Enhancing P2P Systems over Wireless Mesh Networks Marcel Cavalcanti de Castro

Enhancing P2P Systems over Wireless Mesh Networks Marcel Cavalcanti de Castro
Faculty of Economic Sciences, Communication and IT
Computer Science
Marcel Cavalcanti de Castro
Enhancing P2P
Systems over Wireless
Mesh Networks
DISSERTATION
Karlstad University Studies
2011:60
Marcel Cavalcanti de Castro
Enhancing P2P
Systems over Wireless
Mesh Networks
Karlstad University Studies
2011:60
Marcel Cavalcanti de Castro. Enhancing P2P Systems over Wireless Mesh Networks
DISSERTATION
Karlstad University Studies 2011:60
ISSN 1403-8099
ISBN 978-91-7063-398-0
© The author
Distribution:
Karlstad University
Faculty of Economic Sciences, Communication and IT
Computer Science
S-651 88 Karlstad
Sweden
+46 54 700 10 00
www.kau.se
Print: Universitetstryckeriet, Karlstad 2011
A minha esposa
Roberta
e aos meus filhos
Sofia e Matteus
Enhancing P2P Systems over Wireless Mesh Networks
MARCEL CAVALCANTI DE CASTRO
Department of Computer Science, Karlstad University, Sweden
Abstract
Due to its ability to deliver scalable and fault-tolerant solutions, applications based on
the peer-to-peer (P2P) paradigm are used by millions of users on the internet. Recently,
wireless mesh networks (WMNs) have attracted a lot of interest from both academia
and industry, because of their potential to provide flexible and alternative broadband
wireless internet connectivity. However, due to various reasons such as unstable wireless
link characteristics and multi-hop forwarding operation, the performance of current P2P
systems is rather low in WMNs.
This dissertation studies the technological challenges involved while deploying P2P
systems over WMNs. We study the benefits of location-awareness and resource replication to the P2P overlay while targeting efficient resource lookup in WMNs. We further
propose a cross-layer information exchange between the P2P overlay and the WMN in
order to reduce resource lookup delay by augmenting the overlay routing table with physical neighborhood and resource lookup history information.
Aiming to achieve throughput maximization and fairness in P2P systems, we model
the peer selection problem as a mathematical optimization problem by using a set of
mixed integer linear equations. A study of the model reveals the relationship between
peer selection, resource replication and channel assignment on the performance of P2P
systems over WMNs. We extend the model by formulating the P2P download problem as
chunk scheduling problem. As a novelty, we introduce constraints to model the capacity
limitations of the network due to the given routing and channel assignment strategy.
Based on the analysis of the model, we propose a new peer selection algorithm which
incorporates network load information and multi-path routing capability.
By conducting testbed experiments, we evaluate the achievable throughput in multichannel multi-radio WMNs. We show that the adjacent channel interference (ACI) problem in multi-radio systems can be mitigated, making better use of the available spectrum.
Important lessons learned are also outlined in order to design practical channel and channel bandwidth assignment algorithms in multi-channel multi-radio WMNs.
Keywords: peer-to-peer overlay, wireless mesh networks, peer selection, channel assignment, routing, optimization, adjacent channel interference, channel bandwidth adaptation.
Acknowledgments
The five years and a half spent working on my Ph.D. thesis at the Computer Science
department is an experience that I will never forget. This thesis would have never been
completed without the help and support of many persons who contributed to it directly
or indirectly.
First, I would like to thank Prof. Andreas Kassler for giving me the opportunity to
pursue my Ph.D. studies under his supervision, and for providing me with excellent support and dedication during all these years. I am privileged for having him as my supervisor.
I would like to thank Prof. Mario Gerla, Prof. Yevgeni Koucheryavy, and Prof. Evgeny
Osipov for their willingness to be on the examination committee, and Prof. Edmundo
Monteiro for accepting the role of opponent and its comments in a preliminary version
of this thesis.
I would like to thank the European Regional Development Fund through the Interreg IVB project E-CLIC (European Collaborative Innovation Centres for Broadband
Media Services), NEWCOM++ (Network of Excellence in Wireless Communications),
Knowledge Foundation of Sweden, and STINT (stiftelsen för interntionalisering av högre
utbildning och forskning) for the financial support.
I am very grateful for having the opportunity to work with so many wonderful people at the Computer Science department, whom deserve credit for providing such a nice
and friendly work environment. A special thanks to the members of the Distributed
System and Communications Research Group (DISCO) and the co-authors of the publications included in this thesis. It has been a privilege to work with all of you.
Lastly, a huge thank to all of those who supported me in any respect during the
completion of this work. This thesis is dedicated to my wife, Roberta M. Agostini, and
my parents Jazon & Adelcy Castro for supporting and encouraging me to pursue this
degree. Without my wife’s encouragement, I would not have finished the degree.
Karlstad, November 2011
Marcel C. Castro
iii
Contents
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
i
iii
1
Introduction
1.1 Problem Definition and Research Questions . . . . . . . . . . . . . . . . . . .
1.2 Thesis Outline and Contributions . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3 Other Related Papers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1
3
5
9
2
Background
2.1 Peer-to-Peer Systems . . . . . . . . . . . . . . . . . . . .
2.1.1 Centralized . . . . . . . . . . . . . . . . . . . .
2.1.2 Distributed . . . . . . . . . . . . . . . . . . . . .
2.1.3 Hybrid . . . . . . . . . . . . . . . . . . . . . . .
2.2 Wireless Mesh Networks . . . . . . . . . . . . . . . . .
2.2.1 Multi-channel Multi-radio WMNs . . . . . .
2.2.2 Routing . . . . . . . . . . . . . . . . . . . . . . .
2.2.3 Channel Assignment . . . . . . . . . . . . . .
2.2.4 Joint Routing and Channel Assignment . .
2.2.5 Modeling Capacity in WMNs . . . . . . . .
2.2.6 IEEE 802.11 PHY and MAC layers . . . . .
2.3 Peer-to-Peer Systems and Wireless Mesh Networks
2.3.1 Transparent Layer . . . . . . . . . . . . . . . .
2.3.2 Cross-layer Layer . . . . . . . . . . . . . . . . .
2.3.3 Integrated Layer . . . . . . . . . . . . . . . . .
2.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . .
3
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Adapting the P2P Overlay to WMNs
3.1 Our Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Proposed Architecture for P2P Systems over Wireless Mesh Networks . .
3.2.1 Architecture components . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.2 Example of P2P file sharing application: resource store and lookup
phases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
v
11
11
12
12
13
14
15
15
17
19
19
20
22
23
24
25
26
27
27
28
28
31
CONTENTS
3.3
3.4
3.5
3.6
3.7
3.8
3.9
Challenges of Deploying P2P Systems over Wireless Mesh Networks . . .
The Bamboo Overlay Maintenance Cost in WMNs . . . . . . . . . . . . . .
3.4.1 Bamboo DHT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.4.2 Bamboo Overlay Maintenance . . . . . . . . . . . . . . . . . . . . . . .
3.4.3 Experimental Setup and Simulation Results . . . . . . . . . . . . . .
The Benefits of Location Awareness DHT and Resource Replication . . .
3.5.1 Georoy DHT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.5.2 Location-aware versus Location-unaware DHT . . . . . . . . . . . .
3.5.3 Resource Replication . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.5.4 Experimental Setup and Simulation Results . . . . . . . . . . . . . .
Reducing Routing Stretch Through Neighborhood and Cache Information
3.6.1 Experimental Setup and Simulation Results . . . . . . . . . . . . . .
Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Validity of the Results and Limitations . . . . . . . . . . . . . . . . . . . . . . .
33
37
37
38
39
43
44
45
45
46
53
56
58
60
60
4
Peer Selection Problem Formulation in WMNs
4.1 Our Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2 Network Model and Link Capacity Formulation . . . . . . . . . . . . . . . .
4.2.1 Network Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.2 Modeling Link Capacity . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3 Routing and Channel Assignment in Multi-Channel Multi-Radio WMNs
4.4 Peer Selection Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . .
4.4.1 Max Rate Allocation . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.4.2 Minimum Guaranteed Maximum Rate Allocation . . . . . . . . . .
4.4.3 Maximum of Minimum Rate Allocation . . . . . . . . . . . . . . . .
4.4.4 Proportional Fairness . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.5 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.6 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.8 Validity of the Results and Limitations . . . . . . . . . . . . . . . . . . . . . . .
5
Extending the Peer Selection Problem for Multiple Chunks
83
5.1 Our Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
5.2 Chunk-based Peer Selection Problem Formulation . . . . . . . . . . . . . . . 85
5.2.1 Makespan Minimization . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.2.2 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.3 Designing interactions between Peer Selection, Routing and Channel Assignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
5.3.1 BitTorrent Background and Key Mechanisms . . . . . . . . . . . . . 94
5.3.2 Bestpeer Multi-path Peer Selection Proposal . . . . . . . . . . . . . . 96
5.3.3 Experimental Setup and Simulation Results . . . . . . . . . . . . . . 99
5.4 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
vi
63
64
65
65
65
67
70
71
71
72
72
73
80
81
81
CONTENTS
5.5
5.6
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
Validity of the Results and Limitations . . . . . . . . . . . . . . . . . . . . . . . 106
6
Practical Considerations for Channel and Channel Bandwidth Assignment
in WMNs
107
6.1 Our Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
6.2 Testbed Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
6.3 Adjacent Channel Interference . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
6.3.1 ACI Impact on 802.11 PHY and MAC Layer . . . . . . . . . . . . . . 110
6.3.2 Hardware Design Influence to the ACI . . . . . . . . . . . . . . . . . . 113
6.4 ACI Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
6.4.1 Impact of Board Crosstalk, Radiation Leakage and Antenna Engineering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
6.4.2 Impact of ACI under Different Sending Powers . . . . . . . . . . . . 117
6.4.3 Joint Effect of ACI, Channel Heterogeneity and Different PHY
Rates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
6.5 Channel Bandwidth Adaptation . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
6.5.1 Changing Channel Bandwidth using Commodity Wi-Fi Hardware 124
6.5.2 Example of Channel Overlapping for a 20/40 MHz Channel Bandwidth Combination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
6.6 Channel Bandwidth Experimental Results . . . . . . . . . . . . . . . . . . . . 127
6.6.1 Impact of ACI for 20 MHz and 40 MHz Channel Bandwidth . . . 127
6.6.2 Analysis of the Receiver Side B1 . . . . . . . . . . . . . . . . . . . . . . 130
6.7 Lessons Learned for Channel and Channel Bandwidth Assignment in
WMNs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
6.8 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
6.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
6.10 Validity of the Results and Limitations . . . . . . . . . . . . . . . . . . . . . . . 134
7
Conclusions
137
7.1 Reviewing the Achievements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
7.2 Future Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
References
141
Index of References
155
Acronyms
155
vii
List of Figures
2.1
2.2
Wireless Mesh Network Architecture [20] . . . . . . . . . . . . . . . . . . . .
Design choices of P2P and WMN integration . . . . . . . . . . . . . . . . . .
14
23
3.1
3.2
3.3
Architecture: components and interconnections . . . . . . . . . . . . . . . . .
Example of a store and find resource action using DHT functionalities. .
Message sequence diagram for a store and find resource action using put/get
DHT functionalities. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Impact of Bamboo management traffic . . . . . . . . . . . . . . . . . . . . . . .
Percentage of management overhead compared to user traffic . . . . . . . .
Percentage of lookups completed . . . . . . . . . . . . . . . . . . . . . . . . . . .
CDF of lookup delay in the 36 nodes topology . . . . . . . . . . . . . . . . .
Comparison between the number of logical hops in Georoy and Bamboo
in a grid topology. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Comparison between the number of physical hops in Georoy and Bamboo in a grid topology. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Comparison between the average lookup delay in Georoy and Bamboo
in a grid topology. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Comparison between the percentage of lookups completed in Georoy
and Bamboo in a grid topology. . . . . . . . . . . . . . . . . . . . . . . . . . . .
Comparison between the routing stretch factor in Georoy and Bamboo
in a grid topology. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Comparison between the number of logical hops in Georoy and Bamboo
for random topologies. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Comparison between the number of physical hops in Georoy and Bamboo for random topologies. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Comparison between the average lookup delay in Georoy and Bamboo
for random topologies. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Number of logical and physical hops in Bamboo in a grid topology with
225 nodes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Number of logical and physical hops in Georoy in a grid topology with
225 nodes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
29
32
3.4
3.5
3.6
3.7
3.8
3.9
3.10
3.11
3.12
3.13
3.14
3.15
3.16
3.17
ix
33
41
41
42
42
47
47
48
49
50
51
51
52
53
53
LIST OF FIGURES
3.18 Average lookup delay for Georoy and Bamboo in a grid topology with
225 nodes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.19 Example of the one-hop DHT neighborhood and lookup request. Node
000 is requesting a resource located at node 101. The overlay neighborhood, network topology and key lookup process are represented in the
left, middle and right part of the figure respectively. . . . . . . . . . . . . . .
3.20 Average cache size versus node density for standard and cross-layer cache
schemes using 500 and 1000 nodes . . . . . . . . . . . . . . . . . . . . . . . . . .
3.21 Routing stretch versus node density for 1000 nodes . . . . . . . . . . . . . . .
54
55
57
57
Collision domain for link l (1, 2) using single channel WMN . . . . . . . . .
K-Partition channel assignment . . . . . . . . . . . . . . . . . . . . . . . . . . . .
BFS channel assignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Impact of replication for files located at the gateways, two NICs and six
channels. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Impact of replication for files located at the random nodes, two NICs and
six channels. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Throughput versus number of channels . . . . . . . . . . . . . . . . . . . . . .
Impact of number of radios for BFS and random replication . . . . . . . . .
Tradeoff between maximum throughput and fairness for the degree of
replication of two at gateways, two NICs and six channels . . . . . . . . . .
Theoretical and simulation results of the normalized throughput for different degrees of replication at the gateways and single channel assignment
67
69
69
Chaska topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Optimum makespan for the P2P download problem: chunk-based versus
non-chunk-based, assuming single channel WMN deployment. . . . . . . .
5.3 Optimum makespan for different channel assignment strategies and increasing number of seeds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.4 Optimum makespan for K-Partition and BFS channel assignment strategies, 1 seed, 6 channels, 2 NICs . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.5 Optimum makespan versus number of channels for grid topology, 1 seed,
6 leechers, and 2 NICs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.6 Interaction between peer selection, routing and channel assignment . . . .
5.7 Total download time versus number of seeds for different peer selection
schemes using 10 leechers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.8 Total download time versus number of seeds for different peer and path
selection schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.9 CDF of total download time for Chaska topology, 15 seeds and 30 leechers and different peer and path selection schemes . . . . . . . . . . . . . . . .
5.10 Total download time versus number of seeds for channel reassignment . .
89
4.1
4.2
4.3
4.4
4.5
4.6
4.7
4.8
4.9
5.1
5.2
6.1
74
75
77
77
78
79
90
91
92
92
94
100
101
102
103
Testbed experimental setup (a) and development board used (b) . . . . . . 109
x
LIST OF FIGURES
IEEE 802.11a transmit spectrum mask [21] . . . . . . . . . . . . . . . . . . . .
Critical ACI scenarios in multi-radio mesh networks . . . . . . . . . . . . . .
ACI experimental setup: basic setup (upper part) and 2-boxes setup (lower
part) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.5 Impact of board crosstalk, radiation leakage, and antenna engineering
on 802.11a multi-radio performance in terms of normalized aggregated
throughput and PHY rate of 6 Mbit/s . . . . . . . . . . . . . . . . . . . . . . .
6.6 Impact of ACI under different link qualities in terms of normalized aggregated throughput and PHY rate of 6 Mbit/s . . . . . . . . . . . . . . . . .
6.7 Throughput A-C for different channel combinations (20 cm antenna distance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.8 PHY rate with highest throughput A − C for different channel combinations (20 cm antenna distance) . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.9 Throughput versus channel bandwidth in 802.11 networks for different
modulation schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.10 Example of channel overlapping for a 20/40 MHz channel combination
and different channel separation . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.11 Normalized throughput for various interferer and receiver combination
at 20 and 40 MHz . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.12 Normalized throughput of link A− B1 for various interferer and receiver
combination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2
6.3
6.4
xi
111
112
114
116
118
120
123
125
126
129
131
List of Tables
2.1
Overview of IEEE 802.11 physical layer [22] . . . . . . . . . . . . . . . . . . .
21
3.1
3.2
3.3
Bamboo management timers (secs) . . . . . . . . . . . . . . . . . . . . . . . . .
Average routing traffic received per node in the 36 nodes topology . . . . .
Number of dropped packets and correspondent reason in the 36 nodes
topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
40
42
6.1
6.2
42
Measurements parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
Percentage of packets sent at a given PHY rate for the SampleRate algorithm, with and without ACI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
xiii
Chapter 1
Introduction
WMNs have attracted the interest of the research community and industry over the last
decade. WMNs have matured to a point where the wireless mesh technology is been
used as an attractive means to provide connectivity in complement to access as offered
by wireless local area networks (WLANs) or cellular networks. Various qualities of this
paradigm include the combination of low cost, high speed, large coverage and robustness.
WMNs consist of a backbone of quasi-stationary mesh routers forming a multi-hop
wireless network, in which mesh routers wirelessly relay traffic on behalf of others. By
using gateway functionality, mesh routers allow the integration of the wireless mesh network with the internet or existing wireless networks. The WMNs can be visualized as
an integration of two planes where the access plane provides connectivity to the clients
while the forwarding plane relays traffic between the mesh routers [20, 23]. Currently,
mesh routers have been equipped with multiple radios, which allow them to send and
receive on multiple channels in parallel and consequently increase network capacity.
Due to their flexible structure, WMNs possess a wide range of application scenarios. The most common application has been the deployment of wireless community
networks, where users own the mesh routers and form the wireless mesh backbone, enabling access to other users for mutual benefit [24]. WMNs have also been used at control
systems in order to deploy public area surveillance [25] and temporary infrastructure in
disaster and emergency situations[26]. Other applications also considered to WMNs include industrial automation [27], traffic control [28], and sensor monitoring systems
[29].
Allied to the rapid proliferation of wireless technologies, the increasing development
of P2P communication systems in the internet has also gained a lot of attention. Using
the definition by [30], P2P communication refers to technology that enables two or more
peers to collaborate spontaneously in a network of equals (peers) by using appropriate
information and communication systems without the necessity for central coordination.
P2P systems and WMNs share many key characteristics such as self-organization and decentralization due to the common nature of their distributed components designed to
2
Introduction
operate in dynamic network environments and hop-by-hop connection establishment.
The common characteristics shared by both technologies also dictate that P2P systems
and WMNs are faced with the same fundamental challenge, that is to provide connectivity in decentralized environments.
P2P systems represent the next frontier to wireless communication, as wireless users
might expect to use the same services which are already available on the internet, like
popular file-sharing systems such as BitTorrent [31], voice over P2P applications such
as Skype [32], and emerging live or on-demand streaming applications such as PPLive
[33]. In addition to that, P2P systems also represent an important alternative to scale
WMN services, such as mobility and control management [34, 35], network routing
[36], and anonymity [37]. Therefore, deploying P2P systems over WMNs represents
exciting possibilities, but at the same time several challenges need to be investigated.
To leverage distributed resource lookup and peer discovery, P2P systems implement
an abstract overlay network on top of the physical network topology. Generally, the
P2P systems are designed for the wired internet, and rely on the IP routing infrastructure which is resource rich especially in terms of bandwidth availability. As a result,
such overlay networks use a membership management in order to keep the connectivity
among neighboring peers and consequently maintain the P2P overlay network. If deployed over wireless mesh networks, the maintenance of such overlay may require the
exchange of frequent control traffic with overlay neighbors that are physically located
several wireless hops away. This is because no direct mapping exists between the overlay
network and the physical network topology. This brings down the network performance
in WMNs as delay and packet loss increase with the number of wireless hops traversed.
Thus, solutions that adapt P2P overlays to current WMNs conditions are important to
scale the aforementioned services.
Using multiple radios increases the network capacity in WMNs, as multiple orthogonal channels are available and assigned such that the interference in the network is minimized. In addition, finding paths with better channel diversity can leads to increase in
the overall network capacity. As a result, the channel assignment and routing protocol
are quite dependent on each other, as the channel assignment determines the physical
network topology on which the routing protocol works.
Given that P2P systems are deployed over WMNs, and the resource lookup process
is completed by finding the list of peers holding the requested resource, the peer selection
phase takes place. In multi-channel multi-radio WMNs, the peer selection, which considers how to select neighbors from a set of peers, may also be influenced by the channel
assignment and routing as they dictate the network capacity and topology. Thus, an important issue to solve is how to model the achievable performance of the P2P download
process in a WMN given the capacity constraints imposed by the channel assignment
and routing. To address this issue, it is important to study the interactions between the
peer selection, channel assignment and routing layers. By using mathematical optimization models, it is possible to derive such interactions as important network optimization
tasks which combines channel assignment, routing, peer selection strategy, resource replication, and the impact of number of peers, channels and radios can be easily performed
1.1. Problem Definition and Research Questions
3
in order to derive achievable P2P download performance. Despite the use of limited assumptions, the solution of such models are an important asset as it gives more insights
in the theoretical understanding of P2P systems in WMNs, and helps to define novel
peer selection strategies in multi-channel multi-radio WMNs, taking into account their
limitations in terms of resource availability.
Apart from the mathematical models, researchers have also been allowed to undertake
testbed evaluations given the increasingly cheaper and more accessible WMNs technology. The use of testbed evaluations has allowed hands-on experience on implementing
theoretical ideas and also worked as a tool to improve mathematical and simulation models. For example, the increasing deployment of multi-radio WMN testbeds has shown
that in practice the amount of orthogonal channels is reduced as the use of close-by radios operating on adjacent channels causes considerable network interference, known as
the ACI. In addition, the need for spectrum flexibility without further increase in hardware costs has allowed the possibility for adapting the channel bandwidth according to
changing environment conditions such as interference and network traffic demands [22].
However, the consequences of using cheap off-the-shelf hardwares in WMN testbeds also
needs to be considered while analyzing the results derived from the measurements, as
imperfect radio shielding, low quality board design and antenna engineering contribute
to the overall interference in the network, and consequently impact on the achieved network capacity.
Given those considerations, we describe in Section 1.1 the problem definition and
research questions that are investigated in this dissertation. In Section 1.2 we present the
thesis outline and the contributions of this dissertation, together with the clarifications
about the work done in cooperation with other researchers. Other related papers to this
dissertation are presented in Section 1.3.
1.1
Problem Definition and Research Questions
In this dissertation we study:
How to enhance P2P system performance over wireless mesh network environments.
To address this problem, we sub-divide it into three important research questions:
I. How to organize P2P overlay membership in wireless mesh networks in order to provide efficient P2P resource lookup ?
Due to the characteristics of the wireless links and multi-hop forwarding paradigm,
the performance of traditional distributed services, such as P2P systems, is rather
low in wireless mesh networks. To address this question requires first a detailed
analysis of the challenges involved in the deployment of P2P systems in wireless
mesh networks described in Chapter 3. Normally, P2P systems implement an abstract overlay network on top of the physical network topology in order to leverage resource indexing and peer discovery. Since the maintenance of the P2P overlay involves a membership management overhead to the wireless networks, an ac-
4
Introduction
ceptable trade-off is required between the P2P resource lookup performance and
overlay membership management. Targeting efficient resource lookup in wireless
mesh networks, we analyze in Chapter 3 different cross-layer information exchange
strategies among the P2P overlay and the wireless mesh network.
II. How to model the achievable performance of a P2P system in wireless mesh networks
given the capacity constraints imposed by the channel assignment and routing layers ?
Aiming to achieve high performance and fairness in P2P systems over wireless
mesh networks, the peer selection problem is the one that identifies the set of best
peers to be chosen during the resource exchange phase. To address this question,
four different peer selection schemes are formulated in Chapter 4 using a mathematical model that describes a set of linear equations. Numerical results are derived in
order to estimate the achievable capacity of the P2P system given the constraints
imposed by the network capacity. As outlined in Chapter 5, an extension of the
peer selection problem which allows peers to upload resource’s segments among
themselves marks an important step towards the decrease of the amount of time required to disseminate resources in the network. Moreover, the underlying routing
and channel assignment layers in WMNs have a big impact on the P2P download
problem, as the selection of peers having suboptimal paths in the network will impact on resource consumption in all intermediate nodes in WMN scenarios. Therefore, the study of the interactions of those lower layers with the peer selection is
carried out through packet-level simulations. Given the insights gained through
the models and simulations, a novel peer selection algorithm is proposed which
accounts for the path load information and multi-path routing capability.
III. What performance improvements can be achieved using common off-the-shelf multiradio devices and what is the impact of interference on performance ?
The availability of cheap off-the-shelf multi-radio devices have enabled the wide deployment of wireless mesh testbeds. Those testbed deployments have shown that
certain assumptions used in mathematical and simulation models are not always
true in practice. One important example is related to the amount of orthogonal
channels available in multi-channel multi-radio WMNs. Through testbed experiments, we have shown that in practice the amount of orthogonal channels in
WMNs is reduced as the use of close-by radios operating on adjacent frequency
bands causes considerable network interference, known as the ACI. The impact of
ACI on achievable performance is studied in Chapter 6. Moreover, the use of cheap
hardwares implies an additional performance degradation in WMN deployments,
as imperfect radio shielding, bad board design and antenna engineering contribute
to the ACI problem. In addition, the need for spectrum flexibility and high network performance have motivated the use of channel bandwidth adaptation. Thus,
an evaluation of the channel bandwidth adaptation and ACI while targeting performance improvement in multi-channel multi-radio WMN scenarios is carried out
in our studies. Given the knowledge acquired through such testbed experiments,
1.2. Thesis Outline and Contributions
5
important lessons learned can be used to develop better channel and channel bandwidth assignment algorithms.
1.2
Thesis Outline and Contributions
The thesis is organized as follows. In Chapter 2, we introduce the background work
necessary to understand the research presented in this thesis. Thereafter, four chapters
are presented. Each chapter presents a short introduction to the problem discussed along
with the contributions and the detailed description of the performed studies. Related
works and conclusions of the achieved results in the chapter are outlined, followed by
a discussion of the validity of the results and their possible limitations. The concluding
remarks and future directions are then outlined in Chapter 7.
We now outline the individual contributions of each chapter along with the work
done in collaboration.
Chapter 3: Adapting the P2P Overlay to WMNs
Chapter 3 starts by introducing our proposed architecture for P2P systems over wireless
mesh networks, describing the required architectural components and their interactions.
To better describe the challenges involved while deploying P2P systems over wireless
mesh networks, we conduct simulation studies showing the trade-off between P2P resource lookup efficiency and the overlay management overhead. Given those insights, we
analyze the benefits of location-aware distributed hash tables (DHTs) and resource replication in order to increase the overlay lookup efficiency in WMN scenarios. Throughout the simulation results, we have shown that by constructing the overlay DHT using
physical location information at larger topologies contributes to lower lookup delays as
longer physical paths in the network are avoided while routing resource lookup through
the overlay. Moreover, the use of multiple replicas increases the probability of finding the
requested resource within a shorter distance, benefiting both location-aware and locationunaware DHTs. Through interactions with the routing layer, we replace long-range overlay neighbors by physical neighbors, and exploit the resource lookup history through
information caching. By using limited caching allied to neighbor-of-neighbor information provided by the routing layer at mesh routers, we show that the routes selected by
the P2P overlay are almost as efficient as the shortest path routes. These investigations
have previously been published in the following papers.
[1]. Marcel C. Castro, Laura Galluccio, Andreas Kassler and Corrado Rametta. Opportunistic P2P Communications in Delay Tolerant Rural Scenarios. In the EURASIP
Journal on Wireless Communications and Networking, 2011:1–14, 2011.
[2]. Marcel C. Castro, Andreas Kassler, Carla-F. Chiasserini, Claudio Casetti and Ibrahim Korpeoglu. Peer-to-Peer Overlay in Mobile Ad-hoc Networks. In X. Shen, H.
Yu, J. Buford, M. Akon, editors, Handbook of Peer-to-Peer Networking, Springer,
pages 1045–1080, 2010.
6
Introduction
[3]. Marcel C. Castro, Andreas Kassler, Gabriel Kliot, Roy Friedman, Raphäel Kummer, Peter Kropf, Pascal Felber. Minimizing DHT Routing Stretch in MANETs.
In the 9 t h Scandinavian Workshop on Wireless Adhoc Networks (Adhoc’09), Uppsala,
Sweden, 4–5 May, 2009.
Concerning Paper [1], the author of this dissertation shares the development of the original idea, evaluation and analysis of the results with Corrado Rametta. The problem
formulation was refined in collaboration with Laura Galluccio and Andreas Kassler, who
provided valuable insights in the evaluation and writing phases. Concerning Paper [2] ,
the author of this dissertation developed all the original ideas. In Paper [3], the author of
this dissertation shares the development of the original idea with Raphäel Kummer and
Gabriel Kliot. The DHT implementation was carried out by Raphäel Kummer, while
the location-awareness and caching modifications were done in conjunction with the author of this dissertation. The evaluation of the implementation, the experiments, and
paper writing were carried out by the author of this dissertation.
Chapter 4: Peer Selection Problem Formulation in WMNs
Chapter 4 starts the study of the interdependency between the P2P overlay, channel assignment and routing layers in wireless mesh networks. The collision domain model
[38, 39] is used to estimate achievable wireless link capacity in multi-hop communication scenarios. We first present a novel mathematical model which allows to calculate
the achievable performance of a P2P system in wireless mesh networks given the capacity constraints imposed by the channel assignment and routing layers. While previous
models have applied the collision domain model to estimate the achievable throughput
in wireless mesh networks, we are the first who extends this model to include peer selection, channel assignment, routing and resource replication on the problem formulation
for multi-channel multi-radio wireless mesh networks. We devise four peer selection
strategies targeting network throughput maximization and fairness among peers during
the P2P download process. Numerical results on the achievable throughput and fairness
are presented, showing the relationship between peer selection, replica placement and
channel assignment. An analysis of the results shows that with a lower degree of resource
replication and enough channels, it is better to replicate randomly the resource in the
network compared to replication at the gateways. This is due to the fact that the download traffic can be spread out more effectively in the network as more links operating
on orthogonal channels can be used simultaneously. The optimum number of channels
can be identified for a given number of radios, beyond which the throughput can not be
increased by adding more channels. According to the results, the optimum number of
channels depends on the channel assignment, peer selection and number of radios available in the network. Part of these results were included in:
[4]. Marcel C. Castro, Durga M. Prasad, Andreas Kassler, Stefano Avallone. Peer-toPeer Selection and Channel Assignment for Wireless Mesh Networks. In the Inter-
1.2. Thesis Outline and Contributions
7
national Workshop on Network Modeling and Analysis (IWNMA), pages 1–8, Bangalore, India, 2nd January 2011.
The work published in Paper [4] was coauthored with Durga M. Prasad during his
internship at Karlstad University. The author of this dissertation proposed the original
problem formulation and acted as leading author of the paper. The development of the
model in Octave was mostly done by Durga M. Prasad. The numerical results, the validation of the model with the simulator, and the paper writing were carried out by the
author of this dissertation.
Chapter 5: Extending the Peer Selection Problem for Multiple Chunks
Chapter 5 presents an extension of the mathematical model proposed in Chapter 4 by
incorporating the availability of multiple chunks in the peer selection problem formulation. We formulate the P2P download problem as a chunk scheduling problem given the
constraints imposed by the channel assignment and routing protocol on the achievable
capacity. While such chunk-based peer selection scheduling exists, we are the first who
considers channel assignment and routing protocol in the network constraints formulation targeted to multi-channel multi-radio wireless mesh networks. The new chunk-based
peer selection problem, formulated as a mixed integer linear program, allow us to devise
the optimum makespan, that is the minimal time required to disseminate a file among all
peers involved in the P2P download process. Using a numerical analysis, we compare the
chunk based scheduling with the non-chunk based scheduling described in Chapter 4. In
addition to that, a packet-level simulation is carried out in order to further analyze the
interdependency between P2P overlay, channel assignment and routing layers for larger
set of peers, chunks and topologies. A new peer selection algorithm which incorporates
path load information, multi-path routing capability and channel reassignment is proposed while targeting improvements on the P2P download process. An analysis of the
results shows that the combination of the breadth first search (BFS) channel assignment
together with the possibility of mutual resource exchange among peers contributes to a
faster resource dissemination in the context of wireless mesh networks. Moreover, we
show that by using the new peer selection metric which accounts for network path load
together with the availability of multiple paths among peers, we can reduce by up to forty
percent the total download time to disseminate a resource, compared to BitTorrent [31].
Part of these results were included in:
[5]. Marcel C. Castro, Andreas Kassler. On the Interaction Between Peer Selection,
Routing and Channel Assignment. In the 10 t h Scandinavian Workshop on Wireless
Adhoc Networks (Adhoc’11), Stockholm, Sweden, 10–11 May, 2011.
The work presented in Section 5.2 of Chapter 5 was a joint work with Nadia Ulrich.
The author or this dissertation developed the original idea by proposing the extended
model presented. The development of the model in Octave was jointly performed by
both authors. The porting of the model and evaluation carried out in SCIP and CPLEX
8
Introduction
were all carried out by the author of this dissertation. The work presented in Section 5.3
represents the work published in Paper [5]. The author of this dissertation developed
all the original ideas and implementations contained in the aforementioned publication.
Andreas Kassler provided valuable insights during the evaluation and writing phases.
Chapter 6: Practical Considerations for Channel and Channel Bandwidth Assignment in WMNs
In Chapter 6 we focus our attention to the practical considerations for channel and channel bandwidth assignment in multi-channel multi-radio wireless mesh networks. By conducting testbed experiments using off-the-shelf multi-radio devices, we show that in practice the amount of orthogonal channels used by the channel assignment algorithms in
WMNs is reduced as the use of close-by radios operating on adjacent channels causes considerable network interference, known as the ACI. Allied to that, we study the hardware
design problems caused by imperfect radio shielding, low quality board design and antenna engineering, which added to the performance degradation caused by ACI. In order
to increase network capacity and maximize the number of available channels, channel
bandwidth adaptation is used in our wireless mesh testbed. Furthermore, we list important points to be considered in the design of practical channel and channel bandwidth
assignment algorithms. An analysis of the results shows that by using a good hardware
engineering for the node enclosure, board, radio cards, sufficient antenna separation, and
orthogonal polarization it is indeed possible to achieve orthogonal channels using the
IEEE 802.11a standard in multi-radio mesh nodes. Moreover, the results indicate that
joint effects of ACI together with channel heterogeneity need to be considered while targeting network throughput prediction at higher PHY rates. The results also show that
it is better to use wider channel bandwidths if the radio interfaces are operating on adjacent channels as it forces them to share the medium via clear channel assessment (CCA)
mechanism avoiding than the ACI while transmitting on adjacent channels. These investigations are included in:
[6]. Peter Dely, Marcel C. Castro, Sina Soukhakian, Arild Moldsvor, Andreas Kassler.
Practical Considerations for Channel Assignment in Wireless Mesh Networks. In
the IEEE Broadband Wireless Access Workshop, held in conjunction with Globecom
2010, Miami, USA, 6–10 December, 2010.
[7]. Marcel C. Castro, Andreas Kassler, Stefano Avallone. Measuring the Impact of ACI
in Cognitive Multi-Radio Mesh Networks. In the IEEE 72nd Vehicular Technology
Conference (VTC Fall 2010), Ottawa, Canada, 6–9 September, 2010.
The work presented in Sections 6.4.2 and 6.4.3 of Chapter 6 represents the work
published in Papers [6], coauthored with Peter Dely and Sina Soukhakian. The author
of this dissertation proposed the original idea contained in Paper [6]. The experiments
were jointly performed by the authors, while the paper editing was done by Peter Dely
in conjunction with the author of this dissertation. The work presented in Section 6.5
1.3. Other Related Papers
9
represents the work published in Paper [7]. The author of this dissertation developed all
the ideas contained in Paper [7]. Andreas Kassler, Arild Moldsvor and Stefano Avallone
provided valuable insights in the experimental evaluation and writing phases.
1.3
Other Related Papers
The following publications, although not specifically included in the dissertation, contain
material that is related to the contributions of this dissertation:
• Marcel C. Castro, Laura Galluccio, Andreas Kassler, Sergio Palazzo, Corrado Rametta. On the comparison between performance of DHT-based protocols for opportunistic networks. In Proceedings of Future Network and MobileSummit 2010, pages
1–8, Florence, Italy, 16–18 June, 2010.
• Marcel C. Castro, Peter Dely, Andreas Kassler, Francesco Paolo D’Elia, Stefano
Avallone. OLSR and Net-X as a Framework for Channel Assignment Experiments (POSTER). In the 4 t h ACM International Workshop on Wireless Network
Testbeds, Experimental Evaluation and Characterization (WiNTECH), held in conjunction with MobiCom, Beijing, China, 20–25 September, 2009.
• Marcel C. Castro, Peter Dely, Andreas Kassler, Nitin Vaidya. QoS-Aware Channel
Scheduling for Multi-Radio/Multi-Channel Wireless Mesh Networks. In the 4 t h
ACM International Workshop on Wireless Network Testbeds, Experimental Evaluation and Characterization (WiNTECH), held in conjunction with MobiCom, Beijing,
China, 20–25 September, 2009.
• Marcel C. Castro and Andreas J. Kassler. Packet Aggregation for VoIP in Wireless
Meshed Networks. In Y. Koucheryavy, G. Giambene, and D. Staehle, editors, The
Traffic and QoS Management in Wireless Multimedia Networks, chapter Multihop
Wireless Networks, 2008.
• Marcel C. Castro, Eva Villanueva, Iraide Ruiz, Susana Sargento, and Andreas Kassler.
Performance Evaluation of Structured P2P over Wireless Multi-hop Networks. In
the International Conference on Advances in Mesh Networks (MESH 2008), Cap Esterel, France, 25–31 August, 2008.
• Nico Bayer, Marcel C. Castro, Peter Dely, Andreas Kassler, Yevgeni Koucheryavy,
Piotr Mitoraj, and Dirk Staehle. VoIP Service Performance Optimization in PreIEEE 802.11s Wireless Mesh Networks. In the IEEE International Conference on
Circuits & Systems for Communications (ICCSC2008), Shanghai, China, 26–28 May
2008.
• Marcel C. Castro, Peter Dely, Jonas Karlsson, Andreas Kassler. Capacity Increase
for Voice over IP Traffic through Packet Aggregation in Wireless Multihop Mesh
Networks. In the International Workshop on Wireless Ad Hoc, Mesh and Sensor
10
Introduction
Networks (WAMSNET07) , Jeju-Island, Korea, 10–12 December 2007. Best Paper
Award.
• Andreas J. Kassler, Marcel C. Castro, Peter Dely. VoIP Packet Aggregation based
on Link Quality Metric for Multihop Wireless Mesh Networks. In Proceedings of
Future Telecommunications Conference (FTC2007), Beijing, China, 11–12 October
2007.
• Marcel C. Castro, Andreas Kassler. SIP based Service Provisioning for hybrid
MANETs. In Proceedings of International Workshop on Telecommunications (IWT
2007), Santa Rita do Sapucaí, Brazil, 12–15 February, 2007.
• Marcel C. Castro, Andreas Kassler. Challenges of SIP in Internet Connected MANETs. In Proceedings of International Symposium of Wireless Pervasive Computing
(ISWPC 2007) , San Juan, Puerto Rico, 5–7 February, 2007.
• Marcel C. Castro, Andreas Kassler. SIP in hybrid MANETs - A gateway based
approach. In the 4 t h Swedish National Computer Networking Workshop (SNCNW
2006), Luleå, Sweden, 26–27 October, 2006.
• Marcel C. Castro, Andreas Kassler. Optimizing SIP service provisioning in internet connected MANETs - Invited Paper. In the 14 t h International Conference on
Software, Telecommunications and Computer Networks (SoftCOM 2006)- Symposium
on QoS in Wireless Multimedia Networks, Split, Croatia, 29 September – 1 October,
2006.
Chapter 2
Background
The objective of this chapter is to present the background information required to assist
the reader on the understanding of the research presented in this thesis. We first introduce
the P2P system concept and its classification through a set of well-known examples.
Having in mind that P2P communication paradigm is very important in wireless
multi-hop networks, as centralized servers might not be available or located in the internet, we introduce the wireless mesh network architecture. We also describe the benefit of
using multi-channel multi-radio WMNs and important challenges such as routing, channel assignment, wireless capacity modeling, and the IEEE 802.11 physical (PHY) and
medium access control (MAC) layers. An overview on different principles that guide the
various integration and interaction possibilities for peer-to-peer systems in wireless mesh
networks is presented through a set of proposed solutions in the literature.
2.1
Peer-to-Peer Systems
Recently applications based on the P2P communication paradigm are increasing in popularity. There are numerous P2P systems proposed with very different architectures and
protocols. Examples are popular file-sharing applications (e.g., Gnutella [40], and BitTorrent [31]), voice over IP solutions (e.g. Skype [32]) , and P2P video streaming (e.g.
PPLive, SopCast, CoolStreaming [33]).
A P2P system is a collection of autonomous end-system devices called peers that form
a set of interconnections called an overlay to share the peer’s content. Commonly, P2P
architectures can be organized into three main classes: centralized, distributed, and hybrid.
12
2.1.1
Background
Centralized
The centralized architecture was the first class of P2P systems, popularized by Napster
[41] which paved the way for file-sharing distribution. Napster made use of centralized
index server responsible for maintaining the list of connected peers and the files they provide. The central index could be queried by a peer to lookup for the address information
(e.g. IP and port) of all peers sharing the requested file. After obtaining the list of peers
that contained the desired file, the peer established direct connection with the peers, and
downloaded the file without the involvement of the central index server.
The main disadvantage of the centralized architecture is related to its scalability, as the
central index directory represents a single point of failure. Despite its huge number of
users, Napster went out of service not because of technical failure, but because the single
administration was vulnerable to the legal challenges of record companies [42].
2.1.2
Distributed
The main idea of distributed architectures is to establish and maintain the peer index
without using central entities. Common distributed architectures are classified as unstructured and structured.
The unstructured P2P overlays proposes that each peer in the system would be responsible for indexing its own resource. Thus, search operations along the peers are
necessary in order to find the desired resource in the overlay which may take long time
and consume network resource extensively as there is no relation between the overlay
topology and resource location. In Gnutella [40], the search operation is broadcasted in
the network and no guarantee can be given that the resource can be found. Attempts to
decrease the search overhead has been proposed by using expanded time to live (TTL)
rings [43] or random walk searches directed to peers with higher degree of connectivity
[44]. In unstructured approaches, the load on each peer grows linearly with the number
of search operation and the system size.
Since flooding request in the network is costly in terms of scalability, structure P2P
overlays made use of an organized overlay in order to achieve higher flexibility. As a
consequence, DHT based solutions have been proposed. DHTs offer an indexing service
by mapping each resource and each peer storing the resource on a certain key assigned
to a common identifier space via secure hash algorithm (SHA)-1 hash functions [45].
The one-way hash function leads to every peer being responsible for a range of keys and
having a virtual link with a subset of peers. When someone requests a key to a peer, it
compares its own identifier (ID) with the key and, if it falls in its range, it replies to the
requester. Otherwise the request is forwarded to the neighbor whose ID is the closest
with respect to the searched key.
Chord [46], Pastry [47], Tapestry [48], and Viceroy [49] are all solutions that exploit
a DHT approach. They differ in the way they build and maintain the structure of the
logical overlay. For example, Chord uses a logical ring where every node has an assigned
ID and is responsible for all the keys between its ID and its predecessor ID. Moreover, in
2.1. Peer-to-Peer Systems
13
order to speed up the search process, a finger table is used to connect the node to other
nodes in the network.
After Chord, other robust algorithms were proposed like Pastry [47] and Tapestry
[48]. These protocols follow basically the same methodology for the next-hop choice, i.e.
the node with the longest common prefix with respect to the searched key is selected, but
exploit different routing mechanisms in the overlay. Common features of these schemes
are that the size of the routing tables typically increases logarithmically with the size of
the network.
