned earlier is mixed-bias [54]. Mixed-bias is similar
to that of [37] since it segregates the network resources and applies different fairness
techniques to each section of resources. So as the network becomes more and more
congested and busy the fairness enforcement level becomes stricter. For instance, the first
ten percent of the network may use round-robin scheduling to give precedence to greedy
flows while there are plentiful resources and the network is not congested. As the
network becomes more congested the greedy nodes become penalized more so that a
long-term fairness is achieved. In the case of [54], however, instead of biasing against
congestion, it was chosen to bias against distance between nodes. Some other
characteristics which may be biased against include congestion (queue size), link quality,
and transmission rate; however, this was not noted in [54].
14
2.3.7. Centralized and decentralized scheduling
A scheduling algorithm can also be classified based on whether or not they are
centralized, the type of fairness and the metric or mechanisms they use in scheduling. In
[45] there is a comparison between the key features of centralized and distributed
approaches for scheduling. In [45], there are three observations which explain why
distributed scheduling is beneficial: (i) nodes which cannot communicate with the
coordinator cannot communicate at all in centralized schemes (ii) overhead from nodes
communicating with the coordinator is reduced or eliminated in distributed approach and;
(iii) the single point of failure problem is eliminated. For situations where the MRs are
anticipated to be static, or the network size is small, it may be easier and more beneficial
to use centralized scheduling. In contrast, when the MRs are mobile, it may be better to
make use of a distributed approach in case the network becomes partitioned due to
mobility. If reliability is a concern or the network size is large, distributed scheduling
may also be preferred due to increased reliability and lower overhead.
For fair scheduling in WMNs many of the algorithms use the concept of a
bandwidth allocation vector [31] or similar approaches [54] to determine how much of
the network resource is required from a flow, as well as when and how to schedule the
resources. Also fair scheduling algorithms which attempt to avoid collision altogether
make use of a compatibility matrix to determine which nodes can communicate at the
same time without collisions [9, 42]. Table 1 summarizes the three classifications
discussed to allow quick comparison of recently proposed scheduling algorithms for
WMNs. In order to keep the table to a manageable size the following notations were
15
used: RR: Round Robin, PF: Proportional Fairness, M-M: Max-Min and M-B: Mixed-
Bias.
TABLE 1. SUMMARY OF FAIR SCHEDULING ALGORITHMS FOR WMN [19]
Reference Type of Fairness Metric / Mechanism
Salem [42] RR with spatial
re-use
compatibility matrix for collision
avoidance, distributed
Erwu [31] PF bandwidth requirement weight-factor,
distributed
Koutsonikolas [9] RR interference threshold, SINR,
centralized
Gupta [47] PF with service levels access threshold, distributed
Viadya [45] PF Back-off interval, distributed
Sorensen [56] PF, RR TDMA with and without weighting,
distributed
Cao [37] PF when necessary bandwidth-allocation vector,
centralized / distributed
Nandiraju [46] RR uplink-downlink independent uplink & downlink DCF,
centralized
Singh [54] M-B bias weight function
2.4. Analysis of limitations and assumptions in current techniques
In this section we provide an in-depth analysis of some of the assumptions and
limitations of the current approaches for fair scheduling in WMNs. The assumptions we
will focus on are: limited or no mobility for the MRs, fixed topology of MRs and
gateways, the assumption of a single gateway and downlink and uplink equivalence. We
will discuss how older techniques developed for single hop and ad-hoc networks may be
adapted and applied to WMNs to further advance research and improve scheduling.
Finally, there will be a summary table once again to allow for easy comparison between
algorithms in terms of assumptions and possible direction for future research. In all of the
following sections, an overview on the assumption is provided. Second, the importance
of the assumption is identified. Lastly, the consequences of relaxing the assumption in
16
future work would be. For more detail on which techniques make which assumptions see
summary Table 2.
The assumption of limited or no mobility for mesh routers in a WMN is made in
[9,29,31,37,45,46,47,62]. This assumption is made in order to reduce complexities when
developing the initial algorithm. However, if this assumption is relaxed it allows for a
more general solution which is more flexible and useful. Consider for example a WMN
where the MRs are mounted on cars, trains and buses as part of a transit system as shown
in Figure 2. The MRs could provide Internet access to passengers on the transit system,
allow for wireless surveillance systems on the vehicles to keep passengers safe, or to
collect information on the locations of the vehicles to provide estimates on arrival times.
In this system we could still make the assumption that the MRs have more resources
(power resources, processing and memory) compared with the MCs. Such networks
constitute a typical application of future mobile WMNs and could make use of techniques
used in wireless ad hoc and sensor networks as well as those from WLANs for
scheduling and load balancing.
Figure 2 Example of a Mobile WMN for various applications
17
Another assumption is static topology [9,29,31,37,45,46,47,62]. In this case, it is
assumed that the topology of the MRs and gateways is either fixed or rarely changes and
as such can be manually configured. This is contrary to one of the most important
benefits of WMN. A WMN is supposed to be self-configuring, self-healing and flexible
so MRs and gateways should be able to be added / removed and as mentioned above,
mobile. Coincidentally, all of the solutions which assume no mobility also assumed static
topology. This is interesting because it is often much easier to handle adding or removing
nodes than it is to handle moving nodes, since there is no handoff requirement in the first
case.
Again, to keep the scheduling and load balancing algorithms simple it is assumed
that there is only one gateway [42,45,46,47,56,62] in the WMN or even assumed that
there was no gateway at all (the traffic was limited to local network traffic only)
[9,36,37,54]. However, it has been pointed out that one of the main uses for WMN is to
provide internet access with expanded service areas from traditional WLANs so that
means the majority of the traffic flow is between the gateways and the MCs [61]. Having
only one gateway in this scenario is a major bottleneck so the existing solutions should be
extended to be able to support any number of gateways to make a truly scalable WMN.
In [31,35,46], it is explored whether uplink and downlink scheduling can be
treated equally for scheduling and load balancing. The reasoning behind this is that there
could be a flow which makes use of uplink traffic to large extent while hardly requiring
any downlink traffic, so if there are different schedules for both uplink and downlink,
perhaps a higher throughput and greater fairness could be achieved [46]. In
[9,29,36,45,47,61,62] it is just assumed that downlink and uplink are equivalent. In
18
[37,56] only one is dealt with at once (for example just uplink scheduling) [37] leaving
the reader with the assumption that the opposite (for example downlink) [56] may be
dealt with in the exact same manner.
Table 2 summarizes the previous five categories to allow for easy comparison and
identification of areas of future work for scheduling in WMNs. For clarity we define
downlink as DL and uplink as UL in Table 2.
TABLE 2. SUMMARY OF ASSUMPTIONS FOR FAIR SCHEDULING IN WMN.
Assumptions
Reference Limited
Mobility
Fixed
Topology
Single GW DL = UL
Multi-
hop
Salem [42] partial partial yes Yes Yes
Erwu [31] yes Yes yes No Yes
Koutsonikolas [9] yes Yes no GWs Yes Yes
Bejerano [61] partial partial no Yes Yes
Ramachandran [29] yes Yes no Yes Yes
Gupta [47] yes Yes yes Yes No
Viadya [45] yes Yes yes Yes No
Sorensen [56] yes Yes yes DL only No
Cao [37] yes Yes no GWs UL only Yes
Bejerano [62] yes Yes yes Yes No
Nandiraju [46] yes Yes yes no No
Singh [54] yes Yes no GWs yes Yes
Popa [36] no No no GWs yes Yes
2.5. Motivation for cross-layer design in WMNs
As discussed in [15], certain features within WMNs demand cross-layered design
such as advanced antenna technologies, physical layer technologies [22,27,28] and multi-
channel [23] or multi-radio technologies [3,49]. Sometimes it is necessary to have
information from various layers in order to make decisions in higher layers in the
network. This may result in better decision making when it comes to routing, allocating
resources or scheduling. It has been confirmed in [34] that cross-layer design is a
promising technique for performance improvements in WMNs. Several solutions for
19
scheduling in WMNs do not consider cross-layered optimizations or leave this as a future
work [36,42,54]. Additionally, there are also some works that do consider cross-layer
design but are for ad-hoc networks or single-hop wireless local area networks instead.
These solutions could be extended to include WMNs as well and take advantage of the
unique properties that exist. Cross-layering allows network layers which normally are
unable to communicate in the traditional layered network models to share data. This
shared data may allow more intelligent decision making in terms of routing or
scheduling. One innovative technique for achieving cross-layered design is to have a
network status stack which contains information from all of the layers in the traditional
network stack [38]. The traditional stack is then modified to take information out of this
parallel stack in order to make more intelligent decisions.
2.6. Overview of cross-layer architectures
There are several popular architectures which can be used for cross-layer design in
WMNs. In [38], a common stack exists in addition to the traditional layers. The shared
stack is accessible by all layers. Each layer can then use information from this shared
stack to make more informed decisions at various points during communications. The
main advantage of this technique is that it is very extensible compared to other techniques,
and often the traditional layered network stack can remain intact. If there is information in
the dual stack then the protocols can use it. Otherwise, they can fall back to legacy
techniques. The implementation of this technique in practice can vary, from a simple
parallel stack to a database style system.
In contrast to this, there are various highly coupled techniques, which require two
or more layers to communicate directly with one another by modifying the existing
20
layered model. These types of solutions are difficult to extend and maintain and should be
avoided [57]. These types of designs may be useful for cutting edge work that is not
mature, since the development is less planned and can go ahead more rapidly.
Lastly, there are also systems which do not directly communicate, but insert extra
control packets into the system that are meant to be read by the layers and discarded
before the actual communication occurs. This type of system may not be as practical since
extensive modifications may need to be made in order for the layers to understand the new
control packets. Since there is little regard for software engineering concepts in this
approach, future modifications and extensions may become difficult and costly.
Each cross-layer architecture has its benefits and drawbacks, however, the main
priority in cross-layered architectures are primarily: (i) easy maintenance and (ii)
extensibility. Several others confirm this viewpoint [21,38,57,58]. The performance aspect
of the cross-layered approach comes from selecting suitable information and how that
information is used. The architecture should be independent from this. Using an
architecture based on good software engineering principles allows future work that is
increasingly complex and based on even more feedback from all the layers. In this way,
with greater feedback we can achieve better performance and make far better decisions
than we can in a traditional layered design. Without good architecture, however, this
becomes difficult and we are limited to a small amount of feedback.
2.7. Analysis of cross-layer design
Scheduling has been studied in operating systems for user scheduling, in databases
for transaction scheduling. Cross-layered design has been studied to some extent in wired
networks. There is less motivation for cross-layer work in wired networks because the
21
channels are more reliable, and are not prone to radio interference in the way that wireless
networks are. Other the other hand, in wireless networks, cross-layering has been studied
in both ad-hoc networks and single hop wireless local area networks. These solutions
could both be extended and the same techniques could be used in wireless mesh networks.