Several enhancements have been proposed to DHTs in order to improve performance
over resource scarce networks. For example, in Bamboo [50], a proactive overlay management approach intended to react faster to unexpected changes in the underlay network.
While targeting a strict relationship between the overlay and underlay, Georoy [51] have
proposed a location-aware variant of the Viceroy algorithm, where a geographical hash
function is proposed.
2.1.3
Hybrid
Compared to the previous architectures, the hybrid architecture proposes auxiliary mechanisms in order to facilitate the resource location in P2P systems. For example, in hybrid
solutions such as FastTrack [52], a two-tier architecture is proposed in which high capacity peers act as superpeers and forms an upper level which provides resource location
services to the other peers at the lower level. Despite superpeers sharing some similarities with the centralized architectures, two important factors distinguish them from the
central index server. First, a superpeer is not as powerful as the central server as it is in
charge of a subset of peers in the network. Second, a superpeer does not only coordinate the resource location, but it also acts as a normal peer contributing with its own
resources.
BitTorrent [31] introduced the idea of distributing larger files into pieces. By using a
mutual distribution of the pieces between the set of peers, called a swarm, BitTorrent proportionated further scalability during the download phase. However, as in the centralized architecture, the resource index still relies on a centralized entity known as tracker.
Although file download in BitTorrent is decentralized, the central tracker is recognized
as a scalability bottleneck been also susceptible for single control availability [53]. Thus
in case the tracker is inoperable, new users can not be bootstrapped into the network.
To overcome this issue, BitTorrent allows the possibility to have multiple trackers per
file. In case the primary tracker is down or have a long response time, next trackers are
selected.
The use of multiple tracker can increase the reliability, but still depends on the scalability of the centralized trackers. Alternatively, several BitTorrent clients implement a
distributed tracker, also known as trackerless, which serve the same purpose as the central
tracker but implemented through DHT.
14
Background
Internet
Wireless Mesh
backbone
Mesh router
Mesh router
with gateway
Mesh router
with gateway
Wired clients
Sink node
X
Sensor
Access point
X
Mesh router
with gateway/bridge
X
Mesh router
with gateway/bridge
X
Mesh router
with gateway/bridge Wireless
clients
Mesh router
with gateway/bridge
Sensor
networks
Wi-Fi
networks
Base station
Cellular
networks
Base station
WiMAX
networks
Figure 2.1: Wireless Mesh Network Architecture [20]
2.2
Wireless Mesh Networks
WMNs provide a low cost and easy to deploy solution to build broadband wireless access
networks, thereby ensuring connectivity to remote locations. WMNs consist of a backbone of quasi-stationary mesh routers forming a multi-hop wireless networks, in which
mesh routers wirelessly relay traffic on behalf of others.
Figure 2.1 shows a WMN architecture, composed of mesh routers and mesh clients.
Commonly, mesh routers form an infrastructure/backbone in order to provide connectivity to mesh clients. With gateway functionality, mesh routers can also provide internet
access among themselves and to their clients. The mesh clients are static or mobile devices which are connected to the mesh routers. Different from mesh routers that are
connected to the electricity network, wireless mesh clients are more sensitive to power
consumption as they can rely on battery. Typical application scenarios for WMNs are
community networks, city networks, public safety or the extension of WLAN hotspots.
WMNs can concurrently support a variety of wireless radio and access technologies
such as IEEE 802.16 (WiMAX), IEEE 802.11 (WLAN) and IEEE 802.15 (Bluetooth and
Zigbee), thus providing the flexibility to integrate different radio access networks [54].
2.2. Wireless Mesh Networks
15
With the tremendous growth of WLANs and the proliferation of IEEE 802.11 devices,
the IEEE 802.11 wireless technology is the most common technology used on the deployment of WMNs. In this thesis we focus on the use of IEEE 802.11 for deploying
WMNs.
2.2.1
Multi-channel Multi-radio WMNs
Traditionally, WMNs were based on single channel single radio networks. Thus, due the
multi-hop communication characteristic, the throughput capacity of WMNs were highly
limited. It has been shown by [55], that on a string topology of n nodes and using carrier
sense multiple access with collision avoidance (CSMA/CA) based MAC protocol, the
network throughput degraded approximately to 1/n of the channel bandwidth.
Due to the availability of inexpensive off-the-shelf wireless radios, mesh routers have
been recently equipped with multiple radios, allowing them to send and receive on multiple frequency bands in parallel and consequently increase the throughput capacity of
WMNs. By using all available orthogonal channels and maximizing spatial channel reuse,
network traffic can be increased as a larger number of simultaneous transmissions is possible.
While multi-channel multi-radio provides several benefits such as increased throughput capacity, they also face several challenges as routing and channel assignment.
2.2.2
Routing
The main task of routing protocols is to select path(s) between the source node and the
destination node. Due to its common characteristics, several mobile ad-hoc networks
(MANETs) routing protocols have been used in deployments of WMNs. In order to select the best path(s) among all available, the routing protocol make use of routing metrics.
Routing Protocols
The routing protocols are normally categorized into reactive and proactive schemes. In
the reactive scheme, routes are established on-demand when data transfer is requested
from a given source to the destination. By operating on-demand, such scheme reduce
the amount of routing overhead, however the routing establishment phase may incur an
additional delay to the communication. Well-known representatives of such scheme are
dynamic source routing (DSR)[56] and ad-hoc on-demand distance vector (AODV)[57].
In AODV, if a source node wants to transfer data, it first need to check if a valid route
to the destination already exist. In case no valid route exist, the node initiates a route
request (RREQ) message to the destination node broadcasted over the network [57].
The destination, or intermediate node holding already a route to the destination, replies
with a route reply (RREP) message. When link fails, a route error (RERR) message is
generated in order to trigger route repair.
16
Background
The proactive routing scheme is characterized by the periodically routing updates exchanged among the nodes, independent of the data transfer. The proactive scheme incurs
negligible delays to the communication as routes are always available. In order to avoid
excessive routing overhead, the concept of multi point relay (MPR) is introduced by optimized link state routing (OLSR)[58]. The MPRs are selected nodes which forward
broadcast messages during the flooding process. Thus, each node maintains the topology
information about the network through link state information acquired via topology
control (TC) messages generated by the MPRs. Additionally, host and network association (HNA) messages are generated in order to disseminate external routing information
into the network.
In IEEE 802.11s [59], the hybrid wireless mesh protocol (HWMP) is the default routing protocol adopted, and it represents an adaptation of AODV and tree-based routing
at layer 2. The IEEE 802.11s standard also proposes the radio aware optimized link state
routing (RA-OLSR) as an optional routing protocol. The RA-OLSR is an adaptation of
OLSR at layer 2 working with arbitrary routing metrics.
Several benefits arises in case multiple paths are used. The option to allow the use
of multiple paths between source and destination have been proposed as an extension to
the routing protocols described so far. Examples of multi-path routing protocols are the
ad-hoc on-demand multipath distance vector (AOMDV)[60] and the Horizon [61].
As shown in [60], the multi-path routing achieve faster and efficient recovery from
route failures and therefore can guarantee better network resilience if compared to AODV.
Multi-path routing also allows network nodes to perform load balancing among the multiple paths selected, and therefore increase throughput and fairness among network flows
[61].
Routing Metrics
What is common to all routing protocol is the use of a routing metric in order to select
the best route(s) among all available paths in the network. In single channel WMNs, the
minimum hop count is the most common metric as it aims at reducing the amount of
concurrent link access in order to minimize the network bandwidth utilization in the
network. However such simple strategy does not consider important factor inherent to
wireless scenarios, such as interference level, link errors and criticality. Moreover, in
multi-channel multi-radio WMNs the channel diversity is also an important factor that
needs to be taken into account during the path selection.
A variety of routing metrics have been proposed for WMNs, providing a large range
of routing algorithms with different objectives. The expected transmission count (ETX)
is one of the first routing metrics that accounts for wireless link quality. According to
[62], the ETX of a link is the predicted number of data transmission required to send a
packet over that link, including retransmission. ETX has been implemented in various
routing protocols such as AODV, DSR and OLSR.
Since the size of the probe packets is fixed and the loss ratio in wireless networks varies
with packet size, inaccurate prediction are implicit to ETX. Moreover, ETX ignores the
2.2. Wireless Mesh Networks
17
fact that different links can operate over different data rate and therefore consumes different amount of channel time. The expected transmission time (ETT) extends ETX by
applying a packet pair probing technique in order to estimate the link rates.
Given the availability of multiple orthogonal channels, the weighted cumulative ETT
(WCETT) metric proposes an adaptation of ETT accounting for the channel diversity
[62]. The WCETT metric gives a trade-off between throughput and delay as it measures
path latency and bottleneck link along available paths. One limitation of WCETT is
its lack on accounting for inter-flow interference, which may cause path selection with
higher number of interfering neighbors and therefore worst performance.
Metrics such as the metric of interference and channel-switching (MIC)[63], the interference aware (iAWARE)[64] and the metric for interference and channel diversity
(MIND)[65], are an evolution of WCETT as they make use of number of interfering
neighbors, SNR/SINR correlation, and traffic load, respectively, in order to account for
interference.
2.2.3
Channel Assignment
The use of multi-channel multi-radio WMNs opens up the possibility for different channel assignment schemes with the common objective of assignment channels to radio interfaces in order to achieve efficient channel utilization and minimize interference.
Depending on the criteria selected, different classifications of channel assignment
arises. In [66, 67], the channel assignment is classified according to the frequency that
channels are modified, and can be described as static, dynamic, semi-dynamic and hybrid.
Other classifications, e.g. according to the level of centralization and implementation aspects, are also available [68].
Static Channel Assignment
In the static channel assignment, channels are permanently fixed to the radio interfaces.
The most simple example of such assignment is the use of a common channel assignment,
where the radio interfaces of each node are all assigned to the same set of channels [62].
Such approach guarantees similar network connectivity if compared to the single channel
approach.
Dynamic and Semi-dynamic Channel Assignment
On the other hand, the dynamic channel assignment proposes channel changes on-demand,
e.g. on a per-packet basis, and therefore requires frequent channel switching within each
node. According to [69], the channel switching time on current off-the-shelf hardwares
are between 200µs and 20 m s, causing a very high overhead for per-package switched
channel. Moreover, a high level of coordination between the nodes is necessary in order
to assign the channels at the right time and maximize the overall channel utilization.
In the semi-dynamic scheme, channels are re-assigned at larger time intervals, e.g.
several minutes or hours, depending on specific metrics such as network demand and/or
18
Background
interference. A channel assignment scheme that takes into account the network load
demand in order to re-assign channels to radio interfaces is proposed by [70]. Thus,
depending on the node position and traffic load, [70] re-assigns links to channels in order
to accommodate the link expected demands.
By assuming that traffic is commonly target to the internet, several channel assignment strategies make use of a BFS fashion assignment where nodes close to the gateways have higher priority while reassigning channels [70, 71, 72]. By incorporating such
parent-child relationship in the channel reassignment, those strategies avoid non convergence states during the channel reassignment period. However, ripple effects might
occurs as the channel reassignment at a given node might change due to the variations in
one of its neighbor’s channel assignment [67].
While capturing packets from the medium, each node in [71] can measure the number of radio interfaces and the bandwidth consumed by those radios in order to estimate
interference on a given channel. Based on those periodical channel interference information gathered by a central server, [71] proposes a channel re-assignment scheme that
minimize the interference within the wireless mesh network.
Hybrid Channel Assignment
The hybrid channel assignment strategy combines both semi-dynamic and dynamic assignment properties. Net-X [73], a known hybrid approach, applies a semi-dynamic assignment to the its fixed interface, used primarily for receiving data from neighbors, and
a dynamic assignment to the switchable interfaces, used to transmit data to its neighbors.
A local balancing algorithm that looks for the less utilized channel of a given node’s
neighborhood is applied to the fixed interface by assuming similar traffic distribution
over the channels [73]. If a node notices that the number of nodes using the same fixed
channel as itself is large, it can reassign its interface to a less used channel and inform
its neighbors. Based on the observation that traffic demands are mostly not uniformly
distributed in a WMNs, [9] proposes a demand-aware channel assignment which tries to
assure that nodes with high traffic demands are assigned to less loaded channels.
The switchable interface can be used to transmit to neighbors whose fixed interfaces
may potentially be on different channels [73]. Thus, the channel in the switchable interface can be changed frequently without having to inform all neighbors, as it depends
only on the next hop destination of the packet been transmitted.
Since the channel switching time incurs a non-negligible delay, a queuing algorithm to
buffer packets is required, as well as a round robin scheduling policy to transmit buffered
packets in order to reduce frequent switching. In order to consider the requirements
of real time traffic, [10] proposed to replace the round robin scheduler by a quality of
service (QoS) aware scheduler that takes into account packet priorities. By applying the
QoS-aware channel scheduler, [10] reduces the end-to-end delay and network jitter of
voice over IP (VoIP) traffic while still guaranteeing reasonable throughput of non-priority
TCP traffic.
2.2. Wireless Mesh Networks
2.2.4
19
Joint Routing and Channel Assignment
In multi-channel multi-radio WMNs, the problem of routing and channel assignment are
interdependent. The channel assignment determines the set of links sharing the same
channel and consequently the channel capacity and network topology. On the other
hand, the routing impacts the link bandwidth along the selected path, changing also the
interference level among links that share a common channel. Thus, a joint routing and
channel assignment mechanism is desirable in order to maximize network capacity.
Several works in the literature propose optimization frameworks in order to jointly
solve the routing and channel assignment problem, by using linear programming formulations [74, 75] and also heuristic approaches [66, 70, 76, 77].
In [70], the routing and channel assignment problem are solved using an interactive
approach, where an initial link load estimation is used to identify the capacity of a link
based on the set of communication nodes and interference range links. Given the expected load estimation, channel assignment and routing are performed interactively in
order to guarantee the expect traffic demands.
Following a distributed heuristic approach, [77] proposed a joint coordination between channel assignment and routing based on the traffic information measured and
exchanged among two hop neighbors. Since the network global view is not assumed, every node performs locally channel assignment and routing decisions based on the channel
cost metric (CCM) which represents the cost of channel time weighted by channel utilization. In case the measured metric on a given channel is higher than its predefined
threshold, representing a channel overload situation, the node start to check for each
neighbor node and channel, if a channel re-assignment, re-routing, or a combination both
is necessary. Channel re-assignment is negotiated among a node’s two hop neighborhood,
whereas the search for new routes is restricted by only changing the interfaces between
adjacent nodes, while the node sequence of the entire path remains the same.
Moreover, joint optimizations involving other layers have also been proposed, such
as the joint optimization of channel bandwidth adaptation, topology control and routing
by [78] and the channel assignment, routing and rate assignment by [79].
2.2.5
Modeling Capacity in WMNs
In order to model the capacity in wireless networks, two popular wireless interference
models formalized by [80] are used; the physical interference model and the protocol
interference model.
The physical interference model, also known as signal to interference plus noise ratio
(SINR) model, is based on the power capture model which states that a transmission
is correctly received if the SINR at the receiver is greater than a threshold so that the
transmitted signal can be decoded with an acceptable bit error probability. Given that,
[80] derives lower and upper bounds on the wireless capacity.
The known accuracy of the physical interference model comes with its high complexity cost when it involves cross-layer optimizations in a multi-hop network [81]. To
20
Background
work around the complexity of the physical interference model, the protocol interference model, also known as unified disk graph model, has been proposed. The protocol
interference model defines successful communication over a given link, when the receiving node falls inside the transmission range of the transmitting node but also falls outside
the interference ranges of other non-intended transmitters [81].
The protocol interference model is commonly used to represent the interference relationship between pairs of nodes or pairs of links by using graph representation, and assumes a simple deterministic path loss channel model and uniform transmit power [82].
Given the graph representation possibility, graph theory has played an important role
in wireless networks as a way to model connectivity and interference. Inside the scope
of wireless mesh network, it has been particularly used in the channel assignment, and
capacity modeling problems.
In the channel assignment problem, the graph theory can be used as a graph coloring
problem where given a finite number of colors (e.g.: channels) the goal is to assign the
colors to edges while satisfying certain constraints, such as the number of distinct colors
assigned to edges incident in a node can not be greater than the number of radio interfaces
on that node.
Inside the scope of capacity modeling, important works such as [38, 39] make use
of the collision domain model in order to derive the nominal capacity of wireless links.
According to the collision domain model, the capacity of a given link can be determined
by the number of simultaneously active physical links in the transmission range of its
transmitter or its receiver, that are assigned to the same channel. However, the exact
short-term instantaneous link capacity is changing continuously and depends on more
complex propagation and interference phenomenas which are not taken into consideration by the protocol interference model.
2.2.6
IEEE 802.11 PHY and MAC layers
The IEEE 802.11 standard is divided into two main layers; the PHY layer and the common MAC layer [83]. The PHY layer is composed by four technologies; infrared (IR),
frequency hopping spread spectrum (FHSS), direct sequence spread spectrum (DSSS),
and orthogonal frequency division multiplexing (OFDM). Table 2.1 presents a summary
of the most important physical layer features of the IEEE 802.11 standard.
While the 802.11b standard builds upon DSSS to increase the data rate in 2.4 GHz
using complementary code keying (CCK), 802.11a defined a new physical layer in 5 GHz
based on OFDM supporting data rates of up to 54 Mbps. In IEEE 802.11a PHY layer,
64 sub-carriers of 0.3125 MHz forms a 20 MHz channel. 52 out of the 64 sub-carriers are
used for actual transmission consisting of 48 data sub-carriers and 4 pilot sub-carriers. The
pilot signals are used for tracing the frequency offset and phase noise. The short training
symbols and long training symbols, which are located in the preamble at the beginning of
every PHY data packet, are used for signal detection, coarse frequency offset estimation,
time synchronization, and channel estimation. A guard time is attached to each data
OFDM symbol in order to eliminate the inter symbol interference (ISI) introduced by
2.2. Wireless Mesh Networks
21
Table 2.1: Overview of IEEE 802.11 physical layer [22]
PHY technology
Data rates
Frequency band
Channel bandwidth
802.11b
DSSS/CCK
1-11 Mbps
2.4 GHz
25 MHz
802.11a
OFDM
6-54 Mbps
5 GHz
20 MHz
802.11g
OFDM DSSS/CCK
1-54 Mbps
2.4 GHz
25 MHz
802.11n
OFDM
6-600 Mbps
2.4 and 5 GHz
20, 40 MHz
802.11ac
OFDM
1Gbps
5 GHz
80, 160 MHz
the multi-path propagation. In order to combat the fading channel, information bits are
coded and interleaved before they are modulated on sub-carriers.
In 802.11n, the rate increase continues as spatial multiplexing using multiple-in multipleout (MIMO) and 40 MHz channel bandwidth are supported [22]. In MIMO systems, the
combination of the transmit beamforming and receive diversity improve link robustness.
When multiple spatial streams ( defined as streams of bits transmitted over separate spatial dimensions) are used with MIMO, the maximum data rate of the system increases
as a function of the number of independent data streams. By using multi-user MIMO,
256-QAM modulation and larger channel bandwidths of 80 and 160 MHz, the 802.11ac
draft standard target to achieve even higher throughputs beyond 1Gbps in the 5 GHz
band [84].
Regarding the MAC layer, the IEEE 802.11 standard defines two medium access technologies; the point coordination function (PCF) and the distributed coordination function (DCF). Also known as infrastructure mode (basic service set (BSS)), the PCF requires
a central point entity, or access point, responsible for coordinating the access of the stations to the medium. As PCF requires the central control entity, it is seldom used in
WMNs. The DCF, also known as ad-hoc mode (independent basic service set (IBSS)),
uses a distributed coordination function in order to share the medium between multiple
stations. The DCF works as a listen-before-talk scheme, based on the CSMA/CA.
While using DCF, stations operating on the same channel contend for the medium
in order to avoid co-channel interference, which occurs when two nodes transmit on a
given channel within interference range of each other. Thus, before a station transmits
a packet, it listens for activity in the given channel and begins its transmission only if it
finds the channel to be idle. If no transmission is sensed for a specific period of time called
distributed inter frame space (DIFS), the medium is sensed idle and the station is allowed
to transmit. If the channel is busy, the station waits for the medium to become idle,
defers for DIFS, and waits for a random back-off period. If the medium remains idle for
the DIFS deferral and the back-off period, the station can begin the packet transmission.
On the other hand, if another station occupies the medium during the back-off period
of the station, the back-off timer stops and it is resumed after the channel is sensed idle
again.
The acknowledgment (ACK) packet is generated by the receiver station in order to
provide feedback on the correct transmission of a packet. In case the sender station does
22
Background
not receive the ACK response packet in a given time interval, this is an indication that
the packet was not successfully transmitted, either due to collision, poor channel conditions, or high interference during the the packet transmission time interval. In 802.11,
broadcast and multicast packets do not benefit from the reliability proportionated by the
MAC acknowledge mechanism.
In DCF, the carrier sense uses both virtual and physical carrier sense functions in
order to determine the state of the medium. The virtual carrier sense, also known as
network allocation vector (NAV), resides in the MAC and uses reservation information
carried in the duration field of the MAC headers. By setting the time expected to transmit a packet in the duration field, the station can update the NAV of other stations by
informing them how long the medium will be busy. However, to do so the packet must
be successfully demodulated by the neighboring stations.
The physical carrier sense resides in the PHY layer and its known as CCA. The
CCA typically tries to detect the presence of energy at the carrier frequency by using
a threshold-based approach, which indicates signal presence. Alternatively, the channel is
declared busy if a preamble is detected at a certain signal level. The CCA based on energy
detection is less reliable than preamble detection based schema but more energy efficient
because in order to detect a preamble it is required to continually sense the channel.
The distributed channel access method introduced by DCF brings challenges to the
WMNs such as collisions among stations operating at the same channel. The hidden
node problem is a typical example of the lack of collision free operation in DCF. Here,
stations located outside the transmission range of each other do not detect the ongoing
transmission and becomes hidden to each other. Thus, in case such hidden nodes transmit packets simultaneously, the destinations might not be capable of demodulating the
packets correctly. In order to minimize collision, request to send (RTS) and clear to send
(CTS) control frames can be exchanged among transmitting and receiving stations, after
determining that the medium is idle and after any deferrals, prior to data transmission
[83]. However, RTS / CTS is not a complete solution and may decrease throughput even
further since in many cases the reservation of the channel is less efficient than just dealing
with the collision through retransmissions [85], or network coding [86].
2.3
Peer-to-Peer Systems and Wireless Mesh Networks
One of the main differences between peer-to-peer and wireless mesh networks is related
to the level where they operate. P2P is essentially focused on building and maintaining
overlay network connections at the application level, while the main focus of wireless
mesh networks is to provide multi-hop connectivity among wireless nodes at the network
level.
2.3. Peer-to-Peer Systems and Wireless Mesh Networks
P2P Application
P2P Application
Routing Algorithm
Routing Algorithm
IEEE 802.11
IEEE 802.11
Network Interface
Network Interface
(a) Layered design
(b) Cross-layered design
P2P Application
Cross-layer interactions
IPv4/IPv6
IPv4/IPv6
23
IPv4/IPv6
Routing Algorithm with
P2P functionalities
IEEE 802.11
Network Interface
(c) Integrated design
Figure 2.2: Design choices of P2P and WMN integration
Due to the characteristics of multi-hop communication and the low resource availability in such networks, simply deploying a peer-to-peer system as is on top of wireless
mesh network routing layer as shown in Figure 2.2(a), might cause poor performance,
significant message overhead and redundancy in communication.
One alternative for avoiding bad interactions between those layers is the paradigm
of cross-layer design, as shown in Figure 2.2(b). Here, information from, for instance,
the routing or MAC layer is made available at the peer-to-peer layer or vice versa in order to improve the peer-to-peer system performance in wireless mesh networks. As we
present in Section 2.3.2, various approaches implement different cross-layer interactions.
As a result, a cross-layered design could offer a significant performance improvement if
compared to the simple layered design.
Another alternative to increase performance is to integrate the peer-to-peer layer with
the routing layer beyond the strict layering rule, as shown in Figure 2.2(c). Typically new
routing mechanisms such as key-based routing are developed, and try to implement peerto-peer concepts in the routing layer itself. In the next sections we provide an overview
of those three approaches, by giving examples and also trying to evaluate the key features
of each of them.
2.3.1
Transparent Layer
Deploying a peer-to-peer system directly on top of an existing broadcast based routing
protocol in wireless mesh networks does not require any changes to the routing or P2P
overlay layer. For example, if using a DHT approach, every resource name and peer can
be mapped to a key in the identifier space through standard hash algorithms (e.g. SHA-1
). Besides the routing table at the network layer, each node participating in the overlay
also needs to maintain an overlay routing table used to direct lookup queries to an overlay
neighbor peer closer to the requested key. In order to route messages between neighbors
in the overlay, standard routing protocols as described in Section 2.2.2 can be deployed,
which usually acquire topology information using broadcast.
24
2.3.2
Background
Cross-layer Layer
Via a cross-layer communication channel between the network layer and the P2P layer,
the mobile peer-to-peer (MPP) protocol [87] proposed a file sharing system for wireless
multi-hop networks. The MPP protocol stack reuses existing network protocols as much
as possible. For node-to-node communication, the enhanced dynamic source routing
(EDSR) combines Gnutella-style flooding and DSR routing. For the transportation of
user data it uses HTTP over TCP. The mobile peer control protocol (MPCP) is used as
an inter-layer communication channel between the application (MPP) and the network
layer (EDSR). Using the MPCP, the application can register itself in the EDSR layer to
initialize search requests and to process incoming search requests from other nodes.
P2PSI [88] proposes P2P file sharing based on swarm intelligence. Basically, it is an
hybrid push-and-pull system composed by advertisement and discovery processes. In
the advertisement process, each hotspot periodically advertises a seed message containing
digest information about files to be shared within a limited area ,e.g. as determined by the
hop count. In P2PSI, Bloom filter technique [89] is applied as a method for summarizing
the list of shared files. In the discovery process, the node willing to search for a file first
checks if it has cached the desired file information. Otherwise, it generate query messages
forwarded to intermediate nodes in the network based on the ant-based routing (ARA)
[90].
In CrossROAD [91], a cross-layered architecture defines the interactions between
peer-to-peer and routing layer in order to piggyback overlay advertisements into routing messages periodically sent by OLSR. As an specialization of Chord to wireless mesh
networks, MeshChord [92] makes use of the 1-hop broadcast nature of wireless communication while performing key lookup. Following a two-tier architecture for wireless
mesh networks, MeshChord clients rely on mesh routers for looking up for resources
in the overlay. Mesh routers are therefore responsible to forward the resource request
in the DHT overlay according to the rules specified by the Chord protocol, until the resource query can be answered. MeshChord explores location awareness by assigning IDs
to peers according to their coordinates, accomplished by, for example, the use of global
positioning system (GPS) receivers.
Hashline [93] represents another cross-layer approach based on CAN [94] where the
lookup functionality and the routing functionality are unified. Unlike CAN, Hashline
uses a one-dimensional space into which keys and node IDs are mapped. The hashline
is divided hierarchically into segments so that each node is responsible for one segment.
The relationship between segments can be considered as a tree consisting of parents and
children, so that the hashline segment of a parent spans the hashline segments of all its
children. Hence a tree based routing is used during lookup process.
Georoy [51] proposes a location-aware variant of the Viceroy [49] where a unit ring
topology is combined with a butterfly network topology to guarantee lookup performance in wireless mesh networks. As in MeshChord, Georoy also proposes a direct
mapping of the overlay identifier space to the physical location of the nodes by using a
geographic aware hash function. In order to maintain communication between overlay
2.3. Peer-to-Peer Systems and Wireless Mesh Networks
25
neighbors, Georoy make use of AODV routing protocol.
Exploiting also a cross-layer approach, [95] proposes a traffic optimization for BitTorrent in wireless mesh networks through the exchange of network layer information
between the routing agents and the BitTorent tracker via an application-layer traffic optimization (ALTO) based architecture [96]. Given the network layer information acquired
by a number of routing agents co-located within the wireless mesh networks, the ALTO
server is able to order the peers in terms of increasing distance from the asking peer and
therefore provides a better ranking for the peer involved in the download process.
2.3.3
Integrated Layer
ORION [97] makes use of an integrated approach depicted in Figure 2.2(c) by enhancing
a general purpose distributed lookup service with P2P file sharing capability. ORION
applies the integration of Gnutella-style [40] flooding into the AODV [57] routing to
locate requested files in the network. With ORION, each node has a local repository
containing the files that the node is sharing. When a node wants to locate a certain
file, it issues a query message that is broadcasted through the network. After receiving
responses match the description (e.g., file name, key words, etc.) specified in the query,
data request message is sent to the provider(s) using the AODV-style routes discovered
during the search.
The zone-based P2P (ZP2P) [98] tries to reduce the heavy overhead of always broadcasting search requests in the networks by applying the concept of local zones, determined by fixed hop-counts. When a node is interested in a certain object, it will first
check its local cache to see whether any of its zone members can provide the desired object. However, in case the requested object is not available in the node’s own zone, it
will initiate a bordercast of the request through its border nodes, i.e., to those of its zone
members that are exactly k hops away.In case a border node finds that there are no members in its zone that could provide the requested object, it will continue the bordercast
by forwarding the request to its own border nodes. This process continues until either a
predefined TTL expires or the whole network has been searched.
Following the same principle of ORION and ZP2P, the virtual ring routing (VRR)
[36] also developed an integrated approach where peer-to-peer concepts were pushed to
the network layer. However, instead of flooding the network with routing messages,
VRR proposes a new routing protocol based on the DHT concept. Based on Pastry
[47], VRR organizes the nodes into a virtual ring ordered by their identifiers, where each
node maintains a small number of routing paths to its neighbors in the ring. Unlike
routing protocols that forward packets based on destination address, VRR nodes route
packets to destination identifiers (keys) by forwarding them to the next hop towards the
path endpoints whose identifier is numerically closest to the destination identifier from
among all the endpoints in their routing table.
The scalable source routing (SSR) [99] brings the same concept of VRR while trying
to integrate the peer-to-peer overlay into the network layer. But while VRR does not
assume any specific routing protocol integration, SRR combines the DSR [56] in the
26
Background
physical network with Chord routing [46] in the virtual ring formed by the address
space. SSR trades off shortest path for a reduced amount of state information, leading to
less maintenance overhead.
In order to take physical location into consideration, MADPastry is proposed by
[100]. It integrates the application layer Pastry and the AODV routing protocol and
make use of the random landmark concept [101] to create physical clusters where nodes
share a common overlay ID prefix. To avoid fixed landmark nodes, MADPastry divides
the overlay ID space into equal-sized segments making use of a set of landmark keys.
As broadcast imposes network overhead, landmark beacons are only propagated within
the landmark’s own cluster, i.e. beacons are only forwarded by nodes belonging to that
cluster.
BitHoc [102], proposes an enhancement to BitTorrent protocol aiming to minimize
the resource download time in wireless ad-hoc networks. In order to avoid the dependency on the tracker, BitHoc employs a trackerless approach where resource and peers
are discovered through the use of network layer hello message flooded within a given
TTL limit.
2.4
Conclusion
In this chapter, a general overview of P2P systems and wireless mesh networks are given.
We start by introducing a basic classification of P2P systems. Inside the scope of wireless
mesh network, we start by introducing the benefits of using multi-channel multi-radio
systems, and also describe their challenges from the point of view of routing, channel
assignment, wireless capacity modeling, and IEEE 802.11 PHY and MAC layers. The
chapter also gives an overview on different principles that guide the various integration
and interaction possibilities for peer-to-peer systems in wireless mesh networks through
a set of proposed solutions in the literature.
Chapter 3
Adapting the P2P Overlay to
WMNs
Adapting the P2P overlay to wireless mesh network scenarios is the focus of this chapter.
We first present the proposed architecture, followed by the detailed description on the
challenges encountered while deploying P2P systems through DHTs in wireless mesh
networks. The performance penalties of a transparent layered approach are detailed in
Section 3.4 where a packet level performance analysis of Bamboo over wireless mesh
networks has been conducted.
As described in Section 2.3, one alternative to avoid possible bad interactions is by
enabling information exchange between the two layers, through a cross-layer or integrated approach. Here, information from, for instance, the routing or MAC layer is
made available at the overlay or vice versa. As a result, such information exchange can
offer a significant performance improvement if compared to the transparent layered approach. Thus, we described in Section 3.5 a performance comparison between Bamboo, a
layered approach, and Georoy, a cross-layered approach targeted to WMNs, and identify
a trade-off between resource lookup efficiency and algorithm complexity. Additionally,
we make use of neighborhood and caching information in Section 3.6 in order to augment the overlay routing information as a way to further improve the overlay efficiency
in wireless scenarios. Finally, we draw our conclusions and outline the limitations of our
work.
3.1
Our Contribution
The first contribution of this chapter is the proposed architecture for enabling P2P systems over wireless mesh networks, where we described the main components together
with the modules and interactions required. We also present an overview on the challenges encountered while deploying P2P systems over wireless mesh networks. Chal-
28
Adapting the P2P Overlay to WMNs
lenges, such as bandwidth constrains, P2P overlay maintenance and network resilience
are described in detail and possible solutions are outlined. To have a further insight on
the challenges described, we carried out simulation in order to characterize the Bamboo
management and control traffic in wireless mesh environments. Thus, the second contribution of this chapter is the identification of trade-offs when deploying such structured
overlay solution over wireless mesh networks. Recommendations on configuration settings are proposed to balance the P2P lookup efficiency and management traffic overhead
in such scenarios.
There are several examples in the literature where the knowledge on the interactions
between P2P and wireless multi-hop networks can either help to realize more efficient
P2P networks and services on top of mesh networks or lead to the design of better and
more scalable routing protocols [36, 51, 100, 103, 104]. Understanding such interactions
might help to clarify which support among different layers shall be required for scalable
operation of P2P on top of wireless mesh networks. Thus, our third contribution of this
chapter is on testing the benefits of location-awareness and resource replication while
analyzing Bamboo and Georoy over wireless mesh scenarios. We start by analyzing the
impact of location-awareness on different network size in terms of number of physical
and logical hops, lookup delays, routing stretch and number of successfully completed
lookups. While targeting on improving the lookup process, we also study the impact
of resource replication mechanisms on Bamboo and Georoy. Now as multiple resource
replicas are present in the network, the resource availability is not mandated anymore
just by the availability of a single responsible node.
We also explore the benefits of location-awareness by considering the one and two
hop physical neighborhood information from the routing layer in order to augment the
overlay routing table. By replacing long-range overlay neighbors by physical neighbors,
the lookup requests rely mainly now on leafset and physical neighbors. Furthermore,
we exploit the lookup request history through the use of a least recent used (LRU) cache
scheme. The intuition behind this idea is that by overhearing the lookup requests at
the MAC layer, the nodes can enrich their cache entries and increase the overlay lookup
efficient in wireless scenarios.
3.2
Proposed Architecture for P2P Systems over Wireless Mesh Networks
In this section, we describe our proposed architecture for P2P systems over wireless mesh
networks. We make use of a resource store and lookup phases of a file sharing application
to illustrate the benefits of the proposed architecture.
3.2.1
Architecture components
Figure 3.1 depicts the three main components of the proposed architecture: (1) the mesh
router, (2) the mesh client, and (3) the central server. As we explain below, each com-
3.2. Proposed Architecture for P2P Systems over Wireless Mesh Networks
Client
29
Applications
PHY/MAC
802.11
P2P
Mesh Router
PHY/MAC
802.11
Routing
Tracker/DHT
Channel
Assignment
Packet
Forward
Central Server
Network
Channel
Assignment/
Routing
Tracker
control flow
data flow
Figure 3.1: Architecture: components and interconnections
ponent is composed by a set of modules and its control and data interconnections. The
main objectives of the proposed architecture is to provide the required building modules
and interactions to enable efficient P2P systems over wireless mesh networks.
Mesh router: is responsible to form an infrastructure in order to provide connectivity to mesh clients. Via the packet forward module, the mesh router can relay traffic from
mesh clients and others mesh routers. We assume that the mesh router has at least two
network interface cards; one operating in the infrastructure mode in the 2.4 GHz band
(IEEE 802.11b/g) used to communicate with mesh clients, and the other(s) operating in
ad-hoc mode in the 5 GHz band (IEEE 802.11a) used to communicate with mesh routers
forming the backhaul.
The assignment of channels to the network interface cards is realized by the channel
assignment module. We assume that a semi-dynamic channel assignment is applied to the
ad-hoc interfaces. The channel assignment algorithm can operate distributed, through
the information exchange in the mesh router’s two-hop neighborhood or via the central
server. A static channel assignment is applied to the access point interface operating on
the IEEEE 802.11b/g, as we assume the network bottleneck occurs in the backhaul links
30
Adapting the P2P Overlay to WMNs
(links between mesh routers) and not at the access links (links between access point and
mesh clients).
The mesh router is also responsible for establishing routing paths in the network
in order to guarantee end-to-end communication between source and destination nodes.
Mesh routers may also act as internet gateways. Since mesh clients and mesh routers
reside in a private address space, internet gateways perform a network address translation
(NAT) [105] before forwarding the packet to the Internet.
To enable resource sharing among mesh clients, mesh routers make use of the P2P
overlay module that is responsible for maintain the virtual overlay among mesh routers
and also acts as a tracker to the associated mesh clients. By using DHTs, the mesh routers
provide key-based routing and a conventional DHT interface to maintain the overlay.
Thus, distributed applications (e.g. P2P SIP [106], file sharing [31], video streaming
[107, 108]) that have been build using DHTs in the internet can be ported in a straight
forward manner to wireless mesh networks. Via resource replication and cross-layering
information exchange with the routing and 802.11 PHY/MAC module, better lookup
efficiency is achieved as the overlay replaces long-range overlay neighbors by physical
neighbors.
Mesh client: is assumed to be static or mobile devices (e.g. smart phones, laptops,
etc.) that access other mesh clients or the internet via multi-hop communication enabled
by mesh routers. The mobile clients are not part of the wireless mesh topology. We
assume that mesh clients have at least one wireless network interface card operating in
the 2.4 GHz band.
Different application are expected at the mesh clients. We emphasize the use of P2P
file sharing application e.g. BitTorrent [31]. Using a BitTorrent-like approach, the mesh
client can request the desired resource index i.e. the list of peers participating on a given
download process, by using the mesh routers or the central server. Modification to the
P2P application at the mesh clients is not necessary. If the central server is available, it
acts as a tracker by holding the resource indexes available in the network and answering
mesh clients directly upon request. Otherwise, the requests are directed to the mesh
router which act as a decentralized tracker, as it provides DHT functionality.
Central server: is located in the internet and applies monitoring and control functionalities to the wireless mesh network, e.g. DHCP server, central monitoring point,
etc. The central server can act as a BitTorrent tracker, by providing the list of peers
holding the desired resource in the network.
We assume that, if necessary, the central server can perform channel assignment and
routing functionalities on behalf of the mesh routers. In such case, each mesh router periodically collects neighborhood information in terms of neighbors and channels utilized
in its two-hop neighborhood, together with the traffic matrix. Given such global view,
the central server performs the routing and channel assignment that leads to minimum
interference and contention with neighboring nodes. Finally, the output of the channel
assignment and routing algorithms are enforced at the mesh routers. Despite the availability of different proposals to collect and enforce routing and channel assignment decisions
3.2. Proposed Architecture for P2P Systems over Wireless Mesh Networks
31
at the mesh routers [70, 71, 72, 75], we remark that the best solution for implementing
the central server depends on many factors, such as network size, traffic dynamics, number of interface available. We left this aspect unspecified in our work, as we emphasize
in this theses on the interaction benefits between the P2P overlay, routing and channel
assignment modules.
3.2.2
Example of P2P file sharing application: resource store and
lookup phases
We emphasize that different P2P applications can be deployed in a distributed fashion
by using simple put (store) and get (lookup) operations provided by the DHT in our
proposed architecture. We assume that all mesh routers can store (key, value)-pairs on
the DHT and later find values by looking up for keys. The SHA-1 hash function leads
to every mesh router being responsible for a range of keys. Put and get operations are
performed by finding the mesh router responsible for the search key and then sending
it a put or get request. If a mesh router wants to store or lookup for the search key, it
compared its own identifier (ID) with the key and, if it falls in its range it replies to the
requester, otherwise the request is forwarded to the overlay neighbor whose ID is the
closest with respect to the search key, as we illustrate in the example depicted in Figure
3.2.
When DHTs are used for swarm discovery, the address of all peers involved in the
swarm have to be stored for each BitTorrent swarm. This is implemented by using the
mapping of swarms as keys and the mesh routers as values via hashing functions. Each
swarm is mapped to a unique mesh router, known as the responsible peer. We emphasize
that a single DHT is needed in the wireless mesh network irrespectively of the number
of swarms. More information on the overlay routing operation on DHTs is provided in
the upcoming sections.
Figure 3.2 presents the example of resource store and lookup phases in DHT swarm
discovery over wireless mesh networks. In this example we do not make use of the central server, and therefore the tracker functionality is realized by the mesh routers which
provides a hypertext transfer protocol (HTTP) service which responds to HTTP get requests. The mesh router working as a tracker can be used by any associated mesh client
by adding the mesh router address to an existing or newly created swarm.
We represent in Figure 3.2 the mesh routers and swarm ID by three binary digits. In
this example, the mesh routers forms a ring-based overlay network, where mesh clients
Alice and Bob are involved in the download process of file1. We assume that Alice has
initially the whole content, and the mesh router 011 is the responsible peer for the swarm
file1 ( hash(file1) = 011 ) 1 .
We depict in Figure 3.3, the message sequence chart of the example presented in Figure
3.2 where Alice wants to store itself as a swarm member, and Bob wants to retrieve the
1 For sake of simplicity, we assume that the file1 swarm ID and the mesh router 011 ID are mapped to the
same bit sequence 011
32
Adapting the P2P Overlay to WMNs
000
HTTP
get/reply (file1)
111
Alice
001
DHT
put/get (file1)
110
010
Bob
file1
101
011
HTTP
get/reply (file1)
100
Alice
Bob
Figure 3.2: Example of a store and find resource action using DHT functionalities.
peerlist for swarm file1. The swarm discovery starts when Alice has joined download
swarm file1, by publishing the file1 at the mesh router 001 via HTTP get request. As in
BitTorrent, the HTTP request includes information such as the key (swarm ID of file1),
IP address, port, and download status (started/stopped/completed) [31].
The mesh router 001 stores the swarm in the DHT by issuing a put message routed
greedily through the overlay network until it reaches the responsible peer for the swarm,
the mesh router 011. The responsible peer stores the mesh client Alice as a participant
in the swarm and sends an acknowledgment back to the mesh router 001 by issuing a put
reply message. The store phase ends up as Alice receives the HTTP reply from its mesh
router informing about the list of peers participating in the swarm, which in this case has
only Alice.
A similar process is done by Bob, while searching for the list of peers participating
in the swarm file1. Here, Bob’s mesh router 101 starts the lookup request process in the
overlay by inquiring the responsible peer 011 through a get message the actual peer list
for swarm file1. As in the put message, the get message is routed greedily through the
overlay network until it reaches mesh router 011, which generates a get reply holding
the list of peers participating in the swarm. The lookup phases terminates when Bob
receives the list of peers from mesh router 101. Thereafter, Bob starts the download
phase by contacting directly the peers from the peer list obtained, in this case Alice.
3.3. Challenges of Deploying P2P Systems over Wireless Mesh Networks
Alice
Bob
Mesh-001
Mesh-101
33
Mesh-011
HTTP_get (file1)
store
put(file1,Alice)
file1
Alice
put_reply(file1,Alice)
HTTP_reply (file1)
lookup
HTTP_get (file1)
get(file1,Bob)
file1
get_rely(file1,Bob)
Alice
Bob
HTTP_reply (file1)
Figure 3.3: Message sequence diagram for a store and find resource action using put/get
DHT functionalities.
3.3
Challenges of Deploying P2P Systems over Wireless
Mesh Networks
The use of wireless mesh networks for applications that rely on a P2P architecture for
information exchange presents several challenges [2]. Here, not only the wireless nodes
require content delivery but also act as content providers. In addition, wireless users
are expected to use and offer data services in an effective manner, despite the scarcity
of bandwidth inherent to wireless communication and the possibility of intermittent
connectivity due to wireless characteristics and node mobility. Therefore, we list below
some of the technical challenges involved in deploying P2P systems over WMNs.
Bandwidth Constraints
The challenge of introducing P2P concepts in multi-hop networks is that P2P overlays
designed for the wired Internet rely on the IP routing infrastructure, which is resource
rich especially in terms of bandwidth availability. Wireless mesh networks are however
rather limited in bandwidth if compared to wired technologies. Therefore, a high overlay maintenance traffic, e.g. as it is used currently in structured overlay networks, will
lead to scalability problems when legacy P2P systems are used "as-is" in multi-hop environments. One of the main issues is therefore how to efficiently provide the same P2P
services implemented in legacy wired networks in mesh networks, and how to enable
efficient overlay services and applications on the resource constrained wireless environment. We study in Sections 3.5 and 3.6 approaches that try to overcome such challenge
by integrating, or applying cross-layering techniques between the P2P and the WMN.
34
Adapting the P2P Overlay to WMNs
P2P Overlay Maintenance
Keeping the overlay routing table of each node up to date is one of the main tasks of
a P2P system 2 . Efficient routing depends on routing information being current and
consistent. Invalid entries cause unnecessary overhead because of misrouted messages
and suboptimal routing. To avoid these inconsistencies, DHT protocols employ maintenance mechanisms to keep the routing tables up to date. Typically, nodes probe their
neighboring nodes via periodic ping request and response messages to learn whether they
are still available or not. In WMNs, such maintenance traffic further contributes to congestion and collisions. As path failure might lead to topology changes in the routing layer,
there might be potential for misrouted messages in case the overlay and the routing layer
have inconsistent topology information. Also, triggering such maintenance traffic during
network re-routing further contributes to network instability. Thus, a balance between
overlay lookup efficiency and management traffic overhead is studied in Section 3.4.
Network Resiliency
In P2P networks with structured overlay, DHTs are considered to be very resistant
against node failures. Backup and recovery mechanisms, that use distributed redundant
information, ensure that no information is lost if a node suddenly fails. When DHT
protocols are used in an multi-hop environment, resilience is as a very important issue.
The resilience of a DHT determines how much time may pass before expensive recovery
mechanisms have to be evoked.
The quality of connections in wireless mesh networks is dependent on the environment, as link quality variation may occur over time, and in the worst case links can
become temporarily inaccessible. Thus, if the recovery process is started too early an
avoidable overhead is caused if the node becomes accessible again. However, if the topological structure allows the DHT protocol to delay recovery mechanisms without losing
routing capability, these costly recovery measures can be avoided and the maintenance
costs of a DHT can be significantly reduced.
As described in Section 3.4, a compromise made between overlay management traffic
and network congestion is important in order to find a balance between lookup efficiency
and management traffic overhead. We also analyze in Section 3.5 the benefit of resource
replication, and show that by making use of several replicas the resource availability is
not constrained any longer by the availability of a single responsible node. However, the
availability of multiple replicas may be constrained by storage limitations.
Routing Stretch
Unlike the P2P overlay in the internet, where the neighbor is directly reachable using
an underlying routing protocol, in the P2P overlay in WMN scenarios, contacting the
neighbor may require going through multiple wireless hops. For this purpose, a pointer
2 Along
this thesis we make use of structured P2P overlay and DHT interchangeably.
3.3. Challenges of Deploying P2P Systems over Wireless Mesh Networks
35
is maintained for every overlay’s neighbor as a path through the network, consisting of a
set of physical links from the node hosting the pointer to its overlay’s neighbor.
When routing to a destination via DHTs, the node resorts to simple greedy routing: it selects the overlay’s neighbor that makes the most progress in the identifier space,
and then forwards the packet along the pointer. Forwarding along this pointer can be
achieved either through a source route inserted by the sender (e.g. SSR [99]) or through
embedded state in the network in the form of incremental source routes to the overlay
neighbor (e.g. VRR [36]). When the packet reaches the overlay-neighbor, it repeats the
same greedy routing process until the packet makes it all the way to the destination.
Therefore, routing proceeds at two levels: along the overlay from one overlay neighbor
to another, and then from one overlay neighbor to another via hop-by-hop along the
pointer source route.
The ratio between the cost of selected route using the overlay-neighbor to the optimal
shortest path routing through the routing protocol is defined as the routing stretch metric.
Small routing stretch means that the selected route is efficient compared to the shortest
path route, as a smaller number of nodes is involved in the forwarding process, reducing
the contention probability in the network. This is a key quantitative measure of route
quality used by the P2P overlay, and affects global resource consumption, delay, and reliability. Thus, minimizing routing stretch is a critical issue for a multi-hop environment as
both delay and packet loss increase significantly with the growth of the number of hops in
the physical path. In Section 3.6 we study the use of physical neighborhood information
as a way to reduce the routing stretch and speed up resource lookups.
Exploiting Heterogeneity
Another important point while deploying P2P overlay is which nodes should participate
in the overlay given that not all nodes in a network may be overlay members. While
typically nodes in an overlay are initially placed manually, nodes may also dynamically
and automatically decide to join and make services available. This issue may be especially
important in multi-hop environments because overlay participation may be dictated by
topological location which might change over time. Note, that other (e.g., physical)
constraints may drive the decision to participate in the overlay. For example, nodes with
limited power may not wish to act as overlay routers for other nodes.
Query Propagation
The propagation of query messages in the network is a critical aspect of the information
sharing mechanism in P2P networks. Indeed, there are two contrasting requirements
that arise in WMNs. On the one hand, queries for information must be forwarded by
relays until they reach nodes holding such information, and some redundancy in forwarding is necessary to compensate for the unreliable nature of broadcast transmission
of queries (i.e., no acknowledgments). On the other hand, congestion deriving from
excessive spreading of queries and reply duplication must be limited.
36
Adapting the P2P Overlay to WMNs
The simplest solution for query propagation is, of course, plain flooding of requests,
but this is hardly viable in tightly-meshed, bandwidth-hungry wireless networks where
congestion is more than likely. The introduction of a query TTL can shorten the reach
of broadcast queries. A balance should be found between small values of TTL, which
limit the success probability of a query, and query load.
Smarter ways of relaying traffic can also be achieved, for example, by forcing each
relay to wait for a query lag time before rebroadcasting the query. The propagation of a
request can be halted if a node in the neighborhood returns a response in the meantime,
thus making any further query propagation useless. However, issues such as how to select
intermediate nodes for query rebroadcast may turn out to be very important. Information caching can also be used to improve resource availability and consequently reduce
query propagation in the network. We show in Section 3.6, the action of overhearing
queries at the MAC layer enhances the probability to target the right destination node in
wireless mesh scenarios.
Cooperative Content Caching
In purely decentralized overlays, a highly debated issue addresses the most appropriate
caching strategy in an environment where a cache-all-you-see approach is clearly unfeasible but where the availability of sought-after information from nearby nodes is the key
to success.
This issue can be addressed through distributed caching strategies where nodes may
cache highly popular contents that pass by, or record the data path and use it to redirect
future requests. An interesting aspect is also how to minimize data access cost when
network nodes have limited storage capacity. Furthermore, in case different copies of the
same information are injected in the network, maintaining cache consistency among the
different nodes becomes a critical issue.
Inside the scope of cooperative caching in wireless mesh networks, we analyze in
Section 3.6 the possibility to increase the lookup efficiency by exploiting the lookup
request history through the use of a cooperative caching scheme among wireless nodes.
Security
Deploying security mechanisms in P2P networks is also a challenging task due to the
characteristics of P2P paradigm such as anonymity, decentralization, self-organization
and frequent disconnections. Security in P2P over wireless mesh networks is even more
challenging due to node mobility and easy access to wireless channels. Most security solutions require use of public keys for authentication, shared secret establishment, or integrity checking, and hence somehow depend on a public key infrastructure (PKI) [109].
PKI is needed by asymmetric cryptography to establish the validity of the public keys.
For this purpose, PKI stores digital certificates that attach a public key to the name of its
owner by the digital signature of a trusted third party called the Certification Authority
(CA). The management of certificates is a complex duty that requests a substantial infras-
3.4. The Bamboo Overlay Maintenance Cost in WMNs
37
tructure, especially in large-scale applications. Integration of PKI and CAs, or a similar
security infrastructure, into P2P over WMN is challenging due to ad-hoc and infrastructureless nature of the network and lack of centralized entities. Even in P2P networks with
servers (hybrid centralized or partially centralized), these servers usually do not fully control the peer behaviors as much as servers can do in a conventional client-server model.
Thus, the centralized architecture of PKI may introduce several important problems that
contradict with the important characteristics of the P2P networks and WMNs. Additionally, PKI and security services may introduce substantial amount of control traffic
into the network, which means more load to bandwidth-limited wireless channels.
3.4
The Bamboo Overlay Maintenance Cost in WMNs
We present in this section the performance penalties of a transparent layered approach,
exemplified through the deployment of Bamboo DHT over wireless mesh network scenarios. We start by giving a brief background on the Bamboo DHT, followed by a detailed description on its overlay maintenance operation. We finalize the section by presenting our simulation results which show the impact of different overlay management
settings in terms of management overhead, resource lookup completed and latency for
different network size.
3.4.1
Bamboo DHT
Bamboo [50] is inspired by previous DHT schemes such as Pastry [47] and aimed at reducing congestion due to large management traffic. In Bamboo an identifier space ranging
from 0 to 2160 − 1 is used to identify both resources and nodes participating in the overlay. Each node in Bamboo is assigned a 160 bits node identifier (ID) used to indicate the
node’s position in the circular ring-based identifier space. The node identifier is assigned
randomly and uniformly distributed in the identifier space when the node joins the system. The node identifier is generated by hashing the IP address and port of the node
through a SHA-1 hash function [50].
Resources are also assigned to the identifier space and commonly known as keys. The
keys are 160 bit identifiers generated by hashing the resource description (e.g.: file name).
Since the nodes participating in the overlay manage the identifier space, every key needs
to be mapped within nodes that have the closest node identifier. Thus, every key stored
in the Bamboo overlay has a responsible node storing the resource or the pointer to the
node holding the resource.
To maintain the overlay structure, Bamboo uses two sets of neighbor information
at each node: the leafset and the routing table. The leafset consists of successors and
predecessors that are numerically closest in the identifier space. Taken Figure 3.2 as an
example, node 001 has nodes 010 and 000 as its successor and predecessor respectively. In
the next section, we describe how Bamboo maintains its neighborhood set.
At the core of each Bamboo node lies the ability to route a packet based on a key,
38
Adapting the P2P Overlay to WMNs
towards the node in the network that is currently responsible for the packet’s key. This
process is referred to as indirect or key-based routing. To improve lookup performance,
a routing table is used, which is populated with nodes that share a common prefix. In
Bamboo, routing table lookups are ordinary longest prefix matches. In order to guarantee
the O(l o g n) upper
€ bound
Š in the search operation, Bamboo implements a routing table
b
of size l o g2b n × 2 − 1 , where n is the number of nodes in the network and b is a
configuration parameter (e.g. b = 2).
While Bamboo is based on the routing logic of Pastry, the management of overlay
structure is different in the aim of being more scalable in dynamic environments. The
major difference between Pastry and Bamboo is the way they handle overlay management
traffic. In Pastry, the overlay management is initiated when a network change is detected,
while in Bamboo management traffic is periodic, regardless of network status. Thus, by
applying a proactive management approach, Bamboo intends to react faster to unexpected
changes in the underlay network. The approach to use periodic updates has shown to be
beneficial during churn [50], since it does not cause management traffic bursts during
network congestion.
3.4.2
Bamboo Overlay Maintenance
In order for Bamboo to be able to serve requests and maintain a consistent network view
among its nodes, it needs to perform overlay maintenance message exchange between
nodes. Periodic management traffic occurs at all layers of the Bamboo system. Neighbor
ping is generated by every node in order to make sure that the node can still reach its onehop neighbors in the overlay. It is also used to maintain a round trip time (RTT) estimate
used for retransmission timeout calculations. Such timer values are used to derive, if e.g.
members left the overlay or messages need to be re-transmitted. However, re-transmitting
too early will lead to high number of packets in the network. An accurate timeout value
is crucial in order to predict if a packet is lost and needs to be resent along a different path
in the overlay.
To maintain overlay consistency, Bamboo needs to keep leafset information up-todate. In order to propagate leafset changes, every node periodically issues leafset updates
by performing pushing and pulling of leafset information to a random node in its leafset.
By doing that, a Bamboo node avoids the situation where its ID is not present in the
leafset of its neighbors.
As in Pastry, the routing table in Bamboo is used to augment the overlay efficiency
by reducing the number of logical hops that a lookup must traverse. Bamboo considers
that two nodes share the same level when one node contains the other node in its routing
table. Therefore, local routing update is used to exchange the node information in that
level. In case a node gets information about other nodes that fit into the routing table,
Bamboo probes the nodes to test reachability and consequently to get a RTT estimate. A
node is added if it is reachable and fits into an empty field in the routing table. The node
with the lowest latency is chosen, in case many nodes are eligible for a given routing table
entry.
3.4. The Bamboo Overlay Maintenance Cost in WMNs
39
In order to minimize the routing stretch, Bamboo implements the proximity neighbor selection (PNS) scheme called global sampling. Global sampling influences the process of choosing among the potential neighbors for any given routing table entry according to their network latency to the choosing node. Thus, global routing update is used in
order to improve the routing entries at all level. At a high enough rate the global routing update would achieve perfect proximity at the overlay structure, but at the cost of a
very large number of lookups and latency probes. According to [110], a small amount
of global routing updates is necessary to produce considerable improvement in latency.
To guarantee data storage consistency among the leafsets, the data storage update synchronizes the node’s stored keys with a random node selected from its leafset. By comparing the list of stored keys, a node can request from each other the resources that are
missing. The data storage update is also responsible for moving resources that are no
longer within a node’s storage range, by invoking the put message to the node current
responsible for such resource.
In standard configurations, Bamboo optimizes latency. It is important to note that an
optimized routing table does not influence lookup correctness, but only lookup latency
[111]. As wireless networks are rather limited in bandwidth, a balance between overlay
lookup efficiency and management traffic overhead is important [12].
3.4.3
Experimental Setup and Simulation Results
To analyze the performance of Bamboo over wireless mesh networks we show in this
section simulation results using different overlay management settings. As described previously, Bamboo uses proactive management traffic in order to maintain the network
structure. It is expected that the proactive management maintenance introduced by Bamboo should increase network traffic, and consequently as the network grows, congestion
may start to increase.
In order to find a balance between management traffic in the overlay and network
congestion at the underlaying wireless network, three different configurations for Bamboo management traffic are compared in this section; no management, standard management, and custom management. It is important to note that, for all three configuration settings, no exchange of information between the DHT overlay and the underlaying
wireless mesh network layers occurs. Table 3.1 presents the parameters used by each
configuration in Bamboo.
The s t and a r d management configuration used is based on the settings proposed by
[50] and used in internet scenario. By issuing frequent leafset and data storage updates,
the leafset nodes guarantee up-to-date leafset information among its neighbors and also
consistent data storage through the synchronization of stored keys. Neighbor pings are
also generated in order to maintain RTT estimation and one-hop neighbor reachability.
Moreover, by applying routing table updates the standard management configuration can
reduce the amount of logical hops on lookup request by exchanging routing table information. Opposed to that, no management configuration just guarantee one-hop neighbor reachability and RTT estimation among leafset nodes. By applying no management
40
Adapting the P2P Overlay to WMNs
Table 3.1: Bamboo management timers (secs)
Leafset update
Local routing update
Global routing update
Data storage update
Neighbor ping
NO
Standard
Custom
0,5
1
5
10
2
0,5
5
10
20
6
0,5
we expect a drastic decrease in overhead but at the cost of degrading lookup performance.
Thus, a trade-off between management overhead and lookup performance in a wireless mesh network is proposed through the c u s t o m configuration. The idea behind custom configuration is simple, as it tries to reduce the management overhead by increasing
the configuration timeout, but at the same time keep the updates at leafset, data storage
and routing level. Through the comparison of the three configuration settings we envision to find a balance between lookup efficiency and management traffic overhead. Too
frequent management traffic will lead to high overhead in wireless mesh environments
and thus lead to network congestion. No management, on the other hand, will leads to
low lookup efficiency.
The simulations were performed using ns-2 [112] over different scenarios where the
nodes were positioned on a grid at a distance of 200m with 250m of transmission range
and 500m of carrier sense range using two ray ground as radio propagation model. The
transmission rate is set to 11Mbit/s, and the basic rate to 1Mbit/s. The AODV-UU
routing protocol was adopted using default settings proposed by [113]. Simulations were
performed for 60 seconds. No bootstrapping period is required, as we assume that at the
beginning of the simulation the peers have already joined the network, and the overlay
topology is formed. During the experiments, every 2 seconds, each node generates a 500byte put message with a random key to store data in the overlay. All nodes also try to
acquire random selected keys that are located on other nodes generating a 32-byte get
message every 2 seconds. By considering such request rate we are targeting a worst case
scenario where the network is constantly loaded with resource requests.
The total management traffic is defined as the aggregation of the overlay management
traffic (e.g. neighbor ping, leafset update, routing table update, and data storage update)
during the whole simulation time. Here, the routing traffic generated by AODV-UU is
not considered in the total overlay management traffic. We show in Figure 3.4 results for
the total overlay management traffic per number of nodes for the three different scenarios. As expected, the overhead introduced by Bamboo increases with number of nodes,
and is much higher for standard timeout settings compared to no, and custom management. This is mainly due to the aggressiveness of periodic updates required by Bamboo
to monitor the status of other nodes in the overlay and update the overlay data structures. On the other hand, in the case of no management, each node does not generate
3.4. The Bamboo Overlay Maintenance Cost in WMNs
41
periodic updates, but neighbor ping is still performed in order to maintain reachability
and consequently RTT estimation.
1000
160
no
custom
standard
140
no
custom
standard
120
800
Overhead (%)
Total overlay traffic (KBytes)
1200
600
400
100
80
60
40
200
20
0
0
4
9
16
25
Number of nodes
36
49
4
9
16
25
36
Number of nodes
49
Figure 3.4: Impact of Bamboo management Figure 3.5: Percentage of management overtraffic
head compared to user traffic
Figure 3.5 presents the percentage of management overhead compared to the total
user traffic (get, getreply and put messages) in bytes generated over simulation time. It can
be seen that the periodic management traffic used in standard and custom management
imposes a high overhead for the mesh network, and it increases as larger topologies are
used. However, custom management has a smaller impact compared to the standard
management. For no management, the overhead remains constant for different number
of nodes, as every node generates get, put and also neighbor ping messages.
Lookup completed represents the percentage of lookups that are eventually delivered
to the correct responsible node in the overlay. When the number of nodes increases,
network load increases (Figure 3.5) and the number of lookups completed decreases according to Figure 3.6. In the 36 nodes grid, the number of lookups completed is 0.61,
0.41 and 0.19, respectively for no, custom and standard management. The low amount
of lookup completed for higher number of nodes can be explained by the higher percentage of management and routing overhead in order to maintain the overlay structure,
as shown in Figure 3.5 and Table 3.2. The ability to find the destination nodes which
are responsible for the specific keys degrades as management overhead increases network
contention. This results in higher number of packet retransmission and consequently
packet dropped due to network congestion and problems in the routing layer. This can
be seen from Table 3.3, where we analyze the reasons for packet drop in the 36 nodes
topology. In standard management case, most of the dropped packets are due to failure on route establishment (TOUT, TTL and NRTE) which consequently leads to queue
overflow (IFQ). In summary, the percentage of dropped routing traffic are 16.93%, 4.93%
and 1.53%, respectively for standard, custom, and no management.
In order to analyze the performance of the management strategies in relation to
lookup efficiency, we present in Figure 3.7 the cumulative distribution function of success
lookup delay for the 36 nodes topology. The lookup delay is defined as the time elapsed
42
Adapting the P2P Overlay to WMNs
no
custom
standard
1
Probability (lookup delay > x)
Lookup completed
1
0.8
0.6
0.4
0.2
0
0.9
0.8
0.7
0.6
0.5
0.4
no
custom
standard
0.3
0.2
4
9
16
25
36
Number of nodes
49
1
2
3
4
5
6
Lookup delay (s)
7
8
Figure 3.6: Percentage of lookups completed Figure 3.7: CDF of lookup delay in the 36
nodes topology
Table 3.2: Average routing traffic received per node in the 36 nodes topology
Routing traffic (bps)
NO
Custom
Standard
7875.67
6981.16
6322.87
when a node issue a get request in the overlay until the time it receives the get reply message informing about the responsible node of the requested key. The CDF shows that
custom and standard management strategies maintain lower lookup delay performance
for successful gets if compared to no management. It also shows that some lookups take
longer to complete for custom than for standard management. This is due to the aggressiveness of standard management using frequent periodic exchange of leafset table,
routing table and data storage information. This results in fewer hops on average for resource lookups in the overlay. In contrast, for no management scenario, the number of
nodes performing forwarding operations of put messages is higher as no leafset and routTable 3.3: Number of dropped packets and correspondent reason in the 36 nodes topology
Route TTL (TTL)
Route timeout (TOUT)
No route (NRTE)
Node queue full (IFQ)
Total dropped - pkts
Total dropped - %
NO
Custom
Standard
32
186
421
49
688
1.53
32
551
1190
264
2037
4.93
32
3826
2699
694
7251
16.93
3.5. The Benefits of Location Awareness DHT and Resource Replication
43
ing table information are exchanged among nodes. This results in more hops on average
(for example 3.70, 3.85, and 4.85 hops on average, respectively for standard, custom, and
no management in the 36 nodes topology).
In summary, the results presented in this section showed that the lookup efficiency,
translated as the ability to find destination nodes which are responsible for the specific
key information, can be severely degraded as the P2P management overhead increases network contention. Too frequent management traffic led to high overhead in the wireless
mesh environment and thus network contention. On the other hand, no management led
to low lookup efficiency as the lack of information updates among neighbors’ node cause
suboptimal routes to be selected through the overlay. Therefore, a trade-off is suggested
between P2P lookup efficiency and management traffic overhead. It is worth to note that
the settings proposed in the trade-off approach is target to a mesh scenario, where the
goal was to analyze the impact of different settings in the P2P overlay performance and
not to propose operational settings for different scenarios and topologies.
In the evaluation conducted, the churn process is not realized through join/leave
node’s process, but through link quality variability as the wireless links are subjected
to higher contention as management traffic and network topology increase. Thus, in the
scenarios studied, a temporary loss of routing information weakens the correctness and
performance guarantees of the DHT. Node mobility can also augment the churn process,
however it is node considered in our evaluation. We foresee that in case node mobility
is added to the scenario, lower performance of the P2P lookup would be achieved as frequent updates at the overlay in order to cope with the mobility would increase drastically
the management overhead. Given the results obtained, we explore in the next sections
possible adaptations of the P2P overlay in wireless mesh networks in order to improve
resource lookups.
3.5
The Benefits of Location Awareness DHT and Resource Replication
In P2P systems over WMNs, the lookup time is directly proportional to the network
size [1, 12]. As illustrated in Section 3.4, the characteristics of the underlying wireless
mesh network has a great effect on the performance of the overlay as the DHT induces
a constant flow of overlay management and query messages. This is especially the case
when deployed over wireless mesh networks as the lookup message needs to contend
for the medium at each physical hop. Therefore, reducing the total number of physical
hops traveled directly impacts on the achievable performance. Interesting ideas such as
location-aware overlay and resource replication have been proposed in order to reduce
the amount of traversed hops by overlay messages. We present in this section a performance evaluation and comparison study of two P2P resource management approaches in
wireless mesh scenarios, and investigate the benefits of location-awareness and resource
replication while considering the two protocols Bamboo and Georoy.
44
3.5.1
Adapting the P2P Overlay to WMNs
Georoy DHT
The Georoy algorithm [51] is a location-aware variant of the Viceroy algorithm [49],
where a unit ring topology is combined with a butterfly network topology guaranteeing
a lookup performance of O(log n), where n is the number of nodes in the network.
The main target of Georoy is to build an overlay network that can provide accurate
and efficient resource lookup in wireless scenarios, supporting either node mobility and
resource adding or removing. Using a geographic aware hash function, Georoy targets
smaller routing stretch, giving tight relationship between the physical hops traversed during resource lookup and those that would have been traveled going directly to the final
destination using a simple minimum hop count routing.
As a main difference with Bamboo and other well known DHTs such as Chord, is that
Georoy does not use a flat node topology, but employs a two level hierarchy with two
different kinds of nodes: Leaf Peers (LP) which share and request resources by querying
their associated super peers and Super Peers (SP) which provide the distributed resource
catalog and are used by LPs to publish and request resources.
Following a wireless mesh architecture, SPs are mesh routers which are placed in the
network and do not move, while LPs are mesh clients that can move and stay connected
via a handoff mechanism like in cellular networks. In Georoy, the DHT is managed only
by SPs which are also responsible for the overlay construction and maintenance; so the
IDs in the DHT are assigned only to these nodes. When a LP wants to share a resource
it must associate this resource with a key provided by its SP according to a distributed
hash function. Resource IDs are mapped in the same ID space of SPs, i.e. [0,1] so, both
the SPs ID space and the resource keys space are mapped in the same interval [0,1]. Each
resource key is managed by the SP with the smallest ID larger than the key ID so that
each SP is responsible of all IDs between its own one and the one of its predecessor.
In order to provide geographic awareness, a mapping function is proposed which gives
a SP an ID depending on its physical x and y coordinates, e.g. obtained through GPS. To
explain this function we assume that nodes are deployed in a square region of side s , so
all SPs are located in R = [0, s)2 . The mapping function M is defined as follows:
y
+ ∆ ∆s
s2
šy
(s −x)∆
+ ∆ ∆s
s2
( x∆
M (x, y) =
š 
if
if
šy
š ∆y 
∆
is even
is odd
(3.1)
with 0 < ∆ < s.
As in Viceroy’s, four distinct levels are utilized in order to form the butterfly topology. When a node joins the network, it first computes its ID using the function described
in Equation 3.1. Then it chooses a level at random and joins the ring through lookup
predecessor and successor forming unit-ring links. It also establishes level-ring links in
order to form a virtual bi-directional ring between the peers at the same level. After
that, butterfly links are established by maintaining upward and downward connections
among the levels. Summarizing, every peer in the Georoy overlay network has at most
seven outgoing links; two unit rings, two level-ring links, and three butterfly links.
3.5. The Benefits of Location Awareness DHT and Resource Replication
3.5.2
45
Location-aware versus Location-unaware DHT
Most of the DHT solutions, such as Bamboo, are location-unaware DHTs, as they employ a logical overlay which is independent of the existing physical network. Thus, in
general, even if two nodes are physically close in the underlaying wireless network, they
can be far away in the overlay. This can lead to a problem when such overlays are deployed over resource scarce wireless networks such as wireless mesh networks. In such
scenarios, it is crucial to minimize the number of physical hops as this directly impacts
on the achievable delay and packet loss. As we have seen in Section 3.4, Bamboo tries to
minimize the routing stretch by using the proximity neighbor selection scheme through
global routing updates. But the results have shown that at larger topologies such transparent layered approach does not scale due to the mismatch between the overlay and
underlaying topology.
However, there are DHT variants, called location-aware DHTs, that build the overlay
based on physical topology information such as physical node positioning or neighborhood information. Examples of location-aware overlays are Georoy [51], MADPastry
[100], and MeshChord [92]. As discussed in Section 3.5.1, in order to provide a locationaware overlay, a mapping function M is proposed by Georoy which gives to the peers an
identification (ID) that depends on its physical x and y coordinates.
3.5.3
Resource Replication
In WMNs, P2P resources replication can be beneficial as it allows the possibility to balance the network traffic among the nodes. Thus, in case replication is used, the availability of a given resource is not mandated anymore just by the availability of the node
responsible for storing the given resource (here called responsible node), as multiple replicas are present in the network. Thus, if the responsible node crashes, the resource will
still be available through the replicas. In fact, when more copies of a resource are available in the network, it is expected that the resource can be located in the closer proximity
of the requesting node, thus reducing significantly the download time and resource consumption in wireless mesh networks.
In Bamboo, a replication mechanism is already incorporated. Basically, a node holding a given resource also caches it within some of its leafset neighbors. This is done
according to a number of desired replicas. To this purpose, put messages are generated by
the responsible node to select peers among its successors and predecessors in the leafset.
For example, if the desired number of replicas is set to 4, the responsible node generates
4 put messages destined to 2 random successors and 2 random predecessors, achieving
a total of 5 resource copies in the network. The increase of replication comes with the
cost on the amount of overhead in the network. The maximum amount of replicas is
given by the total number of nodes in the leafset (number of successors and predecessors). This means that in the default scenario where the number of leafsets is configured
to 7, a maximum of 15 copies of the resource will be available in the network.
When a lookup for the replicated resource is generated, it will be forwarded through
46
Adapting the P2P Overlay to WMNs
the overlay as usual. If the resource is available at one of the nodes traversed along the path
to the responsible node, this node replies to the lookup before reaching the responsible
node. In case of high load at the responsible node and no replica along the path, the
responsible node can also forward the lookup to another node holding the desired replica.
In case an existing node leaves the system, it takes the data it has stored with it. Therefore,
the redundancy given by the replication strategy also guarantees that the resource will be
still available in the remaining leafset neighbors. As described in Section 3.4.1, in order
to keep the distributed storage consistent, data storage updates are applied by Bamboo
periodically in order to synchronize the stored data within the node’s leafset. For certain
applications, the number of desired replicas can cause large demands for storage space.
This can turn into serious scalability problems when disseminating these replicas to many
nodes in the leafset.
Similar to Bamboo, Georoy also apply replication mechanism in order to increase
resource availability. However, in Georoy, it is the LP who decides on the degree of
resource replication. Different from Bamboo that selects randomly the replica placement
among its leafset, in Georoy replication is done along the visited SP nodes. Therefore,
for every replicated resource at the visited SP, an update at the home SP managing the
range of keys to which the resource belongs is required. Similar to Bamboo, the home
SP (or responsible node in Bamboo) manages the list of all SP where the given resource
is replicated. Resource lookups are forwarded to the home SP, who owns the list of SP
nodes that have the requested resource. Based on the ID of the node who issued the
lookup, the home SP answers with the ID of closest SP to the requester, as closer ID in
the logical space represent also closer physical location. In case the lookup request passes
a visited SP that holds the desired resource, the lookup is answered before reaching the
home SP.
As for Bamboo, the replication in Georoy implies an increase in the rate of availability
of a resource in the network but can cause an increase also in the overhead at network
nodes. Thus, in Georoy, a mechanism to control the number of replicas in the network
is also considered, where older copies of a resource are deleted after a time out so that a
maximum number of replicas for a given resource can be found into the network.
3.5.4
Experimental Setup and Simulation Results
In this section we compare the performance of the two protocols, Bamboo and Georoy
in different conditions. We want to better understand their behavior by means of a comparison using two significant scenarios representing the static backhaul of wireless nodes:
grid and random scenarios. Here, a number of static SP nodes are placed within an area
of a certain size during the simulations. We vary the number of such SP nodes between
25 and 225. As in Section 3.4, the simulations were performed using ns-2, where the
simulation area size ℜ depends on the number of SP nodes (ℜ = NS P · 104 m 2 ). Routing
between the connected nodes uses AODV-UU [113], but different choices are possible.
In the random topology, nodes are thrown randomly in the area. We consider infinite
storage space on the replication nodes. We make such choice since if the storage size
3.5. The Benefits of Location Awareness DHT and Resource Replication
Number of physical hops
Number of logical hops
25
Bamboo
Georoy
10
8
6
4
2
0
47
Bamboo
Georoy
20
15
10
5
0
25
36
49
64
81 100 121 144 169 196 225
Number of nodes
25
36
49
64
81 100 121 144 169 196 225
Number of nodes
Figure 3.8: Comparison between the num- Figure 3.9: Comparison between the number of logical hops in Georoy and Bamboo ber of physical hops in Georoy and Bamboo
in a grid topology.
in a grid topology.
is limited, achievable performance may largely depend on cache replacement strategies,
which is outside the scope of this study. In the random topology case, for each scenario
identified by the number of nodes, we tested 5 different random topologies and for each
topology we performed 100 random lookups. Average values and confidence intervals
when applicable were reported for the following performance metrics being investigated:
• number of logical hops traveled in the overlay network to perform a lookup for a
specific resource
• corresponding number of physical hops traveled in the physical network to perform a lookup for a specific resource as a consequence of the logical path followed
• lookup delay needed for the lookup to reach the node who stores information
about the requested resource. We only consider here correctly completed lookups.
• percentage of lookups correctly completed.
• routing stretch factor, i.e. the ratio between the number of physical hops needed to
complete the lookup as a consequence of the logical hops traversed and the number
of physical hops going end-to-end according to a shortest path approach.
We start the evaluation by focusing on the impact of the network size on the scalability of the lookup procedure. We then evaluate the impact of the replication technique.
Impact of network size in grid and random topologies
In Figure 3.8 we show the number of logical hops traveled when employing the two algorithms. We make use of Bamboo custom management configuration settings described
48
Adapting the P2P Overlay to WMNs
6
Bamboo
Georoy
Average delay [s]
5
4
3
2
1
0
25
36
49
64
81 100 121 144 169 196 225
Number of nodes
Figure 3.10: Comparison between the average lookup delay in Georoy and Bamboo in a
grid topology.
in Section 3.4. By comparing the results we observe that, in general, Bamboo results in
a smaller number of logical hops as compared to Georoy. This is related to the fact that
the amount of overlay routing information used by Bamboo (i.e. leafset and routing table) is higher if compared to Georoy, which limits the number of existing logical links to
7. In addition, by having the possibility to acquired better overlay information through
its routing table and data storage updates, Bamboo increase its probability to reach the
requested resource through a smaller number of logical hops as it makes better progress
while routing packets in the overlay.
It is also important to consider the number of physical hops, as it impacts directly on
the lookup performance. This is because it determines the number of forwarding operations a packet needs to undergo in the wireless mesh network to reach the destination
(i.e. the node holding the resource). Figure 3.9 shows that as the network size grows, the
number of physical hops needed to complete a lookup increases. However, an interesting
observation is that for topologies equal or greater than 169 nodes, the number of physical
hops is in general lower when using Georoy as compared to Bamboo. This is due to the
overlay addressing scheme in Georoy where the logical and physical topologies are tightly
coupled so that the logical path does not differ much from the physical one. In fact, for
large network topologies, the ratio between the physical and logical hops is around 2 for
Georoy and rises to 5 for Bamboo.
Since the formation of the overlay network is independent of the physical location
of the nodes in Bamboo, for larger topologies the probability that a peer selects a close
logical neighbor located far away in the physical topology is higher. This results in longer
routes when topologies are larger. Also, note that the standard deviation on the number
of physical hops is smaller in Georoy compared to Bamboo for larger topologies. This
is again due to the addressing scheme of Bamboo, which randomly selects its ID in the
3.5. The Benefits of Location Awareness DHT and Resource Replication
Bamboo
Georoy
1
Lookup Completed
49
0.8
0.6
0.4
0.2
0
25
36
49
64
81 100 121 144 169 196 225
Number of Nodes
Figure 3.11: Comparison between the percentage of lookups completed in Georoy and
Bamboo in a grid topology.
overlay identifier space.
In wireless mesh networks, the more hops a packet is forwarded, the larger the delay
and, in general, the higher the packet loss probability. This is because at every intermediate node, the packet needs to compete for medium access and collisions due to e.g.
hidden nodes might lead to frequent re-transmissions and consequently high packet loss.
The impact of an increase in the number of physical hops traveled in case of large topologies can be seen in the average lookup delay comparison shown in Figure 3.10. Here, we
can see that for smaller topologies, Bamboo outperforms Georoy as less physical hops
are required. However, due to the location-awareness addressing scheme, the increase
in the number of physical hops is smaller for larger topologies in Georoy, compared to
Bamboo. Therefore, Georoy provides better lookup delays with larger topologies. Interestingly, Georoy shows smaller number of physical hops as compared to Bamboo when
network size is equal or larger than 169 nodes. However, the lookup delay of Georoy is
smaller as compared to Bamboo already at a network size of about 121 nodes. The explanation to this result is related to the increasing overhead imposed by Bamboo to maintain
the overlay topology. Even by having a smaller number of physical hops compared to
Georoy , the amount of traffic generated by the Bamboo overlay maintenance increases
the network contention and consequently the average lookup delay.
Another interesting observation is that the number of successfully completed lookups
decreases as network size increases (see Figure 3.11). By increasing the number of nodes
in the network we also increase the amount of messages exchanged (management traffic required to maintain the overlay plus key lookup requests and replies) among the
nodes and consequently the wireless contention for the medium. Also, when lookup
packets traverse more hops, they need to compete more often for medium access and the
probability to collide due to e.g. hidden nodes is higher. Interestingly, the number of
50
Adapting the P2P Overlay to WMNs
Bamboo
Georoy
5
Routing stretch
4
3
2
1
0
25
36
49
64
81 100 121 144 169 196 225
Number of nodes
Figure 3.12: Comparison between the routing stretch factor in Georoy and Bamboo in a
grid topology.
completed lookup requests is smaller for Bamboo as compared to Georoy, even for small
topologies. This can be attributed to the fact that the management traffic of Bamboo is
significantly higher. Such high management traffic leads to more load and contention
leading to higher chance that the lookup request cannot be completed correctly [12]. In
Bamboo, in this case the lookup request can be re-transmitted by each node up to a limit
of five times until the agent gives up and declares the request as not successful.
The routing stretch factor presented in Figure 3.12 shows that both protocols can
satisfy lookup requests with a limited increase in the number of hops traversed when
compared to the shortest path approach. As we have seen in Figure 3.9, Georoy needs
fewer hops to forward a lookup request to the destination when the network is composed
of 169 nodes or more. Consequently, the routing stretch factor of Georoy is smaller
compared to Bamboo at large network sizes.
When considering the random topologies, similar conclusions can be drawn. However observe that, in the random case, nodes are not distributed on the vertices of a grid,
so physical proximity can help to reduce the number of physical hops (see Figure 3.14)
and thus decrease the average lookup delay as seen in Figure 3.15. In addition, due to
the random nature of the node location, we could observe more clustering of nodes as
compared to a grid scenario. Therefore, as nodes are more close to each other in most of
the area, less physical hops are required, thus implying less delay to complete the lookup
operation. Clearly, due to the randomness in node location, there is more variability in
the number of physical hops and delay. However, Figure 3.13 shows that the logical hops
instead do not vary much as compared to the grid scenario.
3.5. The Benefits of Location Awareness DHT and Resource Replication
Number of physical hops
Number of logical hops
25
Bamboo
Georoy
10
8
6
4
2
0
51
Bamboo
Georoy
20
15
10
5
0
25
36
49
64
81 100 121 144 169 196 225
Number of nodes
25
36
49
64
81 100 121 144 169 196 225
Number of nodes
Figure 3.13: Comparison between the num- Figure 3.14: Comparison between the number of logical hops in Georoy and Bamboo ber of physical hops in Georoy and Bamboo
for random topologies.
for random topologies.
Impact of number of replication for grid topologies
Besides the impact of network size in grid and random topologies, another important
point that we address is to determine the benefit of using a replication mechanisms in
wireless mesh scenarios. We start by looking at the impact of having different number
of replicas to speed up the resource lookup process, as more replicas might increase the
probability to find the requested resource among the neighboring peers easily and consequently reduce the resource lookup time in the network.
In our experiments we considered that both in Georoy and Bamboo each resource
was replicated at 3, 5 or 7 different nodes. We compared with the case where no resource
replication is used and thus just one copy is available. We assume the replicas of the
resource are randomly distributed in Georoy (independently of the LP movement) and
are assigned to random nodes in the leafset in Bamboo. For example, in the case of 3
copies, the responsible node and two random selected nodes among its leafset table are
selected to hold the given resource.
In Figures 3.16 and 3.17 we observe that there is a decrease in the number of logical
and physical hops when replication (3, 5, or 7 copies) is used in comparison with no replication (1 copy). This is because, when increasing the number of replicas, the probability
of finding the resources closer raises as well. The reason why Bamboo benefits more
from the resource replication if compared to Georoy is related to Bamboo’s data storage
updates, which frequently synchronize the stored keys among leafset nodes. Thus, when
using more replicas, the delay to complete a resource lookup can be reduced as seen in
Figure 3.18. Also, consider that in Bamboo there is almost no variation in the number
of logical hops between 3, 5 and 7 copies. The reason for this behavior is the replication mechanism, which in Bamboo disseminates replicas randomly at nodes in the leafset
which are thus very close in the logical space .
52
Adapting the P2P Overlay to WMNs
6
Bamboo
Georoy
Average delay [s]
5
4
3
2
1
0
25
36
49
64
81 100 121 144 169 196 225
Number of nodes
Figure 3.15: Comparison between the average lookup delay in Georoy and Bamboo for
random topologies.
In summary, we considered in this section two efficient P2P schemes for wireless
mesh networks and enhanced them by introducing procedures to allow increasing scalability and reliability by using multiple replicas of the same resource in the network.
Performance results were aimed at comparing the performance of the two algorithms
(Bamboo and Georoy) in wireless mesh networks. For all scenarios studied we have seen
the impact of the number of logical and physical hops on the delay performance of resource lookups. For both protocol, as the network topology increases, the amount of
physical hops required during the forwarding process of resource lookup also increase
impacting the lookup delay.
While both P2P schemes use a ring topology to maintain the overlay neighbor consistent, they apply different routing table strategies in order
lookup perfor€ to improve
Š
mance. In Bamboo, the routing tables of size l o g2b n × 2 b − 1 3 is populated within
the peers sharing a common prefix, and lookups are based on longest prefix matching.
Differently, Georoy applies a butterfly topology where the routing table holds a limited
number of seven outgoing links to the chosen long range peers. We have seen through the
results, that the routing table strategy choice impacts the number of logical hops required
by each lookup. In Bamboo much smaller number of logical hops are achieved due to the
higher number of routing table information available and the frequent updates.
However, we see the benefit of location-awareness introduced by Georoy when analyzing the number of physical hops results. By avoiding longer physical paths while
routing resource lookup through the overlay, Georoy can achieve smaller lookup delays
in case of larger grid topologies. Moreover, we have also seen the benefits of using multiple resource replicas by allowing shorter physical hops and consequently lower lookup
delays to both P2P schemes. In Bamboo, this benefit is larger while compared to Georoy
3 The
configuration parameter b is set to two and n represents the number of nodes in the network.
20
1 copy
3 copies
5 copies
7 copies
Logical and physical number of hops
Logical and physical number of hops
3.6. Reducing Routing Stretch Through Neighborhood and Cache Information
15
10
5
0
20
53
1 copy
3 copies
5 copies
7 copies
15
10
5
0
Logical hops
Physical hops
Number of nodes
Logical hops
Physical hops
Number of nodes
Figure 3.16: Number of logical and physical Figure 3.17: Number of logical and physical
hops in Bamboo in a grid topology with 225 hops in Georoy in a grid topology with 225
nodes.
nodes.
as data storage updates increase the probability to find the requested resource among the
neighboring peers. For the scenarios studied we assume that the content of the resource
available in the network is not modified along the simulation. The amount of resource
change and required consistency in the network is dependent on the application and not
part of the scope of this section.
3.6
Reducing Routing Stretch Through Neighborhood
and Cache Information
Commonly, DHTs make use of routing table information to accelerate lookups. In Bamboo, the routing table is filled with overlay neighbors information divided into different
levels according to its proximity in the overlay identifier space. Overlay neighbors that
are far in the overlay, also known as long range neighbors, are used to quickly route messages to remote locations in the identifier space. As seen in Sections 3.4 and 3.5 given
the limited bandwidth constraint of WMNs, the maintenance of overlay neighbors can
be prohibitively heavy-weight as they could be located several hops away in the physical
wireless topology.