2.7.1. Strengths and weaknesses of cross-layer design
There has been some debate in the past few years on whether or not cross-layered
design in wireless networks is beneficial. In [57] a cautionary perspective on cross-layer
design is presented. It states that the cross-layer approach can cause undesired interactions
and make it difficult for future innovations on the protocol. There are several
considerations that must be taken into account when making use of cross-layer design for
wireless networks. These include avoiding unbridled design that causes “spaghetti code”,
being careful to consider unintended effects from cross-layer optimizations, (for example
timescale separation may be required for some optimizations) and long term architectural
value of the optimizations [57,58]. On the other hand, the performance improvements and
increased ability to make informed decisions based on more data makes cross-layering an
attractive option for use in wireless mesh network scheduling protocols. Cross-layering is
especially useful at the medium access control (MAC) layer which is the layer that decides
when a device has the ability to access the medium. If much information is combined here
to give the MAC layer an informed decision, the network may perform much better. For
instance, if the MAC layer knows information about the link quality, the transmission
speeds available, how much queue space it has left, and its distance from its final
destination, it is able to gauge how much priority its transmission should have. As long as
the system for getting the relevant information to the MAC layer is designed carefully,
22
there is no reason why cross-layering should not be applied. Similarly at the routing layer,
if we know information about the state of the links, we can make a good choice on which
link may be the best option for sending the packet. As mentioned previously, one popular
method for retaining good design and a cross layered approach is by using a parallel stack
which may be accessed by a modified traditional stack [38]. In this way the traditional
layered stack from wired networks is retained and still modular while avoiding code which
is difficult to maintain and improve.
2.7.2. Comparison of cross-layer techniques
Based on literature reviewed, we identified four techniques or methods in which
cross-layered design is applied for fair scheduling in wireless mesh networks. These are
power adaption, rate control and route adaption and network coding. The following
paragraphs will outline some examples of each technique of cross-layered design. The
layers used in each technique will also be identified.
1. Power Adaption: Power adaption is where the power levels of competing
nodes are adjusted to ensure greater fairness between the two. This technique is used in
[23,24]. In addition to ensuring more fairness in regards to scheduling it also allows the
MCs to save power which is especially important in networks where the clients are
restricted in this respect, for example in a wireless sensor mesh network. In [24] the
cross-layering is between the network and link layer while in [23] cross-layering is
between the link layers (MAC) and the physical layer. According to [13], using this
approach “solves the hidden terminal problem without aggravating the exposed terminal
problem.”
23
2. Rate Control: Rate control allows the MRs to control the transfer rates of the
MCs which are associated with them. This technique is useful for scheduling while
retaining respectable throughput. The rates for a given link are raised when the link
quality is higher, however, to ensure fairness it cannot go above a certain threshold where
other MCs are affected. This technique is used in [23, 60]. In [60] cross-layering is
between the transport and link layer. In [13] this type of cross-layering is useful for
increasing network performance because of more accurate link and network conditions
provided to the corresponding congestion (rate) control. In [23] it is between the network,
transport and link layers. One additional interesting example is [27] where a combination
of power control and rate control is used. This protocol is a link (MAC) and physical
layer cross.
3. Route Adaption: Route adaption is used as a congestion avoidance technique
that can also help ensure fairness in the network. This technique works by changing the
routing information for certain flows based on the congestion levels within the path. If a
given path becomes congested, information from lower layers (such as link layer)
informs the higher layers (example network layer) that a new path should be taken. Once
again there must be a threshold which is crossed and only some of the paths should be
informed so that a “ping-pong effect” is avoided where the paths are constantly changing
back and forth from constant congestion. This technique is used in [39, 41]. In [41] the
cross-layering is between the network and link layer and in [39] it is between the
transport and network layer.
4. /etwork Coding: Lastly, network coding is used to allow multiple unicast
transmissions to occur simultaneously. In [28] a concept of poison-antidote is presented
24
where the poison is considered to be the coded bit which is decoded by the antidote. The
cross-layering in this protocol occurs in the link (MAC) and physical layers. For a
summary of all of the above mentioned protocols see Table 3.
TABLE 3. CROSS-LAYERED DESIGN FOR SCHEDULING TECHNIQUES
Reference Technique Layers
J. Tang et. al [22] Power Control Link (MAC), Physical
J. Thomas [24] Power Control Network, Link (MAC)
X. Wang et. al [60] Rate Control Transport, Link
J. Tang et. al [23] Rate Control Transport, Network, Link
K. Karakayali et. al [27] Power / Rate Link (MAC), Physical
M.J. Neely et. al [39] Route Control Transport, Network
M.S. Kuran et. al [41] Route Control Network, Link
K. Li et. al [28] Network Coding Link (MAC), Physical
2.7.3. Cross-layer extension of existing scheduling solutions
There are many solutions already proposed for scheduling in wireless mesh
networks that do not already make use of cross-layered optimizations. Many of these
existing solutions may be altered and extended with cross-layering to provide increased
performance or to make more intelligent decisions in some aspects of the protocol. One
particular example is [42] where a fair scheduling protocol for wireless mesh networks is
proposed. Several assumptions are made including: single gateway, uplink-downlink
equivalence, static topology of MRs and non-mobile MRs. These assumptions could be
relaxed and the cross-layered design approach could be applied resulting in an extended
version of the protocol which improves much over the original. For more examples of
protocols which could be extended with cross-layering, it may be useful to consult [16] or
25
other general surveys on WMNs. Many existing protocols are listed as references which
may benefit from cross-layered design.
The limitations of cross-layering techniques are similar to those of normal
scheduling techniques in wireless mesh networks. The solutions that currently exist
make many assumptions including: single gateways (or no gateways) [12,30], limited or
no mobility of mesh routers [11, 12, 22,23], non-overlapping cellular coverage areas [23],
static topologies [23,11] and uplink-downlink equivalence. All of these assumptions
leave much future work to be done in the area. If these assumptions are relaxed more
general and flexible solutions could be designed. On the other hand as noted in [60], the
complexity of cross-layered design may be the reason why so few solutions have been
extended with cross-layering. This limitation can, however, be solved if the optimality
requirement of the scheduling is relaxed. When this is the case, it is often claimed that a
whole class of relatively simple and efficient scheduling can be implemented in a
distributed fashion.
2.8. Motivation for cross-layer mixed-bias framework
There is much motivation for using a cross-layered mixed-bias framework for
scheduling in wireless mesh networks. The merits of using a cross-layered approach have
been outlined previously in section 2.5 so this section will focus mostly on the mixed-
bias approach. In the original framework by [54], promising results are demonstrated
when compared with other biasing schemes such as proportional fairness and max-min.
In the mixed-bias technique, the amount of hops away from a gateway was biased
against. Rather than naively increasing the biasing factor, the original work proposed a
mixed-biasing technique which strongly biases against the hops for a certain portion of
26
the resources while maintaining a weaker bias for the rest. For example, half the
resources may use the strong biasing scheme while the rest may fall back on proportional
fairness. We propose to take this further and bias against several characteristics at one
time in a combined mixed-biased approach. This way, the network resources will be
further divided into different mixed-bias schemes. This more general framework will
allow for further customization of the network and allow for biasing against many
characteristics at once. This will allow greater performance by taking into consideration
more factors which negatively affect network performance such as poor link quality and
greedy nodes (nodes with high sustained demand). If nodes and flows which have
desirable characteristics as determined by the biasing function, a reduction in the overall
number of retransmits should be achieved. A reduction in retransmission will greatly
increase performance since retransmission and the exponential back-off function can
significantly increase the time it takes a packet to reach its destination [1].
27
Chapter 3
Proposed approaches
In this thesis, we present two approaches for scheduling in WMNs. The first
approach is the fair scheduling approach. This approach makes use of a technique similar
to that of [42] while relaxing several assumptions made in the original work. The key
contributions in this approach are multiple gateways and a distributed clique propagation
algorithm. The emphasis of this approach is fairness, leaving overall throughput as a
secondary goal. This is especially important when trying to provide equal service to all
users within the WMN. However, this approach is not always the best for maximizing
throughput and can waste some of the bandwidth in the network. As an alternative to this
approach we also present the mixed-bias scheduling approach. This approach is a unique
interpretation of the mixed-bias technique [54] except instead of biasing against just one
characteristic of the network like the original work, we propose to bias against several
characteristics so that a balance between throughput and fairness will be provided to all
users.
3.1. System assumptions
While one of the goals of a completely general and flexible wireless mesh
network is to make as few assumptions as possible, this work makes several key
assumptions. These assumptions were mainly made in order to keep the amount of time
for the research to a reasonable length and to keep the complexity of the system to a
manageable level. The assumptions we make in this thesis are common among similar
works, however, our work is unique with respect to some in that we do reduce some
assumptions. One of the main assumptions we make in this work is that there is no
28
mobility in any nodes in the network. Mobility introduces new complexity in that a mesh
client may no longer be associated to the network via a single connection point at a single
MR. This property requires some mechanism to keep track of the point of attachment in
the work for each MC and introduces many new problems due to the increased overhead
in handling this situation. There is also the problem of whether the network or the client
should handle the hand-offs, with each affecting performance differently. In the case
where MRs are mobile, even more problems are introduced. If the MRs are sufficiently
mobile, the network may break apart into partitions and certain regions may be unable to
communicate with the rest of the network.
3.2. Overview of fair scheduling approach
The fair scheduling approach we propose relaxes the assumption from [42] that
the network has only one gateway. We also assume that the requirement and routing
tables are distributed across the mesh routers. In our fair scheduling approach it is
assumed that MRs and GWs are not mobile. Their positions are fixed throughout the
simulation. This assumption is quite common in many of the existing solutions. However,
the goal is that eventually this assumption could be relaxed resulting in the concept of a
Mobile Mesh Network where the MCs and even the MRs are not fixed and the topology
of the network is extremely dynamic. There are many benefits and applications of this
type of network. It could be used in transit systems, military applications or disaster
relief. Rather than having to deal with multiple handoffs of many moving clients, the
moving clients could be associated with a moving MR. This would allow the network to
focus on dealing with only one handoff while all of the MCs associated with the MR
retain their attachment to the network. The scheduling we propose is a spatial time
29
division multiple access (STDMA) scheduling algorithm. This is a good approach
because it eliminates interference by only allowing links which do not interfere transmit
simultaneously. In wireless mesh networks in particular, interference has been identified
as one factor which severely degrades performance and scalability [26,32].
In contrast to most existing solutions, we assume that the network may contain
multiple gateways as shown in Figure 3. This is an important assumption because
limiting the network to one gateway causes an extreme bottleneck. Even if the traffic
within the network is balanced, having only one gateway can decrease the performance of
the network because it remains the only location where traffic can enter or exit the
network. Once the assumption of a single gateway is relaxed, we are then free to apply
gateway load balancing across the multiple gateways similar to [53]. The single gateway
property also implies that the network has a single point of failure.