The upper part of Figure 3.19 presents an example of the Bamboo DHT formation
and lookup request process for a simple wireless mesh network composed by eight nodes.
We present the physical wireless topology in the middle part of the figure, where the links
among the nodes are indicated by arrows connecting the nodes. We represent the ID
number of nodes in the overlay by three binary digits, which correspond to an identifier
space of 23 , 8 nodes. In the left part of the figure, we represent the ring topology created
by Bamboo, where the overlay neighbors of node 000 are shown (red arrows represent
two leafset and gray arrows two long range neighbors). In this example, we assume that
54
Adapting the P2P Overlay to WMNs
4.5
Bamboo
Georoy
4
Average delay [s]
3.5
3
2.5
2
1.5
1
0.5
0
1 copy
3 copies
5 copies
7 copies
Number of nodes
Figure 3.18: Average lookup delay for Georoy and Bamboo in a grid topology with 225
nodes.
node 000 generates a key lookup to a resource that is located at node 101. As illustrated in
the right part of Figure 3.19, in order to reach the destination node 101, the key lookup
needs to pass through node 100, as it is the long prefix match node seen by 000. In this
example, the number of logical hops necessary to achieve the destination is two, while
the number of physical hops is six, as the resource lookup generated by node 000 needs
to be forwarded by node 100 (located 3 hops away in the physical topology from node
000), who forwards it to node 101 (also located 3 hops away in the physical topology
from node 100).
We clearly see in the example that better overlay neighbors are available in case physical neighbors are used in the overlay’s routing table. Thus, we start by proposing the
one-hop DHT approach where each node only needs to maintain its leafset and makes use
of the one-hop physical neighbors in order to augment the lookup process. By maintaining the leafset, we guarantee correctness of the routing process and limited management
overhead. However, the lookup requests rely mainly on the leafsets and physical neighbors. When receiving a lookup request, the node checks among its physical neighbors if
there is a closer node to the requested key compared to its leafset information. If so, this
node will become the next destination of the lookup request. We call this a shortcut link.
If not, the request is forwarded to the next node in the logical path. As a result, at every
step the lookup request gets closer to the node responsible for the key, until this node is
reached.
The lower part of Figure 3.19 presents the modifications done to the DHT in order
to achieve the one-hop DHT approach. As we can see in the example, by using one-hop
physical neighbors at node 000, we can find the destination node 101 within its overlay
neighbors. Thus, reducing the number of logical and physical hops to one. Compared
to the previous example where four (unidirectional) wireless links are used in order to
3.6. Reducing Routing Stretch Through Neighborhood and Cache Information
000
111
55
000
000
001
111
001
011
110
110
010
101
010
110
111
101
011
100
010
101
001
100
100
Key lookup
DHT
000
111
011
000
000
001
111
001
011
110
110
010
101
010
110
111
101
011
100
One-hop DHT
100
010
001
101
011
100
One-hop key lookup
Figure 3.19: Example of the one-hop DHT neighborhood and lookup request. Node
000 is requesting a resource located at node 101. The overlay neighborhood, network
topology and key lookup process are represented in the left, middle and right part of the
figure respectively.
perform the resource lookup, in this example we reduce the number of wireless links
used to one and consequently the amount of traffic contending for channel access inside
the network.
Having more and diverse possibilities to select shortcut links can greatly improve the
lookup efficiency of the overlay. Thus we propose one extension of the above approach
where the visibility of a node can be increased by two-hop neighborhood information,
which we denote as the neighbor-of-neighbor approach (NoN). In wireless networks,
neighbor-of-neighbor information can be achieved via hello (beacon) messages, ameliorating then the lookup efficiency while routing it across the network. As it occurs in the
one-hop approach, when routing the request to a destination via the overlay, the node
resorts to simple greedy routing by selecting the physical neighbor among its neighbors
and neighbor-of-neighbors that makes the most progress in the overlay.
It is straight forward to note that as we increase the visibility of nodes, for example
by enabling the use of n-hop information, better resource lookups can be achieved. The
best case would be the scenario where each node holds the complete network topology
information while performing lookups. However, the cost involved for such case are
prohibitive as the network overhead increases while trying to achieve n-hop neighborhood information. But, we can exploit the lookup request history through the use of a
56
Adapting the P2P Overlay to WMNs
cache scheme without increasing the network overhead costs. As the wireless nodes may
participate in the forwarding process of several lookup requests while been part of the
overlay, it is possible to use such information to augment the ability to find even closer
next logical hops for a specific key lookup request than its one-hop physical neighbors
or neighbor-of-neighbor information. We name this proposal standard cache, where each
node maintains a cache that stores for each observed lookup request the searched key and
the next logical hop neighbor of the lookup request. We assume that such approach is
beneficial at scenarios with low mobility, such as wireless mesh networks, as the lookup
request history is expected to be valid over a longer period of time compared to high
mobile scenarios such as MANETs.
For a given resource lookup load, we expect that the probability of a node to forward
resource requests can decrease as the network becomes dense. In such scenarios, by just
caching the lookup requests that pass through a node may not be sufficient to achieve
better lookup performance. Thus, we propose the use of the cross-layer cache approach
that also takes into account the advantage of the one hop broadcast communication in
the wireless access medium. By overhearing key lookup requests at the MAC layer, a
node can further enrich its cache entries. This solution is particularly adapted to WMNs
as no additional transmissions are required. This however comes at an increased energy
cost, since nodes that constantly overhear all transmissions can not remain in the idle
power-low state if necessary.
3.6.1
Experimental Setup and Simulation Results
To analyze the benefits of neighborhood and cache information in static wireless mesh
network scenarios, the simulator in this section mimics a simple Bamboo DHT overlay
where no management traffic is present. Different number of node density and nodes
are used in the evaluation. The node density in our evaluation represents the number of
one-hop physical neighbors that each node has in the network. For scenarios composed
by 500 and 1000 nodes, we construct network topologies were the number of one-hop
physical neighbors at each node varies between 10, 13, 22 and 40. Two nodes are considered physical neighbors if their distance is smaller than the transmission range limit of
100 meters.
The simulation executes a warm-up phase to reach a steady state prior to issuing the
key lookups. In the warm-up phase, each node generates 20 lookups to randomly selected keys. We use standard AODV-UU as the routing protocol responsible to route key
lookup requests between logical neighbors inside the WMN.
We start this section by comparing the two proposed caching schemes; standard cache
which just account for request lookup information forwarded by a node, and the crosslayer cache, which accounts for the request lookup information forwarded by the node
and also the overheard lookups at its MAC layer.
Figure 3.20 motivates the use of cross-layer cache in comparison to the standard cache.
In this figure we plot the average cache size per node (log scale) versus the node density,
for 500 and 1000 nodes topology and assuming the same amount of resource lookups. It
3.6. Reducing Routing Stretch Through Neighborhood and Cache Information
57
is important to observe that we used infinite cache size per node in order to evaluate the
maximum cache population that a node can acquire in such scenarios. It can be noted that
by just caching lookups which pass through the node (standard cache) does not guarantee
a high number of cache entries. Moreover, the number of entries for the standard cache
decreases for higher network densities (e.g. 22 and 40 node density) as the number of
logical nodes traversed by a key lookup reduces in denser networks. In contrast to that,
by taking the advantage of the wireless broadcast communication, the cross-layer caching
enables nodes with the possibility to overhear key lookup packets and therefore acquire
a higher average cache size per node.
Given those results, we continue our evaluation by comparing the performance benefit of the cross-layer caching scheme together with the one-hop DHT and neighbor-ofneighbor approaches. For the cross-layer cache scheme, we consider two cases: nodes
with limited cache size of 256 entries (Cache) and nodes with infinite cache size (InfinityCache). The limited cache size is implemented based on the Least Recent Used (LRU)
scheme. Routing messages could also be used to enrich the cross-layer caching scheme,
however we do not consider it since we decide to design it independently of the routing
protocol used. The performance metric under investigation is the DHT routing stretch
of key lookups.
10000
one-hop
NoN
NoN+Cache
NoN+InfinityCache
3.5
Routing stretch
Average cache size (log)
4
500 standard
1000 standard
500 cross-layer
1000 cross-layer
1000
3
2.5
2
1.5
1
100
~10
~13
~22
Node density
~40
~10
~13
~22
Node density
~40
Figure 3.20: Average cache size versus node Figure 3.21: Routing stretch versus node
density for standard and cross-layer cache density for 1000 nodes
schemes using 500 and 1000 nodes
Figure 3.21 presents the average routing stretch of four approaches as a function of
node density. The one-hop approach where the nodes in the overlay maintain leafset
and one-hop physical neighbors in order to augment the lookup process. The NoN
approach which enhances the one-hop approach by allowing also neighbor-of-neighbor
information. And two NoN variations making using of cross-layer caching scheme; the
NoN+Cache which uses a LRU cache of size 256, and the NoN+InfinityCache which
uses an infinite cache size. Each result was obtained by averaging over 100 random
key lookups performed after the warm-up period. As expected, the use of neighborof-neighbor (NoN) information compared with the one-hop approach ameliorates the
58
Adapting the P2P Overlay to WMNs
DHT routing stretch, specially in higher density topologies. This is because via neighborof-neighbor information each node can acquire a greater view of the network topology
and therefore select better next logical neighbors. We can see that those results are similar
to the results for larger topologies illustrated in Figure 3.12.
By using caching size of 256 entries allied to neighbor-of-neighbor information (NoN
+ Cache256), the DHT routing stretch can be reduced even further. In this case, the
knowledge acquired by the cache during the warm-up phase and the overheard key lookups
help in the selection of physical nodes closer to the destination, while routing through
the logical neighbors. One problem with limited cache is that the cache may contain
many entries which are not necessarily significant, as the cache entries may not be evenly
mapped along the logical space. By using LRU scheme we assume that the cache entries
are updated according to the newest lookups, however it does not guarantee an optimal
mapping of the logical space as the entries acquired are not evenly distributed in the logical space.
We also verified that using an infinite cache size leads to a DHT routing stretch of
approximately one, which means that the selected route via the overlay is as efficient as
the shortest path route. This is however not surprising, since using an infinite cache and
infinite lifetime for entries means that after enough lookup requests a node will store all
possible node identifiers, thus having the full network membership.
In summary, we have presented in this section an envisioned DHT implementation
which combines a minimal Bamboo DHT structure with a cross-layer cache scheme.
By replacing long range logical neighbors by physical neighbors we limit the overlay
overhead as lookup requests rely mainly on the physical neighbors.
We also see that in case the cross-layer cache scheme is utilized, the possibility to select
better shortcut links increases. However, we foresee that the benefit of caching might
be degraded in high mobility scenarios. Apart from LRU, to look for different cache
management is an interesting aspect as it might help to improve the lookup efficiency and
consequently reduce its average path cost, however such improvements are not considered
in this section.
3.7
Related Work
In this section we discuss the related literature in the field of P2P systems and their adaptations to wireless mesh network scenarios. As we briefly discussed in Chapter 2, several
published works have considered the problem of enabling P2P resource sharing in wireless multi-hop scenarios involving MANETs and wireless mesh networks. The proposed
approaches normally implement interactions between the P2P overlay and the routing
layer by proposing cross-layer or integrated solutions (as seen in Section 2.3), complemented by location and caching information.
The use of location information such as node position and landmarking have been
used as a way to create tight relationship between the logical and the physical topologies.
In MeshChord, an specialization of Chord applied to wireless mesh networks, where
3.7. Related Work
59
the availability of a wireless infrastructure, and the one-hop broadcast nature of wireless
communication are taken into account while performing key lookup [92]. MeshChord
explores location awareness by assigning identifiers to peers according to their physical coordinates through the use of GPS information. As we have also seen through the
Georoy’s results shown in Section 3.5, the use of location awareness tends to scale the
lookup operations under large network scenarios as it rules out the possibility of having
close-by peers in the physical network which are far-away in the overlay.
Targeting also the benefit of location information while constructing the overlay,
MADPastry [100], which integrates Pastry and AODV, proposes the concept of landmark to create physical clusters where nodes share a common overlay identifier [101].
Nodes associate themselves with a landmark node that is currently closest to them (e.g.
as determined by the hop count) by adopting its overlay ID prefix. For that purpose,
landmark nodes periodically send out beacons propagated within the landmark’s own
cluster, i.e. beacons are only forwarded by nodes belonging to that cluster. Hence, if
the requested resource is not found in physical vicinity (inside cluster area), intra-cluster
communication is established via landmark nodes. We emphasize that such optimization
might be useful for popular resources hosted by multiple nodes, as it enables the possibility to find the resource locally. Otherwise, regular network-wide lookups are necessary.
Meanwhile, caching strategies have also been used to augment routing information at
the overlay layer. For example, by overhearing lookup request packets at the MAC layer,
MeshChord can speed up the resource lookup phase as more information is available
during the overlay routing decision. In wireless environments, such caching strategy
is commonly achieved by overhearing at the MAC layer transmitted packets that are
destined to other nodes.
Following the DSR routing concept, in SSR data packets carry routing path information in terms of source address, destination address and source route. By constructing the
route cache, each SSR node contains source routes to the node’s neighbors in the virtual
ring. Besides that, the caches will contain source routes to other destinations also. All
nodes making part of a source route in the cache can be viewed as potential destinations.
When routing a packet, the respective node chooses the (intermediate) destination from
its cache that is physically closest to itself and virtually closest to the final destination of
the packet, by appending the source route from its cache to the packet’s header.
By applying the route cache concept, SSR reduces the routing stretch in the network,
as nodes use their routing caches to prune unnecessarily long source routes ( e.g. routes
containing cycles) or to find a shorter sub-path (short cut) to one of the nodes in the
source route. However, as discussed by [100] and also seen in our results in Section 3.6, to
achieve maximum efficiency while using caching strategies depends upon the availability
of cache entries.
60
3.8
Adapting the P2P Overlay to WMNs
Conclusion
In this chapter we have outlined several ways to adapt P2P overlays to wireless mesh
networks in order to improve the resource lookup process. We first introduce the main
challenges of deploying structured P2P overlays in wireless mesh networks. A packet
level performance analysis of Bamboo overlay maintenance over wireless mesh networks
is conducted. Based on our findings we identify a trade-off between resource lookup
efficiency and management traffic overhead. For the scenarios analyzed, we propose a
recommendation on the overlay configuration settings.
While targeting at interactions among P2P and WMNs, we investigate the benefits
of location-awareness and resource replication by comparing Bamboo and Georoy over
wireless mesh scenarios. We also introduce enhancements to those overlay structures by
adding procedures that allow increasing scalability and reliability by the use of multiple
replicas of the same resource in the network. Furthermore, we have also shown that for
networks with medium to high density, the use of neighborhood information from the
routing layer increases the performance of P2P overlays at WMNs as the overlay routing table is augmented with neighbor and neighbor-of-neighbor information. If lookup
resource history and enough caching is exploited, optimal routing stretch is achieved.
3.9
Validity of the Results and Limitations
We conclude this chapter by discussing the validity of the obtained results and the limitations of the studies presented in the previous sections. Section 3.5 shows the benefits
of applying resource replication in wireless mesh environments. We assume that the
resources are equally important among the nodes. Its is known that for common P2P
applications such as file sharing, the likelihood of equally important resources is limited.
However, we do not explore in this thesis the possible interactions between the resource
popularity and replication mechanisms. We foresee that depending on the scenario envisioned, such interaction would contribute even further to the scalability of resource
availability in wireless mesh scenarios if compared to the homogeneous resource popularity. It is also important to note that the scenarios studied assume that the P2P resource
are located inside the wireless mesh network. As we are going to see in Chapter 4, in case
we also assume resource in the Internet, aspects such as number of gateways available
need to be considered.
Regarding the cross-layer caching proposal, we further remark that in case of scenarios with constrained energy resources, the option to always overhear resource lookups at
the MAC layer to augment the overlay routing table might be inappropriate. For such
scenarios, a trade-off between energy efficiency and routing stretch is important, but not
considered in this study. Moreover, for certain realizations of multi-hop networks such
as MANETs, mobility is an intrinsic characteristic. The node mobility represents an additional problem as it causes frequent join and leave process at the overlay topology and
also out-of-date cache information. However, we emphasize that our scenarios studied
3.9. Validity of the Results and Limitations
61
are mostly target to wireless mesh networks, where nodes involved in the overlay are
quasi-stationary and therefore the churn process is mostly related to wireless link variability through for example interference and hidden nodes, or mesh nodes been powered
up and down. However, the later assumption is not considered here.
Chapter 4
Peer Selection Problem
Formulation in WMNs
In this chapter we focus our attention on the resource exchange phase while deploying
P2P services in wireless mesh networks. We have seen in Chapter 2, that the P2P overlay
adaptation to wireless mesh networks contributes to scalable P2P resource lookup. Given
that the resource lookup is completed and a list of peers participating in the swarm process is available, the resource exchange phase takes place as the peers need to identify the
best set of peers to download the requested resource from.
Therefore, we study in this chapter the peer selection problem formulation in the context of resource replication and channel assignment for wireless mesh networks. Given
the set of peers involved in the download process, the peer selection problem is to identify
the best peers which guarantee the highest throughput and fairness, based on the channel
assignment and replication strategy chosen.
Resource replication has been an effective approach for reducing the overhead and
increasing the accessibility and performance of distributed systems. When deployed over
WMNs, the replication of an object (e.g. movie, music file or a service that is in digital
format) can reduce communication overhead in the network as the object can be retrieved
from a set of peers (replicated candidates) and not just from a single peer. Moreover, the
availability of multiple replicas in the network proportionates the use of load balancing,
as multiple paths to the requested object are available, and also the possibility to find
objects located within a smaller number of hops.
It is important to note that the peer selection problem is also directly related to the
channel assignment and routing problems. To solve the peer selection problem requires
the knowledge of the end-to-end available bandwidth between the peers, which depends
on the selected routing and channel assignment strategy. In addition, solving the routing
problem requires also a solution for the channel assignment, as the channel assignment
algorithm determines the set of links sharing the same channel and accordingly their
64
Peer Selection Problem Formulation in WMNs
available link bandwidth. In the same way, solving the channel assignment requires the
expected users’ download rates on the network links, which is determined by the routing
and peer selection. Added to that, the change of the peer selection results in different
paths been selected and consequently different link loads, which might require different
channel assignment to optimize for capacity. Thus, peer selection, routing and channel
assignment are closely inter-dependent problems that should be jointly solved.
In this chapter we evaluate the interdependency of peer selection, replica placement
and channel assignment approaches for multi-channel multi-radio mesh networks. We
start by presenting in Section 4.1 the contributions of this chapter. We present in Section 4.2 our network model and the proposed link capacity formulation inspired by the
collision domain model. Then we introduce in Section 4.3 the two channel assignments
and routing algorithm studied. As gateways are an important element to bridge the wireless mesh network with the internet, we study two resource replication strategies; at
gateways and at random nodes in the wireless network. In Section 4.4, four peer selection approaches are formulated such that throughput maximization and fairness goals are
achieved. In Section 4.5 we evaluate our proposal through numerical experiments where
the achievable throughput and fairness together with the benefit of increasing the number of channels and radios are analyzed. Finally, we present the related work, draw our
conclusions and outline the limitations of our work.
4.1
Our Contribution
The first contribution of this chapter is on the study of the interdependency of peer
selection, replica placement and channel assignment approaches for multi-channel multiradio wireless mesh networks. We make use of the collision domain model in order to
derive the wireless link capacity. K-Partition and BFS channel assignments are applied in
order to assign channels to the network interfaces of the wireless mesh routers.
We formulate the peer selection problem with a set of linear equations, which allow
to estimate available capacity for a given peer selection, channel assignment and routing approach. Our second contribution of this chapter is on the formulation of four
peer selection schemes targeting throughput maximization and fairness among the users’
download rate. We also evaluate the benefits of replicating the resources at gateways or at
random in the network. While the collision domain model has been applied before to estimate achievable capacity in a mesh network, this work is the first attempt to extend this
concept to include peer selection and replication strategies for multi-channel multi-radio
WMNs. Furthermore, based on four peer selection schemes, two channel assignment
algorithms and two replication strategies, we draw numerical results on the achievable
throughput and fairness and analyze the benefits of increasing the number of channels
and network interface cards (NICs) in WMN scenarios.
4.2. Network Model and Link Capacity Formulation
4.2
65
Network Model and Link Capacity Formulation
In this section we present our network model. We also describe the link capacity formulation based on an extension of the collision domain model for multi-channel multi-radio
wireless mesh network. Several assumptions about the network model and link capacity formulation are made in order to study the peer selection problem in the context of
wireless mesh networks.
4.2.1
Network Model
We consider a multi-channel multi-radio wireless mesh network composed of quasi stationary wireless mesh routers that provides connectivity to mobile clients within their
coverage range. The mesh routers form a mesh network among themselves in order to
relay traffic to and from mesh clients.
Some of the mesh routers can serve as gateways between the WMN and the internet.
The mesh routers are equipped with multiple NICs, which are tuned to a particular
channel. We assume that a channel assignment algorithm determines which radio is tuned
to which channel. For two nodes to communicate, they need to share a common channel.
The channel assignment and routing strategy used are described in Section 4.3.
We model the WMN as an undirected graph G = (V , E). V represents the set of mesh
routers. E represents the set of links, where l (u, v) ∈ E is the link between nodes u and
v such that u and v are within the transmission range of each other and there is at least
one common channel to both. We assume symmetric connectivity, such that l (u, v) ∈ E
if and only if l (v, u) ∈ E. We also assume that a set of orthogonal channels and multiple
NICs per node are available in the network.
The download rate function xi , f , j indicates that peer i is downloading a file f from
peer j . The path between peers i and j participating in the peer selection problem is
denoted by pi , j .
4.2.2
Modeling Link Capacity
In order to derive the capacity of multi-channel multi-radio WMNs, we need to derive the
capacity of each link taking into account interference from parallel packet transmissions
of nodes nearby.
We make use of the concept of collision domains (CDs) introduced by [38] and later
on generalized by [39]. Basically, the collision domain C D u,v of a link l (u, v) is the set
of all the links operating on the same channel which can not be active in parallel to link
l (u, v), as the interference from a transmission on links l (s , t ) would be strong enough
to disturb a parallel transmission on link l (u, v).
According to [38], in case an asymmetric MAC protocol is considered e.g. CSMA/CA, the direction of the links are taken into account when determining the set
of links participating in the collision domain. For example, a transmission on link l (s , t )
blocks the successful transmission on link l (u, v) if link l (s, v) or l (t , u) exist. Thus, for
66
Peer Selection Problem Formulation in WMNs
a successful transmission on link l (u, v), all one hop neighbors of v must not transmit
and all one hop neighbor of u must not receive. In case of a symmetric MAC protocol
e.g. RTS/CTS mechanism of IEEE 802.11, both ends of a transmission are protected, and
a transmission on link l (s , t ) prevents the successful transmission on link l (u, v) in case
at least one of the links l (u, s), l (u, t ), l (v, s) or l (v, t ) exists. It is important to note
that the definition of collision domain does not include the case where small interferences from multiple parallel transmissions sum up and might together disturb a success
transmission on a given link.
The nominal load of a collision domain corresponds to the number of transmission
that take place in the collision domain. We define in Equation 4.1, the number of transmissions n u,v on a given link l (u, v) as the number of all paths for which link l (u, v) is
part of the path pi , j .
n u,v = | { (i , j ) : l (u, v) ∈ pi , j } |
(4.1)
Thus, the number of transmissions in the collision domain C D u,v is given by m u,v ,
the sum of the number of transmissions on each link l (s, t ) which makes part of the
collision domain, represented by Equation 4.2.
X
m u,v =
n s ,t
(4.2)
s,t ∈ C D u,v
Assuming that all links have nominal MAC throughput B, the capacity of the whole
collision domain is also B, as it is composed of all interfering links which thus share the
capacity. The throughput of any download rate traversing at least one link of the collision
domain C D u,v is given by B/m u,v . The collision domain which carries the maximum
load (or forwards the maximum amount of traffic) is known as the bottleneck collision
domain [38]. In general, a network might have several bottleneck collision domains.
In order to better illustrate the concept of the collision domain model, we depict a
simple wireless mesh network scenario in Figure 4.1 composed by eight nodes, where all
links are operating on the same channel. The collision domain of link l (n1, n2), illustrated by the shadow area in Figure 4.1, is composed by five links {l (n2, n3), l (n2, n4),
l (n3, n5), l (n4, n7)) and also l (n1, n2)}. We assume that three peers (n1,n6, and n7) are
involved in the download process of a file f , and the peer selection problem enforces
nodes n6 and n7 to download the file from node n1.
Assuming single path routing, the number of transmissions in the collision domain
of link l (n1, n2), mn1,n2 , is six, which represents the sum of the number of download
rates traversing each link in the collision domain. Here, the download rate function
x(n1, f , n7) traverses three links (l (n1, n2), l (n2, n4), and l (n4, n7)) and the download
rate function x(n1, f , n6) traverses also three links (l (n1, n2), l (n2, n3), and l (n3, n5)).
Since the total load in the collision domain of link l (n1, n2) can at most be the nominal
MAC layer capacity B, we have 3 … xn1, f ,n7 + 3 … xn1, f ,n6 ¶ B as one of the inequalities
generated by the collision domain model.
Note that, this is not necessarily the bottleneck collision domain of the network.
Therefore the above calculation needs to be done for every link in order to find the bot-
4.3. Routing and Channel Assignment in Multi-Channel Multi-Radio WMNs
67
file
n1
n2
P1,9
n4
P1,6
n3
n5
n7
n8
n6
Figure 4.1: Collision domain for link l (1, 2) using single channel WMN
tleneck collision domain. The outcome of the collision domain model is a set of inequalities limiting the achievable capacity. Given that the bottleneck collision domain imposes
the highest load in the network, it can serve to derive the strongest bound on achievable
throughput and thus on nominal network capacity.
4.3
Routing and Channel Assignment in Multi-Channel
Multi-Radio WMNs
We assume that the set of peers involved in the P2P download process is obtained by
the P2P overlay network, as discussed in Section 3.2. The lookup operation by a peer i
requesting a file f returns a set of peers j involved in the download process of such file.
Thus, we assume for each triple (i, f , j ), peer i downloading file f from peer j , that an
end-to-end path is established via the routing protocol. We make use of the Dijkstra’s
algorithm in order to find paths with shortest hop count. Multiple shortest paths might
be available. If so, the routing algorithm selects randomly one shortest path from the list.
Given the set of shortest paths, the channel assignment is executed in order to ensure
that the nodes along the selected paths share a common channel. As discussed in [79],
joint routing and channel assignment strategy is important as uncoordinated decision in
one layer might impact on the other. We assume that the channel assignment algorithm
68
Peer Selection Problem Formulation in WMNs
maintains the network connectivity by ensuring that two neighbors, which need to communicate (i.e. whose link needs to carry traffic) share a common channel.
In order to perform the channel assignment, we assume knowledge of the selected
paths chosen by the routing algorithm. However, at this point, there is no knowledge
about the amount of traffic, or download rates, that can be sustained over the selected
paths. As we are going to discuss in Section 4.4, this is an outcome of the optimization
problem.
We make use of two different channel assignments, K-Partition [70], and BFS [72].
Single channel assignment is also used as a baseline approach to show the advantage of
multi-channel multi-radio in WMN. As in [70], we must ensure that both channel assignments need to satisfy four constraints:
1. The number of distinct channels that can be assigned to a mesh router is limited by
the number of NICs on it,
2. Two mesh routers involved in a link that is expected to carry traffic should be
bound to a common channel,
3. The sum of the expected loads on the links that interfere with one another and that
are assigned to the same channel cannot exceed the nominal MAC layer capacity B,
and
4. The total number of available channels is fixed.
The constraints 1, 2 and 4 are captured by the channel assignment formulation. As
described in Section 4.2.2, the constraint 3 is ensured by the collision domain model.
Following [70], the channel assignment problem is sub-divided into two problems; the
neighbor-to-interface binding and the interface-to-channel binding problems.
In neighbor-to-interface binding, for each node we divide all of its neighbors’ nodes
within its transmission range into k-groups. Here, the number of groups k is given
by the number of available NICs, where each NIC on a node is thus mapped to one
group. Thereafter, each of this neighbors’ nodes also partitions its neighbors’ nodes into
k-groups, while maintaining the grouping done by the first node as a constraint. This
process is repeated until all nodes have partitioned their neighbors. By doing that we ensure that the first constraint is satisfied. Thus, each node has made the decision by which
interfaces it is going to communicate with its neighbors.
In the interface-to-channel binding, channels are assigned to interfaces. First, the interfaces of nodes are divided into groups to ensure the second constraint. Then, each
group is allocated channels making sure that constraint four is not violated. Apart from
constraint 3, all the other constraints need to be fulfilled in order to apply the K-Partition
and BFS channel assignments.
The K-Partition channel assignment does not take into account the path selected by
the routing algorithm. Thus, in the neighbor-to-interface binding, neighbors are bound
in a sequential fashion. In the interface-to-channel binding, we start by allocating the
4.3. Routing and Channel Assignment in Multi-Channel Multi-Radio WMNs
n2
69
n2
P2,3
NIC2
n4
NIC2
NIC1
n1
n3
n5
Figure 4.2: K-Partition channel assignment
NIC1
n1
n4
P4,3
n3
P5,3
n5
Figure 4.3: BFS channel assignment
channels sequentially until the total number of channels is reached and then each interface
is allocated a channel that is least used in its two hop neighborhood. However, as in [70],
the assignment does not take into consideration the traffic load of the links, and thus
assumes that the links in the network have similar amount of load. One simple example
of K-Partition channel assignment is illustrated in Figure 4.2, where in the neighbor-tointerface binding, the neighbors of node 1 are bound sequentially to the two NICs (N I C1
and N I C2 ). Since two channels are available in this example, channel green (straight line)
and red (dashed line), the interface-to-channel binding allocates the two channels to the
two NICs sequentially.
Differently from K-Partition, the BFS channel assignment takes into account the
paths selected by the routing algorithm in order to assign channels to the NICs. Thus, in
order to apply BFS, the set of paths between the peers are assumed to be known a priori.
Another difference to the K-Partition is that in BFS the nodes are traversed in a breadth
first search fashion starting at a pseudo node connected to all source nodes holding a
file. In the neighbor-to-interface binding, neighbors are bound to the interface which is
supporting the least number of download rates. The idea behind this scheme is that each
group mapped to an interface would carry the minimal amount of download rates. But at
this point, the actual download rates are not known, thus the neighbors are bound to the
interface just by considering the number of download rates. In the interface-to-channel
binding, the groups are allocated channels that have the least number of download rates
in the two hop neighborhood.
We illustrate the BFS scheme in Figure 4.3, where three download rates are active in
the network, corresponding to the paths p2,3 , p4,3 , and p5,3 . In this example, node 3 is
70
Peer Selection Problem Formulation in WMNs
the seed holding the file and therefore the BFS channel assignment starts at this node.
Since node 3 has just one physical neighbor (node 1), it selects randomly channel green
to the link with this neighbor. Following the breadth first search fashion, the node 1 is
the next in the channel assignment process. As its N I C1 is already bound to node 3, the
neighbor-to-interface binding forces the other nodes to be bound to N I C2 as it presents
the smaller number of paths. In the interface-to-channel binding, given that N I C1 is
bound to channel green, node 1 make use of channel red at N I C2 , since it represents the
channel with the least number of download rates in its two hop neighborhood.
4.4
Peer Selection Problem Formulation
We assume that the network is composed by a set of peers N and a given set of files F . The
download rate function xi f j is used to represent the download rate at which peer i downloads the file f from peer j . The routing protocol described in Section 4.3 is responsible
for establishing end-to-end paths pi, j between peers i and j involved in the download
rate xi f j . We also assume that the channel assignment algorithms described in Section
4.3 allocate channels to radio interfaces of each node. In case of resource replication, we
assume that the requested file f ( f ∈ F ) can be located at different peers j .
By making those assumptions, we intend to evaluate the interdependency of peer selection, replica placement, routing and channel assignment approaches for multi-channel
multi-radio wireless mesh networks. For the set of download rate functions, the problem is to estimate the maximum throughput and fairness for different peer selection and
replica placement strategies in a wireless mesh network scenario. The peer selection
problem is represented by a set of feasible download rate functions xi f j subject to the
upcoming constraints, and it is the output of the linear programming model.
It is important to note that in this formulation the file f is downloaded from peers j
that holds the entire file. In Chapter 5, we extend this model by allowing the file f to be
divided into smaller pieces, also known as chunks, and we extend it later to allow multiple
chunks to be downloaded simultaneously. Generally, P2P systems implement the chunkbased peer selection scheme, however our formulation proposed in this section is the first
step towards the understanding of the interdependency between peer selection, channel
assignment and routing protocol on the capacity modeling of P2P systems in wireless
mesh networks.
Given that routing paths pi , j are necessary in order to obtain valid download rate
functions xi f j , and the number of transmissions in the collision domain represent the
set of paths traversing each link in the collision domain (as seen in Section 4.2.2), we need
to satisfy the following constraints.
xi f j ≥ 0, ∀i, j ∈ N , ∀ f ∈ F
(4.3)
4.4. Peer Selection Problem Formulation
X
71
n s,t … xi f j ≤ B ∀i , j ∈ N , ∀ f ∈ F , ∀l (u, v) ∈ E
(4.4)
s ,t ∈ C D u,v
Constraint (4.3) guarantees non negative download rate functions. Constraint (4.4)
guarantees that the total download rate in any collision domain does not exceed the nominal MAC capacity. Solving an optimization problem leads then to rate allocations chosen
by the peer selection scheme that can be feasibly scheduled given the routing and channel
assignment and replication strategies.
By making those assumptions we can now formulate the objective function of the
peer selection problem as a set of linear functions of the download rates. Four peer
selection objectives are studied. As we are going to see, besides the basic constraints (4.3)
and (4.4), different objectives lead to additional specific constraints.
4.4.1
Max Rate Allocation
The max rate allocation (MRA) selects the peers for which the sum of user’s download
rates achieved is maximum. Thus, by MRA we can estimate the maximum throughput
achievable. In MRA our objective is:
M a x i mi z e
X
xi f j
(4.5)
i , j ∈N , f ∈F
subject to constraints (4.3) and (4.4). As this approach tries to maximize the aggregated throughput it can result in unfair rate allocations ending up with zero rates for a
few peers.
4.4.2
Minimum Guaranteed Maximum Rate Allocation
The minimum guaranteed maximum rate allocation (MGMRA) tries to achieve a degree
of fairness by finding the maximum aggregated throughput while putting a lower bound
xG on the achievable download rates. In MGMRA our objective is:
M a x i mi z e
X
xi f j
(4.6)
i, j ∈N , f ∈F
xi f j ≥ xG ∀i, j ∈ N , ∀ f ∈ F
(4.7)
Besides the constraints (4.3) and (4.4) we need also to satisfy constraint (4.7). Setting xG = 0 translates to MRA. A higher xG value increases fairness but might result in
allocations that may not be schedulable for a given channel assignment and routing.
72
4.4.3
Peer Selection Problem Formulation in WMNs
Maximum of Minimum Rate Allocation
The maximum of minimum rate allocation (MMRA) selects the peers for which the minimum among the download rates is as high as possible. MMRA is to some extent a fair
selection scheme as it ensures none of the peers are deprived. In MMRA our objective is:
M a x i mi z e x mi n
(4.8)
xi f j ≥ x mi n ∀i, j ∈ N , ∀ f ∈ F
(4.9)
Besides the constraints (4.3) and (4.4), we need also to satisfy constraint (4.9). The
peer selection for which x mi n is the maximum, is the one selected by MMRA. Despite
targeting to achieve the maximum of the minimum download rate, MMRA does not
guarantee complete fairness among all download rate functions.
4.4.4
Proportional Fairness
Assuring fairness may involve several layers, since unfairness occurs mainly in MAC (e.g.
channel access and scheduling) and transport layers (e.g. congestion control). Here, we
use a simple proportional fairness index λ = mi n(x)/ma x(x) ∈ [0, 1] proposed by [114],
where x is the set of download rates. When λ = 0, some user’s download rates are allowed
to starve, whereas for λ = 1 absolute fairness is enforced.
The proportional fairness objective is to select peers for which the throughput is maximum. Thus, in proportional fairness our objective is to:
M a x i mi z e
X
xi f j
(4.10)
i , j ∈N , f ∈F
xi f j ≥ x mi n ∀i, j ∈ N , ∀ f ∈ F
(4.11)
xi f j ≤ x ma x ∀i, j ∈ N , ∀ f ∈ F
(4.12)
x mi n ≥ λ mi n .x ma x
(4.13)
subject to constraints (4.3)-(4.4),(4.11)-(4.13). Proportional fairness ensures fairness in
relation with the maximum download rate by considering constraints (4.11), (4.12) and
(4.13). In constraints (4.11) and (4.12) the minimum and maximum download rates are
4.5. Numerical Results
73
known. Constraint (4.13) guarantees proportional fairness. In order to verify that absolute fairness is achieved, we make use of the Jain’s fairness index [115] in our evaluation,
which is given by the Equation 4.14.
J x1 , x2 , ...., xn
4.5
P 2
xi
=€ P Š
n … xi2
(4.14)
Numerical Results
We use Octave [116] and the GNU linear programming kit (GLPK) function to solve
the linear programming model. We assume nodes are always on and have infinite cache
to store the replicas. If nodes would have limited cache, the cache replacement strategy
has a major impact on achievable capacity which we did not evaluate in this work. The
topology of the WMN is a 7x7 grid of wireless mesh routers. We vary the number of
interface per-node from one to four radios. The number of orthogonal channels is varied
from 1 to 12. Channels are bound to radios during the whole download time according
to the channel assignment algorithm. The nodes are spaced by 100 meters having also a
transmission range of 100 meters.
Three nodes are randomly selected to be gateway nodes connecting the mesh to the
internet. In the model, wireless link capacity is set to 100 units, and is assumed to be constant. We make use of the term seed to define the peers who provide the resource file, and
leecher to define the peers that request the file during the download process. Six leechers
are randomly selected and divided into three sets of two leechers. Each set, compose by
two leechers, requests a different file from the set of files ( f1 , f2 , f3 ). The replica placement mandates which seed nodes hold which files. We considered two replication cases:
at gateway and at random nodes. In the gateway based scheme, when no replication is
used, there is just one file per gateway. The number of leechers and seeds is kept small as
the simulation time grows rapidly with their increase. However, such decision still allows
us to achieve concise results in order to analyze and compare the selected approaches. For
each combination of peer selection, replica placement, routing and channel assignment
algorithm, average and standard deviation of 5 runs are calculated, if not stated otherwise.
Impact of replication
Figures 4.4 and 4.5 show the aggregated throughput of user’s download rates in terms of
degree of replication for gateway and random based replication, respectively. Each of the
six leechers is looking for three distinct files f1 , f2 and f3 . In case of degree of replication
equal to 1 (x-axis), the three files are located at three different gateways, or three randomly
selected nodes. As we increase the amount of replication, additional replicas are spread
at the gateways or at additional random nodes. For example, at degree of replication
equal to 2.3, we have a total of 7 files in the network (e.g. 3 f1 , 2 f2 , and 2 f3 ), spread over
74
Peer Selection Problem Formulation in WMNs
MRA
MGMRA
600
600
C=1
K-Part
BFS
C=1
K-Part
BFS
500
500
400
400
400
400
300
300
Sum of Flow Rates
500
Sum of Flow Rates
500
Sum of Flow Rates
300
300
200
200
200
200
100
100
100
100
0
1
1.5
2
2.5
Degree of Replication
3
0
1
1.5
2
2.5
Degree of Replication
MRA
3
0
1
1.5
2
2.5
Degree of Replication
MGMRA
1.2
3
1
1.2
1.2
K-Part
BFS
C=1
K-Part
BFS
C=1
K-Part
BFS
1
1
0.8
0.8
0.8
0.8
Jain Index
1
Jain Index
1
0.6
0.6
0.6
0.6
0.4
0.4
0.4
0.4
0.2
0.2
0.2
0.2
0
0
1
1.5
2
2.5
Degree of Replication
3
0
1
1.5
2
2.5
Degree of Replication
3
3
Lambda=1
1.2
C=1
K-Part
BFS
1.5
2
2.5
Degree of Replication
MMRA
Jain Index
Sum of Flow Rates
Lambda=1
600
K-Part
BFS
0
Jain Index
MMRA
600
C=1
K-Part
BFS
0
1
1.5
2
2.5
Degree of Replication
3
1
1.5
2
2.5
Degree of Replication
3
Figure 4.4: Impact of replication for files located at the gateways, two NICs and six
channels.
3 gateways or at 7 different nodes for gateway and random replication, respectively. In
order to ensure a fair comparison, as degree of replication increases, additional seeds are
added rather than changing the complete set of seeds, also maintaining the same set of
leechers. At the lower part of Figures 4.4 and 4.5 we also present the Jain’s index for each
peer selection scheme. Here an index of one represents the maximum fairness achievable,
while zero represents total unfairness among leechers’ download rates. The use of multiple channels and multiple interfaces are presented, while using K-Partition and BFS. For
those graphs we make use of two interfaces and six channels, if not mentioned otherwise.
Analyzing Figure 4.4, we can see that for MRA with a single channel the throughput
increases with the degree of replication. However, in terms of fairness, we note that
single channel remains unfair even at higher degree of replication. When using single
channel, capacity is severely limited and the MRA will try to maximize download rates
for a small set of rates leaving not much remaining capacity for the other users, leading to
unfairness. Compared to single channel, K-Partition provides better throughput, leading
also to more fairness among user’s download rates. However, BFS provides the highest
throughput as it leads to the increase of individual download rates, which consequently
4.5. Numerical Results
75
MRA
MGMRA
600
600
C=1
K-Part
BFS
C=1
K-Part
BFS
500
500
400
400
400
400
300
300
Sum of Flow Rates
500
Sum of Flow Rates
500
Sum of Flow Rates
300
300
200
200
200
200
100
100
100
100
0
1
1.5
2
2.5
Degree of Replication
3
0
1
1.5
2
2.5
Degree of Replication
MRA
3
0
1
1.5
2
2.5
Degree of Replication
MGMRA
1.2
3
1
1.2
1.2
C=1
K-Part
BFS
C=1
K-Part
BFS
C=1
K-Part
BFS
1
1
0.8
0.8
0.8
0.8
Jain Index
1
Jain Index
1
0.6
0.6
0.6
0.6
0.4
0.4
0.4
0.4
0.2
0.2
0.2
0.2
0
0
1
1.5
2
2.5
Degree of Replication
3
0
1
1.5
2
2.5
Degree of Replication
3
3
Lambda=1
1.2
C=1
K-Part
BFS
1.5
2
2.5
Degree of Replication
MMRA
Jain Index
Sum of Flow Rates
Lambda=1
600
C=1
K-Part
BFS
0
Jain Index
MMRA
600
C=1
K-Part
BFS
0
1
1.5
2
2.5
Degree of Replication
3
1
1.5
2
2.5
Degree of Replication
3
Figure 4.5: Impact of replication for files located at the random nodes, two NICs and six
channels.
leads to more fairness if compared to K-Partition and single channel cases. This is mainly
due to nature of BFS, that during the channel assignment gives priority for links that
carry traffic.
For MGMRA, we set the minimum download rate to 10% of the maximal achieved
rate. MRA and MGMRA look very similar for BFS and K-Partition, as the minimum
guaranteed download rate of 10% can be achieved without the reduction of other rates.
What is interesting to note is that for MGMRA, the single channel assignment can not
guarantee the minimal download rate of 10%, therefore no results are plotted.
On the other hand, MMRA can guarantee a better degree of fairness among the download rates, even for single channel and lower degree of replication, if compared to MRA
and MGMRA. However, this comes at the expense of slightly smaller aggregated throughput, especially for K-Partition, which is not effective when allocating channels. By comparing the results of MRA with MMRA and proportional fairness for λ = 1, we can see
clearly the trade-off between fairness and maximum achievable throughput.
The same conclusion can be drawn for random replication scenario seen in Figure
4.5. However, we can note that for all peer selection and channel assignment schemes
76
Peer Selection Problem Formulation in WMNs
there is an improvement in throughput if compared to the replication at the gateways.