Figure 3 Wireless Mesh Network (WMN) with multiple gateways [19]
There are two solutions to the single gateway problem. One is to assume that the
gateway always has enough capacity to serve the demand of the network, regardless of its
size. However, this solution does not solve the single point of failure problem. The other
option, which we have chosen in this thesis, is to allow multiple gateways so that the load
30
of the traffic is spread around more evenly. Multiple gateways have also been used in
other WMNs such as MIT Roofnet [17, 50], however, the emphasis was routing and thus
scheduling was not studied.
Another common assumption is uplink-downlink equivalence. It has been shown
experimentally that uplink and downlink should be treated with separate techniques
[31,40,46]. In this thesis we concentrate on uplink scheduling. The downlink portion
scheduling has not been considered within the scope of this thesis since it is a large
problem on its own which could incorporate further enhancements such as multicast
trees.
3.3. Overview of cross-layer mixed-bias scheduling
Our cross-layered mixed-bias approach takes ideas from many existing cross-
layer works. Many scheduling solutions focus on biasing against only one characteristic
of the network which is detrimental to scheduling or throughput. This is often true
regardless of whether the scheduling scheme is hard-fairness, max-min or proportional.
For example [22, 24, 30] focus on preventing links from interfering in order to improve
the scheduling by adjusting the transmission power of the nodes. [54] gives preference to
nodes which are close to each other when scheduling. However, in order for a scheduling
algorithm to be effective and flexible in many situations it may be beneficial to try to
capture many characteristics in one scheduling algorithm. Cross-layering allows the
algorithm to gain information from many layers which normally cannot contribute to
scheduling. This may be one reason why scheduling algorithms only focus on one
characteristic.
31
In our approach, we propose a cross-layer, mixed bias scheduling algorithm for
wireless mesh networks. The cross-layering will provide information on link-quality and
distance between nodes. Link quality will be provided from the physical layer while
distance can be provided in many ways. Distance could be computed by the number of
hops between two points, by measuring the delay or by using real life coordinates if the
nodes are equipped with Global Positioning Systems (GPS). A portion of the scheduling
resources will be biased according to a set of heuristics that penalize nodes for various
“bad behaviors” such as distance from the gateway, overuse of traffic, poor link quality
and so on. Each heuristic will be assigned a different proportion of the network resources
which will be determined experimentally. Another portion of the resources will be left
for absolute fairness in order to ensure that none of the links are starved and that some
minimum level of service is maintained. Then the collective system will be optimized to
produce what we expect to be a high throughput fair scheduling for wireless mesh
networks.
In [54], it is argued that a fraction of the total resource allocation should be
strongly biased against long connections while the rest remains allocated in a fair (max-
min), or weakly biased manner (proportional fairness). In their proposal they make use
of only two different scheduling schemes in their mixed-biasing. The approach that will
be taken in this thesis, however, is that by using more than two different scheduling
approaches, more could provide even greater utilization of network resources. For
example, in the original Mixed-bias approach, only the long connections (as in number of
hops) are biased against. It was demonstrated that even while biasing against the long
connections, the decrease in performance for those connections was similar to
32
proportionally fair solutions. At the same time, because of the biasing, the short
connections benefited greatly. It is also possible, however, to bias against greedy
connections, poor quality connections or any other metric. Then as a whole this entire
system of biases could be optimized together to maximize both the fairness of the
scheduling and the throughput of the network.
In our approach, we make use of mixed-biasing to enhance our previous work
which was an STDMA scheduling for wireless mesh networks [19]. The same approach
can also be implemented in a network using an 802.11 distributed co-ordination function
by giving preference in the MAC layer queues to the flows which have the least bias
against them. The mixed-biasing is expected to improve the throughput of the network
while maintaining some degree of fairness due to some of the network resources being
allocated in a fair manner. As opposed to just biasing against long connections, however,
our approach will bias against other factors using a heuristic approach. In our approach
we will represent the capacity of the network, C as
(Eq. 4)
Where: C is the total capacity of the network
γ1, γ2 … γi … γn are the weight of a given biasing scheme
f1, f2 ... fi … fn are the biasing functions for certain characteristic
∑
∑
=
=
=+++++
=+++++=
n
i
ni
n
i
nnii ffffC
1
21
1
2211
1......
1)......(
γγγγ
γγγγ
33
In this general case, any number of biasing function schemes may be applied
which may result in a fine grained tuning of the network against specific biases. If
starvation is to be avoided and some level of fairness ensured, at least one function must
not bias against any node or flow but give “hard fairness”. Also the associated weighting
scheme for this function must not be zero. The weighting could be either static or
dynamic. For example in a dynamic implementation each weight may be allowed to shift
between pre-determined ranges depending on the utilization of the network. For example
one may let γ1 shift from a weight of 0.25 to 0.5 depending on the utilization of the
network or some other parameter.
34
Chapter 4
Detailed descriptions of the proposed approaches
In this chapter, we will present the more technical details of the two approaches in
general, as well as individual algorithms in particular. In the first section, we will give a
detailed description of the first main contribution of this thesis: A fair scheduling for
WMNs with multiple gateways. Then in each subsection, the individual algorithms and
components which make up this scheme are outlined in detail. These components
include the distributed requirements table, the requirements propagation algorithm, the
clique generation algorithm and the schedule generation algorithm. In the second section
of this chapter, we will discuss in detail the mixed-bias approach in general. This will
then be broken down into each of the mixed-bias techniques that will be combined in the
second main contribution of this thesis: the combined mixed-bias scheduling approach. In
this approach all three of the mixed-bias techniques are combined to form one complex
scheduling for the network which biases against several characteristics which are
detrimental to network performance and fairness.
4.1. Fair scheduling for WMNs with multiple gateways
In this scheme, the approach is made up of several different components. Each of
these components will be outlined in greater detail in the following subsections. This
section will provide a general outline of the fair scheduling approach with multiple
gateways, highlighting the main contributions we have made to this approach.
We have proposed an enhancement of the original fair scheduling approach
proposed by [42] which we call the distributed requirements table. The original work
proposed only a scheduling, however, does not provide a mechanism for maintaining and
35
collecting requirements. The requirements are required for generating the scheduling
since this information tells how busy each link is. Thus we propose a distributed manner
of accomplishing this. Each mesh router keeps track of a local requirement table. In this
requirement table, the demand on each link between the router and a neighbour is kept.
When a new schedule is requested, each gateway asks for the partial requirement tables
from each mesh router associated with it.
The gateway then combines these tables to form one complete requirement table
which it uses to generate cliques and eventually the scheduling. One main difference
from [42] approach is that we assume multiple gateways. This means that each gateway
in the network is responsible for scheduling all of the links which will forward packets
towards it. The single gateway assumption is a significant one for two reasons: (i) The
single gateway causes an extreme bottleneck in the network. All traffic which flows in
and out of the network must use this node and so any scheduling work done in the
network is limited by the single gateway. (ii) Similarly, the single gateway node causes a
single point of failure in the network. If the gateway node is to go down in this scheme,
there is no recovery. When multiple gateways are assumed, the bottleneck is eliminated.
Not all of the traffic is destined to the same node in the network and is spread more
evenly, especially with strategic gateway placement. With a more complex scheme than
we proposed, one could further take advantage of the multiple gateways and perform load
balancing on the multiple gateways so that under-utilized gateways could be taken
advantage of for further performance improvements. Lastly, the single point of failure is
eliminated as well. If one gateway experiences an outage, the network has the ability to
reconfigure itself to forward packets and perform scheduling from another gateway. Once
36
the requirement table is formed, the gateway uses this information along with the clique
information to form a scheduling plan. The clique information is all of the sets of links
which may transmit at the same time without interfering with one-another. The clique
information is generated once before any transmissions occur in the network in a manner
similar to the way neighbours are discovered in [25]. In our system model we assume
static nodes and topology, so no nodes are added or removed and there is no mobility.
Thus we do not need to generate this information more than once in the life of the
simulation. This is important because this operation is very expensive computationally,
because clique enumeration is known to be a difficult problem to compute. If we were to
assume non-static topology, we may have to make an assumption of a certain network
size based on the computational resources of the gateway nodes in the network. Using
both the clique information and the requirement information, we can then determine
which links should be activated together and for how long. A further modification of this
scheme would be to use different characteristics other than demand on a link, for example
the quality of the link and the distance from the gateway could also be taken into account
using a biasing scheme as we have proposed.
4.1.1. Requirement tables
The type of fairness used in this solution is round-robin style with spatial re-use.
We use centralized schedule generation at the gateways which makes use of distributed
routing tables located at the mesh routers. We propose the requirement propagation
algorithm which allows each gateway to distribute the requirements and routing table for
the scheduling into the network. At each mesh router, the path to the gateway is
maintained. In this table, requirements for the links on this path are also maintained. For
37
each client requesting to use this mesh router, each link along the way to the gateway in
the local table is given a requirement. When the gateway signals the start time for new
schedule generation, it requests the local requirement information from all of the mesh
routers which are currently using it as their primary gateway. It then combines the
requirements to help determine the scheduling as shown in Figure 4. In Figure 4, each
mesh router has a local requirement table. This requirement table keeps track of the
requirement for itself and for all the nodes on the path towards the gateway. A
requirement is added when a MC sends data to a MR. At that particular MR, the
requirement is incremented for itself and for all hops to the gateway in its local table
since all of these nodes will have to relay the packet. A single gateway is responsible for
generating the scheduling for all of the nodes which route through it. Then when a new
scheduling must be generated, the gateways request the requirement from each table.
Each gateway then combines the requirement information from each mesh router with the
compatibility matrix. The compatibility matrix represents the links which may transmit
simultaneously without interference and is computed or setup manually once when the
network is setup. The gateway then computes the scheduling. After the scheduling is
computed, START packets are sent to the MRs when they are free to transmit and E(D
packets are sent to the same MRs when their transmission period has ended. This
continues until the end of the current schedule and the process repeats.
In our solution, each gateway is responsible for generating a scheduling for the
mesh routers making use of it to relay packets to the Internet. The schedule generation
algorithm from [42,7] requires that a compatibility matrix be generated for the network
before the algorithm operates. The compatibility matrix is a way of representing which
38
links may be activated simultaneously without interference or collisions at MRs. One
main difference with our approach is the use of distributed requirement tables located at
each of the MRs.
Figure 4 Distributed Requirement Tables Combined at GW [19]
4.1.2. Requirement propagation
The requirement propagation algorithm given in Algorithm 1, allows the gateway
to keep track of the requirements across all of the links. At the MR, a table containing a
partial representation of the network is kept for all of the MRs on the way to the gateway.
When a MC associates with a given MR, the requirement is incremented for all the MRs
along the way to the gateway in the local table. When a new schedule generation is to be
completed, the GW requests for the requirements from all of the MRs and combines the
results from the partial tables to determine which links must be activated and for how
long.
39
Requirement Propagation Algorithm Pseudocode
1. Associate MC with MR
2. Generate a Client Requirement at MR for the MC
3. For each link between MR and GW
- Requirement(current-link) ++
4. For each Hop
- Requirement(current-link) --
5. On Drop: For each link between MR and GW
- Requirement(current-link) --
Algorithm 1 Requirement Propagation [19]
In this scheme, each gateway is responsible for generating the centralized
scheduling for all of the links routing to it. The distribution and coordination of the
scheduling is done through the use of START and E(D packets. The gateway sends a
START packet to the MR when it has scheduled time to send and an E(D packet when it
no longer has permission. It is assumed that these control packets are sent on a different
channel from the data and thus do not interfere with data traffic. At the end of one cycle
of scheduling, the process is repeated with a new scheduling plan being computed and
distributed throughout the network.
The round-robin nature of the scheduling allows the solution to be simple
compared to techniques that include weighting functions. At the same time, when
compared with a naive round-robin technique, less time is wasted waiting for links which
have no traffic to send since time is only allocated to links with requirements. Since we
are concerned with fairness among clients who are paying similarly for equal service, this
solution works well. Many existing solutions make use of similar round-robin style
techniques [9,42,46,59] but none of them use of multiple gateways. Using a single GW
to serve a large mesh network is impractical, however, since it becomes a bottleneck
40
quickly as the network size grows. In [42] the solution was distributed in the sense that
the scheduling had to be spread around the network to all the MRs from the centralized
GW, however, the algorithm presented provided no means for the distribution to be
accomplished. We provide a method for this in our solution.
4.1.3. Clique generation
In order to determine which groups of links should be scheduled together, a
concept of gain which was introduced in [42] is used to select groups of links which have
the greatest load. Gain is defined as the sum of the requirements of all the links minus
the greatest requirement. The scheduling algorithm uses the path and requirements
information to give permission to certain MRs to transfer at the required timeslots. When
the fair scheduling algorithm is enabled, a MR may only send packets when it has
permission to do so. If it does not have permission, it retries until a waiting threshold has
been crossed at which point the packet is dropped. When collision occurs because a
buffer is full the packet is dropped. The performance of the network could be improved
further if a retry or backup mechanism was implemented or if load balancing was applied
at the GWs.
4.1.4. Schedule generation
Scheduling is generated for the all of the mesh routers in the network using the
concept of a compatibility matrix similar to that used in [42,55]. The compatibility
matrix is then used to determine which links can be enabled at the same time without
causing interference. In our network model, this means that the two MRs do not have a
common neighbor and are not neighbors with each other. Due to the positioning of the
41
MRs and the communication ranges, if two MRs are not neighbors and do not share a
common neighbor, they are not close enough to cause interference with each other and
they do not compete for the resources of a common neighbor. This way both may
communicate at the same time. The spatial TDMA scheduling allows multiple links to be
activated at the same time when they do not interfere. So the network can be used far
more efficiently than it could if only one link in the entire network were active [13].
Furthermore, since the algorithm uses the concept of compatibility, no two links are
active that compete for resources so collisions are avoided. The solution presented here is
different from many other TDMA solutions because it only allocates time for links which
actually have requirements associated with them.
4.2. Cross-layer mixed-bias techniques
There are four main techniques we will present in the following sections. The first
technique is similar to that of [54]. This technique uses mixed-biasing to bias against the
distance from the gateways. This technique is important because the farther away from
the GW a MR is, the more hops the packet must traverse to arrive at the GW. This means
the probability of a successful delivery decreases as the MRs are farther away from the
GW. Also the average delay increases. Thus if we allow fewer packets from the farthest
gateways will achieve higher throughput overall. At the same time, as the packets move
closer to the GW successfully, there is a greater chance they will arrive since the closer
MRs are given preference.
The second technique favours MRs which have full queues, and thus biases
against those with empty or near-empty queues. This is important because if we can give
some preference to these routers, perhaps fewer packets will be dropped by reducing the
42
frequency with which the queue is full. By giving preference to full queues, we let the
near empty queues build up and at the same time, allow the extremely full queues to
empty. This results in a balancing of all the queues in the network. If we can reduce the
number of dropped packets, the delay will likely also fall significantly since the overhead
in resending a packet is often quite large. This is especially true in multi-hop networks
such as WMNs since the retransmission control packets must also traverse the multiple
hops.
The third technique biases against poor links. This is important because link
quality may change often depending on objects blocking signals, environmental
conditions (such as weather, temperature etc) that may make certain links in the network
perform better than others. When we bias against these links, we give preference to those
links which are performing well and allow them to transmit more, thus increasing the
overall packet delivery ratio.
The last technique is the most ambitious technique. This technique is the
combined mixed-bias approach. In this approach several mixed-bias techniques are
combined to form one all-encompassing scheduling algorithm which allows preferred
access to only those MRs which exhibit all of the qualities of a preferred MR. The
criteria of which MRs are preferred are completely up to the network administrator and
could even be modified dynamically. For example under certain conditions or
applications, more or less characteristics could be factored into the combined mixed-bias
formula. It is expected that the more characteristics we use at one time, the more
complete of a picture of the network we receive and thus we are able to make a more
informed decision on how to schedule access to the MAC layer.
43
It should be noted that for all of these techniques, there are two different
approaches that could be applied when deciding on transmission scheduling. (i) A
centralized approach that could be taken if the gateway is controlling the scheduling
(such as in the STDMA scheduling from the fair scheduling with multiple gateways
approach) (ii) A distributed technique where each MR itself decides whether to send the
packet at a given time or to hold off on sending and give the opportunity to send to
another MR. This technique relies on the assumption that all MRs in the network will co-
operate, which is not always the case. Despite this weakness, the lack of a central
authority removes the single point of failure problem. We will present some experimental
results from both schemes since there are merits to both approaches. It is up to the
network designer / administrator to decide on which technique is best for the particular
application or network.
4.2.1 Biasing against distance from gateways
In our approach for biasing against distance, we follow the same approach as [54],
which forms the basis for much of this work. In the original approach, there is no
concept of gateways but rather an ad-hoc network with a random source-destination
traffic model. It is claimed that mixed-bias provides increased performance compared to
both max-min and proportional fairness if a stronger bias is applied to nodes which are
far away from each other while at the same time applying a weaker bias so that there is
still limited service to the non-preferred nodes. This is different from other scheduling
algorithms because in other solutions, some nodes are starved while others receive all of
the resources. Avoiding starvation is important problem to consider in WMN design
especially since factors such as interference fading and other factors already contribute to
44
low reliability in WMNs [51]. If the problem of starvation can be reduced, the reliability
of the network will increase significantly. In the mixed-bias approach there are two
different biases, one with a factor of 5 and one with a factor of 2 (which is similar to
proportional fairness). There is also a factor of weighting between the two competing
schemes of 0.5, which splits the amount of resources given to each scheme evenly. We
make use of the same parameter values in our approaches to allow for easy comparison.
In our system model we are applying a similar technique to a WMN which has
random source and one gateway as a destination. This is different from the original
approach because we introduce the concept of a gateway making the network model
closer to that of WMN and less like a wireless ad hoc network. We chose to use the same
biasing parameters for our system; however, this could all be altered resulting in vastly
different results. Experimenting with different parameters could be left up to the network
administrator since it is more dependent on the environment the network resides in and
the applications which are being run on it. We include analysis of the technique presented
by [54] as a verification of their results, and as a benchmark for our own schemes. Eq. 5
shows the mixed-bias equation which is used to assign resources to a given node based
on its distance from the gateway it is routing to [54].
(Eq. 5)
Where R is the resources allocated to the node
d is the distance from the node to the gateway, d > 0
α is the weight for each bias technique, 0 < α < 1
β1, β2 are the biasing constants that determines the strength of each bias, β1, β2 > 0
21
1ββ
αα
ddR
−+=
45
For this technique there are several ways in which the distance between the nodes
could be computed. If each node had GPS capabilities, the distance could be easily
computed whenever necessary by an additional query within a RTS packet. Alternatively,
TTL counts or hop counts from higher layers could be used to determine the distance in
hops between nodes and a table of known distances could be built up such that each node
would know roughly how many hops it is away from the GW.
4.2.2. Biasing against low demand (queue size)
Similarly to the previous technique, we chose to bias against those MRs which
have little data queued to be transmitted. By doing this, the MRs which do have full or
nearly full queues have the opportunity to clear space for incoming traffic rather than
dropping packets. At the same time, those MRs which have lots of room in their queues
have the ability to hold onto them until there is more data to send all at once. Of course
this technique will not work as well when the entire network is completely congested, it
depends on the assumption that at least some of the network has some free resources to
make use of. However, this assumption is not unreasonable since one usually also
assumes that a WMN can re-route packets through various links if one link has problems
due to poor quality or over-use. For this assumption to be correct, this would mean that
there must be some other link which can handle the extra capacity. This means that it is
unlikely many of the solutions we have now would operate well in a WMN which is
congested. Similarly to the distance biasing, we propose to split the resources available to
this scheme in half, and bias proportionally (factor of 2) and strongly (factor of 5) against
the queue size. The queue size that we are referring to for this technique is the MAC layer
46
queue. The formula which represents how resources are allocated with respect to queue
size is seen in Eq. 6.
(Eq. 6)
Where R is the resources allocated to the node
q is the length of the queue, q > 0
α is the weight for each bias technique, 0 < α < 1
β1, β2 are the biasing constants that determines the strength of each bias, β1, β2 > 0
In this technique, we use the inverse of the normal resource allocation function
because we want to give preference to those nodes which have the largest queues. This
means that we waste less time giving resources to those nodes which do not have much to
transfer in a given time. The mixed-bias technique, however, still gives some priority to
those nodes which do have a small amount of traffic, so it will prevent starvation of these
nodes.
4.2.3 Biasing against poor link quality
Again for this approach, we follow the mixed-bias framework proposed by in
[54]. The reason for biasing against poor link quality is again to try to reduce dropped
packets and end-to-end delay. Link quality would be determined using a signal to
interference plus noise ratio (SINR) similar to [7] or packet error rate (PER) from [33].
This value would help us determine how many resources to allocate to a node as shown
in Eq. 7. When we avoid links with poor quality by waiting for the link to hopefully
improve, we allow links which are nearby with higher link quality to transmit more
21
11ββ
αα
qqR
−+=
47
packets. This may give the link the opportunity to improve in quality. If the link doesn’t
improve in quality, at least it is getting fewer opportunities to send packets, so the poor
service the users will experience is related directly to the quality of their own link and
does not negatively affect the rest of the network. In other schemes this might not always
be the case. The poor quality link may try to communicate as if the link were behaving
normally. Again we use the same biasing parameters in this scheme with a proportional
bias (factor of 2) and a strong bias (factor of 5) mixed together.