This is because by randomly selecting the replica placements instead of clustering the
replicas at the gateways, we increase the probability to find disjoint paths in the network
and consequently lowering the collision probability over the wireless links. As throughput increases, more fairness is provided to the download rates as seen by the Jain’s index.
Note, that for the degree of replication of 2.3 or higher at random location, and single
channel assignment, we can satisfy MGMRA, which is not possible for gateway replication, as shown in Figure 4.4, due to higher load around the gateway nodes. Moreover, for
the BFS channel assignment and degree of replication equals to three at random locations,
we can achieve the maximum throughput of 600 for some combinations of leechers and
seeds, as seen by the error bar. This is due to the fact that enough non-interfering disjoint
paths among the peers involved in the download process can be found.
Based on the the results presented, we can conclude that MRA peer selection at random nodes needs to be used if maximum throughput is expected. Moreover, for a degree
of replication greater than two, a reasonable amount of fairness is achieved among the
peers’ download rates. In order to guarantee fairness at lower degree of replication, the
MMRA should be preferred. We can also see that irrespectively of the peer selection and
replica placements, the BFS channel assignment is the one that provide highest network
throughput.
Impact of number of channels and radios
Figures 4.6(a) and 4.6(b) show the throughput in terms of the total number of channels
for the gateway and random replication, respectively. We just plot K-Partition for MRA,
however its behaviour is similar to BFS. We have used the same 7x7 grid topology with
each node having two NICs and two leechers for each of the three files, and files replicated
at three gateways or randomly.
We can see that for all repetitions the throughput increases with the total number
of channels but only to a certain number of channels after which it remains the same.
Adding more channels will not reduce the collisions as the number of interfaces is a
limiting factor.
The trend is the same for all peer selection schemes, however different maximum
throughput can be achieved. For MMRA and proportional fairness (Lambda=1), the
maximum achieved throughput is lower compared to MRA and MGMRA as fairness
among download rates needs to be guaranteed. K-Partition always provides lower throughput compared to BFS, as it tries to balance the channels among node’s interfaces without
considering the amount of download rates carried by each interface.
The optimum number of channels, i.e after which there is no increase in throughput,
is higher for BFS than K-Partition in both replication cases. This shows that as more
channels are made available, BFS makes better use of them by breaking down the bottleneck collision domain compared to K-Partition that distributes them evenly. What is
interesting to note is that for BFS and smaller number of channels (e.g number of channels equal to 2, 3 and 4) it is better to replicate at the gateways compared to random
4.5. Numerical Results
77
MRA
BFS(MGMRA,MMRA,Lambda=1)
600
BFS(MGMRA,MMRA,Lambda=1)
600
MGMRA
MMRA
Lambda=1
600
K-Part
BFS
MGMRA
MMRA
Lambda=1
500
500
400
400
400
400
300
300
Sum of Flow Rates
500
Sum of Flow Rates
500
Sum of Flow Rates
Sum of Flow Rates
MRA
600
K-Part
BFS
300
300
200
200
200
200
100
100
100
100
0
0
2
4
6
8
10
Number of Channels
0
12
2
4
6
8
10
Number of Channels
12
0
2
(a) Replication at the gateways
4
6
8
10
Number of Channels
12
2
4
6
8
10
Number of Channels
12
(b) Random replication
Figure 4.6: Throughput versus number of channels
MGMRA
MMRA
Lambda=1
600
500
500
500
500
400
400
400
400
300
200
300
200
100
100
100
100
2 radios
4 radios
0
12
2 radios
4 radios
0
2
4
6
8
10
Number of Channels
300
200
2 radios
4 radios
0
4
6
8
10
Number of Channels
300
200
2 radios
4 radios
2
Sum of Flow Rates
600
Sum of Flow Rates
600
Sum of Flow Rates
Sum of Flow Rates
MRA
600
12
0
2
4
6
8
10
Number of Channels
12
2
4
6
8
10
Number of Channels
12
Figure 4.7: Impact of number of radios for BFS and random replication
replication. This is due to the fact that the BFS channel assignment starts from a smaller
number of nodes (just at the gateways) meaning that better re-use of channels is possible
compared to random replication where the assignment starts at every random source.
Figure 4.7 presents the impact of number of radios as number of channels increases
for BFS and random replication. For all peer selection schemes the total throughput
increases with the number of radios. The optimum number of channels increases from
5 channels while using 2 radio interfaces to 6 channels while using 4 radio interfaces.
By having more radios and channels, nodes can break the collision domain by assigning
different set of channels and interfaces among its neighbors. MMRA and proportional
fairness have a much higher gain in throughput compared to MRA and MGMRA, due to
the increase in number of radios. Therefore, for the given scenario and using BFS, we can
guarantee a higher degree of fairness among users download rates if using 6 channels and
4 radio interfaces.
78
Peer Selection Problem Formulation in WMNs
Jain index
Sum of flow rates
600
500
400
300
200
100
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
C=1
K-Part
BFS
0
0.2
0.4
0.6
Lambda
0.8
1
Figure 4.8: Tradeoff between maximum throughput and fairness for the degree of replication of two at gateways, two NICs and six channels
Impact of lambda
As describe by [117], a fundamental tussle exist while calculating performance and fairness in P2P systems. Higher fairness can be achieved at the cost of worsening the throughput performance. However, this is important for protocol designers since the realization
of such trade-off needs to be considered according to the objective of the application.
In Figure 4.8 we show such trade-off by varying the proportional fairness factor λ. We
use the degree of replication of two at gateways, two NICs and a total of six channels are
used. As expected, higher λ leads to decrease in throughput for all channel assignment
schemes, due to the reduction of some download rates while enforcing fairness among
them. For all lambda values, the higher throughput is achieved by BFS channel assignment, followed by K-Partition and than single channel assignment.
From Figure 4.8, we can observe that the optimum performance strategy is given
by the highest sum of flow rates, but the lowest fairness index. On the other way, the
optimum fairness strategy is given by the highest Jain’s index, but with the lowest sum
of flow rates. It is interesting to note that for the grid topology studied, with λ = 0.5
a good trade-off is achieved with Jain’s index greater than 0.88 while still guaranteeing
high throughput, especially when using BFS channel assignment. However, the optimum lambda value of 0.5 can not be generalized, as we foresee that the trade-off between
performance and fairness in the system is directly related to the network topology used,
and should also be selected according to the objective of the application.
Normalized throughput
4.5. Numerical Results
79
Model
Simulation
Simulation RTS/CTS
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
1
1.5
2
2.5
3
Degree of replication
Figure 4.9: Theoretical and simulation results of the normalized throughput for different
degrees of replication at the gateways and single channel assignment
Model validation through simulation
To validate our assumptions, we compare the results obtained by the model with ns-2 simulation, using ns2.33 version [112]. The interference model implemented by ns-2 is based
on the capture threshold model, which assumes that a packet is successfully received if its
received signal power is larger than any other packet received simultaneously by at least
the capture threshold of 10 dB. Otherwise, collisions occurs and the packet is dropped.
For both simulation and model, we used a 5x5 grid wireless mesh topology and the 802.11
MAC layer with and without RTS/CTS in the simulator. The RTS/CTS mechanism of
IEEE 802.11 is used in order to achieve symmetric MAC by protecting both nodes involved in a link communication from collisions. The constant bit-rate (CBR) UDP flows
are used, as we do not model the TCP’s congestion control behavior. The basic data rate
is set to 6 Mbit/s and the packet size to 1470 bytes.
For this comparison, we used MRA as the peer selection scheme, nodes operating
in a single channel environment, and replication at the gateways. We make use of this
setup as it simplifies the comparison due to the use of backlogged CBR traffic and single
channel assigned at all links. However, the evaluation for different peer selection schemes,
replications and channel assignments are also possible.
Figure 4.9 compares the theoretical with simulation results for the normalized throughput and different degrees of replication at the gateways. Each point is an average of 10
runs normalized to the highest throughput (degree of replication equals to 3). For each
run, three random leechers are selected among the nodes (excluding the gateways), each
downloading a different file from its respective gateway. For the degree of replication
equal to one, each leecher downloads the file from the gateways holding its requested file.
For the degree of replication equal to three, each leecher has a set of three gateways which
it can choose to download from.
80
Peer Selection Problem Formulation in WMNs
In order to make this comparison, we feed the simulation with the set of leechers,
gateways, and shortest paths calculated by the model. As we can see in Figure 4.9, using
RTS/CTS at the MAC layer protects senders and receivers from collisions. This results is
a close match with the model as the collision domain model assumes that link constraints
are symmetric [38].
However, for asymmetric MAC layer (no RTS/CTS), the model does not match completely the simulation results if compared to the symmetric MAC case. The reason for
that is related to the higher number of active links 1 , as no protection of the transmitter
and the receiver in a link is enforced by the RTS/CTS mechanism. The higher number
of active links impacts the throughput results in two ways. For lower degree of replication, this represents a higher amount of collisions as the traffic is directed to few gateways increasing the number of active links that interfere within each other due to hidden
node problems. However, for higher degree of replication, the active links are disjoint
as more gateways are available, and therefore higher throughput is achieved compared to
the model. Here, the mutual interference of active links are not high enough to cause
collision, and therefore higher throughput can be achieved.
4.6
Related Work
A variety of algorithms and solutions for WMNs have been proposed under the scope
of optimization problems, ranging from gateway placement [118] to multi-layer joint
optimization techniques [119].
Inside the scope of capacity modeling, several works such as [38, 39, 119, 120, 121]
make use of the collision domain model in order to derive the nominal capacity of wireless links. Using genetic algorithms and the collision domain model formulation for
WMNs, [120] focuses on WMN planning process by optimizing the WMN performance
in order to achieve a max-min fair throughput allocation. In [119, 121], mixed integer
linear problem (MILP) formulations are proposed to derive link capacity while guaranteeing max-min fairness objectives. However, those works do not considered multi-channel
multi-radio wireless mesh network scenarios, and therefore the interactions between the
peer selection, routing and channel assignment are not taken into consideration while
solving the optimization problem.
The most related to our work is [122], which proposes a linear programing model
to calculate the aggregated throughput for P2P systems over WMNs. Differently from
[122], which considers only single radio approaches and assumes a perfect schedulingbased MAC using time division multiple access (TDMA), in our work we model the
network using the collision domain concept where channel access is shared among mobile routers on a given channel in a CSMA/CA fashion. In addition to that, we also
study the interaction between peer selection, routing and channel assignment in terms of
throughput and fairness while using multiple channels and multiple radios. Therefore,
1 By
idle.
active links we mean links that are transmitting at the same time, as the transmitters sense the medium
4.7. Conclusion
81
our model is more related to current deployed wireless mesh networks, as it considers
the carrier sense multiple access technology and make use of multi-channel multi-radio
systems.
4.7
Conclusion
In this chapter, we have developed a numerical model to calculate capacity and fairness
for the peer selection problem in multi-channel multi-radio WMN. Based on the collision
domain concept we formulated different peer selection strategies, MRA, MGRA, MMRA
and proportional fairness as a constrained linear program to derive the users’ download
rates. The model is flexible, allowing the evaluation in arbitrary topologies, for different
number of NICs and channels. We evaluated the impact of different channel assignment
schemes, number of radios and channels. Given the set of routing, channel assignment
and peer selection scheme, the model is capable of deriving the users’ download rate and
fairness amongst user flows, for resources replicated at gateways or at random nodes.
We have shown that for lower degrees of replication, random replication is better than
replication at gateway nodes, if enough channels are available. This is because the traffic
can be spread out more effectively in the network as more links can be used simultaneously. For single channel assignment schemes, MGMRA might result in infeasible rates
due to the limited network capacity. While MRA allows to achieve the highest aggregated
throughput, it results in limited fairness. On the other hand, MMRA and proportional
fairness are more fair but suffer in terms of aggregated throughput. The effect of total
number of channels on throughput is also studied. The optimal number of channels can
be identified for a given number of radios, beyond which the throughput cannot be increased by adding more channels. According to the results presented, the optimal number
of channels depends on the channel assignment, peer selection, and number of network
interface cards available. We further show that our model is closely related to the simulation model in case symmetric links are enforced. Moreover, the results show that an
important trade-off exists between throughput and fairness in P2P systems deployed over
WMNs.
4.8
Validity of the Results and Limitations
We conclude this chapter by discussing the validity of the problem formulation and obtained results together with the limitations of the work carried out in this chapter. In
Section 4.2 we describe the collision domain model used to derive the link capacity in
wireless mesh networks. In the formulation we assume that all links have the same nominal MAC throughput, which dictates the capacity of the collision domains. We do not
explore the use of different link data rates by considering the availability of different modulation and coding schemes (MCSs). However, as described in [119], applying different
link rates might contribute even further to the completeness of the collision domain
model formulation. It is also important to note that the definition of collision domain
82
Peer Selection Problem Formulation in WMNs
does not include the case where small interferences from multiple parallel transmissions
sum up and might together disturb a successful transmission.
In Section 4.3, the K-Partition and BFS channel assignments and the Dijkstra’s routing algorithm are considered. We further remark that the use of shortest hop count
paths and the independence of routing and channel assignment in our work are simplified assumptions and better strategies might be available (e.g. jointly considering channel
assignment and routing, and the use of advanced routing metrics). However, we need to
remark that our primarily focus was on the impact of peer selection, replica placement
and channel assignment. The necessity to extend this study to a larger set of topologies
is also an important aspect to be considered. As we will seen in Chapter 5, the use of
different topologies result in different absolute performance values, but similar trends are
presented while analyzing the impact of peer selection and channel assignment.
We foresee that the presented model can be easily extended to more sophisticated
routing and channel assignment schemes. It is also important to note that in our peer
selection problem formulation we do not consider the possibility of peers uploading parts
of its content to other peers while still downloading the file. This possibility is enabled in
Chapter 5, where the file is divided into chunks and the peers involved in the download
process can mutually download chunks from each other. Finally, at this chapter we do not
consider the ACI problem in the channel assignment formulation. This will be further
addressed in Chapter 6.
Chapter 5
Extending the Peer Selection
Problem for Multiple Chunks
By allowing peers to upload and download chunks of the resource among themselves
simultaneously, P2P systems such as the BitTorrent file sharing [31] can proportionate
faster resource dissemination and consequently reduce the network impact of resource
distribution.
While in the previous chapter we have considered that only peers holding the entire
resource file can upload to the others, we now consider a more complicated case where the
resource is slit into multiple chunks, and each chunk can be downloaded independently
by peers while targeting the download of the whole resource. We aim to build a model
which allows us to estimate the minimal time required to disseminate a resource among a
set of peers given the network constraints imposed by the channel assignment and routing
protocol.
Important insights on the achievable throughput and download time given by the
network constraints are derived by the mathematical models. For example, by taking
into account information about the bottleneck links in the link capacity formulation
and the use of the BFS channel assignment, we have achieved considerable performance
improvements for the P2P download process at multi-channel multi-radio WMNs, as
shown in Chapter 4.
Moreover, we have also seen in Section 4.5 that the aggregate capacity and performance of WMNs can be increased by using multiple channels. As a limited number of
channels is available, some links may be assigned to the same channel and might interfere with each other if located close-by. It is known that interference among neighboring
links can reduce their effective performance and potentially cause network congestion.
A typical approach to improve performance in WMN is to exploit the path diversity
in the network [61], as it allows mesh routers to perform load balancing among the
multiple available paths and therefore increase throughput and fairness among network
84
Extending the Peer Selection Problem for Multiple Chunks
flows. WMNs typically provide several paths from source to destination, and in case such
paths are used efficiently more network resources can be available to the users. From the
users point of view, by allowing multiple paths to the P2P applications we can increase
end-to-end download rates and also provide robustness against performance fluctuation
of any single path in the network.
In order to further work on the design of interactions between the peer selection
routing and channel assignment, we also perform packet-level simulations in this chapter
in order to study the P2P download problem in multi-channel multi-radio WMNs. As
discussed in [123], the use of packet-level simulation is an important tool while analyzing
the performance of P2P systems, as it allows us to study the influence of the lower level
protocols and interaction among layers.
Given those considerations, we start the chapter by first giving an outline of our
contributions. Next, we presented the chunk-based peer selection problem formulation
and its numerical results. The simulation studies using BitTorrent are also presented and
a new peer selection algorithm is proposed. Finally, we present the related work to this
chapter, followed by conclusions and limitations of our work.
5.1
Our Contribution
The first contribution of this chapter is the extension of the P2P download problem by incorporating the availability of multiple chunks, and the possibility of peers to upload and
download chunks among themselves simultaneously. This translate into a chunk-based
peer selection scheduling problem, which we model as a MILP given the constraints imposed by the network capacity. We make use of the makespan, the minimal time required
to disseminate a resource file among all peers involved in the download process, where
the main goal is to find the optimum makespan for the chunk-based peer selection problem. While such chunk-based peer selection scheduling exists [124], we are the first who
considers channel assignment and routing protocol in the network constraints formulation target to multi-channel multi-radio wireless mesh networks. Through the numerical
results presented in Section 5.2.2 we can analyze the impact of the chunk-based formulation together with channel assignment algorithms, and compare it with the previous peer
selection model proposed in Chapter 4.
The other contribution of this chapter is the further design on the interactions between peer selection, routing and channel assignment layers through simulation studies. Based on those simulations, we design a novel peer selection algorithm which incorporates a path load metric, multi-path capability, and channel reassignment along the
download process. We compare the proposed peer selection algorithm with BitTorrent
through simulation while using different network topologies. Our results indicates that
the BitTorrent tit-for-tat peer selection algorithm adapts well as the number of seeds increases due to the fact that it selects the peers who upload chunks to it with the highest
rate during the download process. Further improvements on the total download time can
be achieve if the minimum hop count and channel diversity metrics are used for the peer
5.2. Chunk-based Peer Selection Problem Formulation
85
selection in multi-channel multi-radio WMNs. By applying an interference metric that
accounts for the path load in the network has proven to be the best alternative if allied to
the availability of multiple paths in wireless mesh network scenarios. The use of multiple
paths given by the routing layer and the BFS channel reassignment have shown to be non
orthogonal solutions for the peer selection as the traffic spread over a larger amount of
paths/links in the network makes difficult to prioritize links during the channel reassignment. Moreover, the P2P traffic in the network changes in a smaller time scale compared
to the channel reassignment, as the download of chunks are subdivided in several parallel
block downloads.
5.2
Chunk-based Peer Selection Problem Formulation
In this section we extend the model presented in Chapter 4 by assuming that the requested resource represented by a given file f can be divided into multiple chunks. Thus,
we propose an extended peer selection problem, named as chunk-based peer selection
problem.
We assume that the network is composed by a set of peers N , a set of files F , and a set
of chunks M per file. In the chunk-based peer selection problem formulation we assume
that the file f shared by the P2P system is subdivided in m ≥ 1 (m ∈ M ) equally sized
chunks and n ≥ 2 (n ∈ N ) peers are interested in downloading the file f ( f ∈ F ). By
considering such chunk-based behavior, a leecher that has downloaded a given chunk m,
can provide it to other leechers at the same time it is downloading other chunks from
different sources.
We are now developing a mathematical model using a MILP formulation which allows us to estimate the minimal time required to disseminate a resource file among a set of
peers, the makespan, given the network constraints imposed by the channel assignment
and routing protocol. In order to solve the problem using mathematical programming
techniques, we need to make use of the following assumptions. The download time required to distribute a file among the peers are divided into finite time slots (T ∈ N). In
each time slot interval, the peers involved in the download process select chunks to be
downloaded from other peers which own the requested chunk already. Within each time
slot ∆ t ∈ R+ , t ∈ {1, . . . , T }, we assume that the rate at which the peers download the
chunk is constant. Moreover, we assume that peers that have already downloaded the
whole file stay in the network during the whole makespan. Thus, the makespan required
to disseminate a given file represents the sum of all time slots used during the download
process.
We make use of the index k ∈ M := {1, . . . , m} to identify the chunks, and indexes
i, j ∈ N := {1, . . . , n} to identify the peers. At the beginning of the download process, a
peer i can own a subset M i ⊆ M of chunks, which can also be empty in case of a leecher
peer. Or it can own all M chunks, as it might represent the seed peer holding the complete
file.
We make use of the boolean variable mi k ∈ {0, 1} to indicate if peer i holds the chunk
86
Extending the Peer Selection Problem for Multiple Chunks
k (k ∈ M i ). We also assume that each peer has the capability to perform multiple downloads and uploads simultaneously, but it does not download parts of one chunk from
different sources.
Thus, the chunk-based peer selection problem can be described by a set of peers N , a
set of chunks M and the initial ownership of chunks M i . In order to model the chunkbased peer selection problem for wired networks, [124] makes use of uplink and downlink capacities, as they assume that the bottleneck occurs in the access link which connected the peers to the core network. While this model may hold for typical wired access
networks, where e.g. clients are connected through asymmetric digital subscriber line
(ADSL) links to a high capacity core network, the situation in WMNs is different. Here,
typically the wireless mesh network is the bottleneck as it handles the traffic originated
by the mesh clients and directed to other mesh clients or to the internet. This requires to
extend the model of [124] in order to identify WMN resource limitations as constraints
to the problem formulation. Thus, we substitute the uplink and downlink capacities
constraints by the set of network paths and its bottleneck collision domain constraints,
as explained later on in this section.
We assume that the wireless mesh network provides connectivity between all pairs
of peers. As the bottleneck occurs at the mesh network, we can simplify the mode and
assume that the peers involved in the download process are the mesh routers (N ⊂ V ),
which assist their mobile clients in the download process. We make use of the set of
inequalities given by the collision domain model described in Section 4.2.2 in order to
derive the link capacity in the network. We make use of chunks per unit of time to
represent the nominal MAC throughput as we assume that all chunks have equal size.
The download rate functions xi f j described in Section 4.2 are modified in order to
allow multiple chunks k in the download problem formulation. Thus, the new download
rate functions xi k j t represents the download rate at which peer i downloads the chunk k
from peer j at the time interval t .
Besides the download rate functions variables xi k j t , two other variables are necessary
to describe the chunk-based peer selection problem formulation; the source selection bi k j ,
and the availability functions ak j t .
The source selection variable bi k j , is assumed to be a binary variable ( bi k j ∈ {0, 1}),
and indicates if peer i downloads chunk k from peer j . The availability function ak j t , is
also assumed to be a binary variable (ak j t ∈ {0, 1}), and indicates if chunk k is available at
peer j at time t .
Given the download rate function, source selection, and availability function variables, a valid solution for the chunk-based peer selection problem can be expressed in
terms of a P2P chunk-based scheduling problem. In order to ensure feasibility of the
chunk-based peer selection problem, the triple (x, b , a) needs to satisfy the following constraints.
X
j ∈N
bi k j = 1 − mi k ∀i ∈ N , ∀k ∈ M .
(5.1)
5.2. Chunk-based Peer Selection Problem Formulation
X
xi k j t = bi k j ∀i, j ∈ N , ∀k ∈ M .
87
(5.2)
t ∈T
xi k j t ≤ ak j t ≤
t −1
XX
η∈N τ=1
x j kητ + m j k
(5.3)
∀i, j ∈ N , ∀k ∈ M , ∀t ∈ T .
By applying Constraint (5.1) we guarantee that peers which do not own a chunk
initially are forced to download it. It also guarantees that a peer can not download parts of
one chunk from different peers. In Constraint (5.2) we guarantee that a peer completely
downloads the chunks that it does not own initially. In Constraint (5.3), we guarantee
that chunk k is downloaded by peer i from peer j at interval t only if chunk k is available
at peer j at time t − 1.
So far, the Constraints (5.1)-(5.3) do not dictate the network constraint. To obtain
feasible download rate functions, we need also to guarantee network connectivity among
the peers involved in the download process, achieved via routing and channel assignment.
As in Chapter 4, we make use of the three channel assignment schemes (BFS, K-Partition
and single channel) and the shortest path routing via Dijkstra algorithm.
Given the download rate functions xi k j t , and the routing and channel assignment,
a single path route pi , j is selected between peers i and j in order to download chunk
k at time interval t . Assuming the network bottleneck at the back-haul of the wireless
mesh network, and the number of transmission in the collision domain of link (u, v)
representing the set of paths pi , j traversing each link (s, t ) in the collision domain, as
describe in Section 4.2.2, the following constraint needs to be satisfied.
X
n s,t … xi k j t ≤ B∆ t
s ,t ∈ C D u,v
(5.4)
∀i, j ∈ N , ∀k ∈ M , ∀t ∈ T , ∀l (u, v) ∈ E
Constraint (5.4) guarantees that within any given time interval ∆ t (t ∈ T ), the sum of
the download rate functions in any collision domain does not exceed the nominal MAC
capacity B.
With the constraints (5.1)-(5.4) described, we can now formulate the selection of an
optimal P2P schedule as a MILP problem, given that the objective function is formulated
as a linear function of the introduced variables.
5.2.1
Makespan Minimization
Given that the constraints describe are already formulated as a MILP problem, we make
use of the makespan minimization as the objective function of our chunk-based peer
88
Extending the Peer Selection Problem for Multiple Chunks
selection problem. Thus, our objective is to minimize the time required to disseminate
the file f with k chunks given the constraints.
By considering the download rate function as piecewise continuous, we assume a finite number T of discontinuities, where peers are allowed to change their download rate.
Thus, in order to formulate the objective function, an upper bound for the number of
time intervals T is needed, during which the download rates of all peers remains constant.
According to Theorem 1 of [124], the upper bound for the number of time intervals
is given by the number of peers and number of chunks, by applying T = n m. Since in
our chunk-based problem formulation we have the dependency of the network capacity
given by the collision domain model (which was not present in [124]), by using the upper
bound T = n m might not represent the optimum makespan for our problem formulation, but it still represents a reasonable assumption. Given those considerations, we can
formulate the optimum makespan download problem as:
mi n
X
∆t
(5.5)
t ∈T
subject to the constraints (5.1)-(5.4), and given that T = n m.
The number of variables and constraints in our problem formulation grows exponentially with the number of time intervals, number of peers and number of chunks. It is
important to note that the formulation of the chunk-based peer selection problem as a
MILP does not give insight about the complexity of the problem. Thus, an analysis and
design of efficient algorithms to solve the MILP is necessary, but outside the scope of this
work.
5.2.2
Numerical Results
We use Octave [116] to formulate our MILP problem. In order to solve it we use the
solving constraint integer programs (SCIP), an open source MILP solver that provides
better heuristics and consequently faster solutions compared to GLPK. We also assume
that nodes are always on, no mobility, and a 7x7 grid and chaska topologies. We analyze
the performance assuming one versus two radio interfaces per mesh router. The number
of orthogonal channels is varied from 1 to 12. Channels are bound to radios during the
whole download time according to the channel assignment algorithm described in Section 4.3. As a consequence, there is no channel reassignment during the whole makespan
duration. For the 7x7 grid topology, the nodes are spaced by 100 meters having also a
transmission range of 100 meters, and carrier sense of 200 meters. The chaska topology,
shown in Figure 5.1 is a city-wide mesh network deployed in Chaska, MN, USA, and
composed by 196 wireless mesh nodes. The shaded circles in Figure 5.1 represents the
coverage of the wireless nodes. We extract a representative sub-topology consisting of 49
nodes for our evaluation, with transmission range of 150 meters, and carrier sense of 300
meters.
5.2. Chunk-based Peer Selection Problem Formulation
89
Figure 5.1: Chaska topology
We assume that the file disseminated among the peers has a size of 10 MBytes. In
addition, wireless links in the mesh backbone are assumed to use the IEEE 802.11a basic
PHY rate of 6 Mbit/s, giving a nominal MAC throughput of 0.075 chunks per second.
Two to seven random leechers are selected in order to download the same file f . We also
varied the number of seeds holding the file f from one to three. As the complexity of
the problem increases exponentially with the number of chunks and peers involved in
the download process, we make use of single chunks in order to obtain feasible solutions
at reasonable time. However, we argue that such decision still allows us to achieve upper
bound results in order to analyze and compare the chunk-based peer selection problem
formulation with the peer selection problem formulation described in Chapter 4, here
called non-chunk-based formulation. For each combination of number of leechers, number of seeds, routing and channel assignment algorithm, average and standard deviation
of 10 runs are calculated, if not stated otherwise. Each run is composed by a different set
of peers randomly chosen.
Impact of chunk-based formulation
Figure 5.2 shows the optimum makespan obtained by the chunk-based peer selection
problem compared to non-chunk-based peer selection problem for single channel scenarios. In order to obtain the later one, we make sure that the source selection variables
bi k j = 0, ∀ j ∈ R, where R is the set of leecher peers involved in the download process
(R ⊂ N ). By doing so, we force the peers to select the chunks only from the seed nodes.
In Figure 5.2(a) we start by varying the number of randomly selected leechers requesting the file f from a randomly selected seed. For both peer selection cases, by increasing
the number of leechers we increase the optimum makespan as more leechers are down-
90
Extending the Peer Selection Problem for Multiple Chunks
400
Optimum Makespan [s]
350
Optimum Makespan [s]
350
non-chunk-based
chunk-based
300
250
200
150
non-chunk-based
chunk-based
300
250
200
150
100
100
50
50
2
3
4
5
6
Number of Leechers (7x7 grid - 1 seed)
7
(a) Varying number of leechers
1
1.5
2
2.5
3
Number of Seeds (7x7 grid - 6 leechers)
(b) Varying number of seeds
Figure 5.2: Optimum makespan for the P2P download problem: chunk-based versus
non-chunk-based, assuming single channel WMN deployment.
loading the file simultaneously. Therefore, more links are active in the network increasing the load in the single channel collision domain, thus reducing the download capacity.
Even by considering the use of single chunk, the results provided by the chunk-based
formulation is more efficient if compared to the non-chunk based formulation. This is
due to the fact that leechers which already finished the download of the single chunk can
contribute with other leechers, increasing then the set of peers holding the chunk and
consequently the option to select peers with less loaded paths.
In Figure 5.2(b) we keep the number of selected leechers to six and vary the number
of selected seeds from one to three. By increasing the number of seeds we decrease the
makespan, as more peers holding the requested file are available initially. The non-chunk
based formulation obtains a more significant decrease in makespan, as the larger number
of seeds enables the possibility of load distribution among the seeds. An interesting point
to conclude is that for the given network scenario and the non-chunk-based formulation,
it is necessary to have three seeds in order to achieve similar makespan values obtained
for the chunk-based formulation using one seed. Again, this is because once a leecher has
downloaded the file it acts as a seed for other peers leading to more options to download
from. When peers are spread out in the network, this leads to spatial reuse and increase
in capacity leading to smaller makespan for the chunk-based formulation.
Impact of channel assignment schemes
In order to analyze the impact of the different channel assignments for both problem
formulations we start by showing in Figure 5.3 the optimum makespan for the BFS, Kpartition and single channel for six leechers, while varying the number of seeds. Here,
K-Partition and BFS make use of six channels and two NICs.
On overall for the non-chunk-based (Figure 5.3(a)) and chunk-based (Figure 5.3(b))
formulations, increasing the number of seeds available reduces the makespan for all chan-
5.2. Chunk-based Peer Selection Problem Formulation
110
60
250
200
150
100
50
90
80
70
55
1 1.5 2 2.5 3
60
50
40
30
20
C=1
K-Part
BFS
1
1.5
2
2.5
3
Number of Seeds (7x7 grid - 6 leechers)
(a) non-chunk based formulation
Optimum Makespan [s]
Optimum Makespan [s]
100
50
45
40
C=1
K-Part
BFS
91
140
120
100
80
60
40
20
1 1.5 2 2.5 3
35
30
25
20
1
1.5
2
2.5
3
Number of Seeds (7x7 grid - 6 leechers)
(b) chunk based formulation
Figure 5.3: Optimum makespan for different channel assignment strategies and increasing number of seeds
nel assignment strategies. Moreover, the BFS presents the best makespan over all number
of seeds, if compared to the K-Partition and single channel cases. This is due to fact that
during the channel assignment, BFS can give priority to network links that carry download rates between peers. It is important to note that the same trend was also depicted in
Figures 4.4 and 4.5 at Chapter 4. The larger error bars presented in Figure 5.3 are related
to the randomly set of peers selected in the topology studied, which implies on different channel assignment and routing along the different runs, and consequently different
performance.
To compare the channel assignment gain at both problem formulations, we depict
in Figure 5.4 the optimum makespan for K-Partition and BFS channel assignment over
different number of leechers, one seed, six channels and two NICs. Figure 5.4 presents
results for the 7x7 grid and reduced chaska topology, both composed by 49 wireless mesh
routers. For lower number of leechers, there is not much difference between both problem formulations as the number of links involved in the download process is small compared to the number of channels available. As we increase the number of leechers we
clearly see the benefit of the chunk-based peer selection, that instead of downloading the
file from seeds, also tries to download it from leechers. This leads to almost constant
download times despite the larger aggregate download volume. Here, the channel assignment helps as the assignment of multiple channels to the wireless links proportionate
better spatial reuse, and consequently parallel downloads over different channels.
While comparing BFS with K-Partition using both formulations, we see that BFS
provides the lowest makespan as it gives priority to the links that carry traffic while
assigning channels. The benefit of BFS compared to K-Partition is more pronounced at
the reduced chaska topology, as its is composed by a sparse topology with lower number
of links compared to the grid topology.
In order to see if the number of channels impacts the performance of the channel
92
Extending the Peer Selection Problem for Multiple Chunks
180
140
K-Part
BFS
K-Part (chunk-based)
BFS (chunk-based)
160
Optimum makespan [s]
Optimum makespan [s]
180
K-Part
BFS
K-Part (chunk-based)
BFS (chunk-based)
160
120
100
80
60
40
20
140
120
100
80
60
40
20
0
0
2
3
4
5
6
Number of Leechers (7x7 grid - 1 seed)
(a) grid topology
7
2
3
4
5
6
Number of Leechers (chaska - 1 seed)
7
(b) reduced chaska topology
Figure 5.4: Optimum makespan for K-Partition and BFS channel assignment strategies,
1 seed, 6 channels, 2 NICs
180
Optimum makespan [s]
160
BFS
K-Part(chunk-based)
BFS(chunk-based)
140
120
100
80
60
40
20
2
4
6
8
10
12
Number of Channels (7x7 grid - 1 seed / 6 leechers)
Figure 5.5: Optimum makespan versus number of channels for grid topology, 1 seed, 6
leechers, and 2 NICs
assignments, we present in Figure 5.5 the optimum makespan versus number of channels for one seed, 6 leechers and two NICs using the grid topology. For both channel
assignment schemes, we see that there is an optimum number of channels after which no
further decrease on the makespan is achieved. Despite not using the same set of peers involved in the download process, we see that the trend of the optimum number of channels
for the non-chunk-based formulation shown in Figure 5.5 matches with the MRA results
presented in Figure 4.6. We foresee that the optimum number of channels in relation to
the makespan is directly related to the number of radios and network topology. However, it is interesting to note that over all ranges of number of channels, the chunk-based
formulation is always beneficial compared to the non-chunk-based formulation.
Based on the results presented, we can summarize that the combination of a good
channel assignment scheme in multi-channel multi-radio WMNs, together with the pos-
5.3. Designing interactions between Peer Selection, Routing and Channel Assignment93
sibility of mutual resource exchange among peers contributes to a faster resource dissemination in the context of wireless mesh networks. This is mainly due to the fact that the
higher availability of peers holding the desired resource proportionated by the mutual
exchange of chunks, allied to the larger spatial reuse given by the channel assignment,
contribute to a large set of possible peers and paths to be selected during the download
process. To further study the peer selection problem for larger number of peers and
topologies, we carry in the next section a packet-level simulation under different peer
selection heuristics using BFS channel assignment. In order to evaluate the benefit of
multi-path routing on the P2P download problem, we enable the possibility of multiple
shortest paths among peers during the peer selection as an important feature to balance
the load in the network and consequently decrease the download time.
5.3
Designing interactions between Peer Selection, Routing and Channel Assignment
It is known through the literature that the integration between channel assignment and
routing leads to performance improvement in the context of WMNs [74, 75]. The channel assignment determines the set of links sharing the same channel and consequently the
link capacity and network topology. However, the routing impacts the link load along
the selected path, changing also the interference level among links that share a common
channel. Inside the scope of P2P download problem, the outcome of the peer selection
layer influences both lower layers, as it increases the load and interference along the path
to the selected peer. Thus, the interaction between those three layers outlined in Figure
5.6 becomes important as an uncoordinated decision in one layer might impact on the
others.
Throughout the results presented so far, we have also seen some benefits given by the
interaction between those layers. In Chapter 3, we saw the benefits of location-aware
DHTs and resource replication (Section 3.5), and routing layer information (Section 3.6)
used by the P2P overlay layer to speed up the resource lookup phase, reducing the average
lookup delay and network routing stretch. Regarding the resource exchange phase, the
analytical models presented in Chapter 4 and Section 5.2 gave us an insight on the achievable throughput and download time given the network constraints imposed by different
channel assignments and the routing protocol for the P2P download process. The models
presented so far make use of important interactions as they take into account information about the bottleneck links in the link capacity formulation and the use of the BFS
channel assignment that takes into account the number of download rates per link given
by the routing protocol while assigning channels to radio interfaces.
Whereas analytical models are important to understand the behavior of the studied
problem, they are commonly based on simplifying assumptions. In order to validate
such models, flow level simulations are applied, as in [125] and [117], by taking only the
access link bandwidth into account and modeling the data transfer as flows. As shown
by [123], the flow-level simulation overestimates the performance of P2P systems, as
94
Extending the Peer Selection Problem for Multiple Chunks
peer demands
Peer Selection
influences
path load & interference
influences
peer selection
Routing
influences
interference level
influences
topology & capacity
Channel Assignment
underlay
traffic
Figure 5.6: Interaction between peer selection, routing and channel assignment
important characteristics of packet-level simulation, such as the influence of the lower
level protocols and interaction among layers, are ignored.
Thus, to further explore the peer selection, channel assignment and routing problems,
and their interactions, we carry out packet-level simulations in order to answer important questions, such as, "does a channel reassignment during the download process helps
to increase the P2P performance?", or "do the interactions between peer selection, channel assignment and routing bring similar benefits in different topologies?", which were
left unanswered due to the complexity of solving the models for larger set of peers and
chunks. Moreover, the underlying WMN has a big impact on the P2P download problem, as the selection of peers having suboptimal paths in the network will cause resource
consumption in all intermediate nodes in WMN scenarios. Therefore, a new peer selection metric that takes into account information from routing and channel assignment
such as path load information and channel diversity along paths is also proposed.
For the packet-level simulation, we make the choice to use BitTorrent protocol at the
P2P layer as it is in widespread use and it is one of the protocols responsible for a large
portion of todays internet traffic [126]. Therefore, we start by describing the BitTorrent
background and its key mechanisms. Later on, we introduce our peer selection proposal
and present the experimental setup and simulation results obtained.
5.3.1
BitTorrent Background and Key Mechanisms
BitTorrent [31] is a common P2P application specially designed to efficiently distribute
large amounts of data over the internet by forming a mesh-based overlay called swarm for
each shared file. The BitTorrent architecture consist of a central resource index, known
as tracker, that peers connected to when downloading a torrent file. The torrent file is
a metadata file used to identify the resource to be shared by the system. The torrent
file contains information about the resource, its length, name, hashing information, and
uniform resource locator (URL) of a tracker.
5.3. Designing interactions between Peer Selection, Routing and CA
95
The search process to acquire the torrent file is normally done out-of-band. For example, web search can be executed to find the torrent of the required resource. Thereafter,
the peer that wants to download the file contacts directly the tracker to determine the list
of peers that are also involved in the given torrent download process. Thereafter, peers
directly communicate with each other using the list of peers obtained from the tracker.
The peers involved in the download process are divided into seeds, which are peers
that have the complete file, and leechers, which are peers that are still downloading the
file. By breaking the file into chunks, also known as pieces, the BitTorrent allows leechers
to download different pieces from the seed, and also exchange pieces among themselves
in order to obtain the ones they are missing.
In order to inform other peers about the available chunks, each peer tries to establish connections via handshake message by adding them to its neighbors set in case the
contacted peers accept the connection. The peers inform their neighbors about which
chunks of the file they hold via bitfield messages. By doing that, peers can signal their
interest to the peers holding the desired chunk via interested message.
During the actual download process, the chunks are subdivided into blocks (16 KBytes)
and downloaded in parallel TCP sessions, known as pipelining. Normally, the BitTorrent
allows the pipelining of five active block requests per selected peer [31].
In order to achieve fairness among peers and avoid free-riders, i.e. peers that do not
contribute to uploading, BitTorrent applies a choking algorithm for choosing which
peers to download and upload. The choking algorithm is commonly known as tit-fortat strategy, and it is executed periodically every 10 seconds. Once the choking period
expires, the local peer selects the three neighbors which uploads to him at the highest
data rate. This strategy provides incentives for peers to contribute upload bandwidth to
the swarm. Moreover, every 30 seconds the local peer selects another random peer in order to upload to. This strategy allows a peer to discover better unused connections than
the ones been currently used, and it is known as optimistic unchoking. Thus, each local
peer can upload to up to four unchoked peers in parallel, while the rest of the interested
neighbors are kept choked.
When unchoked, a peer needs to decide which chunk to download using the chunk
selection strategy. According to [31], a good chunk selection strategy guarantees download performance. An inefficient strategy would end up in peers having identical (easily
available) chunks, and none of the missing ones. BitTorrent applies at the beginning of
the download the random first chunk strategy, where a random chunk is selected to be
downloaded. Thereafter, the rarest chunk first strategy is used allowing most common
chunks to be downloaded at the end. In order to avoid bad performance at the end of the
download process, the endgame mode strategy is applied during the download of the last
chunks by sending block request to all peers available.
Despite having a decentralized resource distribution, the central tracker is a recognized bottleneck in the BitTorrent systems, been also susceptible to single control availability [53]. Thus in case the tracker is down, new users can not be bootstrapped into
the network. To overcome this issue, BitTorrent allows the possibility to have multiple tracker addresses in a torrent file. In case the primary tracker is down or has a long
96
Extending the Peer Selection Problem for Multiple Chunks
response time, next trackers from the list are selected.
The use of multiple tracker can increase the reliability, but the system still relies on
the scalability of the centralized trackers. Alternatively, several BitTorrent clients can
implement a distributed tracker, also known as trackerless mode, which serve the same
purpose as the central tracker. However, the clients are organized using DHT as described
in our proposed architecture in Section 3.2.
5.3.2
Bestpeer Multi-path Peer Selection Proposal
We can divide the resource exchange in BitTorrent into upload and download phases.
The set of peers involved in those phases are dictated by the BitTorrent peer selection
algorithm. In the upload phase, a local peer selects a set of four peers to upload the
requested chunk to. In the download phase, the local unchoked peer interested in a given
chunk holds a list of possible peers that it can download the chunk from. In BitTorrent,
both phases are dictated by the tit-for-tat and optimistic unchoking strategies, as the list
of possible peers to download from is build upon the received unchoked messages.
We foresee that a good peer selection in WMNs environments is the one that take into
consideration information from the underlying network. Given the results presented in
Sections 4.5 and 5.2.2, we have seen that the use of BFS channel assignment in multichannel multi-radio WMNs helps to reduce the P2P download time. The idea behind our
new peer selection proposal, the bestpeer multi-path peer selection algorithm, is to enrich
the peer selection mechanism utilized by BitTorrent taken into consideration important
characteristics from the WMNs such as path load information and multi-path routing
capability.
Thus, we propose a slight modification in the download phase of the BitTorrent peer
selection, by considering routing and channel assignment information to select the best
peer from the list of possible peers. We call this new heuristic as the bestpeer multi-path
peer selection, and it should represent the peer with the highest path capacity to the
local peer in the wireless mesh network. In order to better understand our proposed
algorithm, we start by describing the contributions of the peer selection metrics, multipath routing, and channel reassignment to the realization of the bestpeer multi-path peer
selection algorithm.
Peer selection metrics
According to [127], a good peer selection metric for BitTorrent over single channel
WMNs is the one that take into account the distance between the peers in terms of number of hops. The hopcount peer selection metric selects peers that have the smallest
number of hops to the local peer, in order to reduce the number of links traversed in the
network [127]. The number of hops to a peer is given by the Dijkstra algorithm which
calculates the shortest hop path between source and destination peers.
In multi-channel multi-radio scenarios such metric is very simplistic since it does not
take into consideration channel diversity and link load. In order to account for channel
5.3. Designing interactions between Peer Selection, Routing and CA
97
diversity, routing metrics such as MIC [63], iAWARE[64] and MIND[65], introduced
briefly in Section 2.2.2, make use of the channel switching cost metric, here named as the
weight metric. The weight metric is defined by Equation 5.6, and accounts for intra-flow
interference as it gives paths with consecutive links using the same channel higher weights
than paths that alternate their channel assignments, essentially favoring paths with more
diversified channel assignments. To calculate the weight metric, we assume that the path
is given by the routing layer, and the information about which channels are assigned to
which links on the path is given by the channel assignment layer.
w e i g h t ( p) =
w1
w2
if
if
C H ( p r e v(l )) 6= C H (l )
C H ( p r e v(l )) = C H (l )
(5.6)
0 ≤ w1 ≤ w2
The degree of interference which a given link is subjected also depends on the interference generated by other interfering links, known as inter-flow interference. To account
for inter-flow interference, MIC make use of the number of neighbors that a given mesh
router interferes with when it transmits using a common channel. By just accounting
for the number of interfering neighbors, the metric favors links incident on mesh routers
with less number of interfering neighbors irrespective of whether the neighbor causes
any interference or not. As shown by [64], this results at paths been selected along the
boundary of the network where nodes have less number of neighbors and find longer
paths.
Thus, the degree of interference also depends on the amount of traffic load generated
by the interfering links. To account for this information, we propose the use of a new
metric, here called bestpeer metric. Since different links may carry different amount
of traffic, the number of active flows going through the interfering links can be used as
an indication of how much interference a given link is subjected to [128]. Thus, the
bestpeer metric calculates the amount of active flows at a given link and at all other links
within its interference range working at the same channel, in order to roughly estimate
the amount of interference that a new flow will experience by selecting this link. The
bestpeer metric on a given path is therefore the accumulated interference at every link on
the path. For the peer selection problem studied, every block been downloaded over a
given link represents an active flow on that link. We assume that the information about
the number of ongoing blocks been downloaded at a given link is given by a centralized
entity, which is responsible for the routing establishment and channel assignment. The
scalability of such solutions in larger WMNs might not be feasible, however the study of
such metric give us important insights on the benefits of considering routing and channel
assignment information during peer selection at multi-channel multi-radio WMNs.
Multi-path routing and channel assignment
Allied to the necessity to find better peer selection metrics, the use of multiple paths between source and destination by the routing protocol has shown to be beneficial in multi-
98
Extending the Peer Selection Problem for Multiple Chunks
channel multi-radio WMNs as it allows mesh routers to perform load balancing among
the multiple available paths, and therefore increase throughput and fairness among network flows [61]. Thus, different from the single path routing that considers a randomly
selected shortest path among source and destination peer, we make use of the multi-path
routing responsible for finding all potentially available shortest hop paths in the network.
The idea behind the use of multi-path routing is to enrich the peer selection by allowing
also multiple paths possibilities during the download phase of the BitTorrent protocol.
In order to address the question if a channel reassignment during the download process helps on the P2P download problem, we assume that channels can be reassigned at
a certain periodicity using the BFS algorithm during the file download process. Here,
we make use of the BFS algorithm presented in Section 4.3. Since, links are assigned to
channels according to their link load in a BFS fashion, we assume that the load on a given
link is obtained through the number of ongoing block requests passing through that link.
The proposed peer selection algorithm
The proposed peer selection algorithm is sketched in code listing 1. The input of the
algorithm is the list of possible peers which hold the desired chunk, and which have
unchoked the local peer. The algorithm outputs the best peer from where the local peer
can download the desired chunk. As in BitTorrent, we assume that chunks are divided
into several blocks and the maximum number of ongoing pipeline requests per peer is
commonly set to five.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
Input: peerlist
Output: block request to bestpeer
Interaction:
while has block to download do
if single-path then
pathlist ← getSinglePathList(peerlist) ;
else
pathlist ← getMultiPathList(peerlist) ;
end
pathlist ← OrderPathList(pathlist) ;
for each p in pathlist do
while number-requested-blocks(p) < allowed pipeline request do
bestpeer ← getBestPath(p) ;
makeBlockRequest(p) ;
number-requested-blocks(p)++ ;
end
end
end
Algorithm 1: Modified peer selection algorithm
We consider that the routing algorithm provides multiple shortest paths between the
5.3. Designing interactions between Peer Selection, Routing and CA
99
peers. Thus, the first part (lines 3-7) determines the path list to be used in case of single
path or multi-path. In the second part (line 8), the path list is ordered according to the
peer selection metric (OrderPathList) described.
Thereafter, in the last part (lines 9-15), the local peer checks if the amount of ongoing
block requests to each peer composing the path list is smaller than the number of allowed
pipeline requests. In case there is room for new block requests, the best path is selected
according to the metric and a block request to the peer holding this path is issued. Following that, the number of ongoing requested blocks to the peer is updated. Since we
do not modify the upload phase of BitTorrent, we assume that the BitTorrent tit-for-tat
algorithm is used by the uploading peer while selecting the set of four unchoked peers.
Therefore, the overload of an uploading peer is a function of the download rate of the
selected unchocked peers. In summary, given the list of peers holding the desired chunk,
the modified peer selection algorithm returns the best peer and path to be used at the
next block request according to the metric selected.
5.3.3
Experimental Setup and Simulation Results
In this section we analyze the two peer selection heuristics, BitTorrent and the proposed
bestpeer multi-path, using a packet-level simulation. The ns-2 simulator [112] is used,
where the BitTorrent implemented by [123] is modified to achieve the bestpeer multipath peer selection algorithm. Dijkstra routing algorithm is used in order to establish all
potential multiple shortest hop paths among peers. A periodic BFS channel assignment
is used to evaluate the benefit of channel reassignment during the download process.
In order to illustrate the benefits on the interaction between peer selection, routing
and channel assignment, we make use of two different network topologies, the 7x7 grid
topology and the chaska topology. Different from Section 5.2.2, where a reduced chaska
topology composed of 49 nodes was used, in this section we make use of all 196 nodes
depicted in Figure 5.1. The transmission rate is set to 6 Mbit/s, the basic rate of 802.11a,
and RTS/CTS is enabled. Three network interfaces and 12 orthogonal channels are used
if not otherwise mentioned.
We consider that during a swarm, the peers download a file of size 10 MBytes generated from a short low quality video clip, and a file of 100 MBytes generated from an
example TV show of about 21 minutes in medium quality. The file is divided into chunks
of 256 KBytes and every chunk is divided into blocks of 16 KBytes. Despite been a configurable parameter in BitTorrent, the chunk size of 256 KBytes is commonly used as
a recommended value. Smaller chunk sizes might lead to higher seeding opportunities,
but it incurs a large connection negotiation overhead among peers. At the same time,
larger chunks might represent a smaller overhead, but at the cost of lowering the seeding
opportunities among peers.
Using both topologies, we vary the number of seeds and leechers in order to analyze
the impact of the two peer selection algorithms in terms of total download time. For each
combination of number of leechers, number of seeds, peer selection, routing and channel
assignment algorithm, average and standard deviation of 10 different sets of randomly
100
Extending the Peer Selection Problem for Multiple Chunks
Total Download Time (sec)
1100
bittorrent
hopcount
weight
bestpeer
1000
900
800
700
600
500
400
300
1
2
3
4
5
6
7
Number of Seeds
8
9
10
Figure 5.7: Total download time versus number of seeds for different peer selection
schemes using 10 leechers
located leechers and seeds are calculated.
Peer selection contribution
We start our evaluation by considering the impact of the peer selection metric separately
from the multi-path routing and channel reassignment. Thus, we make use of single path
routing and no channel reassignment during the download process. Here, the BFS channel assignment is invoked just at the beginning of the simulation, before the download
process starts.
We analyze the four different peer selection metrics described in the previous section; BitTorrent tit-for-tat, hopcount, weight and bestpeer metric. The standard BitTorrent metric takes into consideration the tit-for-tat algorithm that preferentially uploads
to peers who upload blocks to it with the highest rates [31]. The hopcount metric selects peers that have the smallest number of hops to the local peer, in order to reduce the
number of links traversed in the network [127]. In the weight metric, we considered the
channel diversity along the path together with hopcount metric. And finally the bestpeer
metric, which represents the accumulated interference level along a given path.
Figure 5.7 presents the results for total download time of a 10 MBytes file size versus
number of seeds and the four peer selection strategies using the grid topology. We note
that for all strategies, the increase in number of seeds decreases the total download time.
This is basically due to the increase on the number of possible peers that a leecher can
download from. It is interesting to note that, the BitTorrent tit-for-tat algorithm adapts
well as the number of seeds increases. This is also valid for larger file size of 100 MBytes.
This is due to the fact that at every choking period, a local peer can adapt its list of
unchoked peers by selecting the peers that have uploaded to him with the highest rate.
The hopcount and weight metrics improve the download time compared to BitTorrent, as they consider path length and channel diversity information. However, we can
see that the benefits brought by hop count and weight metrics are not as good if compared
5.3. Designing interactions between Peer Selection, Routing and CA
300
260
Total Download Time (sec)
Total Download Time (sec)
1100
bittorrent - single path
bittorrent - multi-path
bestpeer - multi-path
280
240
220
200
180
160
140
120
100
80
1
2
3
4
5
6
7
Number of Seeds
(a) grid topology, 4 leechers
8
9
10
101
bittorrent - single path
bittorrent - multi-path
bestpeer - multi-path
1000
900
800
700
600
500
400
300
1
2
3
4
5
6
7
Number of Seeds
8
9
10
(b) grid topology, 10 leechers
Figure 5.8: Total download time versus number of seeds for different peer and path
selection schemes
to bestpeer. One of the reasons is that in such dense topology, by selecting nearby peers
holding paths with good channel diversity (as achieved through hopcount and weight
metric), we end up clustering the peer selection to a smaller set of peers and therefore
high chunk request load on a given set of paths. Therefore, in order to avoid such problem, a metric that accounts for the path load is very important while selecting peers. By
applying the bestpeer metric, we can further reduce the total download time as peers
with lower path’s interference are selected, distributing the load evenly in the network
over different paths.
Multiple path contribution
In this simulation we make use of multi-path capability provided by the routing algorithm while establishing paths between peers. By calculating all possible shortest hop
paths between peers involved in the download process, the multi-path routing enables
the local peer to increase its path list. By doing that, peers have the option to spread the
downloading of the blocks over multiple paths and consequently balancing even more
the load on the network.
Figures 5.8(a) and 5.8(b) present the total download time results versus number of
seeds, for grid topology using scenarios with four and ten leechers, respectively. We start
by noting that while using BitTorrent with multiple paths, we can decrease the download
time if compared to single path BitTorrent. This is due to the fact that the ongoing block
requests are spread over the available multiple paths to the peer selected by the BitTorrent
algorithm.
In case bestpeer selection metric and multi-path is used, we can decrease even further
the total download time. This is due to the fact that we increase even further the number
of possible paths to be used, as we considered the best peers and paths during the peer selection process. For example, by using bestpeer selection metric and multi-path in Figure
102
Extending the Peer Selection Problem for Multiple Chunks
1
0.9
0.8
0.7
CDF
0.6
0.5
0.4
0.3
0.2
bittorrent - single path
bittorrent - multi-path
bestpeer - multi-path
0.1
0
0
50
100
150
200
Download Time [s]
250
300
Figure 5.9: CDF of total download time for Chaska topology, 15 seeds and 30 leechers
and different peer and path selection schemes
5.8(b), we can achieve a 32% reduction in the average total download time compared to
BitTorrent with single-path for the scenarios with 10 leechers.
We also carry the same evaluation using the complete chaska topology illustrated in
Figure 5.1, where 15 seeds and 30 randomly selected leechers, and a 10 MBytes file is used
in the swarm. Here, we make use of the 15 gateways, present in the chaska topology, to
work as seeds over all runs. As in the grid topology an average and standard deviation of
ten runs with randomly selected leechers are presented.
Figure 5.9 shows the cumulative distribution function of leechers’ downloading time
for BitTorrent and bestpeer. As we have seen for grid topologies, the use of multiple
paths helps the peer selection to reduce the average total download time. However, by
looking at the distribution of download time per leecher in Figure 5.9, we note that the
benefit of BitTorrent multi-path is reduced in the chaska topology if compared to the grid
topologies. This is because in chaska topology there are less number of multiple shortest
paths available among peers, compared to the grid topology. The bestpeer multi-path
results outperform BitTorrent, as it reduces by half the download time for 50 % of the
leechers involved in the download process. Here, such improvement is obtained due
to the larger peer and path diversity, leading to less network congestion compared to
BitTorrent and consequently lower download time. Moreover, the total time required to
disseminate the file over all leechers is reduced by 40 % compared to the BitTorrent.
Channel assignment contribution
Finally, in this simulation we add the capability of channel reassignment during the
download process, analyzing the contribution of peer selection, routing and channel assignment together. The channel assignment frequency during the download process is
regulated by a timeout function, where our objective is to analyze the benefit of channel
reassignment on reducing the total download time.
During the simulation, when a channel assignment timeout occurs, all links are re-
5.3. Designing interactions between Peer Selection, Routing and CA
800
ca 0 - bittorrent - single path
ca 0 - bittorrent - multi-path
ca 0 - bestpeer - multi-path
ca 10 - bittorrent - multi-path
1000
900
Total Download Time (sec)
Total Download Time (sec)
1100
800
700
600
500
400
300
103
ca 0 - bestpeer - multi-path
ca 5 - bestpeer - multi-path
ca 10 - bestpeer - multi-path
ca 20 - bestpeer - multi-path
ca 60 - bestpeer - multi-path
700
600
500
400
300
200
100
0
1
2
3
4
5
6
7
Number of Seeds
8
9
10
1
2
3
4
5
6
7
Number of Seeds
8
9
10
(a) grid topology, 10 leechers, BitTorrent channel reas- (b) grid topology, 10 leechers, bestpeer channel reassignment
signment
Figure 5.10: Total download time versus number of seeds for channel reassignment
assigned channels according to the BFS channel assignment. As the BFS assigns channels to links according to the link load information (i.e. number of ongoing block been
downloaded in the link), we assume that a central sever entity in the network holds the
information about the number of ongoing blocks been downloaded per link. We also
assume that the server has a global view of the network connectivity graph. Giving the
network topology, and the number of ongoing block downloads per link, the server can
re-assign channels to the links starting from the most loaded links in a BFS fashion.
Figure 5.10(a) presents the results for channel reassignment timeout every 10 seconds
(ca 10) for BitTorrent compared to no channel reassignment (ca 0). Here the grid topology with 10 leechers is used. We make the choice of 10 seconds timeout interval intentionally, as it is the timeout value used by the BitTorrent choking algorithm. We have
also used lower timeout values, but no better improvement is achieved as channel reassignment between choking intervals degrades the performance of ongoing blocks been
downloaded.
While comparing BitTorrent results, we see that there is a performance improvement
by using multi-path and channel reassignment. The reason behind such improvement
is the additional capacity of the channel assignment to re-assign channels over the paths
selected by the BitTorrent peer selection scheme. Here, since the list of peers selected by
BitTorrent is smaller if compared to bestpeer, the channel reassignment can make better
use of the multiple available channels at a smaller number of links, increasing then the
path performance. However, by using bestpeer multi-path and no channel reassignment,
we can still perform better than BitTorrent with channel reassignment for larger number
of seeds, as we increase the number of possible peers and paths available during the peer
selection phase.
In order to see the benefit of channel re-assignemnt and best peer selection, we plot
in Figure 5.10(b) the average total download time for 5, 10, 20 and 60 seconds channel
reassignment timeout values compared to no channel reassignment. The results in Fig-
104
Extending the Peer Selection Problem for Multiple Chunks
ure 5.10(b) shows that channel reassignment brings lower benefits to bestpeer multi-path
compared to no channel reassignment. We see two reasons behind this. The first reason
is that the traffic in the network changes in a smaller time scale compared to the channel
reassignment. Every time that a block of 16 KBytes is downloaded, a different path/peer
can be selected by the local peer and therefore a different path load on the set of possible
paths might be expected. One would expect that the best reassignment timeout would be
the amount of time required to download a block. That would be right in case of a single leecher. However, for multiple leechers, the time at which blocks are downloaded is
completely different among the leechers, which would require a very small reassignment
timeout impossible to achieve, as it translates into significant network overhead and instability. Another reason is that by using bestpeer multi-path, the ongoing download
traffic is spread over a large amount of paths/links in the network, making it difficult to
prioritize links during the channel reassignment.
Based on the results presented, we can conclude that in order to achieve P2P download performance improvements at wireless mesh scenarios, it is beneficial to make use
of network information during the peer selection decision. We have seen that metrics
that account for path load information can reduce further the P2P download time for
the studied scenarios. Moreover, by enabling multi-path routing, we can achieve load balancing in the network as P2P download rates are spread over multiple paths. However,
we conclude that the channel reassignment strategy used does not bring further benefits
to the total download time if compared to no channel reassignment, given the topology
used.
5.4
Related Work
In this section we study the related work to the P2P problem formulation and performance evaluation in the context of wireless mesh networks. We start by discussing the
related work to the P2P problem formulation, followed latter by the discussion on related
work to the performance evaluation of P2P systems.
In [125], a MILP problem solution is proposed where uplink capacities are constrained while downlink capacities are unlimited, as it assumes that the underlaying network is not overloaded. Based on the simultaneous send/receive version of the broadcasting problem developed by [129], [125] also assumes that each uploading peer shares its
available uplink capacity equally among the open connections (fair sharing), making than
a simplified analysis. In addition to the MILP formulation, [125] also provides an analytical expression for the makespan using fluid limit, i.e. a file is subdivided into infinitely
many chunks, and symmetric up/down link capacities.
In [124], uplink and downlink capacities are assumed to be constrained. Opposed
to [125], it does not assume fair sharing in the uplink capacity of a peer, as it may lead
to suboptimal makespan. [124] is the most related work to ours, as we extended its
MILP problem formulation to the context of wireless mesh networks. Different from
[124, 125], we assume that the network bottleneck occurs at the wireless mesh network
5.5. Conclusion
105
backhaul as the wireless capacity among mesh routers is limited by important factors
such as the number of channels and NICs, channel bandwidth and network interference.
From the point of view of performance evaluation of P2P systems, [123] conducts a
comparison of packet-level simulation versus flow-level simulation and analytical model
of the BitTorrent protocol. [123] shows that the analytical and flow-level simulation
results presents a closer match, but at the cost of overestimating the BitTorrent protocol
performance. However, through the packet-level simulation, important characteristics
such as the influence of the lower level protocols, are taken into consideration while
analyzing the downloading performance of BitTorrent.
Using packet-level simulations, [95, 127] propose BitTorrent optimizations in order
to scale P2P resource sharing in wireless mesh networks. By using an ALTO-based architecture, [95] proposes a BitTorrent tracker modification which orders the peer list
based on network layer information acquired via WMN’s routing agents. By using ETX
link metric and Dijkstra algorithm, the tracker sorts the peer list based on the shortest
distances from the local peer. In contrast, our peer selection metric make use of path
load information given by the routing and channel assignment layers in multi-channel
multi-radio WMNs scenarios.
Focusing on reducing routing overhead and overlay disconnections, [127] studies the
impact of limiting the scope of resource sharing, e.g. in number of physical and logical
hops, on the download time and connectivity. By organizing peers in a minimum spanning tree and limiting the hop count distance between peers, [127] has shown that in
a single radio scenario it is possible to achieve lower P2P download time compared to
BitTorrent. However, as we show in our results, minimum hop count metric is not sufficient to achieve good performance in multi-channel multi-radio scenarios, and therefore
better peer selection metrics which account for channel diversity, path load information
and interference, have been developed by us.
5.5
Conclusion
In this chapter, we have extended the peer selection model by incorporating the availability of multiple chunks during the download process. The chunk-based peer selection
problem is formulated as a MILP problem, where our main goal is to find the minimal
time to disseminate the file, also known as the optimum makespan. Our model assumes
that the bottleneck of all paths between the peers are at the back-haul of the wireless mesh
network. The wireless link capacity is formulated using the collision domain model,
where different numbers of channels and network interfaces are used. Given the set of
leechers, seeds, chunks, routing and channel assignment, the model is capable to calculate
the optimum makespan necessary to download a given file in a wireless mesh scenario.
The numerical results presented show that the chunk-based peer selection formulation is
much more efficient than the non-chunk-based peer selection formulation as it gives to
the leechers a larger set of possible peers to download the chunk from, and consequently
the option to select peers with less loaded paths.
106
Extending the Peer Selection Problem for Multiple Chunks
By using a packet-level BitTorrent simulation model, we further analyze the benefits
of cooperation between peer selection, routing and channel assignment. The simulation
results shows that the tit-for-tat strategy of BitTorrent adapts well with the increase in
the number of seeds and leechers in wireless mesh networks. However, we also show that
improvements can be achieved in case better metrics are used in the peer selection phase.
We propose a modification of the BitTorrent peer selection where peers are selected according to their path load. By balancing the block requests over multiple available paths,
we can further improve the download performance as a better load balance is achieved
in the network. Using the bestpeer multi-path peer selection, we have seen that the total
time required to disseminate the file over all leechers is reduced by 40 % compared to
the BitTorrent. Finally, channel reassignment have also proven its importance as it copes
with the traffic load changes generated by the P2P chunk-based peer selection problem.
5.6
Validity of the Results and Limitations
We conclude this chapter by assessing the validity of the results obtained, together with
the limitations of our assumptions while studying the P2P download problem. As the
model presented in this chapter is an extension of the peer selection model of Chapter 4,
the limitations with respect to the collision domain model are also expected here.
transport control protocol (TCP) has been used as the transport protocol for several P2P applications, such as BitTorrent. The TCP traffic is elastic in nature and its
congestion control mechanism adapts the download rate to the available capacity in the
network. Due to the complexity involved to represent those aspects in the model, we
simplify our model and assume that the download rate functions can be changed between
the time intervals.
The proposed chunk-based peer selection formulation requires information about the
link capacity in the wireless mesh network, as well as the chunk availability at the beginning of the download process. Furthermore, the resulting scheduling given by the
solution of our model needs to be communicated to all participants peers.
In the simulation studies carried in Section 5.3, we assume a central entity responsible for the routing establishment and channel assignment in the wireless mesh network.
However, we leave to future work the proof of scalability of such central entity to large
scale wireless mesh networks. Here, the cost involved to exchange message information
(e.g. link load, routing and channel assignment) between the mesh routers and the central entity in such centralized architecture, or the use of a distributed architecture, are not
addressed in our work.
Chapter 6
Practical Considerations for
Channel and Channel
Bandwidth Assignment in
WMNs
In this chapter we focus our attention to the practical considerations for channel and
channel bandwidth assignment in multi-channel multi-radio wireless mesh networks. We
have seen in Chapters 4 and 5 that the use of multi-radio systems may improve the network capacity if multiple orthogonal channels are available and assigned such that the
interference in the network is minimized.
Common assumptions for designing channel assignment schemes for IEEE 802.11
mesh networks are channel orthogonality, meaning that the channels do not overlap
and can be used simultaneously, and homogeneity, meaning that channels behave uniformly in a given environment. Along with the increased deployment of wireless mesh
networks, it has been recognized that those assumptions do not hold in reality. For example, when two transmitters operate on adjacent 802.11a channels, their imperfect RF
filters allow portions of the signal to leak outside their band, resulting in considerable
interference, known as ACI.
Moreover, [130] and [131] have shown through extensive measurements in 802.11based wireless mesh networks that channels are not homogeneous. Their experiments
show that for a single link different channels may provide different link qualities depending on the propagation environment. Thus, it is necessary to select properly the set of
right channels for the links.
Wireless devices can make use of many variables to reduce interference and consequently improve communication, such as channel reassignment, transmission power, and
108
Practical Considerations for Channel and Bandwidth Assignment in WMNs
modulation/coding adaptation. Most recently, the channel bandwidth can also be varied
in the 802.11 standard to better adapt the spectrum resource usage to changing environmental conditions, such as interference, and network traffic demands.
Having those challenges in mind, we present in this chapter an experimental study
that evaluates the impact of ACI and channel bandwidth adaptation to multi-channel
multi-radio mesh systems. We start by showing the contribution of the hardware design
to the ACI problem together with possible engineering solutions. Throughout testbed
experiments, we also investigate the relationship between ACI, channel heterogeneity,
PHY rates, and channel bandwidth adaptation to the achievable network throughput.
Important lessons learned are listed in order to help in the design of channel and channel
bandwidth assignment algorithms. Finally, we draw our conclusions and outline the
limitations of our work, presenting also important related works.
6.1
Our Contribution
The contributions of this chapter are two-fold; i) on the impact of ACI at multi-radio
mesh systems, and ii) on the benefit of channel bandwidth adaptation in wireless mesh
network scenarios.
Thus, the first contribution of this chapter is on the evaluation of the board crosstalk,
radiation leakage, and antenna engineering to the ACI problem at multi-radio mesh
systems. We show through several experiments using our multi-radio mesh testbed,
KAUMesh [132], that by using a good hardware engineering for the node enclosure,
board, radio cards, sufficient antenna separation, and orthogonal polarization it is indeed
possible to achieve orthogonal channels using the IEEE 802.11a standard in multi-radio
mesh nodes. We also show through the joint effects of ACI and channel heterogeneity,
that channel separation alone can not be used as an estimation factor for throughput
prediction at higher PHY rates.
The second contribution of this chapter is on the evaluation of channel bandwidth
adaptation and ACI while targeting network throughput improvements for multi-radio
mesh systems. We show that different channel bandwidths can be used to reduce ACI and
also fulfill different traffic demand requirements among wireless links. Added to that, we
also list important points to be considered for designing practical channel and channel
bandwidth assignment algorithms.
6.2
Testbed Experimental Setup
The experiments were performed using our indoor multi-channel multi-radio wireless
mesh tested deployed at Karlstad University [132]. The KAUMesh testbed consists of
Cambria GW2358-4 boards based on Intel IXP435 XScale CPU, running Linux 2.6.22.
Every node is equipped with three mini-PCI R52 Mikrotik 802.11a/b/g cards using
Atheros AR5414 chipset and madwifi 0.9.4 driver. Two interfaces are used for the mesh
backhaul operating in IEEE 802.11a mode. The third interface is used for client access
6.3. Adjacent Channel Interference
109
in 802.11b/g mode and was disabled for our experiments. By using the 5 GHz band in
the backhaul, we avoid interference with the campus WLAN network which uses the 2.4
GHz band. We also make sure that no one else is transmitting in the 5 GHz band by
monitoring the channels used during the experiments.
We used a simple chain scenario composed of three nodes A, B and C mounted in the
ceiling of a corridor at the Computer Science department. The three nodes are placed in
line-of-sight with no obstacles and no interference from other sources operating in the
802.11a band. The distance between the nodes (dA,B and dB,C in Figure 6.1(a)) was fixed
to 10 meters each, and a default antenna separation (d 0 ) of 20 cm was used at node B.
A
dA,B
d’
dB,C
UDP
B
UDP
(a) Testbed setup
C
(b)
Figure 6.1: Testbed experimental setup (a) and development board used (b)
At the experiments we have varied the channel assigned to the links A-B and B-C using
the lower and middle 5GHz U-NII bands for indoor use, which accommodate eight IEEE
802.11a channels. In the IEEE 802.11a, the channels are commonly numbered as 36, 40,
44, 48, 52, 56, 60 and 64. UDP traffic is generated in unicast and multicast mode by using
the iperf [133] and mgen [134] tools. As shown later in Figure 6.9, in case no interference
is present, the maximum throughput achieved by the links operating at the basic 6 Mbit/s
PHY rate (BPSK modulation) and using 20 MHz channels is 5.5 Mbit/s due to the PHY
and MAC overhead.
Along the experiments, different transmission powers, antenna separation and polarization, PHY modulations and transmission modes are used. We briefly summarize the
parameters used throughout our experiments in Table 6.1.
6.3
Adjacent Channel Interference
As in many wireless technologies, the IEEE 802.11 standard splits its available spectrum
band into sub-bands called channels. The available spectrum band depends on specific
country regulation bodies. For example, in USA the Federal Communications Commission (FCC) divides the 5Ghz band into three frequency bands. The lower and middle
U-NII bands (indoor use) accommodate eight channels in a total bandwidth of 200 MHz.
The upper U-NII band (outdoor use) accommodates four channels in 100 MHz bandwidth. The channels’ center frequency operating in those bands are spaced by 20 MHz.
110
Practical Considerations for Channel and Bandwidth Assignment in WMNs
Table 6.1: Measurements parameters
Parameters
Frequency operation
Transmission power
Gain of antennas
Antenna separation (d 0 )
Antenna polarization
Antenna diversity
Receive sensitivity
MAC layer protocol
Modulation (PHY)
RTS/CTS
Packet size
Channel bandwidth
Transmission mode
Configuration
5 GHz (ch. 36-64)
3dBm - 14dBm
3.63 dBi
20, 50 and 100 cm
horizontal-horizontal and
horizontal-vertical
Disabled
-88dBm (6Mbit/s) and -71dBm (54Mbit/s)
IEEE 802.11a
OFDM: 6Mbit/s (BPSK), 36Mbit/s
(16-QAM), and 54Mbit/s (64-QAM)
Disabled
1470 Bytes
20 and 40Mhz
Unicast and Multicast
To ensure that devices operating on different channels do not mutually interfere with
each other, the IEEE 802.11 standard defines the so-called spectrum mask, limiting transmitters emitted power in frequency spectrum. The transmit spectrum mask defines the
maximal allowed power spectral density in dBr, i.e., dB relative to the maximum spectral density of a signal. As the IEEE 802.11 standard does not specify a transmit filter
function, filtering is implied in the spectrum mask.
Figure 6.2 depicts the transmit spectrum mask of an 802.11a channel. Besides the
limitation from the center frequency (fc) of +/- 11 MHz, the signal energy of a 20 MHz
channel bandwidth is spread across a 60 MHz (+/- 30 MHz) band around the center frequency. The potential interference from an adjacent channel, spaced at -20 MHz from the
center frequency, is indicated by the dashed gray line. Such considerable overlapping of
the spectrum mask of a 20 MHz neighboring carrier may lead to an essential interference
between radios transmitting on adjacent channels in close physical proximity, known as
ACI.
6.3.1
ACI Impact on 802.11 PHY and MAC Layer
As discussed in Section 2.2.6, the 802.11 MAC protocol uses two coordination functions:
the DCF and the PCF [83]. In 802.11-based wireless mesh networks, the DCF is the
standard MAC coordination function used for sharing the wireless medium, based on
fully cooperation and trust between stations [59]. DCF works as a listen-before-talk
scheme, based on CSMA/CA. Therefore, before a 802.11 node transmits a packet, it
listens for activity in the given channel and begins its transmission only if it finds the
channel to be idle. If the channel is busy, it defers transmission to a later time.
We can divide the impact of ACI in IEEE 802.11 multi-channel multi-radio wireless
6.3. Adjacent Channel Interference
111
Figure 6.2: IEEE 802.11a transmit spectrum mask [21]
mesh networks into two main parts; the impact on the CCA mechanism at the MAC
layer, and on the packet capture mechanism at the PHY layer [135].
In the first case, the impact on the CCA mechanism at the MAC layer is caused as
the out-of-channel signal generated by the interferer radio in a nearby channel leaks into
the band of the interfered radio, e.g. overlapping region between the two 20 MHz channels in Figure 6.2. Due to the close proximity between the radio interfaces in multi-radio
mesh nodes, the generated interfering energy triggers spuriously the CCA policy reducing medium access opportunities of the interfered radio. In such cases the ACI cannot
be handled explicitly by the channel contention techniques, such as RTS/CTS, as the
interfering network interface cards 1 operate on different channels.
In the second case, the packet capturing mechanism of the interfered radio is impacted
as the ACI generated by adjacent channels contributes to the background noise reducing
the SINR, and consequently increasing the bit-error rate. Therefore, depending on the
power of such interference, it may cause cyclic redundancy check (CRC) errors either
in the preamble header or, if the preamble is received correctly, in the data payload.
Naturally, this leads to packet corruption, which in turn leads to high packet loss [136].
Figure 6.3 depicts three important examples of the ACI impact on multi-radio mesh
systems. For those examples, three nodes referred as A, B and C are involved in the
communication process. We assume that there are two distinct links, link A−B1 and link
B2 − C operating on 802.11a adjacent channels. Thus, node B network interface cards B1
and B2 are assigned to adjacent channels separated by 20 MHz, as shown in Figure 6.2.
1 We
make use of radio interface and interface card terms interchangeably along the text.
112
Practical Considerations for Channel and Bandwidth Assignment in WMNs
Channel Busy
A
B1
B2
C
(a) Nearby transmitters scenario
A
B1
B2
C
(b) Nearby receivers scenario
A
B1
B2
C
(c) Receiver and transmitter scenario
Figure 6.3: Critical ACI scenarios in multi-radio mesh networks
The first example is illustrated in Figure 6.3(a) and presents the case of two nearby
transmitter, where the two radio interfaces are placed at node B. Here, the overlapping
ACI of transmitter B2 causes spurious carrier sense at transmitter B1 . Given that the radio
interfaces make use of the listen-before-talk scheme, they are only allowed to transmit
packets if the medium is sensed idle. However, the ACI generated by radio B2 , leaking
over to the channel where radio B1 is tunned, can trigger the CCA mechanism in radio
B1 which reports the channel as busy and defers its transmission unnecessarily. This is
a common example of the impact of ACI on the CCA mechanism located at the MAC
layer.
The second example is illustrated by Figure 6.3(b) and presents the case of two receiving radio interfaces in close proximity. In this case, the ACI causes a variant of the well
known hidden-node problem. Here, since the transmitting radios A and C are operating
on adjacent channels at a reasonable distance, they can not sense each other’s transmission. However, at the radio interfaces B1 and B2 , both received signals are strong enough
to interferer in each other, causing erroneous packets as the quality of the desired signal
at each receiver is degraded. Different from the well-known hidden-node problem which
is caused by co-channel interference, here the ACI is the factor that contributes to the
6.3. Adjacent Channel Interference
113
degradation of the channel quality of the adjacent channel. This problem can not be
resolved with RTS/CTS mechanism since the links are operating in different channels.
The last example, illustrated by Figure 6.3(c), represents the case where the radio
interface B1 acts as a receiver while the radio interface B2 acts as a transmitter, both tuned
to different channels. This example represents clearly the impact of ACI at the packet
capturing mechanism at the PHY layer, as the weak receiving signal at the receiver radio
B1 gets corrupted by the strong signal of the nearby transmitter B2 . We make use of
this scenario in our experiments as it represents a common case in wireless multi-hop
communication as the mesh routers are responsible to forward packet on behalf of other
mesh routers.
6.3.2
Hardware Design Influence to the ACI
Commonly, in multi-channel multi-radio networks, interference problems related to the
hardware design are also mapped inside the broad scope of the ACI. We divide the interference caused by the hardware design problem into three parts; board crosstalk, radiation leakage, and antenna engineering. We emphasize that hardware-related interference
can be mitigated through a good hardware design.
Generally, the crosstalk phenomenon is caused by electromagnetic interaction between disturbed and disturbing circuits. Circuit topologies, impedance levels, physical
layout, and IC technology all play critical roles in crosstalk strength. A simple simultaneous activation of multiple radios, operating in monitor (passive) mode and tuned
to an orthogonal channel, inside the same device may lead to significant degradation in
throughput performance, as shown by [137]. Through the results presented in Section
6.4.1 we show that good RF-design practices for multi-radio systems should be considered
in order to reduce the penalty from board crosstalk.
Moreover, the over-the-air interference due to imperfect shielding of the Wi-Fi cards
also causes interference in multi-radio systems, known as radiation leakage. The chipset,
cabling, connectors and the antennas are the main source of the radiation. Thus, it is
important to provide good shielding among the RF components while designing a multichannel multi-radio hardware system.
By applying antenna engineering techniques, we can customize antennas to reduce the
amount of mutually radiated signal by exploiting different ways that signals attenuate.
A good antenna engineering turns to be an important factor while trying to mitigate
ACI. For example, several studies, such as [138], have shown by physically separating
antennas we can considerably reduce the amount of interference as the signal’s path loss
is increased. Antenna polarization is also an important aspect to be considered. Typically,
antennas have at least one orientation with strong negative gain (called the null region of
the antenna), which can be used to polarize antennas accordingly. Thus, a polarization
mismatching between antennas located at the same node might help to add an extra loss
to the interference signal.
114
Practical Considerations for Channel and Bandwidth Assignment in WMNs
dA,B
d’
dB,C
A
UDP
B
UDP
C
A
UDP
UDP
C
B1
B2
Figure 6.4: ACI experimental setup: basic setup (upper part) and 2-boxes setup (lower
part)
6.4
ACI Experimental Results
In this section we present the ACI measurement results performed in our KAUMesh
testbed. We start by discriminating clearly in Section 6.4.1 the hardware design influences
added up to the ACI problem in multi-radio systems.
While targeting multi-hop scenarios, we analyze in Section 6.4.2 the impact of ACI
under different sending powers. Thereafter, we present in Section 6.4.3 the joint effects
of ACI, channel heterogeneity and different PHY rates at the network throughput.
6.4.1
Impact of Board Crosstalk, Radiation Leakage and Antenna
Engineering
We start the testbed experiments by comparing the impact of board crosstalk, radiation
leakage, and antenna engineering to the ACI problem discussed in Section 6.3.2.
In order to analyze the impact of board crosstalk and radiation leakage we make use
of two different setups. The first one is the basic setup, and represents the scenario where
board crosstalk and radiation leakage is presented at node B in the upper part of Figure 6.4. Here, each mesh router has an enclosure containing the two radios tunned to
different channels.
The second setup, named as 2-boxes setup, represents the scenario where board crosstalk
and radiation leakage is not present, as we make use of two physical nodes in order to
represent node B. By using two separate physical nodes, nodes B1 and B2 , we avoid board
crosstalk as the two NICs involved are sitting in two different circuit boards in different
enclosures. The radiation leakage produced by cabling and connectors inside the enclosure is reduced considerable as the enclosure of the nodes (shown in Figure 6.1(b)) works
as a shield reducing the interference between the two NICs. The transmission power at
the NICs are fixed to 14 dBm.
In order to isolate the three different ACI scenarios illustrated in Figure 6.3, we make
use of backlogged unidirectional multicast traffic where no MAC acknowledgement is
generated in the opposite direction. By doing so, we can separate the receiver and trans-
6.4. ACI Experimental Results
115
mitter scenario illustrated in Figure 6.3(c), from the nearby transmitter and receiver scenarios as no MAC acknowledgments are generated by the receivers B1 and C. Hence, in
this experiment UDP packets are generated from node A to B and from node B to C
using the basic PHY rate (6Mbit/s, BPSK modulation).
Given the two setups, we compare in Figure 6.5 five different scenarios in respect
to the normalized aggregated throughput of the flows A→B and B→ C. The results are
normalized to the maximum aggregated throughput achieved of 10903 kbit/s. In the
x-axis we show the channel separation between the assigned channels to links A-B and
B-C. The channel on link A-B was fixed to 36 and the channel on link B-C varied from
channels 36 to 64, achieving channel separation values from 0 (both links on channel 36)
to 7 (link A-B on channel 36 and link B-C on channel 64). For each point in Figure 6.5, we
also plot error bars representing the standard deviation of ten experiment results. Each
experiment has a duration of 25 seconds.
Basic setup
We make use of the basic setup to show the benefits of using antenna polarization in order
to mitigate the ACI problem in multi-radio systems. We assume that channel orthogonality between channels used at the links A-B and B-C is achieved when the normalized aggregated throughput of one is achieved, as it is the maximum throughput achieved by each
link independently operating at the basic PHY rate. The h o r i z on t a l − h o r i z on t a l
scenario represents the case where both antennas at node B use the horizontal linear polarization. By changing the transmitting antenna at node B to the vertical polarization,
we arrive at the h o r i z on t a l − v e r t i ca l scenario. It is important to note that both scenarios have also the effect of board crosstalk and radiation leakage, as the basic setup with
two NICs operating at node B, and antenna separation of 20 cm is used.
As we can see from Figure 6.5, when the receiving radio in node B is tuned to channel 36, we need to have a channel separation of four channels (ch=52) in order to have
fully orthogonal channels for the horizontal-horizontal scenario. When both links are
using channel 36, they share the same channel according to the basic DCF function and
therefore the maximum aggregated throughput achieved is the throughput of a single
channel at basic PHY rate. However, when link B-C uses channel separation of one,
the aggregated throughput for the horizontal-horizontal scenario is even reduced as the
ACI impacts the throughput of both links. This is mainly due to the fact that in node B,
the out-of-channel signal generated by the transmitting radio leaks into the band of the
receiving radio (see the illustration of the overlapping region between the two 20 MHz
channels in Figure 6.2), generating high interfering energy in the overlapping area which
impacts the PHY packet capturing mechanism and consequently increases the packet
loss, as described in Section 6.3.1.
However, for the horizontal-vertical scenario the aggregated throughput performance
is improved compared to the horizontal-horizontal scenario. This is due to the additional
attenuation of around 20 dBi when antennas in the transmitting and receiving radios at
node B use orthogonal polarization. By changing their polarization we try to lie the an-
116
Practical Considerations for Channel and Bandwidth Assignment in WMNs
tennas in the "null" region of each others [138]. However, such orthogonal polarization
does not completely eliminate the interference among close-by channels as other effects
such as board cross-talk and radiation leakage limit the performance.
1.0
0.9
board cross-talk
Normalized throughput
0.8
0.7
antenna distance
0.6
0.5
antenna polarization
0.4
horizontal-horizontal
0.3
horizontal-vertical
0.2
2-boxes, d'=20cm
2-boxes, d'=50cm
0.1
2-boxes, d'=100cm
0.0
0
1
2
3
4
5
6
7
Channel separation
Figure 6.5: Impact of board crosstalk, radiation leakage, and antenna engineering on
802.11a multi-radio performance in terms of normalized aggregated throughput and
PHY rate of 6 Mbit/s
2-boxes setup
In order to quantify the impact of board cross talk and radiation leakage, we make use
of the 2-boxes setup where node B is replaced by two separate nodes while still using
orthogonal antenna polarization and antenna distance of 20 cm. As the radios are now
in two different enclosures, we avoid board crosstalk and radiation leakage that occurred
before inside node B. The approach of using two separate nodes to avoid board crosstalk
and radiation leakage was also used before by [135, 139, 140, 141].
By comparing the horizontal-vertical scenario with the 2-boxes scenario in Figure
6.5, both using 20 cm of antenna separation, we can see that the amount of interference
between adjacent channels can be further reduced if board crosstalk and radiation leakage
are eliminated. Thus, in case the hardware design problems described in Section 6.3.2
are excluded, we can achieve channel orthogonality with a channel separation of two
6.4. ACI Experimental Results
117
channels (ch=44). However, the improvement for channel separation of one channel is
not large as the antennas operate close to each other at a 20 cm distance.
We analyze in the last two scenario the impact of antenna separation for the 2-boxes
setup. As we can see in Figure 6.5, increasing the physical distance between the antennas
helps to counteract the ACI problem. For the antenna separation of 100 cm the transmitting node B2 is far enough separated from the receiving node B1 given orthogonal
channels. By separating the antennas at 50 and 100 cm, we could reduce the amount of
interference by approximately 3.2 and 12 dB, respectively. Therefore, due to the additional path-loss attenuation, the out-of-channel signal caused by the transmitting radio
B2 on adjacent channels is now weak enough at the receiving radio B1 , resulting in low
interference if compared to antenna distances of 20 and 50 cm. Despite been beneficial,
the use of larger antenna distances might not be feasible, as it might depend in practice
on the environment where the nodes need to be deployed.