(Eq. 7)
Where R is the resources allocated to the node
q is the quality of the link
α is the weight for each bias technique, 0 < α < 1
β1, β2 are the biasing constants that determines the strength of each bias, β1, β2 > 0
4.2.4 Combined cross-layer mixed-bias scheduling
Lastly, our combined mixed-bias technique is novel. It takes all of the mixed-bias
techniques outlined previously and combines them into an all-encompassing technique
which considers multiple characteristics of the WMN at once. This technique provides a
more complete snapshot of the network at a given moment for the MR to make a more
informed decision on its state. Rather than just looking at one characteristic as [54] did
the combination of characteristics provides a more general solution which is more
flexible and useful in a wide variety of environments and applications. This combined
21
1ββ
αα
qqR
−+=
48
mixed-bias approach is similar to the cross-layer solution from [8] except our approach is
for scheduling instead of routing. Also our approach uses mixed-biasing which is quite
different from the simpler metrics used in [8]. The key to the combined mixed-bias
technique is providing a fraction of the resources to each of the mixed-bias techniques.
For example, half the resources could be assigned to the distance technique while a
quarter could be assigned to queue length and the last quarter to link quality. The
network designer or administrator could specify these quantities manually. In a more
complex network, these parameters could dynamically change depending on which
applications are being run on the network at a given time or what the environment around
the network is like. The combined mixed-bias scheme is shown in Eq. 8.
(Eq. 8)
Where R is the resources allocated to the node
R1 is the resources calculated using Eq. 1
γ1 is the weight of scheme 1 in the combined biasing
R2 is the resources calculated using Eq. 2
γ2 is the weight of scheme 2 in the combined biasing
R3 is the resources calculated using Eq. 3
γ3 is the weight of scheme 3 in the combined biasing
γ1, γ2, γ3 > 0
γ1 + γ2 + γ3 = 1
332211 RRRR γγγ ++=
49
Chapter 5
Performance evaluations of the proposed approaches
In this chapter, we will evaluate the performance of the proposed approaches.
Each approach is compared to existing solutions or a baseline approach in order to gauge
the performance appropriately. For the case of fair scheduling with multiple gateways,
since our approach is most similar to that of [42], we compare our solution with multiple
gateways against their solution with a single gateway. Additionally, as a baseline we
compare the two approaches to the approach without fair scheduling.
In the case of the mixed-bias approach, we compare our solutions to [54] since
this approach is closest to ours. In their approach, they bias against only one
characteristic of the network: distance. We also provide an evaluation of this same
technique so that comparisons may be drawn appropriately. The main difference between
our evaluation and that in [54] is evaluation was done using the Network Simulation 3
(NS3) tool rather than Matlab. The NS3 simulation environment more closely models a
wireless network environment than Matlab which is a more general purpose simulation
tool. This allows our results to provide additional verification to those found in [54]. In
addition to this, we compare our proposed mixed-biasing techniques against queue length
as well as a combined mixed-bias approach which combines biasing against both queue
length and distance from the gateway. Our approach is a more generalized one that may
be applied to wide variety of unwanted characteristics in WMNs while making use of the
original mixed-bias framework.
50
5.1. Comparison of simulation environments
There are many options available for performing experimental analysis on
protocols for WMNs. In to select a suitable environment for evaluating the proposed
methods we will present a brief comparison of three of the choices.
5.1.1. Custom C/C++ simulation environment
The C/C++ Simulation environment was used for all of the experiments in the fair
scheduling approach. It was also used for the multiple gateway experiments in the mixed-
bias approach. The reason for this is because the C/C++ simulation environment was
designed to be a multiple gateway environment from the beginning. The NS3 tool which
was used for the remaining experiments was not able to support multiple gateways very
easily. Since the C/C++ simulation environment was designed to support multiple
gateways from the start, it was chosen for the multiple gateway experiment for the
mixed-bias scheme as well.
5.1.2. Network simulation 2
When deciding on simulation environments to use for implementing the
experiments, we reviewed Network Simulation 2 (NS2) [43] extensively since in the past
this has been the standard tool used for wireless network simulation. One main advantage
of using NS2 is the large number of protocols and modules which have been created over
the years. For instance it supports popular routing protocols such as ad-hoc on-demand
distance vector routing (AODV) [4] and dynamic distance-sequenced distance-vector
routing (DSDV) [5]. However, there are several problems with NS2 that make other
options more attractive. NS2 was designed originally to be a wired network simulation.
All of the wireless capabilities were added during later updates to the software and
51
because of this, much of the code for wireless is not consistent. In many cases, modules
have been written by many people and in many different styles and coding standards.
When trying to extend this work, it takes a great deal of time to understand how the
different objects interact. Furthermore, NS2 uses two different languages. The main
simulation program is written in the C programming language, however, the scenarios
which define the environment, nodes, applications and other entities is written in the
TCL scripting language. There is a complex interface between these two languages
which makes using NS2 even less intuitive to use and extend. This was especially true in
modifying the MAC layer, which is where much of our extensions are applied.
5.1.3. Network simulation 3
NS3 [44] is the successor to the most popular network simulation tool for wireless
networks, NS2. NS3 has several advantages over ns2. It uses a single language, C++,
rather than the dual language C and TCL environment which NS2 used. This makes
understanding much simpler than NS2 and requires less of a learning curve. Additionally,
NS3 has an automatic documentation tool which is very useful in learning the
environment. On the other hand, NS3 is still fairly new and does not have near the
amount of online support and resources that NS2 has built over the years. For instance the
only wireless mesh routing protocol available is optimized link state routing (OLSR)
[48]. However, the core development team is eager to help and answer any questions
through the NS3 mailing list online. NS3 has also been designed from the ground up to
be a wireless and wired simulation tool while NS2 was only designed originally for wired
networks. This alone is a key advantage in using NS3 since the code for wireless
52
simulation is more intuitive and requires fewer workarounds for simulating wireless
networks.
5.2 Performance metrics
Since both experiments made use of the same performance metrics, we will
discuss these metrics together before going into detail on each experiment. In our
experiments, we analyse the performance of the proposed approaches with respect to two
metrics. The first metric is packet delivery ratio (PDR). Packet delivery is an important
metric in determining the performance of a scheme because it gives an idea of how many
packets are making it from source to destination. It is a measure of how many packets
have made the complete trip, and does not include packets which have made it part of the
way and have been dropped or lost. We define packet delivery ratio as shown in Eq. 9.
(Eq. 9)
Where: PDR is the packet delivery ratio
Preceived is the number of packets received at the destination GW
Psent is the number of packets sent from the source MR
m is the number of GWs
n is the number of MRs
The second metric we use to evaluate the performance of a scheme is average
end-to-end delay. This metric is also important in gauging performance because the PDR
∑
∑=
n
sent
m
received
P
P
PDR
1
1
53
is not meaningful without it. It does not matter how many packets arrive successfully at
the end destination if they take a very long time to get there. The average end-to-end
delay is computed by adding the delay from source to destination along the path a given
packet takes from source to destination. Similarly, to PDR, this statistic does not include
the delays for packets which have been dropped along their path from source to
destination since we are only concerned with packets which arrive successfully. We
define average end-to-end delay as shown in Eq. 10.:
(Eq.10)
Where: Delay is the average end-to-end delay
d is the delay a successfully received packet has experienced
n is the number of successfully received packets
5.3. Performance evaluation of fair scheduling with multiple gateways
In order to evaluate our proposed fair scheduling with multiple gateways, we
compare our approach against two other techniques. As a baseline, we compare against a
network which makes no used of scheduling at all. In this case, any node is free to
transmit whenever it has a packet in its queue. This approach is similar to that of the
ALOHA protocol which was used in the earliest wireless networks. In order to provide a
more realistic comparison we also compared our approach to that of [42] with a single
gateway. Our approach also uses a similar technique to this except it allows for multiple
gateways. The main purpose of this experiment is to prove that using multiple gateways
n
d
Delay
n
i∑= 1
54
is extremely beneficial and should always be considered when designing WMN
architectures.
5.3.1. Simulation environment
Currently the performance evaluation was carried out using simulation
experiments. The simulation focuses on packets transmissions from MRs to GWs. MCs
are generated at the start of the simulation and are randomly distributed within the
simulation environment. Each MC is associated with the closest MR and each MR routes
its packets to the closest GW. This means that any packets that experience a collision at
the association stage are not counted in the reported results. We consider this problem
separate from the one we are trying to address in this thesis. In this thesis we are
concerned with fair scheduling among the MRs. The control packets for distributing the
scheduling are assumed to be sent on another channel and thus do not impact the
performance of the network. Additionally, the simulation environment acts as an
omniscient observer in that it performs the scheduling and distributes in to the gateways.
In a real-world implementation this would need to either be performed through a
centralized GW or via some kind of distributed GW solution. The interference model
assumes that two nodes interfere if they are within range and transmitting at the same
time or if there is a buffer collision. When interference occurs, retransmission is allowed
until a threshold timeout is reached.
The performance of the fair scheduling was evaluated using two simulation
parameters. Our solution is compared to three different models: (i) model with no fair
scheduling and a single gateway, (ii) no fair scheduling and multiple gateways and (iii)
fair scheduling with a single gateway. We also separately compare the fair scheduling
55
with a single gateway [42] to that with multiple gateways to emphasize the benefit of
using multiple gateways. Uplink traffic only is considered for these results since we
consider downlink scheduling a separate problem which can take advantage of caching
and multicasting to yield further improvements. The results shown in the figures are
averaged over 10 simulation runs.
5.3.2. Performance metrics
This simulation study uses two performance metrics. The first metric is average
packet delivery ratio. It is computed as the ratio of the total number of packets delivered
to the total number of packets sent. The second metric used in the simulation is the
average delay. It measures the time taken by a packet to reach its destination. These
metrics can help to gauge the performance of the protocol effectively. Detailed
information including how each metric is calculated is shown in Section 5.2.
5.3.3. Simulation parameters
In order to keep the scheduling algorithm simple, it is often assumed that there is
one or no gateways [42,45,46,54,55] in the WMN. However, one of the main uses for
WMN is to provide Internet access with expanded service areas from traditional WLANs
and hence the majority of the traffic flow is between the gateways and the MCs [61] via
MRs. Having only one gateway in this scenario is a major bottleneck so the existing
solutions should be extended to be able to support any number of gateways to make a
truly scalable WMN.
There are several parameters used in this simulation. The main parameter settings
are summarized in Table 4. The two main parameters varied during the simulation were
the number of mesh routers and the number of gateways. The retry threshold is used
56
when a collision occurs either from interference or buffer overflow. The packet is
allowed to be retransmitted unless the retry threshold has expired. The retry threshold can
be adjusted depending on the network conditions.
TABLE 4 FAIR SCHEDULING SIMULATION PARAMETERS
Parameter Value
Environment Dimensions 1000m x 1000m
Node Range 250m
Number of Mesh Routers 10 to 50
Number of Mesh Clients 250
Number of Gateways 1 to 5
Mean Packet Arrival 0.01 s
Mean Hop Delay 0.01 s
Retry Threshold 0.01 s
5.3.4. Effects of number of mesh routers
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
10 15 20 25 30 35 40 45 50
Mesh Routers
Avera
ge P
acket
Deli
very
Rati
o
Single GW, No FS
5 GWs, No FS
Single GW, FS
5 GWs, FS
Figure 5 Average Packet Delivery Ratio with Varying Mesh Routers
Figure 5 shows the average packet delivery ratio as a function of the number of
mesh routers in the network. Results are plotted for the case with a single gateway and
five gateways for both fair scheduling and no scheduling. As the network size increases,
the difference between the techniques becomes more pronounced. The cases with
multiple gateways have the highest packet delivery ratio. The results show that a single
57
gateway with no scheduling performs poorly delivering only 30% of the packets
successfully to the Internet. It is interesting to note that using multiple gateways without
fair scheduling can actually perform better than fair scheduling with a single gateway as
can be seen in Figure 5. This demonstrates how important it is to consider the case of
multiple gateways. The purpose of the comparison between multiple and single gateways
in this figure is to highlight the importance of relaxing the assumption of single gateways
and to compare our approach to that of [42]. As expected, multiple gateways yield higher
delivery ratio for all network sizes. This is likely because on average there are fewer hops
between any given MR and its GW. This is important because each hop increases the
likelihood of encountering a MR that is busy which could result in packet loss at the
worst or delay at best.
0
0.005
0.01
0.015
0.02
0.025
0.03
10 15 20 25 30 35 40 45 50
Mesh Routers
Avera
ge D
ela
y (
s)
Single GW, No FS
5 GWs, No FS
Single GW, FS
5 GWs, FS
Figure 6 Average Delay with Varying Mesh Routers
In Figure 6, average delay as a function of the number of mesh routers is shown.
The case of a single gateway without fair scheduling is again the worst case. The results
are similar to that of Figure 4 in that multiple gateways have the greatest performance in
terms of average delay as well. With small network sizes (under about 30 nodes) all four
techniques perform similarly. However, when the network size increases our approach
58
(fair scheduling with multiple gateways) is the best approach in terms of average delay.
Again the case with multiple gateways performs better in terms of delay. As in the
previous figure, this is likely due to the lower average hops any given node must take to
get a gateway. The delay is not only accumulated because of a greater amount of hops in
the single GW case but also because of the time spent waiting for a free buffer.
5.3.5. Effects of number of gateways
Figure 7 shows the average packet delivery ratio as a function of the number of
gateways in the network. These results were compiled with 50 mesh routers because
larger network sizes are affected by a lack of gateways the most. This is reinforced by the
results in Figure 6 which show a large difference between the performance with 1 and 5
gateways with both fair scheduling and no scheduling. Moreover, the results show
performance improvement with fair scheduling with respect to average packet delivery
ratio.
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1 2 3 4 5
Gateways
Ave
rag
e P
acke
t D
eli
very
Rati
o
Fair Scheduling
No Fair Scheduling
Figure 7 Average Packet Delivery Ratio with Varying Gateways
Similarly, Figure 8 shows the average delay as a function of the number of
gateways with a network size of 50. In this case, the results of fair scheduling can best be
59
seen with a single gateway. This is likely because as each additional gateway is added,
both the fair scheduling and the no scheduling cases benefit significantly by easing
congestion on the single bottleneck gateway.
0
0.005
0.01
0.015
0.02
0.025
0.03
1 2 3 4 5
Gateways
Av
era
ge
Dela
y (
s)
Fair Scheduling
No Fair Scheduling
Figure 8 Average Delay with Varying Gateways
5.3.6. Summary of fair scheduling results
In summary, the fair scheduling algorithm we have proposed performs well with
respect to packet delivery ratio when compared to a solution without fair scheduling.
When compared with the solution without fair scheduling to the fair scheduling solution
with respect to delay, however, our approach performed worse. This leads us to believe
that a trade-off between packet delivery ratio and delay exists when it comes to fair
scheduling algorithms. Depending on the amount of delay, it may be worthwhile to
introduce some extra delay into the transmission if it can avoid costly retransmissions
which significantly increase the overall time. There are several factors which may have
caused higher delay in the fair scheduling approach. The most significant factor being
the amount of time that must be sacrificed in order to ensure there are no collisions in the
medium. Lastly, compared to an approach with a single gateway, our approach with
60
multiple gateways performed better, validating the observation that a single gateway can
severely restrict the overall throughput of the network.
5.4. Performance evaluation of cross-layer mixed-bias scheduling
In this section and the following subsections we will discuss the performance of
cross-layer mixed-bias scheduling techniques. These techniques are based on the work
by [54], however, we generalize the idea proposed in the original work. In the original
work, only one network characteristic (distance) is biased against. In our work we also
propose to bias against any characteristic which may affect the throughput or delay in a
negative manner. Additionally, we propose that a combined approach can then use
mixed-biasing against all of these undesirable characteristics with a defined weighting for
each technique so that the only the most desirable flows are given preference to transmit
at a given time. A desirable flow can depend on the application of the network and can be
adjusted based on which characteristics the network administrator chooses to bias against.
In our performance evaluation we evaluate standard 802.11 ad-hoc MAC layer against
three mixed-bias techniques. The first technique is mixed-bias against distance (hops
from gateway) which is comparable to the approach given by [54]. The second technique
is mixed-bias against queue length and the third is a combined mixed-bias approach
which is mixed-bias against both distance and queue length. The routing algorithm used
in each case is the OLSR routing algorithm. This is one of the only routing algorithms
which has been implemented and supports wireless multi-hop networks in NS3 at the
time of writing.
61
5.4.1. Simulation environment
The simulation experiments on the cross-layer mixed-bias scheduling were
performed using two different simulation environments: the NS3 simulation environment
and an extended version of the C/C++ simulation environment used in the Section 5.3.1.
There are several reasons for making this choice. The C/C++ simulation environment
allows us to compare the mixed-bias approach with respect to the fair scheduling
approach. It also allows us to use the mixed-bias technique in an STDMA MAC layer
model. In this model, the scheduling is done via the hybrid approach. The gateway
decides on the final scheduling provided to the MRs based on distributed information
provided to it from the MRs it is serving.
In NS3, we are able to evaluate the mixed-bias approach in a more realistic
802.11 MAC layer model. This environment uses the fully distributed model of
scheduling. Each MR itself decides whether it should transmit or hold off and let another
MR transmit based on the values of the characteristics it is using to evaluate its own state.
In the NS3 environment since MAC, inference and other models are more realistic, we
can focus more on how the approaches would behave in a real network. At the same time
the C/C++ environment also provides valuable insight since it allows us to evaluate fair
scheduling approaches in terms of the mixed-bias approaches.
5.4.2. Performance metrics
In the mixed-bias experiments we evaluate each approach based on two
performance metrics. More detailed information on these two metrics is available in
Section 5.2. Here we focus on how the metrics are used specifically for the mixed-bias
experiments. The first metric is average packet delivery ratio. This metric allows us to
62
determine how many packets are successfully being delivered end-to-end in the network.
It is important to know this ratio because when packets are not being delivered
successfully, the retransmit operation is extremely costly. This is especially true if the
packet must be retransmitted across several hops as each hops accumulates delay
additively across as it traverses the network on its way from source to gateway.
Additionally, we evaluate the performance of the approaches based on the average
end-to-end delay for each packet. Like the previous metric, this metric is important as it
gives an idea of how long a packet is taking to traverse the network. In networks where
delay is important such as those being used for voice, video games and other multimedia
applications this metric is extremely useful. The Internet is increasingly being used for
all of these types of applications more and more so this metric is especially important.
5.4.3. Simulation parameters
In each scenario we compare the approaches with a varying number of mesh
routers. This is an important parameter because it helps to tell if a given approach is
scalable. This has been identified in [16] as one of the main goals of WMN protocols
since it is often a severe limitation. Similarly, we evaluate the approach against the
number of flows in the network. For this experiment we define a flow as traffic between
a given MR and the GW in the network. Since we have no concept of MCs, this flow
could consist of many clients connected to a given MR which are all being routed to the
GW at a given time. If a MR does not have a flow originating from it at a given time, it is
only being used as a forwarding node. In other words, the more flows in a given
scenario, the greater the utilization of the network. This is also important in determining
how scalable a protocol is. This parameter also gives us a
63
n idea of how the network performs as it becomes congested. We keep the
number of gateways constant for these simulations since the goal of this experiment is not
to verify the benefits of using multiple gateways. We assume that like the previous
experiment with multiple gateways, increased performance could be obtained using
multiple gateways. A summary of the main NS3 parameters is shown in Table 5. The
simulation parameters for the C/C++ simulation are similar to those in Table 4 from
Section 5.3.3.
TABLE 5. MIXED-BIAS NS3 SIMULATION PARAMETERS
Parameter Value
Environment Dimensions 1000m x 1000m
Node Range 150m
Number of Mesh Routers 10 to 30
Number of Traffic Flows 2 to 5
Mean Packet Arrival 0.01 s
5.4.4. Effects of number of mesh routers
Figure 9 shows the effect of the number of MRs on the packet delivery ratio in a
WMN with two traffic flows towards a single GW. In these results all of the mixed-bias
techniques perform fairly similarly. One important result is that all of the mixed-bias
approaches perform better than the 802.11 MAC distributed co-ordination function.
Another important result is that in most cases, mixed-bias based on queue size performed
the best. This result was somewhat surprising and may be because of the biasing
parameters that were selected for the experiments. Since the mixed-bias distance
technique often performed the worst of the mixed-bias approaches, it likely dragged
down the performance of the combined approach. This suggests that different weighting
64
between the two approaches may yield greater combined performance. In all of the
results presented the resources assigned to each scheme in the combined approach was
split evenly. Tuning the biasing parameters is expected to result in greater performance
for the combined mixed-biasing scheme; however, these experiments are beyond the
scope of this thesis. The main purpose of these results was to verify that the combined
mixed-biasing approach performs at least as well or better than other mixed-biasing
schemes and in some situations this was the case.
0.4
0.45
0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
10 15 20 25 30
Number of MRs
Packet
Deli
very
Rati
o
M-B Distance
Combined M-B
M-B Queue Size
802.11
Figure 9 Packet Delivery Ratio, Two Flows
In Figure 10, the effects of varying the number of MRs on the average end-to-end
delay is evaluated. In this experiement the network had two flows and a single GW. The
highest delay is the worst performing, and in this case the best performing approach was
the combined mixed-bias approach. This is an important result because, while the
combined approach may not always result in the highest packet delivery ratio (Figure 9)
65
it does give a low delay. This tradeoff may be worthwhile, however, the time it takes for
retransmissions is also important in the case where packets are dropped. The difference
in delay between some solutions is significant. For example with 25 MRs, there is double
the delay using the 802.11 approach compared with the combined mixed-bias approach.
0
0.005
0.01
0.015
0.02
0.025
0.03
0.035
10 15 20 25 30
Number of MRs
Avera
ge E
nd
-To
-En
d D
ela
y
M-B Distance
Combined M-B
M-B Queue Size
802.11
Figure 10 Average End-To-End Delay, Two Flows
Figure 11 shows the effect of varying MRs with respect to average end-to-end
delay. In this experiment there are five flows and a single gateway. One interesting
result here is the mixed-bias distance approach [54]. When the network size grows large
(30 MRs), the delay increases dramatically compared to the other solutions. This means
for larger networks, the approach in [54] alone may not be suitable. This is where the
combined mixed-bias approach becomes an attractive option. The extremely poor delay
result for the mixed-bias distance with 30 MRs also suggests that in this situation, when
66
using the mixed-bias combined approach it may be best to assign more resources to the
queue-size rather than an even split. Here we still have comparable packet delivery ratios
while still retaining low delay.