It is important to note that the absolute reductions in throughput presented here are
hardware dependent and should not be taken as universal. As we are going to see in
the results presented at Section 6.4.3, different vendors of cards and boards may provide
different levels of shielding or emit less radiation. The general pattern, however, should
be consistent across all types of wireless cards.
In summary, using a good hardware engineering for the enclosure, board, radio cards,
etc. antenna separation of around 1 m and orthogonal polarization it is indeed possible to achieve orthogonal channels using a channel separation of one in multi-channel
multi-radio mesh nodes. Moreover, the presented results help to provide important information for channel assignment algorithms for wireless mesh networks. For example,
if the board crosstalk is low and orthogonal antenna polarization can be applied, a channel separation of two is enough to maintain channel orthogonality and maximize the
network throughput.
6.4.2
Impact of ACI under Different Sending Powers
Wireless link performance highly depends on channel quality, which usually exhibits
great variability. Common sources of variations include user mobility, environment
changes, hardware, and interference. The rapidly varying channel condition together
with the ACI in multi-radio systems lead to network performance degradation.
To understand the relationship between ACI and link quality on the achievable throughput, we vary the sending power of node A in order to emulate different link qualities. We
use the 2-boxes setup and default antenna separation of 20 cm used in the previous section
where backlogged UDP traffic in multicast mode is transmitted from nodes A to B and
from B to C at the PHY rate of 6 Mbit/s.
By varying the transmission power at node A we simulated different qualities for the
link A − B1 . The transmission power is chosen such that link A − B1 always obtains full
throughput, if link B2 − C is not transmitting simultaneously. Through the use of a
spectrum analyzer we also make sure that the variations of transmission power levels set
(from 14 dBm to 5 dBm) is the expected ones.
118
Practical Considerations for Channel and Bandwidth Assignment in WMNs
Normalized throughput
1
0.9
0.8
0.7
14dBm
12dBm
10dBm
8dBm
6dBm
5dBm
0.6
0.5
0.4
0
1
2
3
4
5
Channel separation
6
7
Figure 6.6: Impact of ACI under different link qualities in terms of normalized aggregated throughput and PHY rate of 6 Mbit/s
The channel on link A − B1 was fixed to 36 and the channel on link B2 − C varied
from channels 36 to 64. Since the transmission power of the interferer B2 remains fixed
at 14 dBm, the level of interference at the receiver B1 is kept the same, while the received
signal strength of link A − B1 is reduced by decreasing the sending power of transmitter
A.
Figure 6.6 depicts the normalized aggregated throughput for different sending powers.
The values are normalized to 10903 kbit/s, the maximum aggregated throughput value
of all tests. The figure shows that if the signal quality of link A − B1 is strong enough
(transmission power at node A at 14 dBm), we achieve channel orthogonality while using
a channel separation of two (ch 36 and 44). This results is similar to the result presented
in Figure 6.5 for the 2-boxes scenario and default antenna separation of 20 cm.
However, with a reduced sending power of 5 dBm, the signal on link A − B1 is weak,
which gives channel orthogonality only with a channel separation of four (ch 36 and 52).
If we analyze the throughput on both links separately, we can see that link B2 − C always
achieves full throughput, while the throughput on link A − B1 decreases when reducing
the sending power. The difference in throughput between the two links is due to the use
of two independent multicast flows. Here, the throughput degradation occurs on link
A − B1 as the ACI contributes to the interference level at the receiver B1 , leading to a low
SINR and consequently high packet loss.
In summary, we noted that the link quality is an important aspect to be considered
while analyzing the ACI in multi-radio mesh systems. We foresee that additional techniques such as power control can be used to mitigate the ACI on weak links. However,
6.4. ACI Experimental Results
119
such solutions may require the use of optimization techniques in order to guarantee network connectivity at acceptable interference levels, which is outside the scope of this
thesis.
6.4.3
Joint Effect of ACI, Channel Heterogeneity and Different PHY
Rates
Given the results presented and the fact that different channels exhibit different link qualities depending on the propagation environment, we analyze in this section the joint effects of ACI, channel heterogeneity and different PHY rates at multi-channel multi-radio
wireless mesh networks.
We make use of the basic setup, horizontal-vertical polarization, and default antenna
separation of 20 cm. Since in a typical wireless mesh networks, the use of unicast traffic
traversing several hops with different PHY rates is a common scenario, we make use of
a backlogged unicast UDP traffic transmitted from nodes A to C, via node B, at fixed
PHY rates of 6 Mbit/s, 36 Mbit/s, 54 Mbit/s, and using the SampleRate rate adaptation
algorithm [142]. The transmission power of the radio interfaces are set to 14 dBm and
packet size of 1470 bytes is used.
The idea of the experiment is to analyze the network throughput given a different
combination of channels, channel separation, and PHY rates between links A − B and
B − C . Figure 6.7 shows the normalized throughput of the UDP stream from node A
to C via B under different channel assignments using a heat map. The values are normalized to the maximum value obtained of all measurements on the particular PHY rate
while sending UDP traffic from node A to C (5226 kbit/s at 6 Mbit/s, 26063 kbit/s at
36 Mbit/s, 35474 kbit/s at 54 Mbit/s, and 33611 kbit/s at SampleRate). We first describe
the results for the fixed PHY rate of 6, 36 and 54 Mbit/s, and latter we describe the results
for the SampleRate algorithm [142].
Fixed PHY Rate
As expected, for the case where the links operate in the same channel (e.g. channel combination 36/36, 40/40, etc.), the channel capacity is shared among both links via the CCA
mechanism and a throughput between 49% and 52% is achieved.
With a 6 Mbit/s PHY rate illustrated in Figure 6.7(a), a channel separation of two or
three is sufficient to achieve 100% throughput in almost all cases. For channel separation
equal to one or two, the throughput from node A to C can be in some cases lower if compared to the results with no channel separation, where both links share the same channel.
Note that this results is different to the results presented in Figure 6.6 for 14 dBm seding
power. With the use of unicast traffic scenario (A→C), two additional factors contribute
to the lower throughput performance. First, unicast traffic uses MAC-layer ACKs, which
can trigger additional carrier sensing and leads to more collision opportunities of DATA
and ACK transmissions. Second, the unicast traffic spans over two hops (A-B-C) and
Practical Considerations for Channel and Bandwidth Assignment in WMNs
Channel Link A-B
44
48
52
56
60
36
64
60
64
97%
96%
96%
95%
96%
95%
40 36%
51%
41%
96%
96%
96%
96%
97%
44 95%
33%
50%
23%
43%
75%
94%
84%
48 99%
95%
40
36
52%
53%
48% 100% 100% 97% 100% 100%
36 51%
40
41%
51%
46% 100% 100% 100% 100% 100%
44
60%
32%
52%
35%
55%
91%
99% 100%
48
99%
55%
44%
52%
41%
51%
99% 100%
52 100% 100% 75%
43%
51%
29% 100% 100%
56 100% 100% 100% 58%
35%
51%
48%
63%
60 100% 100% 100% 100% 46%
49%
52%
64 100%
52%
38%
99%
98%
99%
50%
40
38%
26%
49%
33%
53%
96%
52 100% 44%
95%
31%
50%
36%
96%
96%
56 99%
42%
96%
94%
34%
50%
39%
94%
46%
60 100% 89%
96%
96%
36%
44%
50%
37%
52%
64 97%
88%
95%
51%
37%
38%
50%
95%
(a) 6 Mbit/s PHY
40
(b) 36 Mbit/s PHY
Channel Link A-B
44
48
52
56
64
60
64
53%
65%
36 50%
34%
59%
91%
62%
46%
61%
78%
40 33%
44%
36%
96%
75%
53%
75%
86%
44 51%
30%
48%
38%
41%
45%
49%
62%
48 75%
41%
38%
51%
38%
41%
68%
92%
52 88%
30%
40%
37%
49%
22%
37%
67%
43%
36 49%
34%
42%
72%
45%
49%
40 45%
51%
43%
86%
59%
66%
57%
82%
44 43%
33%
51%
45%
43%
42%
47%
51%
48 58%
34%
50%
51%
50%
44%
59%
44%
52 65%
38%
40%
46%
51%
42%
43%
45%
56 48%
35%
44%
45%
36
40
Channel Link A-B
44
48
52
56
60
Channel Link B-C
Channel Link B-C
36
Channel Link A-B
44
48
52
56
45%
36
Channel Link B-C
Channel Link B-C
120
44%
51%
49%
45%
56 82%
34%
76%
56%
46%
22%
48%
60 100% 86%
99% 100% 99%
45%
50%
47%
60 98%
77%
96%
99% 100% 27%
46%
34%
64 96%
96%
88%
28%
51%
64 74%
71%
73%
90%
21%
46%
90%
96%
96%
(c) 54 Mbit/s PHY
62%
40%
(d) Autorate (SampleRate)
Figure 6.7: Throughput A-C for different channel combinations (20 cm antenna distance
thereby the end-to-end throughput is dependent on the packet loss probabilities of both
links.
By looking at Figures 6.7(b) and 6.7(c) we note that as the PHY rate increases, the results get less predictable. For example, at 54 Mbit/s PHY rate illustrated in Figure 6.7(c),
even a channel separation of seven does not necessarily give full throughput. In order to
better understand the behavior of those results, we make use of the coefficient of variation (CV) metric, which represents for a given channel separation the ratio of the standard
deviation to the mean. At 6 Mbit/s PHY rate the CV for possible channel assignments at
links A − B and B − C with channel separation of four (e.g.: 36/52, 40/56, 44/60, 48/64)
is 0.004. This indicates that irrespectively of the channel used, all configurations achieve
approximately the same throughput and the performance is very predictable.
However, for the same channel separation, the CVs for 36 Mbit/s and 54 Mbit/s are
0.21 and 0.39, implying a higher variability and therefore less predictability. It is important to note that at higher PHY rates, the modulation and coding schemes require a
higher SINR for successful decoding packets, leaving thus a smaller safety margin. Since
the SINR can be low due to the weak signal on some channels and the additional ACI,
6.4. ACI Experimental Results
121
decoding at high PHY rates might fail, while it is still successful for the same channel
combination on a lower PHY rate.
It is noteworthy that the results in Figure 6.7 represent a snapshot of the achievable
throughput under the current channel conditions only. Repeating the measurements several times showed that at 6 Mbit/s PHY rate the results were stable over time. However
for 36 and 54 Mbit/s PHY rate, the throughput of certain channel combinations varies,
so that a certain combination yields high throughput at one time, but low throughput
another instance of time.
By analyzing the received signal strength for all channels on the involved links, we
note two important factors that might cause such lack of predictability at higher PHY
rates. The first is related to the channel quality variability over time, where a sharp
transition region of 1-2 dB in signal-to-noise ratio (SNR) causes link change from highly
reliable to extremely lossy[143]. The second factor is related to the difference in signal
strength between the channels at the same link, which added to the ACI, presents higher
throughput degradation at weaker links.
To analyze those two factor, we carried out received signal strength indication (RSSI)
measurements at the involved links during several days. We could see, for example, that
during the night the signal strength of the strongest (ch 48) and the weakest channel (ch
44) on link A−B differs on average by 7 dB, while during the day the difference can be up
to 14 dB. Many possible factors cause the observed channel variability, including external
interference, different output power at radio transmitters, and multipath [6]. Since no
interference from other 802.11a networks were present, the frequency-selective fading
due to multipath and the different output powers on different channels show to be the
major factor of high RSSI variability in our experiment.
At the same modulation and coding scheme, the multipath causes some subcarriers to
work better than others, which added to the overall signal strength given by RSSI and the
ACI, affects the packet delivery. Moreover, with difference of up to 3 dB in the output
power distribution of different channels, which occurs randomly along different cards
(even from same manufacturer [6]), we can conclude that the channel separation alone
can not be used as an estimation factor for throughput prediction at higher PHY rates.
Thus, more information on the link quality is necessary to achieve such conclusions
which can include probing links at different PHY rates or using channel information
at the OFDM subcarrier level via channel state information (CSI) [143]. In contrast to
[144], we do not find that the channels on the border of a band have a slightly lower
output power than the one in the middle of a band.
SampleRate Auto Rate Algorithm
Many existing wireless mesh deployments use auto rate algorithms, which measure PHY
and MAC layer parameters such as RSSI or the frame retry-count and aim to select
the PHY rate which maximizes the throughput. We studied the performance of the
wide-spread SampleRate (default in MadWIFI)[142]. SampleRate counts frame retransmissions and reduces or increases the PHY rate after successive erroneous or successful
122
Practical Considerations for Channel and Bandwidth Assignment in WMNs
Table 6.2: Percentage of packets sent at a given PHY rate for the SampleRate algorithm,
with and without ACI
PHY rate (Mbit/s)
6
9
12
18
24
36
48
54
Traffic A-B and B-C
link A − B link B − C
0.0
0.0
0.0
1.2
1.2
1.2
17.8
78.6
0.0
0.0
0.0
1.2
1.2
1.2
37.7
58.7
Traffic A-C
link A − B link B − C
0.3
0.0
0.5
4.0
17.7
17.8
45.2
14.5
0.0
0.0
0.0
1.4
1.4
10.7
75.2
11.3
transmissions. The previous results showed, that with some channel combinations, lower
PHY rates can result in a higher end-to-end throughput.
Comparing Figure 6.7(d) with Figures 6.7(a), 6.7(b), and 6.7(c), we see that the use of
the SampleRate rate adaptation algorithm does not lead to better throughput predictability if compared to the use of fixed PHY rates. To better understand such behavior, we
analyze the channel combination 52/64 in more detail while using the SampleRate algorithm. We start by checking which PHY rates are selected for each link without the
contribution of the ACI. This is realized by measuring the throughput of link A − B
without any interfering traffic at link B − C . Thereafter, we measure the throughput of
link B − C without any interfering traffic at link A − B. We present those measurement
results, represented as Traffic A− B and B − C in Table 6.2, by showing the percentage of
the packets sent at each PHY rate available. Table 6.2 also presents the percentage of the
packets sent at each PHY rate for the results presented in Figure 6.7(d) (Traffic A − C ).
It is interesting to note that for the measurements without the contribution of ACI
(Traffic A − B and B − C ), the SampleRate algorithm selects the PHY rates 48 Mbit/s or
54 Mbit/s for more than 96% of the frames at links A − B and B − C . The resulting measured throughputs are 25.2 Mbit/s and 31.5 Mbit/s for link A−B and B −C , respectively.
However, when both links are active at the same time due to the traffic been sent from
node A to C via node B (Traffic A − C ), only 59% of the packets on link A − B and 86%
on link B − C are transmitted using the PHY rates 48 Mbit/s or 54 Mbit/s. According to
Figure 6.7(d), the throughput achieved between nodes A−C for the channel combination
52/64 is 20.8 Mbit/s, throughput that is 17% lower if compared to the 25.2 Mbit/s at link
A − B without interfering traffic at link B − C .
The main reason for such lower performance of the SampleRate rate adaptation algorithm is due to the influence of ACI, which increase the number of erroneous packets at
link A − B, leading the rate selection algorithm to use a lower PHY rate on average on
the links.
Channel Link B-C
6.5. Channel Bandwidth Adaptation
123
Channel Link A-B
44
48
52
56
60
64
54
36
A
36
36
36
A
54
54
A
A
36
A
A
36
54
54
54
54
36
36
36
36
A
54
54
54
54
36
A
36
40
36
54
40
54
44
48
52
A
54
36
54
54
54
36
36
56
A
54
A
36
54
54
54
36
60
54
54
54
54
54
54
54
54
64
54
54
54
54
54
54
54
54
Figure 6.8: PHY rate with highest throughput A− C for different channel combinations
(20 cm antenna distance)
To summarize, we present in Figure 6.8 the results showing which rate setting (PHY
rates of 6 Mbit/s, 36 Mbit/s, and 54 Mbit/s, or SampleRate Auto Rate) provide the highest throughput according to the different channel combinations at links A−B and B −C .
This figure is generated by comparing the achieved throughput A − C of each channel
combination and rate setting presented in Figures 6.7(a), 6.7(b), 6.7(c), and 6.7(d). As we
can see, the 54 Mbit/s PHY rate provides the highest throughput for a larger number of
channel combinations, followed by the 36 Mbit/s PHY rate and SampleRate Auto Rate.
6.5
Channel Bandwidth Adaptation
Wireless devices can make use of many variables to reduce interference and consequently
improve communication, such as channel reassignment, transmission power adaptation,
and change on modulation. Most recently, the channel bandwidth can also be varied in
the 802.11 standard to better adapt the spectrum resource usage to changing environmental conditions, such as interference, and network traffic demands.
With free spectrum and relatively no increase in hardware cost, doubling the channel
bandwidth is the simplest and most effective way to increase network throughput [22].
In multi-radio mesh networks, the use of channel bandwidth adaptation would allow
meshed nodes to tune to any available frequency portion, adapt the bandwidth of the
spectrum in use to match the available frequency space and capacity demands, and send
and receive in parallel. Such flexibility have been standardized in IEEE 802.11n [145],
where the 40 MHz channel bandwidth 2 is also available at similar hardware cost.
The usage of 40 MHz channel bandwidth leads to questions such as coexistence and
interoperability between 20 MHz and 40 MHz channel bandwidths. In addition, high
2 We refer to channel bandwidth as the bandwidth of the spectrum over which the radio transmits (and
receives) its signals.
124
Practical Considerations for Channel and Bandwidth Assignment in WMNs
density networks need to consider the higher frequency re-use in 40 MHz compared to
20 MHz channels. As illustrated by [146], channel orthogonality is not possible if 40
MHz channels are used in the 2.4 GHz band. However, the 5 GHz bands have more
spectrum available and can easily accommodate multiple 40 MHz channels.
However, when multi-radio systems use adaptive channel bandwidth for their radio
interfaces, it is crucial to study the impact of ACI between e.g. channels having 20 and
40 MHz bandwidth. To analyze the impact of ACI in such dynamic spectrum access
scenario, where multi-radio nodes tune their radios to different channels and bandwidths,
we start this section by providing the software modifications needed to achieve different
channel bandwidth using IEEE 802.11a network interface cards. Thereafter, we present
our experimental results were a combination of 20 and 40 MHz channel bandwidth is
used on our multi-channel multi-radio mesh testbed.
The main goal of this section is twofold. First, to provide insights on the impact of
ACI in multi-radio systems using different channels and channel bandwidths. Second, to
evaluate the benefits of adapting channel bandwidth and hardware engineering in order to
mitigate ACI. Furthermore, this section also gives an understanding of practical issues to
consider in spectrum selection mechanisms for multi-channel multi-radio wireless mesh
networks.
6.5.1
Changing Channel Bandwidth using Commodity Wi-Fi Hardware
Using commodity Wi-Fi hardware and pure software modifications, [147] enabled the
communication on 5, 10 and 40 MHz channels in addition to the standard 20 MHz for
the 2.4 GHz band. We have applied those modifications for the 5 GHz band using IEEE
802.11a cards, having the benefit of the larger spectrum and therefore better channel
separation.
To achieve 5, 10 and 40 MHz channel bandwidth, we have changed the frequency
of the reference clock that drives the phase locked loop (PLL) by modifying the AR5KPHY-PLL hardware register on commodity Atheros-based cards [148]. On such cards,
the frequency synthesizer is implemented using the PLL. A frequency divider determines
the center frequency to be used, and the reference clock frequency used by the PLL determines the channel bandwidth. The reference clock is also used by the baseband/MAC
processor. Thus, a modification of the clock rate impacts also the 802.11 timing parameters, which requires device driver software changes to adapt, for example, the short inter
frame space (SIFS) and the slot duration parameters. Those modifications effectively
change the symbol duration, which allows to generate signals for four channel bandwidths of 5, 10, 20, and 40 MHz.
Different from IEEE 802.11n [145], we do not change the channels’ center frequency
while using 40 MHz channel bandwidth. Thus irrespective of the its channel bandwidth
(e.g. 5, 10, 20 or 40 MHz), the 802.11a channels’s center frequency are used. It is worth
noting that in case inter-operation with 802.11n nodes operating in 40 MHz channel is
required, changes to the 40 MHz center frequency is necessary in our approach. More-
6.5. Channel Bandwidth Adaptation
52
27
125
317
417
17
97
51
42
41
32
31
2
1
67897
6789
56789
Figure 6.9: Throughput versus channel bandwidth in 802.11 networks for different modulation schemes
over, the 802.11n standard defines a set of signaling to establish device capabilities and
ensure interoperability [22]. Thus in order establish links using 40 MHz channels with
802.11n devices, it is also required to make modifications in our approach to enable PHY
and MAC 802.11n high throughput (HT) capabilities.
One of the most promising benefits of channel bandwidth adaptation is the improvement of the throughput, as different channel bandwidths offer best throughput under
different radio frequency conditions. According to Shannon’s capacity formula, capacity of a communication channel is proportional to its bandwidth. Thus, by doubling
the channel bandwidth we should expect double the capacity. However, this two-fold
increase in capacity is not achievable for higher modulation PHY rates due to the overhead in the 802.11 MAC, as some overheads are fixed in terms of absolute time [147]. As
an example, in 802.11a the slot-time is 9 µs, which becomes a relatively high overhead
for wider channel bandwidths. This is shown in Figure 6.9, which presents the throughput for different channel bandwidths and two PHY rates (6 Mbit/s and 36 Mbit/s). For
each channel bandwidth and PHY rate combination, we run UDP unicast and multicast
traffic for a period of 25 seconds. The error bars in this figure represent the standard
deviation of the experiments done on a single link in our KAUMesh testbed described in
Section 6.2.
We note that multicast traffic has a slightly better throughput in comparison with
unicast traffic, as no MAC acknowledgment is generated by the receiver. At 6 Mbit/s
PHY rate (binary phase-shift keying (BPSK)), for the 20 MHz and 40 MHz channel bandwidth, the average throughput achievable while using multicast is 5.45 and 10.63 Mbit/s,
respectively. However, while using unicast, the average throughput achievable is 5.22 and
126
Practical Considerations for Channel and Bandwidth Assignment in WMNs
...
...
-40
-20
fc
+20
+60 MHz
+40
...
...
-40
(a) No channel separation
-20
fc
+20
+40 +60 MHz
(b) Channel separation of one
...
...
-40
-20
fc
+20
+40 +60 MHz
(c) Channel separation of two
Figure 6.10: Example of channel overlapping for a 20/40 MHz channel combination and
different channel separation
10.27 Mbit/s for the 20 MHz and 40 MHz channel bandwidth respectively. By comparing the unicast results presented in Figure 6.9, we can see that the use of wider channel
bandwidth (from 20 MHz to 40 MHz) represents an increase factor of 1.96 at 6 Mbit/s
PHY rate, compared to the increase factor of 1.5 while using the 36 Mbit/s PHY rate
(16-QAM).
6.5.2
Example of Channel Overlapping for a 20/40 MHz Channel
Bandwidth Combination
As illustrated in Figure 6.2, the spectrum mask overlapping is worse if an adjacent 40 MHz
channel is used in comparison to a 20 MHz channel. Here, beside the limitation from the
center frequency (fc) of +/- 22 MHz, the signal energy of a 40 MHz channel bandwidth
is spread across a 120 MHz (+/- 60 MHz) band around the center frequency. Therefore,
the potential interference from a 40 MHz adjacent channel spaced at +20 MHz from fc
in Figure 6.2 is considerably higher compared to a 20 MHz adjacent channel spaced at 20
MHz from fc.
In order to better illustrate the impact of ACI when a combination of 20/40 MHz
channel bandwidth is used, we depict in Figure 6.10 three examples of the overlapping
region among 20 and 40 MHz channel bandwidths combination.
The first example, depicted in Figure 6.10(a), represents the case where both radio interfaces operate at the same channel (no channel separation) and different channel bandwidths. Here, the transmission masks of both channels are completely overlapped, allowing both radios to sense each other.
The second example, depicted in Figure 6.10(b), represents the case where the radio
6.6. Channel Bandwidth Experimental Results
127
interfaces are operating in adjacent channels spaced by 20 MHz, giving a channel separation of one (e.g. channel 36 and 40 in IEEE 802.11a). As we are going to see in the results
presented below, the 10 MHz spectrum mask overlap at level of 0 dBr to approximately
-24 dBr is sufficient to allow both nodes to sense each other via the CCA mechanism and
share the medium.
The third example, depicted in Figure 6.10(c), represents the case for channel separation of two (e.g. channel 36 and 44), where the radio interfaces can not sense each other
but the overlapping is high enough to cause interference. It is worth noting that the spectrum mask overlap region caused by a 40 MHz channel into a 20MHz channel is higher
if compared to the overlap of a 20 MHz channel into a 40 MHz channel.
6.6
Channel Bandwidth Experimental Results
In this section we present the measurement results performed in our KAUMesh testbed
while using channel bandwidth adaptation. We start by quantifying the impact of ACI
on the aggregated network throughput while varying the set of channels and channel
bandwidths, using our experimental setup presented in Section 6.2. Thereafter, we focus
on the analysis of the throughput at the receiver node B1 , while varying the channel
separation and antenna distance, using different channel bandwidth combinations.
6.6.1
Impact of ACI for 20 MHz and 40 MHz Channel Bandwidth
To quantify the impact of ACI at varying channel and channel bandwidth, we generate
four different scenarios using a combination of 20 and 40 MHz channels. The first scenario, the 20R-20T, represents the case when the receiver and the interferer radios at node
B are both using 20 MHz channels. The second scenario, 40R-20T, represents the case
when the receiver is using a 40 MHz channel and the interferer is using a 20 MHz channel. The third scenario, 20R-40T, represents the case when the receiver is using a 20 MHz
channel and the interferer is using a 40 MHz channel. And the last scenario, 40R-40T,
represents the case when both receiver and interferer are using 40 MHz channels.
The intuition behind those four distinct scenarios is to characterize real use cases
present in multi-channel multi-radio wireless mesh networks with capability to adapt
channel and channel bandwidth. For example, the 20R-40T scenario represents a case
where one radio interface in a mesh node uses a 20 MHz channel to receive data from
upstream nodes (e.g. node A) while the other radio interface uses a 40 MHz channel to
forward those packets plus the locally generated ones on an adjacent channel towards a
gateway (e.g. node C).
We make use in this section of the basic and 2-boxes setups depicted in Figure 6.4,
and antenna separation of 20, 50 and 100 cm at node B. The link A − B is set to the fixed
channel (ch 36), and we vary the channel separation of the link B − C from 0 to 7 (ch
36-64). Backlogged unidirectional multicast traffic is used between nodes A → B and
B → C . The transmission power of 14 dBm is chosen such that the two links always
128
Practical Considerations for Channel and Bandwidth Assignment in WMNs
obtain full throughput, if not transmitting simultaneously. We conduct the experiment
making use of the 6 Mbit/s PHY rate due to the lower impact of the MAC overhead at
40 MHz channel bandwidth and higher throughput predictability.
Figure 6.11 presents the impact of ACI as a function of normalized aggregated throughput of links A−B and B −C for the four different scenarios using basic and 2-boxes setups
with antenna separation of 20 cm (d’=20 cm) and 100 cm (d’=100 cm). The throughput
is normalized to 21259 kbit/s, the maximum value obtained of all measurements when
both links operate independently using 40 MHz channels.
In overall the normalized aggregated throughput is greater for the 40R-40T scenario
compared to the other scenarios, as more spectrum is available. If we assume no interference between the two links (links operating with channel separation greater than three),
the normalized aggregated throughput achieved are 0.50, 0.76, 0.76, and 1 for the 20R20T, 40R-20T, 20R-40R, and 40R-40T scenarios, respectively.
By analyzing the results of Figures 6.11(a) and 6.11(b), we can see that the amount of
interference between adjacent channels can be different according to the different scenarios, setup and antenna separation.
Figure 6.11(a) shows the contribution of the board crosstalk and radiation leakage
to the ACI by comparing the results of the basic setup with the 2-boxes setup using the
default antenna separation of 20 cm. Here, the 20R-20T results, presented before in Figure
6.6, represents the case where channel bandwidth adaptation is not present. First of all,
we see that the contribution of the board crosstalk and radiation leakage to the ACI is
present at all channel bandwidth combinations, leading to an additional performance
degradation where channel orthogonality is only achieved with a channel separation of
five for the 40R-40T scenario and 3 for the 20R-20T, 40R-20T and 20R-40T scenarios.
Most of the degradation generated by the board crosstalk and radiation leakage is
seen by comparing the throughput results of both setups and channel separation of two
at Figure 6.11(a). For the scenarios where the radio B2 operates in 20 MHz channel (20R20T and 40R-20T) and the board crosstalk and radiation leakage is eliminated, we can
achieve orthogonality at channel separation of two. However, for the scenarios where
the radio B2 operates in 40 MHz channel (20R-40T and 40R-40T), this is not possible as
the wider 40 MHz spectrum mask generates a higher signal overlapping in the frequency
band of node B1 , as exemplified in Figure 6.10(c).
To verify the benefits of antenna engineering, we plot in Figure 6.11(b) the normalized throughput for the given four scenarios using the 2-boxes setup and antenna separation of 20 and and 100 cm. Without board crosstalk and radiation leakage, we clearly see
that the scenarios with 40 MHz channel bandwidth at the interferer B2 produce a higher
amount of ACI compared to the 20 MHz. The roll-off from the 40 MHz spectrum mask
is at level of 0 dBr to approximately -24 dBr within a 20 MHz channel for a channel separation of one (see Figure 6.2). Thus, for channel separation of one and 40 MHz channel
bandwidth (40R-20T, 20R-40T and 40R-40T), links A−B1 and B2 −C share the medium as
they can sense each other through the partially overlapped spectrum illustrated in Figure
6.10(b).
Through Figure 6.11(b) we see that by increasing the physical distance between the
6.6. Channel Bandwidth Experimental Results
129
Normalized Throughput
basic setup
2-boxes setup
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
40R-40T
20R-40T
40R-20T
20R-20T 0
1
2
6
5
4
3
Channel Separation
(Link B2-C)
7
(a) Hardware design influence
Normalized Throughput
2-boxes, d‘=20cm
2-boxes, d‘=100cm
basic, d‘=20cm
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
40R-40T
20R-40T
40R-20T
20R-20T 0
1
2
6
5
4
3
Channel Separation
(Link B2-C)
7
(b) Antenna engineering influence
Figure 6.11: Normalized throughput for various interferer and receiver combination at
20 and 40 MHz
130
Practical Considerations for Channel and Bandwidth Assignment in WMNs
antennas helps to counteract the ACI problem both at 20 MHz and 40 MHz channels.
For the antenna separation of 100 cm the interferer node B2 is far enough separated from
the receiver B1 , and we can now achieve channel orthogonality with a channel separation
of one for the 20R-20T scenario, and a channel separation of two for the remaining ones.
Therefore, due to the additional path-loss attenuation, the out-of-channel signal caused
by the interferer B2 on adjacent channels is now weak enough at the receiver B1 , resulting
in low interference if compared to antenna distances of 20 cm.
It is important to note that regardless of the antenna separation, the receiver B1 does
not negatively influence the interferer B2 as the link B2 − C always receives full throughput, except for the cases where the receiver and the interferer overlaps its spectrum e.g.
channel separation of zero for all scenarios and channel separation of one for the scenarios
40R-20T, 20R-40T and 40R-40T. To see this, and also to quantify the degradation at the
receiver network interface, we focus our attention in the next section to the normalized
throughput at B1 using the four scenarios.
6.6.2
Analysis of the Receiver Side B1
In this section we measure the interference caused by the radio interface B2 at the receiver
B1 for different channel separation and antenna distance while using the four different
scenarios and the 2-boxes setup. The results are presented in Figure 6.12 in terms of normalized throughput to the maximum value obtained of all measurements and 6 Mbit/s
PHY rate (5451 kbit/s and 10631 kbit/s for 20 MHz and 40 MHz channel bandwidths).
As seen before, when the receiver and the interferer are using the same channel (no
channel separation), they equally share the channel according to the basic DCF function.
This is clearly seen in the scenarios 20R-20T and 40R-40T presented in Figures 6.12(a)
and 6.12(d), respectively. However, for different channel bandwidths, illustrated at the
scenarios 40R-20T and 20R-40T in Figures 6.12(b) and 6.12(c), the radio interfaces still
share equally the channel but the radio interface with 40 MHz channel bandwidth obtains
a higher throughput compared to the 20 MHz as it can transmit more data in its wider
channel bandwidth.
The same happens for the channel separation of one and the scenarios with 40 MHz
channel bandwidth (40R-20T, 20R-40T, and 40R-40T). As seen in the example provided
by Figure 6.10(b), the wider 40 MHz channel bandwidth makes use of part of its adjacent channel spectrum, causing both radios to sense each other via the CCA mechanism.
Here, the antenna separation does not bring any benefit as the spectrum overlapped forces
the sender A to share the medium with the sender B2 .
For channel separation of two, there is still enough interference caused by the radio B2
operating at 40 MHz (20R-40T and 40R-40T) and antenna separation of 20 cm. However,
by putting the antennas separated at 50 cm or 100 cm apart between radios B1 and B2 ,
we can completely mitigate the ACI since the B2 signal suffers path-loss attenuation of
approximately 3.2 and 12 dB, respectively. What is also interesting to note is that for the
scenario 40R-20T and antenna separation of 20 cm, no impact of ACI is present due to the
lower amount of overlapping caused by the 20 MHz interferer in the 40 MHz receiver.
6.6. Channel Bandwidth Experimental Results
1
Normalized Throughput
Normalized Throughput
1
131
0.8
0.6
0.4
100cm
50cm
20cm
20cm(1box)
0.2
0
0
1
2
3
4
5
6
Channel Separation
0.8
0.6
0.4
0.2
100cm
50cm
20cm
0
7
0
(a) 20R-20T
2
3
4
5
6
Channel Separation
7
(b) 40R-20T
1
Normalized Throughput
1
Normalized Throughput
1
0.8
0.6
0.4
0.2
100cm
50cm
20cm
0
0
1
2
3
4
5
6
Channel Separation
(c) 20R-40T
0.8
0.6
0.4
0.2
100cm
50cm
20cm
0
7
0
1
2
3
4
5
6
Channel Separation
7
(d) 40R-40T
Figure 6.12: Normalized throughput of link A − B1 for various interferer and receiver
combination
In summary, we can state that by using a combination of 20 and 40 MHz channel
bandwidths in multi-radio scenarios we can achieve a better network throughput as links
with higher traffic demands can operate on wider channels. The results also indicate that
it is better to use 40 MHz channel bandwidth if the radio interfaces are operating on adjacent channels (scenarios 20R-40T and 40T-20R at channel separation of one). The use
of wider channel bandwidths enforces adjacent channels to share the medium via CCA
mechanism avoiding than the ACI while transmitting on adjacent channels. Furthermore, if antenna separation is applied, further improvements on network throughput
can be achieved as we increase the path-loss attenuation to the interferer signal, and thus
mitigate the ACI completely.
132
Practical Considerations for Channel and Bandwidth Assignment in WMNs
6.7
Lessons Learned for Channel and Channel Bandwidth
Assignment in WMNs
The measurements presented showed that many of the assumptions underlying current
channel assignment algorithms for multi-channel multi-radio wireless mesh networks are
questionable in practical deployments. From our results we give the following remarks
for designing a practical channel and channel bandwidth assignment scheme in such networks:
• A good hardware design and antenna engineering are very important on multiradio systems, as it avoids board crosstalk, radiation leakage, and also contributes to
the attenuation of interfering signals generated by close-by transmitters operating
on adjacent channels.
• Channel separation alone is not a good measure of performance: As shown in Figure 6.7 the channel separation is not sufficient to predict the impact of ACI. This
is in contrast to [149], where a simple interference factor based on the spectrum
overlap area of the receiver filter masks is used. In addition to the spectrum overlap, the SNR of the involved links, measured in our experiment through the RSSI,
plays an equally important role. Higher PHY rates are in general more susceptible to ACI and channel heterogeneity, becoming much less predictable in terms
of network throughput. Channel assignment and rate adaptation are two related
problems, which should be jointly addressed.
• Channels are not homogeneous: For a given link, the RSSI on two channels (even
neighboring ones) can differ by as much as 14 dB in our measurements. Also, for
two different links, the best channels may differ. The only way to obtain the RSSI
of a link on a specific channel is to measure it. Due to differences in propagation,
fading and radio transmitters output power, it is not trivial to estimate the RSSI of
a channel by measuring other channels only.
• A joint channel assignment, channel bandwidth adaptation, and power control
technique can be used to mitigate the ACI. However, such solutions may require
solving very hard optimization problems which need to ensure network connectivity and might depend on the traffic flows.
• Throughput at 6 Mbit/s PHY rate is more predictable and less susceptible to environmental factors. Low PHY rates are preferable to be used when a controlled evaluation of channel and channel bandwidth assignment schemes is performed. However, when operating commercial mesh networks fixing the PHY rate to 6 Mbit/s
might result in unnecessarily low performance. This also motivates the design of
joint channel assignment and rate adaption schemes.
• Channel quality fluctuates over time: Due to the varying nature of the channel,
a channel assignment which is performing well now, might perform poor later,
6.8. Related Work
133
leading to two options: either to optimize the network for a worst case or to assign
the channels to give optimum throughput for the current situation. Thus, frequent
channel assignment leads to signaling overhead and impact on routing.
• Current ACI models are insufficient: Current models of ACI (e.g.: [149], [135]
and [146]), which model the ACI as percentage of area overlap in the spectrum
masks of the cards do not capture the complexity of the problem. Building channel assignment schemes on top of such models will lead to sub-optimal channel
assignments.
• Channel bandwidth adaptation as another degree of freedom: Higher throughput
per link can be achieved as more spectrum is available at wider channel bandwidths.
Furthermore, the use of channel bandwidth adaptation can be used to fulfill different traffic demand requirements among wireless links in multi-channel multi-radio
wireless mesh networks. ACI mitigation can be achieved if a combination of 20
and 40 MHz bandwidths and enough antenna separation are used by close-by radio
interfaces operating on adjacent channels, as it forces the adjacent channels to share
the medium.
6.8
Related Work
A large part of proposed multi-channel multi-radio protocols assume the existence of several non-overlapping channels, e.g. 3 for IEEE 802.11b/g and 11 to 13 for IEEE 802.11a.
On the other hand, the use of large channel separation turns also to be a common approach for channel assignment in diverse wireless technologies (e.g. cellular and IEEE
802.11 networks). While such approach may result in lower ACI, it leads to a considerable amount of spectrum not been used, lower spectrum efficiency, and potential waste
of capacity.
[149] has shown that using overlapping channels might be useful, if enough physical distance between the radios is available . Based on the spectrum overlap area of the
receiver filter masks, an Interference (I)-factor is developed to quantify the amount of
interference over adjacent channels. The I-factor is then used to develop a model which
predicts the packet error rate for a given physical distance between nodes operating on
different overlapping channels. [135] makes use of the I-factor model in order to quantify the ACI effect in 802.11a networks. However, throughout our experiments we show
that the I-factor model does not capture the complexity of the problem as important
factors, such as the ACI impact at the CCA mechanism in the MAC layer and channel
heterogeneity, are not considered.
Related works to ours are also presented by [138], [140] and [139], quantifying the
effect of ACI on multi-radio systems using experimental measurements. [140] makes
use of broadcast traffic in order to isolate different traffic pattern (packets and MAC acknowledgments) of receive and transmit side while analyzing the ACI on the 2.4 GHz
and 5 GHz bands. As also stated by [138], the performance degradation is significant
134
Practical Considerations for Channel and Bandwidth Assignment in WMNs
for the scenario where forwarding nodes transmit and receive simultaneously at nearby
channels, since the weak receiving signal at the receiver radio gets corrupted by the strong
signal of the nearby transmitter. However, none of those works investigate the contribution of the hardware design, and joint effects of channel heterogeneity and higher PHY
rates to the ACI problem.
Within the scope of dynamic spectrum access, [147] was the first to explore channel bandwidth adaptation in IEEE 802.11 networks throughout software modifications.
[150] calculate the required guard band between adjacent channels in order to avoid ACI.
[146] has extended the I-factor model to take into account the spectrum overlap in the
2.4 GHz band between 20 MHz and 40 MHz channel bandwidth with 802.11n devices.
However, an experimental evaluation of the I-factor concept for multi-radio systems is
not available. Different from [146, 147], in our work we make use of 20 MHz and 40
MHz channel bandwidth in the 5 GHz band using 802.11a devices, and quantify the
achievable channel orthogonality on multi-radio systems for different channel and channel bandwidth combination, and antenna separation.
6.9
Conclusion
In this chapter we have investigated the impact of adjacent channel interference on network throughput for multi-channel multi-radio wireless mesh networks. We have shown
that by using good hardware and antenna engineering it is indeed possible to better utilize the available spectrum, e.g. achieving orthogonal channels using a channel separation
of one at multi-radio systems. However, for scenarios with higher PHY rates, channel
separation alone can not be used as an estimation factor for throughput, as the channel
heterogeneity also contributes to the ACI problem.
We have also shown that a variation of the bandwidth a channel can operate brings
benefits to multi-channel multi-radio systems, as it allows meshed nodes to tune to any
available frequency portion and adapt the bandwidth of the spectrum in use to match
the available frequency space and capacity demands, while at the same time sending and
receiving in parallel in different frequency portions. Within such scenario, we have evaluated the channel bandwidth adaptation and ACI while targeting improvements in network throughput for multi-radio mesh systems, and also listed important points to be
considered in the design of channel and channel bandwidth assignment algorithms.
6.10
Validity of the Results and Limitations
We conclude this chapter by assessing the validity of the results obtained, together with
the limitations of our assumptions while studying the impact of ACI on the channel
and channel bandwidth assignment problem in multi-radio systems. In Section 6.4.1 we
have shown the impact of the hardware design and antenna engineering in multi-channel
multi-radio systems in terms of network throughput. We emphasize that the absolute
reductions in throughput presented are hardware dependent and should not be taken as
6.10. Validity of the Results and Limitations
135
universal. Different vendors of cards and boards may provide different levels of shielding
or emit less radiation. However, the general pattern is consistent to the related work to
ours.
At higher PHY rates, the channel separation can not be used alone to predict throughput, as the channel quality and ACI plays an equally important role. As a common techniques to access channel quality, the RSSI is used in Section 6.4.3 in order to explain
the joint effects of ACI, channel heterogeneity and different PHY rates. However, better techniques to measure channel quality, such as CSI [143], are necessary as the RSSI
poorly reflect the frequency selective fading of 802.11 channels, making it hard to predict network throughput for a given channel assignment combination in multi-channel
multi-radio mesh networks.
To achieve different channel bandwidths, we have applied softwares modifications
at our IEEE 802.11a network cards. However, different from IEEE 802.11n, we keep
the channels’ center frequency of the IEEE 802.11a standard while adapting the channel
bandwidth. Thus, no coexistence or interoperability with the IEEE 802.11n 40 MHz
channel bandwidth is possible.
Chapter 7
Conclusions
In this chapter we review the research questions and evaluate the contributions of this
dissertation, presenting the concluding remarks. We conclude this chapter and the dissertation describing the future directions for the present research.
7.1
Reviewing the Achievements
In this dissertation, we have proposed ideas on how to enhance P2P system performance
over wireless mesh networks. As a disruptive technology, P2P systems creates significant
opportunities and challenges for wireless mesh networks. P2P systems and wireless mesh
networks share key characteristics such as their independence of dedicated infrastructure
and centralized control. However, such characteristics may represent their weakness, as
new important challenges, such as how to organize P2P overlay membership and how to
leverage high performance in wireless mesh network environments, must be addressed.
To address those challenges we formulated three research questions. Below, we revisit our
research questions and assess our achievements in each of them.
I. How to organize P2P overlay membership in wireless mesh networks in order to provide efficient P2P resource lookup ?
Addressing such research question required the use of cross-layer information exchange between the P2P overlay system and the wireless mesh network. In order to
address the overlay membership organization in wireless mesh networks, we outlined in
Chapter 3 several approaches to adapt the P2P overlay to wireless mesh networks while
targeting improvements at the resource lookup process. By deploying the Bamboo DHT
transparently in static wireless mesh networks, we identify a trade-off between resource
lookup efficiency and overlay management overhead. Through the use of location-aware
DHT and resource replication, we have shown that we can achieve smaller resource
138
Conclusions
lookup delays at larger topologies, as we avoid longer physical paths while routing resource lookups through the overlay, and increase the probability to find the requested
resource among the neighboring peers. In scenarios with medium to high density, we
show that by using neighborhood and enough caching information we can increase the
performance of P2P overlays at wireless mesh networks as the overlay routing table can
be augmented with neighbor, neighbor-of-neighbor information and lookup resource history.