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.1
10 15 20 25 30
Number of MRs
Avera
ge E
nd
-To
-En
d D
ela
y
M-B Distance
Combined M-B
M-B Queue Size
802.11
Figure 11 Average End-To-End Delay, Five Flows
5.4.5. Effects of number of gateways
In Figure 12, the effect of varying gateways against packet delivery is shown.
These results were obtained using the C/C++ simulation environment so they are slightly
different from those in the previous figures which used NS3. Additionally, the results use
the Fair Scheduling approach from Section 5.3 as a baseline rather than the 802.11
approach since it was not implemented in the C/C++ environment. In these results we
can see that again the mixed-bias approaches perform the best, however, not by very
much. The fair scheduling approach performs similarly to the mixed-bias approach but
still slightly fewer packets are delivered. When there are fewer gateways the combined
67
mixed-bias approach performs slightly worse than the other two mixed-bias approaches
which suggests a different combination of weightings may be more effective. For
example with one gateway, the best performing mixed-bias approach was the queue size
so in this situation it may be best to give more resources to this approach rather than an
equal split.
0.4
0.5
0.6
0.7
0.8
0.9
1
1 1.5 2 2.5 3 3.5 4 4.5 5
Number of GWs
Pa
ck
et
De
liv
ery
Ra
tio
M-B Distance
Combined M-B
M-B Queue Size
Fair Scheduling
Figure 12 Average Packet Delivery Ratio, 250 Clients, 50 MR
In Figure 13 we compare the effect of varying gateways to the average end-to-end
delay. Again this result was obtained using the C/C++ simulation. One interesting result
is the combined mixed-bias approach, which performs poorly compared to the others
when there is only one GW. This result suggests that the combined mixed-bias approach
is affected by the bottleneck introduced when using a single GW, while the other
approaches are not affected as much.
68
0
0.002
0.004
0.006
0.008
0.01
0.012
1 1.5 2 2.5 3 3.5 4 4.5 5
Number of GWs
Av
era
ge
En
d-T
o-E
nd
De
lay
M-B Distance
Combined M-B
M-B Queue Size
Fair Scheduling
Figure 13 Average End-To-End Delay, 250 Clients, 50 MR
5.4.6. Effects of number of traffic flows
In Figure 14 we compare the effect of varying the number of traffic flows on the
average PDR. These results were obtained with a network size of 30 MRs and a single
GW. As the number of flows increase the performance of the network in all cases drops
significantly. In almost all cases, the mixed-bias approaches perform as well or better
than the 802.11 approach. The exception is with four flows where the 802.11 approach
performs slightly better than the mixed-bias approaches. The reason for this may be that
a maximum is being reached in the 802.11 approach. When the number of flows is
increased again to five flows the PDR drops significantly which indicates that a
maximum may have been reached with the 802.11 approach.
69
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
2 2.5 3 3.5 4 4.5 5
Number of Flows
Pa
ck
et
De
liv
ery
Ra
tio
M-B Distance
Combined M-B
M-B Queue Size
802.11
Figure 14 – Packet Delivery Ratio, Varying Flows
5.4.7. Summary of cross-layer mixed-bias results
In summary, the mixed-bias approaches we proposed performed well with respect
to the performance metrics. Often the mixed-bias approaches performed as well or better
than the IEEE 802.11 MAC in terms of packet delivery ratio. In terms of average end-to-
end delay the mixed-bias approaches consistently performed better. Another interesting
result is that the fair scheduling approach from Section 5.3 was outperformed by the
mixed-bias approach. These results provide validation that the mixed-bias approach is a
promising approach to explore in greater depth. Particular the combined mixed-bias
approach which performed as well as all of the other mixed-bias approaches. With
further tuning of the biasing parameters it is expected that even better performance can be
obtained.
70
Chapter 6
Conclusions and future work
This chapter will give concluding statements regarding the research presented as
well as give direction on what can be done in the future as possible extensions of this
work.
6.1. Conclusions
In this thesis, two scheduling schemes were proposed for wireless mesh networks.
The first scheme was a fair scheduling approach with multiple gateways. In this scheme
we proposed distributed requirement and routing tables along with a requirement
propagation algorithm which allowed scheduling information to be collected at multiple
gateways. The second scheme was a collection of cross-layer mixed-bias scheduling
techniques. In these approaches mixed-biasing against various characteristics was
evaluated along with a novel combined scheme which made use of multiple
characteristics at once. For both schemes performance evaluation was carried out using
simulation. Each scheme was compared against existing approaches.
The fair scheduling approach was compared against existing work with a single
gateway and the results showed that using multiple gateways does indeed perform better
than a single gateway in terms of packet delivery ratio and end-to-end delay. When
compared against a solution that did not use any fair scheduling, the fair scheduling
approach also performed better in terms of packet delivery ratio, however, in terms of
delay it performed worse. This was likely due to the delay introduced by waiting to
transmit only when no interference will be encountered. Moreover, due to the higher
71
packet delivery ratio, more packets were dropped in the non fair scheduling approach
causing lower delay overall.
The mixed-bias approaches we proposed were compared against IEEE 802.11
MAC in order to show that the mixed-bias approach is indeed a worthwhile scheduling
approach for wireless mesh networks. Furthermore, we compared the original mixed-
bias distance approach against 802.11 MAC and our own approaches. All the
experiments for mixed-bias evaluation were evaluated in terms of packet delivery ratio
and average end-to-end delay. The results provided verification for the original work
since the mixed-bias distance approach performed better than 802.11 MAC in almost
every case. Our proposed mixed-bias queue size approach performed better than the
mixed-bias distance approach in most of the comparison scenarios. Additionally, the new
mixed-bias combined approach performed at least as well as the other mixed-bias
approaches in almost every case which showed that it is indeed a promising technique
that may be further explored in order to increase performance further.
6.2. Future work
Since scheduling in multi-hop networks such as wireless mesh networks is a
challenging problem, there are a number of possibilities for future research work which
could be extended from the work presented in this thesis.
The first is studying and experimenting with different biasing parameters. It is
currently beyond the scope of this thesis since we were aiming to show that combined
mixed-biasing is a promising framework for scheduling in WMNs. However, the
performance of the scheduling could be improved or degraded depending on the
application of the network, the environment and the selection of the biasing parameters.
72
Further study in this area could help to develop a more comprehensive understanding of
how the values of these biasing parameters affect the network and how the parameters
can be adjusted to cope with changing environmental conditions and application
scenarios.
The second area for future work is to experiment with the approaches presented in
this thesis in a test-bed environment. There are many examples of this type of research in
literature [2,50]. It is often difficult to predict how a protocol or algorithm will perform
with real hardware, even with the most sophisticated network simulation tools. At the
same time, however, the simulation studies are effective way to obtain a starting point for
work in the test-bed environments as they are flexible and allow changing simulation
parameters easily during the experiments.
The third possible area for future research is designing scalable a communication
architecture and protocols for supporting full mobility including all entities such as mesh
clients, mesh routers and gateways. This would make the wireless mesh network more
flexible and allow for greater deployments and applications. At the same time, certain
features would still give it advantages over mobile ad hoc networks, such as MRs with
greater resources to be exploited.
73
REFERENCES
[1] A. Acharya, A. Misra, S. Bansal, “High-performance architectures for IP-based
multihop 802.11 networks,” in IEEE Wireless Communications, vol. 10, pp 22-28
[2] A. Zimmermann, M. Gunes, M. Wenig, U. Meis, J. Ritzerfeld, “How to study
wireless mesh networks: a hybrid testbed approach,” in Proc. of 21st IEEE Int. Conf.
on Advanced Information (etworking and Applications (AI(A 2007), pp. 853-860,
May 2007.
[3] B. Aoun, R. Boutaba, G Kenward, “Analysis of capacity improvements in multi-radio
mesh networks,” in Proc. of 63rd IEEE Vehicular Technology Conf. (VTC 2006), vol.
2, pp. 543-547, May 2006.
[4] C.E. Perkins, E.M. Royer, “Ad-hoc on-demand distance vector routing,” in Proc. of
2nd IEEE Workshop on Mobile Computing Systems and Applications (WMCSA
1999), pp. 90-100, February 1999.
[5] C.E. Perkins, P. Bhagwat, “Highly dynamic destination-sequenced distance-vector
routing (DSDV) for mobile computers,” in ACM SIGCOMM Computer
Communication Review, vol. 24, pp. 234-244. October 1994.
[6] C. Zhu, M.J. Lee, T. Saadawi, “On the route discovery latency of wireless mesh
networks,” in IEEE Consumer Communications and (etworking Conference (CC(C
2005), pp. 19-23, January 2005.
[7] D. Chafekar, V.S.A.Kumar, M.V. Marathe, S. Parthesarathy, A. Srinivasan, “Cross-
layer latency minimization in wireless networks with SINR constraints,” in Proc. of
8th Int. Symposium on Mobile ad hoc networking and computing, pp. 110-119, 2007.
[8] D. Guo, J. Li, M. Song, J. Song, “A novel cross-layer routing algorithm in wireless
mesh networks,” in Proc. of 2nd Int. Conf. on Pervasive Computing and Applications
(ICPCA 2007), pp. 262-266, July 2007.
[9] D. Koutsonikolas, S. M. Das, Y.C. Hu, “An interference-aware fair scheduling for
multicast in wireless mesh networks,” in Journal of Parallel and Distributed
Computing vol. 68, pp. 372-386, March 2008.
[10] D.S.J. De Couto, D. Aguayo, J. Bicket, R. Morris, “A high-throughput path metric for
multi-hop wireless routing,” in Proc. of 9th Int. Conf. on Mobile Computing and
(etworking, pp. 134-146, 2003.
[11] F. Foukalas, V. Gazis, N. Alonistioti, “Cross-layer design proposals for wireless
mobile networks: a survey and taxonomy,” in IEEE Communications Surveys and
Tutorials, vol. 10, pp. 70-85, 2008.
[12] G. Thamilarasu, R. Sridhar, “Exploring cross-layer techniques for security:
challenges and opportunities in wireless networks,” in IEEE Military
Communications Conference (MILCOM 2007), pp. 1-6, October 2007. [13] H.T. Cheng, H. Jiang, W. Zhuang, “Distributed medium access control for wireless
mesh networks,” Wireless Communications and Mobile Computing vol. 6, pp. 845-
864, 2006.
[14] H-Y. Wei, S. Ganguly, R. Izmailov, Z.J. Haas, “Interference-aware IEEE 802.16
WiMax mesh networks,” in Proc. of IEEE 61st Vehicular Technology Conf. (VTC
2005), vol. 5, pp. 3102-3106, June 2005.
74
[15] I.F. Akyildiz, X. Wang, “Cross-layer design in wireless mesh networks,” in IEEE
Transactions of Vehicular Technology, vol. 7, pp. 1061-1076, March, 2008.
[16] I.F. Akyildiz, X. Wang, W. Wang, “Wireless mesh networks: a survey,” Computer
(etworks vol. 47, pp. 445-487, 2005.
[17] J. Bicket, D. Aguayo, S. Biswas, R. Morris, “Architecture and evaluation of an
unplanned 802.11b mesh network,” in Proc. of 11th Int. Conf. on Mobile Computing
and (etworking, pp. 31-42, 2005.
[18] J. Eriksson, S. Agrawal, P. Bahl, J. Padhye, “Feasibility study of mesh networks for
all-wireless offices,” in Proc. of 4th Int. Conf. on Mobile Systems, Applications and
Services (MobiSys 2006), pp. 69-82, 2006.
[19] J.B. Ernst, M.K. Denko, “Fair scheduling for wireless mesh betworks with multiple
gateways”, in Proc. of 23rd IEEE Int. Conf. on Advanced Information (etworking
and Applications (AI(A 2009),May 2009.
[20] J.B. Ernst, M.K. Denko, “A review of cross-layer design for scheduling in wireless
mesh networks,” in Proc. of IEEE Int. Conf. on Science and Technology for
Humanity (STH 2009), September 2009.
[21] J.H. Li, W. Peng, R. Levy, A. Staikos, M. Chiang, “On systemic cross-layer design
for ad-hoc networks,” in Military Communications Conference (MILCOM 2008), pp.
1-7, November 2008 [22] J. Tang, G. Xue, C. Chandler, W. Zhang, “Link scheduling with power control for
throughput enhancment in multihop wireless networks,” in IEEE Transactions on
Vehicular Technology, Vol. 55, pp. 733-742, May 2006.
[23] J. Tang, G. Xue, W. Zhang, “Cross-layer design for end-to-end throughput and
fairness enhancement in multi-channel wireless mesh networks,” in IEEE Transaction
on Wireless Communications, Vol. 6, 10 pp 3482-3486, October 2007.
[24] J. Thomas, “Cross-layer scheduling and routing for unstructured and quasi-structured
wireless networks,” in Proc. of IEEE Military Communications Conf. (MILCOM
2005), vol. 3, pp. 1602-1608, 2005.
[25] J. Zheng, M.J. Lee, “A resource efficient and scalable wireless mesh routing
protocol,” in Ad Hoc (etworks, vol. 5, pp. 704-718, 2007.
[26] K. Jain, J. Padhye, V.N. Padmanabhan, L. Qiu, “Impact of Interference on multi-hop
wireless network performance,” in Proc. of the 9th Int. Conf. on Mobile Computing
and (etworking, pp 66-80, 2003.
[27] K. Karakayali, J.H. Kang, M. Kodialam, K. Balachandran, “Cross-layer Optimization
for OFDMA-based wireless mesh backhaul networks,” in proceedings of (WC(C
2007), 2007.
[28] K. Li, X. Wang, “Cross-layer design of wireless mesh networks with network
coding,” in IEEE Transactions on Mobile Computing, April, 2008. [29] K.N. Ramachandran, M.M. Buddhikot, G. Chandranmenon, S. Miller, E.M. Belding-
Royer, K.C. Almeroth, “On the design and implementation of infrastructure mesh
networks,” in Proc. of the 1st IEEE Workshop on Wireless Mesh (etworks (WiMesh
2005), September 2005.
75
[30] K. Wang, C.F. Chiasserini, R.R. Rao, J.G. Proakis, “A distributed joint scheduling
and power control algorithm for multicasting in wireless ad hoc networks,” in Proc.
of IEEE Int. Conf. on Communications, pp. 725-731, 2003.
[31] L. Erwu, J. Shan, S. Gang, G. Luoning, “Fair scheduling in wireless multi-hop self-
backhaul networks,” in Proc. of the Advanced Int. Conf. on Telecommunications and
Int. Conf. on Internet and Web Applications and Services (AICT/ICIW 2006), pp. 96,
February 2006.
[32] L. Huang, T.-H. Lai, “On the scalability of IEEE 802.11 ad hoc networks,” in Proc.
of 3rd ACM Int. Symposium on Mobile ad hoc networking and computing, pp. 173-
182, 2002.
[33] L. Iannone, K. Kabassanov, S. Fdida, “Evaluating a cross-layer approach for routing
in wireless mesh networks,” in Telecommunication Systems Journal, vol. 31, pp. 173-
193. March 2006.
[34] L, Iannone, K. Kabassanov, S. Fdida, “The real gain of cross-layer routing in wireless
mesh networks,” in Int. Symposium on Mobile and Ad Hoc (etworking and
Computing, pp. 15-22, May 2006. [35] L. Ma, M.K. Denko, “A routing metric for load-balancing in wireless mesh
networks,” in Proc. of 21st Int. Conf. on Advanced Information (etworking and
Applications Workshops (AI(AW 2007), vol. 2, pp. 409-414, May 2007.
[36] L.Popa, A. Rostamizadeh, R.M. Karp, C. Papadimitriou, I. Stoica, “Balancing traffic
load in wireless networks with curveball routing,” in The 8th Annual Int. Symposium
on Mobile Ad Hoc (etworking and Computing (MobiHoc 2007), pp. 170-179,
September 2007.
[37] M. Cao, V. Raghunathan, P.R. Kumar, “A tractable algorithm for fair and efficient
uplink scheduling of multi-hop wimax mesh networks,” in Proc. of 2nd IEEE
Workshop on Wireless Mesh (etworks (WiMesh 2006), September 2006.
[38] M. Conti, G. Maselli, G. Turi, S. Giordano, “Cross-layering in mobile ad hoc network
design,” Computer vol 37, pp. 48-51, 2004.
[39] M.J. Neely, R. Urgaonkar, “Cross-layer adaptive control for wireless mesh
networks,” in Ad Hoc (etworks vol. 5, pp. 719-743, 2007. [40] M.K. Denko, L. Ma, “Uplink scheduling in wireless mesh networks,” in Proc. of
GLOBECOM Workshops, pp. 1-6, November 2008.
[41] M.S. Kuran, G. Gur, T, Tugcu, F, Alagoz, “Cross-layer routing-scheduling in IEEE
802.16 mesh networks,” in Proc. of (Mobilware 2008), February 2008.
[42] N.B. Salem, J.-P. Hubaux, “A fair scheduling for wireless mesh networks,” in Proc.
of 1st IEEE Workshop on Wireless Mesh (etworks (WiMesh 2005), 2005.
[43] Network Simulator 2 [Online] http://www.isi.edu/nsnam/ns/
[44] Network Simulator 3 [Online] http://www.nsnam.org/
[45] N.H. Viadya, P. Bahl, S. Gupta, “Distributed fair scheduling in a wireless lan,” in
Proceedings of 6th Int. ACM Conf. on Mobile Computing and (etworking (MobiCom
2000), pp. 167-178, August 2000.
[46] N.S.P. Nandiraju, H. Gossain, D. Cavalcanti, K.R. Chowdhury, D.P. Agrawal,
“Achieving fairness in wireless LANs by enhanced IEEE 802.11 DCF,” in Proc. of
76
IEEE Int. Conf. on Wireless and Mobile Computing, (etworking and
Communications (WiMob 2006), pp. 132-139, 2006.
[47] P. Gupta, Y. Sankarasubramaniam, A. Stolyar, “Random-access scheduling with
service differentiation in wireless networks,” in Proc. of the 24th Joint Conf. of IEEE
Computer and Comm. Societies (I(FOCOM 2005), vol. 3, pp. 1815-1825, March
2005.
[48] P. Jacquet, P. Muhlethaler, T. Clausen, A. Laouiti, A. Qayyum, L. Viennot,
“Optimized link state routing protocol for ad hoc networks,” in Proc. of IEEE
Technology for the 21st Century, pp. 62-68. 2001.
[49] R. Draves, J. Padhye, B. Zill, “Routing in multi-radio, multi-hop wireless mesh
networks,” in Proc. of 10th Conf. on Mobile Computing and (etworking, pp. 114-
128, 2004.
[50] Roofnet [Online] http://pdos.csail.mit.edu/roofnet/doku.php
[51] S. Dominiak, N. Bayer, J. Habermann, V. Rakocevic, X. Bangnan, “Reliability
analysis of IEEE 802.16 mesh networks,” in Proc. of 2nd Int. Workshop on
Broadband Convergence (etworks, pp. 1-12, May 2007.
[52] S. Khan, A.A. Pirzada, M. Portmann, “Performance comparison of reactive routing
protocols for hybrid wireless mesh networks,” in Proc. of 2nd Int. Conf. on Wireless
Broadband and Ultra Wideband Communications, pp. 1-6, 2007.
[53] S.M. Das, H. Pucha, Y.C Hu, “Mitigating the gateway bottleneck via transparent
cooperative caching in wireless mesh network,” in Ad hoc networks, vol. 5, pp. 680-
703, August 2007.
[54] S. Singh, U. Madhow, E.M. Belding, “Beyond proportional fairness: a resource
biasing framework for shaping throughput profiles in multihop wireless networks,” in
Proc. of the 27th IEEE Conference on Computer Communications (I(FOCOM 2008),
pp. 2396-2404, April 2008.
[55] S. Nelson and L. Kleinrock, “Spatial TDMA: a collision-free multihop channel access
protocol,” IEEE Transactions on Communications, vol. 33, pp. 934-944, September
1985.
[56] T.B. Sorensen, M.R. Pons, "Performance evaluation of proportional fair scheduling
algorithms with measured channels," in Proc. of the 62nd IEEE Vehicular
Technology Conference (VTC 2005), pp. 2580-2585, September 2005.
[57] V. Kawadia, P.R. Kumar, “A cautionary perspective on cross-layer design,” in IEEE
Wireless Communications, pp. 3-11, 2005. [58] V. Srivastava, M. Motani, “Cross-layer design: a survey and the road ahead,” in IEEE
Communications Magazine, vol. 43, pp. 112-119, December 2005. [59] X. Lin, N.B. Shroff, R. Shrikant, “A tutorial on cross-layer optimization in wireless
networks,” IEEE Journal on Selected Areas in Communications vol. 24, pp. 1452-
1463, August 2006.
[60] X. Wang, K. Kar, “Cross-layer rate control for end-to-end proportional fairness in
wireless networks with random access,” in Proc. of (MobiHoc 2005), 2005. [61] Y. Bejerano, S.J. Han, A. Kumar, "Efficient load-balancing routing for wireless mesh
networks," in The Int. Journal of Computer and Telecommunications (etworking
vol. 51, pp. 2450-2466, July 2007.
77
[62] Y. Bejerano, S.J. Han, L. Li, “Fairness and load balancing in wireless LANs using
association control,” in Proc. of the 10th Int. Conf. on Mobile Computing and
(etworking (MobiCom 2004), pp. 315-329, 2004.