II. How to model the achievable performance of a P2P system in wireless mesh networks
given the capacity constraints imposed by the channel assignment and routing layers ?
The second research question refers to the development of the proposed mathematical models which allowed us to calculate the achievable performance of a P2P system in
wireless mesh networks given the network capacity constraints. To leverage high performance in wireless mesh network environments, we study in Chapters 4 and 5 the benefits
proportionated by the interaction between P2P overlay, routing and channel assignment
layers. A numerical peer selection model is derived in order to achieve throughput maximization and fairness for P2P systems deployed in multi-channel multi-radio wireless
mesh networks. Through the numerical results presented in Chapter 4, we have shown
the relationship between resource replication and channel assignment schemes, showing
that an optimum number of channel exist in order to achieve maximum throughput and
fairness, which depends on the peer selection and channel assignment scheme selected.
An extension of the peer selection model which incorporates the segmentation of the
requested resource into multiple chunks is proposed in Chapter 5. The numerical results
presented have shown that the chunk-based peer selection formulation enhances the P2P
download process by reducing the amount of time necessary to disseminate the resource
in the network, as it proportionate to the requesting peer a large set of possible peers to
download the resource from, and consequently the option to select peers with less loaded
paths. A packet-level simulation also proves those findings and shows that further improvements in the P2P download process can be achieved if interactions with the channel
assignment and routing are available, allowing load balancing in the network through the
use of multiple paths and channel reassignment.
III. What performance improvements can be achieved using common off-the-shelf multiradio devices and what is the impact of interference on performance ?
Through the numerical and simulation results we have seen that multi-radio systems
bring great improvements to wireless mesh network capacity if multiple orthogonal channels are available and assigned such that the interference in the network is minimized.
However, as shown in Chapter 6, in practice the amount of orthogonal channels is reduced as the use of close-by network interface cards operating on adjacent frequency
bands causes high ACI. In addition, channel separation alone can not be used as an estimation factor for throughput, as the channel heterogeneity also contributes to the ACI
problem at higher PHY rates. To address those issues, we have shown in Chapter 6 that
7.2. Future Works
139
by using a good hardware design and antenna engineering in multi-radio systems the ACI
problem can be mitigated, making than a better use of the available spectrum. By allowing the adaptation of the channel bandwidth and using a combination of 20 and 40 MHz
channels, we have shown that better network throughput can be achieved as links with
higher traffic demands can operate with wider channel bandwidths. Moreover, the use
of wider channel bandwidths at adjacent channels is beneficial, as it forces the adjacent
channels to share the medium avoiding than the ACI problem. Giving those results, we
outline a list of important points to be considered in the design of channel and channel
bandwidth assignment algorithms for multi-channel multi-radio wireless mesh networks.
7.2
Future Works
In this section we discuss the directions for future works. While studying the resource
replication benefits for the resource lookup performance in Chapter 3 and the peer selection formulation in Chapters 4 and 5, we have assumed that the resources are equally
important among the nodes in the network. However, for some classes of P2P applications, such as file sharing, the likelihood of equally important resources is limited.
Therefore it might be interesting to explore the interaction between resource popularity
and the replication mechanism. We foresee that depending on the scenario envisioned,
such interaction would contribute even further to the scalability of resource availability
in wireless mesh networks. Another important point that also deserve a further investigation is how to better select and manage cache entries targeting good overlay coverage
and finite memory consumption in Chapter 3. Possible strategies include, but are not
limited to, caching according to harmonic distribution i.e. inversely proportional to the
logical distance in the overlay, or trade-off between the logical quality of the shortcut and
its physical cost.
Based on the chunk-based peer selection problem formulated as a MILP program, and
presented in Chapter 5, the optimum makespan for a resource dissemination in wireless
mesh network environments can be derived based on the set of peers, chunks, channel
assignment and routing. However, we have seen that the complexity of the problem increases for larger scenarios involving several peers and chunks. Therefore, it might be
interesting to arrive in heuristics which involves lower complexity and are close to the
optimum to achieve good solutions in an affordable time. Moreover, the analyze of how
close to the optimum are the existing protocols like BitTorrent is also of our interest.
Regarding the interaction between peer selection, channel assignment and routing layers, it should be interesting to develop further studies in order to select good techniques
to obtain network information e.g. channel utilization and routing path capacity, in a
distributed fashion while using multi-channel multi-radio wireless mesh networks.
The experimental evaluation of ACI and channel bandwidth adaptation conducted
in Chapter 6 has derived important lessons learned for channel and channel bandwidth
assignment in wireless mesh networks. Given those outcomes, it might be interesting
to incorporate them on the analytical and simulation models proposed on the previous
140
Conclusions
chapters. Moreover, more ACI experimental measurements at higher PHY rates could
be done to try to characterize different environments and narrow the worst case scenario
for better optimization purposes.
References
[1] Marcel C. Castro, Laura Galluccio, Andreas Kassler, and Corrado Rametta. Opportunistic P2P Communications in Delay-Tolerant Rural Scenarios. EURASIP
Journal on Wireless Communications and Networking, 2011:1–14, 2011.
[2] M.C. Castro, A.J. Kassler, C.F. Chiasserini, C. Casetti, and I. Korpeoglu. Peerto-peer overlay in mobile ad-hoc networks. Handbook of Peer-to-Peer Networking,
pages 1045–1080, 2010.
[3] Marcel C Castro, Andreas Kassler, Gabriel Kliot, Raphäel Kummer, Roy Friedman, Peter Kropf, and Pascal Felber. Minimizing DHT Routing Stretch in MANETs Extended Abstract. In 9th Scandinavian Workshop on Wireless Adhoc Networks (Adhoc’09), pages 1–3, Uppsala, Sweden, 2009.
[4] Marcel C Castro, Durga M Prasad, Andreas Kassler, and Stefano Avallone. Peerto-Peer Selection and Channel Assignment for Wireless Mesh Networks. In The
International Workshop on Network Modeling and Analysis (IWNMA), pages 1–8,
Bangalore, India, 2011.
[5] Marcel C Castro and Andreas Kassler. On the Interaction Between Peer Selection ,
Routing and Channel Assignment. 10th Scandinavian Workshop on Wireless Adhoc
Networks (Adhoc’11), pages 1–2, 2011.
[6] Peter Dely, Marcel C. Castro, Sina Soukhakian, Arild Moldsvor, and Andreas
Kassler. Practical Considerations for Channel Assignment in Wireless Mesh Networks. In IEEE Broadband Wireless Access Workshop, held in conjunction with
Globecom, volume 1, Miami, USA, 2010.
[7] Marcel C Castro, Andreas Kassler, and Stefano Avallone. Measuring the Impact of
ACI in Cognitive Multi-Radio Mesh Networks. In IEEE 72nd Vehicular Technology
Conference (VTC Fall 2010), Ottawa, Canada, 2010.
[8] Marcel C Castro, Laura Galluccio, Andreas Kassler, Sergio Palazzo, and Corrado
Rametta. On the comparison between performance of DHT-based protocols for
opportunistic. In Proceedings of Future Network and MobileSummit, pages 1–8,
Florence, Italy, 2010.
142
REFERENCES
[9] Marcel C Castro, Peter Dely, Andreas J, and Francesco Paolo D Elia. OLSR and
Net-X as a Framework for Channel Assignment Experiments - Poster. In 4th ACM
International Workshop on Wireless Network Testbeds, Experimental Evaluation and
Characterization (WiNTECH), held in conjunction with MobiCom, Beijing, China,
2009.
[10] Marcel C. Castro, Peter Dely, Andreas J. Kassler, and Nitin H. Vaidya. QoSAware Channel Scheduling for Multi-Radio/Multi-Channel Wireless Mesh Networks. 4th ACM International Workshop on Wireless Network Testbeds, Experimental Evaluation and Characterization (WiNTECH), held in conjunction with MobiCom, page 11, 2009.
[11] Marcel C. Castro and Andreas J. Kassler. Packet Aggregation for VoIP in Wireless
Meshed Networks. In D. Staehle et al. Y. Koucheryavy, G. Giambene, editor, In
Traffic and QoS Management in Wireless Multimedia Networks, chapter Multihop
W. 978-0-387-85572-1, 2008.
[12] Marcel C. Castro, Eva Villanueva, Iraide Ruiz, Susana Sargento, and Andreas J.
Kassler. Performance Evaluation of Structured P2P over Wireless Multi-hop Networks. International Conference on Advances in Mesh Networks (MESH 2008), pages
796–801, 2008.
[13] Nico Bayer, Marcel Cavalcanti de Castro, Peter Dely, Andreas Kassler, Yevgeni
Koucheryavy, Piotr Mitoraj, and Dirk Staehle. VoIP Service Performance Optimization in Pre-IEEE 802.11S Wireless Mesh Networks. 2008 4th IEEE International Conference on Circuits and Systems for Communications, pages 75–79, May
2008.
[14] Marcel C. Castro, Peter Dely, Jonas Karlsson, and Andreas Kassler. Capacity Increase for Voice over IP Traffic through Packet Aggregation in Wireless Multihop
Mesh Networks. In International Workshop on Wireless Ad Hoc, Mesh and Sensor
Networks (WAMSNET07), pages 350–355, Jeju-Island, Korea, 2007. Ieee.
[15] Andreas Kassler, Marcel Castro, and Peter Dely. VoIP Packet Aggregation based
on Link Quality Metric for Multihop Wireless Mesh Networks. In Proc. Future
Telecommunications Conference (FTC2007), Beijing, China, 2007.
[16] Marcel C Castro and Andreas J Kassler. SIP based Service Provisioning for hybrid
MANETs. In Proceedings of International Workshop on Telecommunications, Santa
Rita do Sapucaí, Brazil, 2007.
[17] Marcel Castro and Andreas Kassler. Challenges of SIP in Internet Connected MANETs. 2007 2nd International Symposium on Wireless Pervasive Computing, 2007.
[18] Marcel Cavalcanti De Castro and Andreas J Kassler. SIP in hybrid MANETs âĂŞ
A gateway based approach. In Proceedings of Swedish National Computer Networking Workshop (SNCNW), pages 3–6, Luleå, Sweden, 2006.
REFERENCES
143
[19] Marcel C. Castro and Andreas J. Kassler. Optimizing SIP Service Provisioning
in Internet Connected MANETs. 2006 International Conference on Software in
Telecommunications and Computer Networks, pages 86–90, September 2006.
[20] Ian F Akyildiz, Xudong Wang, and Weilin Wang. Wireless mesh networks: a
survey. Computer Networks, 47(4):445–487, March 2005.
[21] IEEE Std 802.11a 1999(R2003). Part 11 : Wireless LAN Medium Access Control
( MAC ) and Physical Layer ( PHY ) specifications High-speed Physical Layer in
the 5 GHz Band. Technical Report June, 2003.
[22] E. Perahia and R. Stacey. Next generation wireless LANs: throughput, robustness,
and reliability in 802.11 n. Cambridge University Press, 2008.
[23] P. Pathak and Rudra Dutta. A Survey of Network Design Problems and Joint
Design Approaches in Wireless Mesh Networks. IEEE Commun. Surveys Tuts,
13(3):396–428, 2010.
[24] Panayotis Antoniadis, Benedicte Lee Grand, Anna Satsiou, Leandros Tassiulas,
Rui Aguiar, João Paulo Barraca, and Susana Sargento. Community Building over
Neighborhood Wireless Mesh Networks. IEEE Society and Technology Special issue
on Potentials and Limits of Cooperation in Wireless Communications, 27(1):48–56,
2008.
[25] Francesco Licandro and Giovanni Schembra. Wireless Mesh Networks to Support
Video Surveillance: Architecture, Protocol, and Implementation Issues. EURASIP
Journal on Wireless Communications and Networking, 2007:1–14, 2007.
[26] Marius Portmann and Asad Amir Pirzada. Wireless Mesh Networks for Public
Safety and Crisis Management Applications. IEEE Internet Computing, 12(1):18–
25, 2008.
[27] Khaldoun Al Agha, Marc-Henry Bertin, Tuan Dang, Alexandre Guitton, Pascale
Minet, Thierry Val, and J B Viollet. Which Wireless Technology for Industrial
Wireless Sensor Networks? The Development of OCARI Technology. IEEE
Transactions on Industrial Electronics, 56(10):4266–4278, 2009.
[28] Kun-Chan Lan, Zhe Wang, Mahbub Hassan, Tim Moors, Rodney Berriman, Lavy
Libman, Maximilian Ott, Bjorn Landfeldt, and Zainab Zaidi. Experiences in deploying a wireless mesh network testbed for traffic control. ACM SIGCOMM Computer Communication Review, 37(5):17, 2007.
[29] Daniel Wu and Prasant Mohapatra. QuRiNet: A wide-area wireless mesh testbed
for research and experimental evaluations. Second International Conference on
COMmunication Systems and NETworks (COMSNETS), pages 1–10, January 2010.
144
REFERENCES
[30] D Schoder and K Fischbach. Peer-to-peer prospects. Communications of the ACM,
46(2):27–29, 2003.
[31] Bram Cohen. Incentives Build Robustness in BitTorrent. 2003.
[32] Skype: Free internet telephony that just works. http://www.skype.com/, 2011.
[33] P. Yum. Coolstreaming/DONet: a data-driven overlay network for peer-to-peer
live media streaming. Proceedings IEEE 24th Annual Joint Conference of the IEEE
Computer and Communications Societies., 3:2102–2111.
[34] A. Houyou, H. De Meer, and Moritz Esterhazy. P2p-based mobility management
for heterogeneous wireless networks and mesh networks. Wireless Systems and
Network Architectures in Next Generation Internet, pages 226–241, 2006.
[35] Roberto Riggio. JANUS: A Framework for Distributed Management of Wireless Mesh Networks. 2007 3rd International Conference on Testbeds and Research
Infrastructure for the Development of Networks and Communities, pages 1–7, 2009.
[36] Matthew Caesar, Miguel Castro, E.B. Nightingale, G. O’Shea, and Antony Rowstron. Virtual ring routing: network routing inspired by DHTs. ACM SIGCOMM
Computer Communication Review, 36(4):351–362, 2006.
[37] J Sun, C Zhang, and Y Fang. A Security Architecture Achieving Anonymity and
Traceability in Wireless Mesh Networks. Ieee Transactions On Dependable And
Secure Computing, 8(2):1687–1695, 2008.
[38] J Jun and M.L. Sichitiu. The nominal capacity of wireless mesh networks. Wireless
Communications, IEEE, 10(5):8–14, 2003.
[39] N Akhtar and K Moessner. On the nominal capacity of multi-radio multi-channel
wireless mesh networks. Computer Communications, 31(8):1475–1483, 2008.
[40] M. Ripeanu. Peer-to-peer architecture case study: Gnutella network. In Peer-toPeer Computing, 2001. Proceedings. First International Conference on, pages 99–100.
IEEE, 2001.
[41] Napster, http://www.napster.com/.
[42] J Risson and T Moors. Survey of research towards robust peer-to-peer networks:
Search methods. Computer Networks, 50(17):3485–3521, December 2006.
[43] B. Yang and H. Garcia-Molina. Efficient search in peer-to-peer networks. In Proceedings of the International Conference on Distributed Computing Systems (ICDCS).
Citeseer, 2002.
[44] Lada A Adamic, Rajan M. Lukose, Amit R. Puniyani, and Bernardo A Huberman.
Search in power-law networks. Physical Review E, 64(4):1–17, September 2001.
REFERENCES
145
[45] Secure Hash Standard. Technical report, NIST, US Department of Commerce,
1995.
[46] I. Stoica, R. Morris, D. Liben-Nowell, D.R. Karger, M.F. Kaashoek, F. Dabek,
and H. Balakrishnan. Chord: a scalable peer-to-peer lookup protocol for internet
applications. IEEE/ACM Transactions on Networking, 11(1):17–32, February 2003.
[47] Antony Rowstron and Peter Druschel. Pastry: Scalable, distributed object location and routing for large-scale peer-to-peer systems. ACM/IFIP Middleware,
(November 2001), 2001.
[48] B.Y. Zhao, L. Huang, J. Stribling, S.C. Rhea, a.D. Joseph, and J.D. Kubiatowicz.
Tapestry: A Resilient Global-Scale Overlay for Service Deployment. IEEE Journal
on Selected Areas in Communications, 22(1):41–53, January 2004.
[49] Dahlia Malkhi, M. Naor, and David Ratajczak. Viceroy: A scalable and dynamic
emulation of the butterfly. In Proceedings of the twenty-first annual symposium on
Principles of distributed computing, pages 183–192. ACM, 2002.
[50] Sean Rhea, Dennis Geels, Timothy Roscoe, and John Kubiatowicz. Handling
churn in a DHT. In Proceedings of the annual conference on USENIX Annual Technical Conference, number June, pages 10–10. USENIX Association, 2004.
[51] L Galluccio, G Morabito, S Palazzo, M Pellegrini, M Renda, and P Santi. Georoy:
A location-aware enhancement to Viceroy peer-to-peer algorithmâŸE.
˛ Computer
Networks, 51(8):1998–2014, June 2007.
[52] J. Crowcroft, M. Pias, R. Sharma, and S. Lim. A survey and comparison of peer-topeer overlay network schemes. IEEE Communications Surveys & Tutorials, 7(2):72–
93, 2005.
[53] Jarret Falkner, Michael Piatek, J.P. John, Arvind Krishnamurthy, and Thomas
Anderson. Profiling a million user DHT. In Proceedings of the 7th ACM SIGCOMM
conference on Internet measurement, pages 129–134, New York, USA, 2007. ACM.
[54] Y. Zhang and H. Hu. Wireless mesh networking: architectures, protocols and standards. Auerbach Pub, 2006.
[55] Jinyang Li, Charles Blake, Douglas S.J. De Couto, Hu Imm Lee, and Robert Morris. Capacity of Ad Hoc wireless networks. Proceedings of the 7th annual international conference on Mobile computing and networking - MobiCom ’01, (1):61–69,
2001.
[56] David B Johnson and David A Maltz. Dynamic Source Routing in Ad Hoc Wireless Networks. Mobile Computing, 353:153–181, 1996.
146
REFERENCES
[57] C.E. Perkins and E.M. Royer. Ad-hoc on-demand distance vector routing. In
IEEE Workshop on Mobile Computer Systems and Applications, volume 6, page 90.
Published by the IEEE Computer Society, June 1999.
[58] P Jacquet, P Muhlethaler, and T Clausen. Optimized link state routing protocol
for ad hoc networks. IEEE International Multi Topic Conference, 2001.
[59] J Camp and E Knightly. The IEEE 802.11s Extended Service Set Mesh Networking
Standard. IEEE Communications Magazine, 46(8):120–126, 2008.
[60] Mahesh K Marina and Samir R Das. On-demand multipath distance vector routing
in ad hoc networks. 29th Annual IEEE International Conference on Local Computer
Networks, 11-14 Nov.:14–23, 2001.
[61] B. Radunović, Christos Gkantsidis, Dinan Gunawardena, and Peter Key. Horizon:
Balancing TCP over multiple paths in wireless mesh network. In Proceedings of
the 14th ACM international conference on Mobile computing and networking, pages
247–258. ACM, 2008.
[62] R Draves, J. Padhye, and B Zill. Routing in multi-radio, multi-hop wireless mesh
networks. In Proceedings of the 10th annual international conference on Mobile
computing and networking, pages 114–128. ACM, 2004.
[63] Yaling Yang, Jun Wang, and Robin Kravets. Interference-aware load balancing for
multihop wireless networks. Technical report, University of Illinois at UrbanaChampaign, 2005.
[64] Anand Prabhu Subramanian, Milind M Buddhikot, and Scott Miller. Interference aware routing in multi-radio wireless mesh networks. 2nd IEEE Workshop on
Wireless Mesh Networks, pages 55–63, 2006.
[65] V. Borges, Daniel Pereira, Marilia Curado, and Edmundo Monteiro. Routing metric for interference and channel diversity in multi-radio wireless mesh networks.
Ad-Hoc, Mobile and Wireless Networks, pages 55–68, 2009.
[66] Habiba Skalli, Samik Ghosh, Sajal Das, and Luciano Lenzini. Channel Assignment Strategies for Multiradio Wireless Mesh Networks: Issues and Solutions.
IEEE Communications Magazine, 45(11):86–95, November 2007.
[67] J Crichigno, M Wu, and W Shu. Protocols and architectures for channel assignment in wireless mesh networks. Ad Hoc Networks, 6(7):1051–1077, September
2008.
[68] Weisheng Si, Selvadurai Selvakennedy, and Albert Y. Zomaya. An overview of
Channel Assignment methods for multi-radio multi-channel wireless mesh networks. Journal of Parallel and Distributed Computing, 70(5):505–524, May 2010.
REFERENCES
147
[69] Arunesh Mishra, Vivek Shrivastava, Dheeraj Agrawal, Suman Banerjee, and Samrat Ganguly. Distributed channel management in uncoordinated wireless environments. Proceedings of the 12th annual international conference on Mobile computing
and networking - MobiCom ’06, page 170, 2006.
[70] Ashish Raniwala, Kartik Gopalan, and Tzi-cker Chiueh. Centralized channel assignment and routing algorithms for multi-channel wireless mesh networks. ACM
SIGMOBILE Mobile Computing and Communications Review, 8(2):50, April 2004.
[71] K. N. Ramachandran, E. M. Belding, K. C. Almeroth, and M. M. Buddhikot.
Interference-Aware Channel Assignment in Multi-Radio Wireless Mesh Networks. In IEEE International Conference on Computer Communications (INFOCOM), pages 1–12. IEEE, 2006.
[72] Daniel Wu and Prasant Mohapatra. From Theory to Practice: Evaluating Static
Channel Assignments on a Wireless Mesh Network. IEEE International Conference on Computer Communications (INFOCOM), pages 1–5, March 2010.
[73] Chandrakanth Chereddi, Pradeep Kyasanur, and Nitin H. Vaidya. Design and
implementation of a multi-channel multi-interface network. Proceedings of the second international workshop on Multi-hop ad hoc networks: from theory to reality REALMAN ’06, page 23, 2006.
[74] M. Alicherry, R. Bhatia, and L.e. Li. Joint channel assignment and routing for
throughput optimization in multiradio wireless mesh networks. IEEE Journal on
Selected Areas in Communications, 24(11):1960–1971, November 2006.
[75] A. Mohsenian-rad and Vincent S. Wong. Joint logical topology design, interface
assignment, channel allocation, and routing for multi-channel wireless mesh networks. IEEE Transactions on Wireless Communications, 6(12):4432–4440, December 2007.
[76] M.K. Marina, S.R. Das, and A.P. Subramanian. A topology control approach
for utilizing multiple channels in multi-radio wireless mesh networks. Computer
Networks, 54(2):241–256, 2010.
[77] H. Wu, F. Yang, K. Tan, J. Chen, Q. Zhang, and Z. Zhang. Distributed Channel Assignment and Routing in Multiradio Multichannel Multihop Wireless Networks. IEEE Journal on Selected Areas in Communications, 24(11):1972–1983,
November 2006.
[78] Li Li and Chunyuan Zhang. Joint Channel Width Adaptation, Topology Control,
and Routing for Multi-Radio Multi-Channel Wireless Mesh Networks. 6th IEEE
Consumer Communications and Networking Conference, 2:1–5, January 2009.
148
REFERENCES
[79] Stefano Avallone, I.F. Akyildiz, and Giorgio Ventre. A Channel and Rate Assignment Algorithm and a Layer-2.5 Forwarding Paradigm for Multi-Radio Wireless
Mesh Networks. IEEE/ACM Transactions on Networking, 17(1):267–280, February
2009.
[80] P. Gupta and P.R. Kumar. The capacity of wireless networks. Information Theory,
IEEE Transactions on, 46(2):388–404, March 2000.
[81] Y Shi, YT Hou, and J Liu. How to correctly use the protocol interference model
for multi-hop wireless networks. on Mobile ad hoc networking, 2009.
[82] Paulo Cardieri. Modeling Interference in Wireless Ad Hoc Networks. IEEE Communications Surveys & Tutorials, 12(4):551–572, 2010.
[83] ANSI/IEEE Std 802.11. Information technology -Telecommunications and information exchange between systems - Local and metropolitan area networks- Specific requirements - Part 11 : Wireless LAN Medium Access Control ( MAC ) and
Physical Layer ( PHY ) Specifications. Technical report, 1999.
[84] RV Nee. Breaking the Gigabit-per-second barrier with 802.11 AC. Wireless Communications, IEEE, 18(2):4–4, 2011.
[85] Kaixin Xu, Mario Gerla, and Sang Bae. How effective is the IEEE 802.11 RTS/CTS handshake in ad hoc networks. IEEE Global Telecommunications Conference (GLOBECOM 02), 1:72–76, 2002.
[86] Shyamnath Gollakota and Dina Katabi. Zigzag decoding: combating hidden terminals in wireless networks. SIGCOMM Comput Commun Rev, 38(4):159–170,
2008.
[87] I. Gruber, R. Schollmeier, and W. Kellerer. Performance evaluation of the mobile
peer-to-peer service. In IEEE International Symposium on Cluster Computing and
the Grid (CCGrid), pages 363–371. IEEE, 2004.
[88] Cheng-Chang Hoh and Ren-Hung Hwang. P2P File Sharing System over
MANET based on Swarm Intelligence: A Cross-Layer Design. IEEE Wireless
Communications and Networking Conference, pages 2674–2679, 2007.
[89] Burton H. Bloom. Space/time trade-offs in hash coding with allowable errors.
Communications of the ACM, 13(7):422–426, July 1970.
[90] M. Gunes, U. Sorges, and I. Bouazizi. ARA-the ant-colony based routing algorithm for MANETs. In Parallel Processing Workshops, 2002. Proceedings. International Conference on, pages 79–85. IEEE, 2002.
[91] F. Delmastro. From Pastry to CrossROAD: CROSS-Layer Ring Overlay for AD
Hoc Networks. Third IEEE International Conference on Pervasive Computing and
Communications Workshops, pages 60–64, 2005.
REFERENCES
149
[92] Simone Burresi, Claudia Canali, M. Elena Renda, and Paolo Santi. MeshChord:
A Location-Aware, Cross-Layer Specialization of Chord for Wireless Mesh Networks (concise contribution). Sixth Annual IEEE International Conference on Pervasive Computing and Communications (PerCom), pages 206–212, March 2008.
[93] H Sozer, M Tekkalmaz, and I Korpeoglu. A peer-to-peer file search and download
protocol for wireless ad-hoc networks. Computer Communications, 32(1):41–50,
2009.
[94] Sylvia Ratnasamy, Paul Francis, Mark Handley, Richard Karp, and Scott Schenker.
A scalable content-addressable network. ACM SIGCOMM Computer Communication Review, 31(4):161–172, October 2001.
[95] Francesco Paolo D Elia, Giovanni Di Stasi, Stefano Avallone, Roberto Canonico,
and Via Claudio. Bittorrent traffic optimization in Wireless Mesh Networks with
ALTO service. In HotMesh, 2011.
[96] E Burger and J. Seedorf. Application-Layer Traffic Optimization (ALTO) Problem Statement. In RFC 5693, IETF, pages 1–14, 2009.
[97] A. Klemm, C. Lindemann, and O.P. Waldhorst. A special-purpose peer-to-peer file
sharing system for mobile ad hoc networks. IEEE 58th Vehicular Technology Conference. VTC 2003-Fall (IEEE Cat. No.03CH37484), pages 2758–2763 Vol.4, 2003.
[98] Wolfgang Kellerer. Proactive search routing for mobile peer-to-peer networks:
Zone-based P2P. 5th Workshop on Applications and Services in Wireless Networks
(ASWN), 2005.
[99] Thomas Fuhrmann, Pengfei Di, Kendy Kutzner, and Curt Cramer. Pushing
Chord into the Underlay : Scalable Routing for Hybrid MANETs. Technical
report, Universität Karlsruhe, 2006.
[100] Thomas Zahn and J. Schiller. MADPastry: A DHT substrate for practicably
sized MANETs. In 5th Workshop on Applications and Services in Wireless Networks
(ASWN), 2005.
[101] R Winter, T Zahn, and J Schiller. Random landmarking in mobile, topologyaware peer-to-peer networks. Proceedings 10th IEEE International Workshop on
Future Trends of Distributed Computing Systems (FTDCS), pages 319–324, 2004.
[102] Amir Krifa, Mohamed Karim Sbai, Chadi Barakat, and Thierry Turletti. BitHoc:
A content sharing application for wireless ad hoc networks. IEEE International
Conference on Pervasive Computing and Communications, pages 25–27, 2009.
[103] H. Pucha, S.M. Das, and Y.C. Hu. Ekta: An Efficient DHT Substrate for Distributed Applications in Mobile Ad Hoc Networks. Sixth IEEE Workshop on Mobile Computing Systems and Applications, (Wmcsa):163–173, 2004.
150
REFERENCES
[104] Thomas Fuhrmann. Performance of scalable source routing in hybrid MANETs.
2007 Fourth Annual Conference on Wireless on Demand Network Systems and Services, pages 122–129, January 2007.
[105] Milind Buddhikot, Adiseshu Hari, Kundan Singh, and Scott Miller. MobileNAT:
A New Technique for Mobility Across Heterogeneous Address Spaces. Mobile
Networks and Applications, 10(3):289–302, June 2005.
[106] K.K. Dhara and S. Baset. Dynamic Peer-To-Peer Overlays for Voice Systems.
Fourth Annual IEEE International Conference on Pervasive Computing and Communications Workshops (PERCOMW’06), pages 151–156.
[107] W.-P.K. Yiu and S.-H.G. Chan. VMesh: Distributed Segment Storage for Peer-toPeer Interactive Video Streaming. IEEE Journal on Selected Areas in Communications, 25(9):1717–1731, December 2007.
[108] Abhishek Bhattacharya, Zhenyu Yang, and Shiyun Zhang. Temporal-DHT and Its
Application in P2P-VoD Systems. IEEE International Symposium on Multimedia,
pages 81–88, December 2010.
[109] Murat Karakaya, Ibrahim Korpeoglu, and O. Ulusoy. Free riding in peer-to-peer
networks. Internet Computing, IEEE, 13(2):92–98, March 2009.
[110] Sean Christopher Rhea. Open DHT: A public DHT Service. PhD thesis, University
of California at Berkeley, 2005.
[111] Sean Rhea, Brighten Godfrey, Brad Karp, John Kubiatowicz, Sylvia Ratnasamy,
Scott Shenker, Ion Stoica, and Harlan Yu. OpenDHT: A public DHT service
and its uses. In Proceedings of the 2005 conference on Applications, technologies,
architectures, and protocols for computer communications, pages 73–84. ACM, 2005.
[112] The network simulator ns2. http://www.isi.edu/nsnam/ns.
[113] Björn Wiberg. Porting AODV-UU implementation to ns-2 and Enabling Trace-based
Simulation. PhD thesis, Uppsala University, 2002.
[114] Juho Määttä and Timo Bräysy. A Novel Approach to Fair Routing in Wireless
Mesh Networks. EURASIP Journal on Wireless Communications and Networking,
2009:1–14, 2009.
[115] Dah-Ming Chiu and Raj Jain. Analysis of the increase and decrease algorithms
for congestion avoidance in computer networks. Computer Networks and ISDN
Systems, 17(1):1–14, June 1989.
[116] Octave. http://www.gnu.org/software/octave. GNU Octave.
REFERENCES
151
[117] Bin Fan, John C. S. Lui, and Dah-Ming Chiu. The Design Trade-Offs of
BitTorrent-Like File Sharing Protocols. IEEE/ACM Transactions on Networking,
17(2):365–376, April 2009.
[118] Bing He, Bin Xie, and Dharma P Agrawal. Internet Gateway Deployment Optimization in a Multi-Channel Multi-Radio Wireless Mesh Network. IEEE Wireless
Communications and Networking Conference, pages 2259–2264, 2008.
[119] Michal Pioro, Mateusz Zotkiewicz, Barbara Staehle, Dirk Staehle, and Di Yuan.
On max-min fair flow optimization in wireless mesh networks. Ad Hoc Networks,
May 2011.
[120] Rastin Pries, Dirk Staehle, Marieta Stoykova, Barbara Staehle, and Phuoc TranGia. A genetic approach for wireless mesh network planning and optimization.
Proceedings of the 2009 International Conference on Wireless Communications and
Mobile Computing Connecting the World Wirelessly - IWCMC ’09, page 1422, 2009.
[121] Dirk Staehle, Barbara Staehle, and Rastin Pries. Max-Min Fair Throughput in
Multi-Gateway Multi-Rate Mesh Networks. IEEE 71st Vehicular Technology Conference, pages 1–5, 2010.
[122] Mohammad Zulhasnine, Changcheng Huang, and Anand Srinivasan. Favorable
Peer Supported Throughput Optimization in Wireless Mesh Network. IEEE
Global Telecommunications Conference (GLOBECOM), pages 1–5, 2010.
[123] Kolja Eger, Tobias Hoß feld, Andreas Binzenh, and Gerald Kunzmann. Efficient
Simulation of Large-Scale P2P Networks : Packet-level vs . Flow-level Simulations.
Methodology, pages 9–15, 2007.
[124] Konstantin Miller and Adam Wolisz. Transport Optimization in Peer-to-Peer Networks. Proc. of the 19th Euromicro International Conference on Parallel, Distributed
and Network-Based Computing (PDP 2011), 2011.
[125] Jochen Mundinger, Richard Weber, and Gideon Weiss. Optimal scheduling of
peer-to-peer file dissemination. Journal of Scheduling, 11(2):105–120, August 2007.
[126] Internet Phenomena. Fall 2010 Global Internet Phenomena Report Same Destination , Different Routes : Converged Subscriber Expectations. Europe, 2010.
[127] Mohamed Karim Sbai, Emna Salhi, and Chadi Barakat. P2P content sharing in
spontaneous multi-hop wireless networks. Second International Conference on
COMmunication Systems and NETworks (COMSNETS), pages 1–10, January 2010.
[128] A Raniwala and T. Chiueh. Architecture and algorithms for an IEEE 802.11based multi-channel wireless mesh network. In IEEE International Conference on
Computer Communications (INFOCOM), volume 3, pages 2223–2234. IEEE, 2005.
152
REFERENCES
[129] Oh-Heum Kwon and Kyung-Yong Chwa. Multiple message broadcasting in communication networks. Networks, 26(4):253–261, December 1995.
[130] A Subramanian, Jing Cao, and Chul Sung. Understanding channel and interface
heterogeneity in multi-channel multi-radio wireless mesh networks. Passive and
Active Network, 2009.
[131] Mathias Kurth, Anatolij Zubow, and Jens-Peter Redlich. Multi-channel link-level
measurements in 802.11 mesh networks. Proceeding of the 2006 international conference on Communications and mobile computing - IWCMC ’06, page 937, 2006.
[132] Peter Dely and Andreas Kassler. Kaumesh a multi-radio multi-channel mesh
testbed. In Proceedings of 9th Scandinavian Workshop on Wireless Ad-hoc & Sensor Networks, 2009.
[133] Ajay Tirumala, Feng Qin, Jon Dugan, Jim Ferguson, and Kevin Gibbs. Iperf The TCP/UDP bandwidth measurement tool, 2004.
[134] Naval Research Laboratory.
MGEN:
http://pf.itd.nrl.navy.mil/mgen/mgen.html.
The
Multi-Generator.
[135] Vangelis Angelakis, Stefanos Papadakis, Vasilios Siris, Apostolos Traganitis, and
Networks Lab. Adjacent Channel Interference in 802.11a: Modeling and Testbed
Validation. In IEEE Radio and Wireless Symposium (RWS2008), number 2, Orlando, Florida, 2008.
[136] Ramakrishna Gummadi, David Wetherall, Ben Greenstein, and Srinivasan Seshan.
Understanding and mitigating the impact of RF interference on 802.11 networks.
ACM SIGCOMM Computer Communication Review, 37(4):385, October 2007.
[137] J Robinson, K Papagiannaki, C Diot, X Guo, and L Krishnamurthy. Experimenting with a multi-radio mesh networking testbed. In International workshop on
Wireless Network Measurements (WiNMee), pages 1–6. Citeseer, 2005.
[138] CM Cheng, PH Hsiao, and HT Kung. Adjacent Channel Interference in Dualradio 802.11a Nodes and Its Impact on Multi-hop Networking. IEEE Global
Telecommunications Conference (GLOBECOM), (Section II), 2006.
[139] Sebastian Robitzsch, Christian Niephaus, John Fitzpatrick, and Mathias
Kretschmer. Measurements and Evaluations for an IEEE 802.11a Based CarrierGrade Multi-radio Wireless Mesh Network Deployment. Fifth International Conference on Wireless and Mobile Communications, pages 272–278, 2009.
[140] Jens Nachtigall, Anatolij Zubow, and Jens-Peter Redlich. The Impact of Adjacent
Channel Interference in Multi-Radio Systems using IEEE 802.11. International
Wireless Communications and Mobile Computing Conference, pages 874–881, August 2008.
REFERENCES
153
[141] D Valerio, F Ricciato, and P Fuxjaeger. On the feasibility of IEEE 802.11 multichannel multi-hop mesh networks. Computer Communications, 31(8):1484–1496,
May 2008.
[142] John C Bicket. Bit-rate Selection in Wireless Networks. PhD thesis, MIT, 2005.
[143] Daniel Halperin, Wenjun Hu, and Anmol Sheth. Predictable 802.11 packet delivery from wireless channel measurements. ACM SIGCOMM Computer, 2010.
[144] Vijay Raman. Dealing with Adjacent Channel Interference effects in Multi-channel,
Multi-interface wireless networks. PhD thesis, University of Illinois at UrbanaChampaign, 2008.
[145] IEE E Computer Society. IEEE 802.11n. Part11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications. Amendment 5: Enhancements for Higher Throughput. Technical report, 2009.
[146] Vivek Shrivastava, Shravan Rayanchu, Jongwoon Yoonj, and Suman Banerjee.
802.11N Under the Microscope. Proceedings of the 8th ACM SIGCOMM conference on Internet measurement conference - IMC ’08, page 105, 2008.
[147] Ranveer Chandra, Ratul Mahajan, Thomas Moscibroda, Ramya Raghavendra, and
Paramvir Bahl. A case for adapting channel width in wireless networks. ACM
SIGCOMM Computer Communication Review, 38(4):135, October 2008.
[148] Madwifi.
Atheros
Registers.
project.org/wiki/DevDocs/AtherosRegisters., 2011.
http://madwifi-
[149] Arunesh Mishra, Vivek Shrivastava, Suman Banerjee, and William Arbaugh. Partially overlapped channels not considered harmful. ACM SIGMETRICS Performance Evaluation Review, 34(1):63, June 2006.
[150] Supratim Deb, V. Srinivasan, and R. Maheshwari. Dynamic spectrum access in
DTV whitespaces: design rules, architecture and algorithms. In Proceedings of the
15th annual international conference on Mobile computing and networking, pages
1–12. ACM, 2009.
Acronyms
ACI
ACK
ADSL
ALTO
AODV
AOMDV
ARA
BFS
BPSK
BSS
CBR
CCA
CCK
CCM
CD
CRC
CSI
CSMA/CA
CTS
DCF
DHT
DIFS
DSR
DSSS
EDSR
ETT
ETX
FCC
FHSS
GLPK
adjacent channel interference.
acknowledgment.
asymmetric digital subscriber line.
application-layer traffic optimization.
ad-hoc on-demand distance vector.
ad-hoc on-demand multipath distance vector.
ant-based routing.
breadth first search.
binary phase-shift keying.
basic service set.
constant bit-rate.
clear channel assessment.
complementary code keying.
channel cost metric.
collision domain.
cyclic redundancy check.
channel state information.
carrier sense multiple access with collision avoidance.
clear to send.
distributed coordination function.
distributed hash table.
distributed inter frame space.
dynamic source routing.
direct sequence spread spectrum.
enhanced dynamic source routing.
expected transmission time.
expected transmission count.
Federal Communications Commission.
frequency hopping spread spectrum.
GNU linear programming kit.
156
Acronyms
GPS
HNA
HT
HTTP
HWMP
iAWARE
IBSS
ID
IR
ISI
LRU
MAC
MANET
MCS
MGMRA
MIC
MILP
MIMO
MIND
MMRA
MPCP
MPP
MPR
MRA
NAT
NAV
NIC
OFDM
OLSR
P2P
PCF
PHY
PLL
PNS
QoS
RA-OLSR
RERR
RREP
RREQ
RSSI
RTS
RTT
SCIP
global positioning system.
host and network association.
high throughput.
hypertext transfer protocol.
hybrid wireless mesh protocol.
interference aware.
independent basic service set.
identifier.
infrared.
inter symbol interference.
least recent used.
medium access control.
mobile ad-hoc network.
modulation and coding scheme.
minimum guaranteed maximum rate allocation.
metric of interference and channel-switching.
mixed integer linear problem.
multiple-in multiple-out.
metric for interference and channel diversity.
maximum of minimum rate allocation.
mobile peer control protocol.
mobile peer-to-peer.
multi point relay.
max rate allocation.
network address translation.
network allocation vector.
network interface card.
orthogonal frequency division multiplexing.
optimized link state routing.
peer-to-peer.
point coordination function.
physical.
phase locked loop.
proximity neighbor selection.
quality of service.
radio aware optimized link state routing.
route error.
route reply.
route request.
received signal strength indication.
request to send.
round trip time.
solving constraint integer programs.
Acronyms
157
SHA
SIFS
SINR
SNR
SSR
TC
TCP
TDMA
TTL
URL
VoIP
VRR
WCETT
WLAN
WMN
ZP2P
secure hash algorithm.
short inter frame space.
signal to interference plus noise ratio.
signal-to-noise ratio.
scalable source routing.
topology control.
transport control protocol.
time division multiple access.
time to live.
uniform resource locator.
voice over IP.
virtual ring routing.
weighted cumulative ETT.
wireless local area network.
wireless mesh network.
zone-based P2P.
Enhancing P2P Systems over
Wireless Mesh Networks
Due to its ability to deliver scalable and fault-tolerant solutions, applications based on
the peer-to-peer (P2P) paradigm are used by millions of users on the internet. Recently,
wireless mesh networks (WMNs) have attracted a lot of interest from both academia and
industry, because of their potential to provide flexible and alternative broadband wireless
internet connectivity. However, due to various reasons such as unstable wireless link characteristics and multi-hop forwarding operation, the performance of current P2P systems
is rather low in WMNs.
This dissertation studies the technological challenges involved while deploying P2P systems over WMNs. We study the benefits of location-awareness and resource replication
to the P2P overlay while targeting efficient resource lookup in WMNs. We further propose a cross-layer information exchange between the P2P overlay and the WMN in order
to reduce resource lookup delay by augmenting the overlay routing table with physical
neighborhood and resource lookup history information.
Aiming to achieve throughput maximization and fairness in P2P systems, we model the
peer selection problem as a mathematical optimization problem by using a set of mixed
integer linear equations. A study of the model reveals the relationship between peer selection, resource replication and channel assignment on the performance of P2P systems
over WMNs. We extend the model by formulating the P2P download problem as chunk
scheduling problem. As a novelty, we introduce constraints to model the capacity limitations of the network due to the given routing and channel assignment strategy. Based on
the analysis of the model, we propose a new peer selection algorithm which incorporates
network load information and multi-path routing capability.
By conducting testbed experiments, we evaluate the achievable throughput in multi-channel multi-radio WMNs. We show that the adjacent channel interference (ACI) problem
in multi-radio systems can be mitigated, making better use of the available spectrum. Important lessons learned are also outlined in order to design practical channel and channel
bandwidth assignment algorithms in multi-channel multi-radio WMNs.
Karlstad University Studies
ISSN 1403-8099
ISBN 978-91-7063-398-0
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement