GATEWAY PLACEMENT AND FAULT TOLERANCE IN QOS AWARE WIRELESS MESH NETWORKS

Transcription

1 GATEWAY PLACEMENT AND FAULT TOLERANCE IN QOS AWARE WIRELESS MESH NETWORKS A dissertation submitted to Kent State University in partial fulfillment of the requirements for the degree of Doctor of Philosophy by Yasir Drabu Dec 2010

2 Dissertation written by Yasir Drabu B.S., Bangalore University, 1998 M.S., Kent State University, 2003 Ph.D., Kent State University, 2010 Approved by Dr. Hassan Peyravi, Chair, Doctoral Dissertation Committee Dr. Javed E. Khan, Member, Doctoral Dissertation Committee Dr. Feodor F. Dragan, Member, Doctoral Dissertation Committee Dr. Kazim Khan, Member, Doctoral Dissertation Committee Accepted by Dr. Jonathan I. Maletic, Chair, Department of Computer Science Dr. Timothy S. Moerland, Dean, College of Arts and Sciences ii

3 TABLE OF CONTENTS LIST OF FIGURES vii LIST OF TABLES ix Acknowledgments xii Dedication xiii 1 Introduction Motivation Classification of Wireless Mesh Networks Wireless Mesh Networks (WMN) Architecture Components of WMNs WMNs Compared to Mobile Ad Hoc Networks (MANETs) WMNs Compared to Sensor Networks Advantages of WMNs Self organizing and self configuring Low deployment costs Increased reliability Scalability Interoperability Applications of WMNs iii

4 1.5.1 Broadband Wireless Access Industrial Applications Future Applications: Research Contributions Structure of this Dissertation Background and Related work Introduction WMN Standardization IEEE s WiFi Mesh IEEE Bluetooth IEEE Zigbee IEEE WiMAX WMN Deployments Academic Testbeds Commercial Implementations Problems and Challenges in WMNs Phycial Layer Issues Medium Access Layer Issues Transport Layer Issues Network Layer Issues Topological and Deployment Issues Gateway Placement in WMNs iv

5 3.1 Introduction Background and Related Work System Model and Assumptions Clustering with QoS Constraints Gateway Placement Algorithms Integer Linear Program Formulation Split-Merge-Shift Algorithm (SMS) Algorithm Illustration Algorithm Analysis Performance Summary Fault Tolerance in Cellular WMNs Background and Related Work An Augmented Tri-Sectorized System with Directional Antennas Sectorized cells directional antennas Honeycomb Networks Routing in Honeycomb Mesh QoS Routing Fault Tolerant Routing Fault Models Comparision of Routing Schemes Fault Tolerance in WMNs v

6 5.1 Background and Related Work System Model Network and QoS Model Fault Model Fault Recovery Topology Control Fault Recovery Algorithm Algorithm Evaluation Simulation Results Summary Falut-Tolerance Provsioning Through Cluster Overlapping in General WMNs Introduction Research Issues Preliminaries Planning with Disjoint Clustering Analysis Planning with Joint Clustering Simulation Results Summary Conclusion and Future Work BIBLIOGRAPHY vi

7 LIST OF FIGURES 1.1 Classification of Wireless Networks Infrastructure mesh network Survey of broadband access in rural and urban areas [47] Worldwide Internet usage peneration for 2010 (Courtesy: Internet World Stats) Roofnet testbed setup at Kent State University, Dept. of Computer Science Effect of distance on throughput, Roofnet testbed Effect of hops on throughput, Roofnet testbed Simulation setup to study effect of hop distance Load vs. throughput for varying hop distance Relay Load Flows Algorithm overview Shift operation overview Effect of S q on gateway count Effect of L q on gateway count Effect of S q on cluster size variation (a) Omni directional antennas, (b) Tri-sectorized antenna Transforming a Regular Cellular structure to a Trisectional Directional Cellular System vii

8 4.18 Tri-sectored node with three directional antennas and one omni directional antenna Brick network representation of a honeycomb network Detours, d and d, around faults Routing around a link failure A clustered wireless mesh with the overlay cluster graph Fault recovery algorithm working with three cases Effect of fault location on recovery probability (50 node network) Load Vs Throughput with two faults and CBR traffic Load Vs Delay Load Vs Jitter Maximum throughput performance across different traffic loads A linear multi-hop network A 100-node mesh network Initial clustering with h = 2 and I c = Inter-cluster distance Final disjoint clustering, 2 h 4, I c = Joint clustering with R = 2, I c = h = 1, I c = h = 1, I c = h = 2, I c = Overlapping clusters simulation setup Load vs. throughput with and without overlapping viii

9 LIST OF TABLES 1.1 Network architectures Broadband Access Technology ix

10 Acronyms 3G 4G AODV BSS DOCSIS DSL DSR ESS HWMP IP Kbps LAN LQSR MAC MAN MANET Mbps MCL Third Generation Wireless Network Fourth Generation Wireless Network Ad Hoc On-Demand Distance Vector Basic Service Set, as defined in IEEE family of protocols Data Over Cable Service Interface Specification Digital Subscriber Line Dynamic Source Routing Extended Service Set, as defined in s draft IEEE standard Hybrid Wireless Mesh Protocol Internet Protocol Kilo Bits per Second Local Area Network Link Quality Source Routing Medium Access Control Metropolitan Area Network Mobile Ad-hoc Network Mega Bits Per Second Mesh Connectivity Layer x

11 MIMO MTM ORBIT PAN PDA PTM PTP QoS SDR STA SNR TCP WAN Wifi WMN WWW Multiple Input Multiple Output Multipoint to Multipoint Open-Access Research Testbed for Next-Generation Wireless Networks Personal Area Network Personal Digital Assistants Point to Multipoint Point to Point Quality of Service Software Defined Radio Mesh Station, as defined in s draft IEEE standard Signal to Noise Ratio Transport Control Protocol Wide Area Network Wireless Fidelity, refers to the family of protocols Wireless Mesh Networks World Wide Web xi

12 Acknowledgments I would like to thank my supervisor Dr. Hassan Peyravi for his guidance and support, during my research at the Internet Engineering Lab at Kent State University. He has inspired me to think critically and help me hone my analytical abilities. His hands on approach towards experimentation and simulation formed a corner stone of my research and have saved me countless hours in evaluating our algorithms and protocols. I would also like to thank my committee members, Dr. Javed Khan and Dr. Feador Dragan and Dr. Kazim Khan for their feedback. I would also like to thank my parents and my wife, Sabah Drabu, who have been a constant source of moral support and constant encouragement. I would also like to extend my special thanks to Sue Peti and Marcy Curtiss for their help throughout the years as a graduate student. I am grateful to the Internet2 foundation and the Ohio Board of Regents for supporting parts of this research. xii

13 To Sabah and Murad xiii

14 Abstract Wireless Mesh Networks (WMN s), in the form of WiFi (802.11x) or WiMax (802.16x), or their integrations have been proposed as an effective communication alternative for ubiquitous last mile wireless broadband access. They can be viewed as a hybrid between traditional cellular, point-to-point wireless systems, and ad-hoc networks. They offer more flexibility, mobility, coverage and expandability compared to their traditional counterparts at the expense of complex architecture and deployment structure. Though WMNs hold great promise in abetting network ubiquity, there still remain several challenges in the design and development of WMN s to support diverse services with different quality of service (QoS) requirements and large scale deployment. The focus of this dissertation is to address some of the core issues that directly affect the QoS in terms of delay, throughput, and fault tolerance. First we look at the deployment problem of the placement of wired gateways. This aspect of WMNs has a significant impact on the network s throughput performance, cost and capacity to satisfying the quality of service requirements. In the context of gateway placement, the QoS is influenced by the number of gateways, the number of nodes served by each gateway, the location of the gateways, and the relay load on each wireless router. While finding an optimal solution to simultaneously satisfy all the above constraints is known to be an N P-hard problem, near optimal solutions can be found within the feasibility region in polynomial time using various heuristic methods. In the initial part of this dissertation, we first present a near optimal heuristics algorithm for gateway

15 2 placement that facilitates QoS provisioning and fault tolerance in WMNs. We then investigate fault tolerance and recovery problems in WMNs. We present a fault recovery algorithm that can exploit the known geometry of a regular cellular mesh network. While keeping the QoS metrics intact, we consider a post-deployment fault recovery algorithm and pre-deployment fault tolerance planning.

16 CHAPTER 1 Introduction 1.1 Motivation One of the ultimate goals in the field of networking is to have ubiquitous broadband access for all, i.e. any time, any place access with satisfactory Quality of Service (QoS). Asymmetric and symmetric Digital Subscriber Line (xdsl) and Data Over Cable Service Interface Specification (DOCSIS) have fueled broadband access to many densely populated areas - residential as well as businesses. However their cost of deployment and the nature of wired medium limits their coverage and access to a broader spectrum of people. Wireless network systems, by virtue of their medium can provide mobility and wider coverage. Their usage is well established in narrow-band access systems, however their foray into broadband is relatively new. Wireless networks have changed the way we work and play. They have changed every facet of our lives to make it easier to communicate with our peers. Wireless networks tend to form a natural extension to the way we communicate making them more imbued in our daily lives. Researchers are pushing the envelope by improving the performance of traditional cellular networks. Great strides are are being made in bandwidth capacity in 3G and 4G networks. Today individuals and businesses use mobile devices like laptops, Personal Digital Assistants (PDAs), smart phones, etc, to send text, audio and video messages. Many use it to listen to streaming music and watch Video on Demand (VoD). As technology progresses we will see a greater demand for smarter and more efficient 1

17 2 devices coupled with much faster wireless networks to address and aid the human appetite for information, communication and entertainment. In spite of all these advances, current wireless networks are marred with limited coverage, low throughput, unreliable connections, security concerns, and power limitation. Therefore, it is vital to build an economic, scalable and fault tolerant wireless system that can meet the exponentially growing demand. While improving existing wireless networks, new architectures like multi-point to multi-point wireless networks are being actively investigated and are providing solutions to some very challenging problems like the economics of last mile connectivity, community and neighborhood networks and other interesting applications. In this research, we focus on the deployment issues and challenges of wireless mesh networks.we specifically look into the pre-deployment planning in term of optimal placement of gateways and QoS provisioning, and post-deployment operation in terms of fault-tolerance and network connectivity. 1.2 Classification of Wireless Mesh Networks Wireless networks can be classified based on several criteria - like size, topology, power levels, applications or protocols used. In the context of this dissertation we classify WMNs based on the kind of connectivity between the various network elements. Broadly speaking, current wireless networks are either Point to Point (PTP) or Point to Multi- Point (PTM) networks. The complete taxonomy of this classification is show in the Figure 1.1. PTP networks are highly reliable but have low adaptability and they are not scalable.

18 3 Wireless Networks Single Hop (PTM) Point To Point (PTP) Multi-hop (MTM) Infrastructure-based (hub&spoke) Infrastructure-less (ad-hoc) Infrastructure-based (Hybrid) Infrastructure-less (MANET) Bluetooth Cellular Networks (GSM/CDMA) Wireless Sensor Networks Wireless Mesh Networks Car-to-car Networks (VANETs) Figure 1.1: Classification of Wireless Networks PTM are moderately scalable, but they have low reliability and adaptability. To overcome these limitations Multi-point to Multi-point (MTM) networks are evolving rapidly and they are offering features that provide high reliability, adaptability and scalability to accommodate a large number of users. The characteristics of each network are tabulated in Table 1.1. Table 1.1: Network architectures Topology Reliability Adaptability Scalability Routing Point-to-Point High Low None None Point-to-Multi-point Low Low Moderate Moderate Multi-point-to-Multi-point High High High High MTM wireless networks seem like a natural fit for wide area coverage. Using multiple hops provides increased coverage without the need for increasing the transmission power. As the number of nodes in the network increases the transmission power needed for each interface can be reduced. In this model if P t is the transmission power and d is

19 the transmission radius, P t d α, where 2 α 4. According to this model if the transmission distance (d) is decreased by a factor of 2, the transmission power (P ) will 4 decrease by a factor of 4 or 16. In this research we focus on a special class of PTP network called Wireless Mesh Networks. MTM wireless networks can also be realized using today s standard commodity based wireless networking equipment like the IEEE family of protocols [1]. These type of networks are called community mesh networks [12]. Projects like the Roofnet [17] address the setup and routing problems of wireless mesh networks based on b/g. Roofnet s proactive routing protocol probes the network for link quality and topology changes to provide the best possible route. Most of these commodity wireless technologies operate using a single radio and a shared channel. Most of the early protocols proposed were for Point-to-point wireless networks, that essentially extended the range of wired networks. However in the mid 1990s, with the commercialization of wireless radios, multiple hop wireless networks started to garner interest in solving many interesting problems like the last mile connectivity and ad-hoc networking. For example, community mesh networks like [33, 51] are working towards provide co-operative wireless broadband access to its users in their community. While the network formed using such technologies provide acceptable performance on smaller networks, they do not scale well for larger networks which have higher node density and multiple hops.

20 5 1.3 Wireless Mesh Networks (WMN) Architecture WMNs are multi-hop infrastructure based wireless networks that are interconnected by a set of relatively stationary wired gateways connected to the Internet. The routers that relay traffic and the client may or may not be mobile. Most of the traffic in a WMN flows from the client to the gateways. So, the traffic pattern may be asymmetric. A typical WMN is shown in Figure 1.2. Figure 1.2: Infrastructure mesh network Components of WMNs As mentioned, a wireless mesh network consist of three types of nodes: WMN Clients: These are the end-user devices like PDAs, laptops, smart phones, etc, that can access the network for using applications like , web surfing, VoIP, and alike. These devices are assumed to have limited power, mobile, having none

21 6 or limited routing capabilities, and may or may not be always connected to the network. Mobile Ad-hoc Networks (MANETs) can be assumed to be special case of WMNs that are formed purely by WMN clients. WMN Routers: These network elements are primarily responsible for routing traffic in the network. Traffic does not originate or terminate at a router. The routers are characterized by limited mobility and relatively high reliability. Compared with a conventional wireless routers, a wireless mesh router can achieve the same coverage with much lower transmission power consumption through multi-hop communications. Additionally, the Medium Access Control (MAC) protocol in a mesh router supports multiple-channels and multiple interfaces to enable scalability in a multi-hop mesh environment. With the capability of self-organization and selfconfiguration, WMNs can be deployed incrementally, one node at a time. WMN Gateways: This is a router which has direct access to the wired infrastructure (i.e. Internet). Most of the client nodes on a WMN, communicate with the wired infrastructure [5] for information deliverted over the World Wide Web (WWW), s, audio and video. Since they have multiple interfaces (wired and wireless), the gateways are typically more expensive, both to install and operate. Typically they are fewer in number and their placement has a significant impact the performance of the network. The distributed nature of WMN infrastructure in terms of control and coordination and multi-hop and multi-point connectivity enables development of self-configuration and self-healing capabilities. They do not have a centralize controller like today s cellular

22 7 networks, therefore no central point of failure or bottleneck exists. Further since they can communicate wirelessly they are more economical to deploy. In terms of requirements and design principals, WMNs are ideal to meet the growing demands of the wireless community. However, fundamental technical challenges need to be addressed before they become commercially viable. There is a lot of commercial and academic interest in this area due to the impact this technology can have on the economics of access networks. WMNs also have the potential to bring much larger throughput to the last mile in sparsely populated areas and rural areas which have limited economic resources WMNs Compared to Mobile Ad Hoc Networks (MANETs) Wireless Mesh Networks share a lot of properties with MANETs. The most important characteristic is the that nodes are connected without any wired infrastructure and routing is done across multiple hops in both networks. However, the similarity ends there. Mobile Ad hoc Network (MANET) assume all nodes have similar functionality in terms of packet forwarding and other node attributes like power, mobility and reliability. On the other hand, WMNs have different network elements like client, wireless router and gateways with wide range of practical applications. Thus ad hoc networks can be considered as a special case of WMNs where only the client nodes are connected. MANET may or may not depend on wired infrastructure, where as mesh networks have wired gateways that they rely on. Further, the traffic in a MANET is between users where as in WMNs the traffic is between the client/user and the gateway. Last, but not

23 8 least, MANNET nodes typically have higher mobility in comparison to WMN routers WMNs Compared to Sensor Networks Sensor networks and WMNs are both multi-hop networks with intermediate nodes acting as routers. Both direct traffic from the client (users or sensors) towards a gateway or data acquisition device. However, there are significant difference mainly due to the intent and applications. Typically, sensor networks have low bandwidth in order of tens of Kilo Bits Per Second (Kbps) whereas WMNs typically are designed for higher bandwidth in excess of 1 Mbps. Sensor network nodes are constrained by power, hence power efficiency is a big design consideration. However, in WMNs power efficiently, while recommended, is not a design constraint. Finally, due to the fact that sensor networks collect data within their fixed sensory radius, they are normally stationary. In the case of WMNs the client nodes can be highly mobile. 1.4 Advantages of WMNs Self organizing and self configuring While dependant on the implementation and the protocols used, the ultimate goal of WMNs is to be self healing and self configuring. This reduces the setup time and maintenance cost. This also allows the network service providers the ability to change and adapt the network to meet the demands of the end users. Wireless mesh nodes are easy to install and un-install, making the network extremely adaptable and expandable as more or less coverage is needed.

24 Low deployment costs Since WMNs use wireless only routers, deploying them over larger areas of coverage is cheaper compared to single hop cellular networks. This is primarily due to fewer wired routers/access points that are more expensive to install and maintain. This coupled with easier maintenance leads to a lower operation cost Increased reliability WMNs have multiple paths between any given source and destination node. This allows for alternate routing in case the currently used route fails. Alternate routes can also serve as a means to balance the network load. If part of a network becomes congested, a communication pair can chose an alternate path thus minimizing bottlenecks. Load balancing via alternate routing, if implemented in WMNs, can significantly increase network reliability Scalability Unlike traditional wireless networks, as the number of nodes increases in a WMN, the greater its transmission capacity become. This is mainly achieved by better load balancing traffic along alternate routes. In some configurations, WMNs allow local networks to run faster, because local packets don t have to travel back to a central server. Obviously this is limited by the configuration and protocols that manage the contention domain of the medium.

25 Interoperability WMNs can work with existing WiFi and WiMaX standards, thus making them very attractive for incremental deployment and reuse of existing infrastructure. The IEEE set of standards actually has protocols defined that enables it to be configured as a WMN. Futher s, as discussed in Section defines protocols that will enable mesh networking on existing WiFi protocols. While it is necessary to improve protocols across most layers of the network stack, - most of the improvements can augment the current standards to maintain interoperability. 1.5 Applications of WMNs Broadband Wireless Access Broadband access has become critical for today s information economy. It enables applications like video on demand, video telephony, online-gaming and telecommunications. Each new application has a significant impact on quality of life. Telecommuting in particular can lead to increased productivity due to the time saved traveling back and forth from the place of work. It also reduces traffic on the roads during busy hours, thus having a positive impact on the environment. Studies have shown that people with broadband access versus those that use dial-up/narrow-band are more likely to publish content online and participate in social and community based activities [48]. Present day back-haul bandwidth is very large, in the order of 10 6 Gbps, primarily due to the capacity of optical fiber cables. However providing access to this large bandwidth to the end user needs expensive termination into each and every home and office. The current last mile technologies are highlighted in Table 1.2.

26 11 Technology Advantage Disadvantage Fiber to customer Very high bandwidth Very expensive - infrastructure costs and expensive Customer Premise equipment Cable/DOCSIS Uses existing infrastructure Shared bandwidth limits quality of service, reliability and scalability xdsl Uses existing copper Distance for Central office, un-bundling the local loop, not all copper is up to specification Satellite Access anywhere High investment cost, shared bandwidth, higher latency Wireless Low cost and fast to deploy Needs line of sight Table 1.2: Broadband Access Technology It is clear that wired access like Cable and DSL are economically feasible only in urban and sub-urban areas with a reasonably high population density. Rural area have limited coverage using wireless technologies like satellite and cellular networks. Satellite access is expensive and has much higher latency due to the distance between the end client and the satellite. In the case of cellular networks the towers are expensive to install and operate. As can be seen in Figure 1.3, rural areas in the US, have a much lower broadband access penetration - primarily due to the lack of service providers and the higher cost of the service itself. Further the problem of access is exacerbated, when considering access disparity at a global level as shown in Figure 1.4. Inadequate Internet access in rural areas and developing countries creates a greater digital divide between well connected towns and cities and rural areas, thus limits social communication and puts rural businesses at an disadvantage [36]. To foster the wider adoption of Internet access, WMNs offer a novel and cost effective

27 12 Figure 13: Main Reason for No High-Speed Internet Use at Home, Rural/Urban, 2009 Can Use Somewhere Else 3.6% Rural: 10.9 Million Households Not Available 11.1% No Computer or Computer Inadequate 16.3% Lack of Skill 2.3% Don't Need/Not Interested 37.7% Can Use Somewhere Else 4.7% Urban: 32.1 Million Households Not Available 1.1% No Computer or Computer Inadequate 19.0% Lack of Skill 3.2% Don't Need/Not Interested 38.1% Other 6.4% Too Expensive 22.3% Other 6.2% Too Expensive 27.6% Figure 1.3: Survey of broadband access in rural and urban areas [47] When other types of non-use are examined, however, the rankings can and do change. For example, respondents who do not use the Internet anywhere ranked the value proposition significantly higher than affordability. Figure 14: Main Reason Given for No Internet Use at Any Location, 2009 No Computer or Computer Inadequate 22.3% Lack of Skill 4.3% Too Expensive 18.6% Other 5.5% Not Available 0.7% Can Use Somewhere Else 1.4% Don t Need/ Not Interested 47.2% This contrasts with the category of households that do not access the Internet at home, which rated cost as the clear-cut top concern. Figure 1.4: Worldwide Internet usage peneration for 2010 (Courtesy: Internet World Stats) 13

28 alternative in areas where cable TV or DSL lines aren t available. They can be deployed 13 quickly and the service provider can see a quick return on investment. Other than network service providers, community mesh networks like [61] are also ways to provide broadband access to a local community. City wide municipal area networks that provide free broadband access to its residents are also become more feasible due to the lower deployment and operational costs of mesh networks. Many such networks already exist, and more are on the way Industrial Applications Healthcare: Many hospitals are spread out through clusters of densely constructed buildings that were not built with computer networks in mind. Wireless mesh nodes can sneak around corners and send signals short distances through thick glass to ensure access in every operating room, lab and office. The ability to connect to the network is crucial as more doctors and care-givers maintain and update patient information test results, medical history, even insurance information on portable electronic devices carried from room to room. Hospitality: High-speed Internet connectivity at hotels and resorts has become the rule, not the exception. Wireless mesh networks are quick and easy to set up indoors and outdoors without having to remodel existing structures or disrupt business. Temporary Venues: Construction sites can capitalize on the easy set-up and removal of wireless mesh networks. Architects and engineers can stay wired to the office, and Ethernet-powered surveillance cameras can decrease theft and vandalism. Mesh nodes can be moved around and supplemented as the construction project progresses.

29 14 Other temporary venues like street fairs, outdoor concerts and political rallies can set-up and tear down wireless mesh networks in minutes. Warehouses: There is simply no effective way to keep track of stock and shipping logistics without the types of Ethernet-enabled handheld scanners used in modern warehouses. Wireless mesh networks can ensure connectivity throughout a huge warehouse structure with little effort Future Applications: The U.S. military, which helped develop wireless mesh technology, foresees a day when thousands of microchip-size mesh nodes can be dropped onto a battlefield to set up instant scouting and surveillance networks. Information can be routed to both ground troops and headquarter personnel 1.6 Research Contributions In this research we have focused on the deployment, fault tolerance and quality of service issues in WMNs. We first investigated and developed a new heuristic which is able to places wired gateways subject to multiple constraints on delay and relay load. We then studied and developed routing and fault recovery algorithms in WMNs with a known brick topology. We also developed a probabilistic fault recover algorithm in more generalized WMNs while maintaining the QoS. The contribution of this dissertation is three-fold in both pre-deployment and postdeployment of wireless mesh networks under some QoS constraints. First, we investigate the problem of optimal placement of gateways in wireless mesh network with some QoS constraints including hop-count delays, relay capacity, and gateway (cluster) capacity.

30 15 We introduce a pre-deployment clustering algorithm for wireless mesh network that incorporates the QoS constraints into the clustering technique, and then incorporates links and node faults in post-deployment while keeping the QoS metrics intact. We further analyses the clustering approach in terms of its complexity and its limitation. This result has been published in [24] and [22, 25]. Second, we introduce a new pre-deployment near optimal gateway placement by finding the maximal clique among all disjoint clusters, while keeping the hop count constant to control the delay. In order to support fault tolerance for distant nodes from a cluster head (gateway), we allowed inter-cluster nodes to join neighboring clusters. We also considered overlapping clustering in which a node dynamically seeks membership in neighbor clusters. This results has been submitted in [26]. Third, we studied the problem of fault-tolerance has been investigated in the context or wireless cellular mesh network. This result has been published in [23] 1.7 Structure of this Dissertation In chapter 2, we discuss wireless mesh networks in depth. We see what makes them unique. We look at some of standardization efforts, review the enabling technologies and the various open research issue in deploying WMNs. We also look at some of the academic and commertial implementations. In chapter 3 we investigate one of the fundamental problems of placement of wired gateways in WMN deployment. We propose a multi-constraint QoS aware linear equation and present a heuristic to solve the equations. our algorithm to those proposed in literation. We also compare the performance of We show through simulation that our approach produces better placement of gateways in terms of the number of gateways

31 16 and overall network performance. Using wired gateways, typical deployments of WMNs use an overlay spanning tree that is associated with a wired gateway to deliver the mesh traffic. While this approach makes Quality of Service (QoS) and deployment more attainable, it introduces sparseness in the routing paths and that limits the alternate paths available to recover from failure. In chapter 4 we look a a special case of wireless Mesh networks, that make use of directional antennas to form a network topology with a known geometric properties. We propose and design a fault-tolerant WMN based on a hexagonal topology with multiradio and directional antennas. For this model, we introduce an addressing and routing scheme that simplifies the network operations. Further, we extend the routing approach to cope with one or multiple network link failure, as link failure is common in wireless networks. To address this, we exploit the regularity and multi-path characteristics of an augmented tri-sectioned hexagonal system to route around link or node failures. In chapter 5 we investigate fault tolerance in a WMN without any assumption about the topological properties. As mentioned, the overlay of spanning trees introduces sparseness in a WNM. We propose a reactive fault recovery algorithm that works on these overlay based WMNs. The algorithm attempts to recover from failed links or nodes by making minimally localized path and clustering changes to the network topology, while maintaining the initial QoS constraints. Through simulation, we show that using our approach makes WMNs more resilient to faults and gives better throughput/delay performance when compared to similar fault scenarios in a non-overlay WMN. In chapter 6 we a pre-deployment fault tolerance approach using clustering with overlapping. Based on our simulations we observe that nodes that are further way from

32 17 the gateway have are more likely to suffer a link failure. This is due to the increase number of hops and a larger contention domain. We present algorithms that allow for the boundary nodes to be members of adjacent clusters. This allows them to form alternate paths in case of congestion or link failure to their primary cluster gateway. Finally in Chapter 7 we summarize our finding and discuss ideas on how to extend this research.

33 CHAPTER 2 Background and Related work 2.1 Introduction As we have discussed in Chapter 1, the demand for ubiquitous and seamless broadband wireless Internet access has been the major driving force behind the development of multi-hop wireless mesh networks (WMNs). In this chapter we discuss the standardization efforts, academic testbeads and commercial deployments, and the problems and challenges associated with the performance of WMNs. While packet radio networks [6] have been extensively researched in different contexts for several decades, wireless packet networks pose several challenges to deploy, mainly due to the unreliable medium, limitation of radio technology, and power management. The hidden and exposed nodes problems, initially identified in [64], have been addressed by a 3-way hand shake in the IEEE family of protocols. Most early architectures were based on point-to-multipoint communications. In these networks, a wireless base station (access point) provides connectivity to a wired backbone network and served as an extension to the wired infrastructure. However, in the recent years, due to the advances in radio technology and lower communication cost, wireless mesh networks have been investigated as an alternative for point-to-multipoint and multipoint-to-multipoint systems. 18

34 WMN Standardization IEEE s WiFi Mesh The family of standards (802.11a/b/g) where intended to extend the range and connection flexibility of wired networks. Consequently, they were designed only for single hop operations. Networks based on this standard did not have any aspect that would enable the working of multi-hop networks. To overcome this limitation many companies like developed their own WMN solutions due to the flexibility and cost effectiveness WMNs offer compared to typical WiFi networks. Though most of these products enable the end client to connect using standard MAC, they are not interoperable. To fulfill the need and extend the coverage of , the work group was formed in The goal of the s standard include: Increasing range/coverage of existing networks Increases throughput and performance Increased flexibility in deployment and use Provide seamless security Maintain backward compatibility with the existing standards The s standard defines an Extended Service Set (ESS), and a Wireless Distribution System (WDS) that provides a protocol for auto-configuring paths between APs. over self-configuring multi-hop topologies in a WDS to support both broadcast/multicast and unicast traffic s has three key components:

35 20 1. Mesh Portal(MP): MP acts as a gateway/bridge to external networks. 2. Mesh STA (station): STA relays frames hop-by-hop in a router-like fashion. 3. Mesh AP (Access Point): AP provides relaying functions as well as the connectivity services for clients. The s protocols defines how the APs can discover the Mesh station (STA). It further defines an extensible path selection framework that allows routing between various STAs using Hybrid Wireless Mesh Protocol (HWMP) which is a combination of Ad hoc On Demand Distance Vector (AODV) [50] and tree-based proactive routing algorithms s also includes mechanisms to provide deterministic network access, a framework for congestion control, power saving and Quality of Service. As of July 2010, the standard is in a 6th version of the draft and the task group is working towards finalizing the standard by early 2011 [2] IEEE Bluetooth Bluetooth was developed as a replacement for wires and initially was more of a wire replacement technology and not a real networking standard. It was then standardized by the IEEE task group to form a Personal Area Network (PAN). The standard define the MAC and PHY protocols that target a bit rate of up to 1Mbps. This standard is not suitable for mesh networking due to low bandwidth and limited hardware support. However it does have a provision for defining multi-hop scatternets that allow for creating small mesh network that allow multiple devices within a very short range connect to each

36 21 other IEEE Zigbee Zigbee was initially proposed by Motorola as a way to support a class of sensory networks that have multi-month to multi-year lifetime using small batteries. This standard is used to create low data rate (20-250Kbps) sensory or Personal Area Networks. This standard supports mesh topology by defining a coordinator that is responsible for setting up the multi-hop network. This standard is very suitable for setting up sensor mesh networks IEEE WiMAX a is a wireless communications specification for metropolitan area networks (MANs). It was approved in January 2003 and released in April 2003 as part of a set of standards known as or WiMax. The standards complement the older (WiFi) family of specifications. The a standard was developed for wireless MANs operating on licensed and unlicensed radio-frequency (RF) bands between 2 GHz and 11 GHz, at data speeds of up to 75 megabits per second (Mbps), with low latency and efficient use of spectrum space. The security of is enhanced by encryption features. Forward error correction (FEC) and space/time coding optimize accuracy under marginal signal conditions. The maximum range can be extended to approximately 30 miles (48 kilometers) with some sacrifice in throughput. The a specification is ideally suited for advanced communications methods such as voice over IP (VoIP) and prioritized data traffic.

37 WMN Deployments Several academic institutes have created created mesh network testb beds. Additionally, comapanies have created software and equipment for commercial deployments. We list some of the most notable testbeds and companies below Academic Testbeds Roofnet (MIT) Roofnet is an experimental b/g mesh network in development at MIT which provides broadband Internet access to users in Cambridge, MA. There are currently around 40 active nodes in the network. Roofnet defines a routing and MAC protocols that enable based nodes to work in a multi-hop environment. The protocols include link-level measurements of , finding high-throughput routes in the face of lossy links, adaptive bit-rate selection. The software is available for public usage and new protocols are being developed that take advantage of radio s properties. We have also used this for our experimental test bed that is covered in Section Emulab (University of Utah) Emulab is a network test bed, setup initial at the University of Utah. The name Emulab refers both to a facility and to a software system. It provides three experimental environments - simulated, emulated, and infrastructure. Emulab unifies all of these environments under a common user interface, and integrates them into a common framework. This framework provides abstractions, services, and namespaces common to all, such as allocation and naming of nodes and links [63]. Currently the test bed provides access to hundreds of PCs (168 as of this writing), several with wireless NICs ( a/b/g) and a

38 wide-area network nodes geographically distributed across approximately 30 sites. 23 ORBIT (Rutgers Winlab) The ORBIT (Open-Access Research Testbed for Next-Generation Wireless Networks) radio grid emulator is an indoor wireless network test bed. ORBIT is a two-tier laboratory emulator/field trial network test bed designed to achieve reproducibility of experimentation, while also supporting evaluation of protocols and applications in real-world settings. The project was started in September 2003 under the NSF Network Research Test beds (NRT) program that was a collaborative effort of Rutgers, Columbia, Princeton, Lucent Bell Labs, Thomson and IBM Research. It currently has a wireless system of over 400 nodes supporting x set of protocols. Technology For All(TFA) (Rice University) The TFA project aims to develop fundamental information technology advances that address the unique needs of underserved communities and developing regions [62]. Since 2004, the project has operated a research test bed in an under-resourced community in Houston s East End. The network serves over 4,000 community residents via fully programmable network nodes. Community residents can access the network via legacy devices or custom instrumented mobile phones. The test bed jointly serves as a community resource and a platform for test-driving the above research advances Commercial Implementations The flexibility of deployment and the business opportunity, many companies have developed WMN solutions by extending standard protocols. Most of these are proprietary

39 24 extensions and are not interoperable between two different implementations. Tropos Networks Tropos is based in Sunnyvale, California and focuses on providing broadband mesh networks for cities and providing distributed area networks for building and utilities. They have over 27 patents that are productized as wireless mesh router hardware and a Tropos Mesh Operating System. The operating system creates a self-organizing and self-healing wireless mesh topology and intelligently selects the optimum path through the network. It leverages the router s on-board intelligence to monitor and maximize performance, minimizing network congestion and adapting in real-time basis to interference and other variables common in an outdoor environment. Currently, Tropos has several network deployments including Avista Utilities, Oklahoma City wireless network and Google s network in Mountain View, California. BelAir Networks BelAir is based in Ontario, Canada and has developed a patented WMN architecture. They provide b coverage for large zones - like large hotels, municipalities, etc. BelAir use wireless mesh routers (like BelAir 200) that have three radios and eight fixed directional antennas. They have implemented dynamic transmission power control and use the directional antennas to communicate with other mesh routers. They also use cross layer optimization to improve routing, by getting congestion and latency feedback from the PHY layer. They have many small and large mesh network deployments in educational institutes, municipal governments, hotels and other wireless internet access providers.

40 25 Firetide Firetide is a provider of hardware and software for wireless infrastructure mesh networks, based in Los Gatos, California. They have a software called HotView Pro that centrally manages the mesh routes that are based on their hardware platform and Mobility Controller that enables mesh networks to work on mobile units like mass transport. They also have developed a proprietary AutoMesh routing protocol that not only provides routing but also load balancing and congestion control. They have installations in airports like Singapore International Airport, mass transit system like Seoul Metropolitan Subway and educational institutions like California State University, Long Beach. Mesh Dynamics Mesh Dynamics is based in Santa Clara, California and has developed software and hardware to deploy wireless mesh networks. Their networks are compatible with a/b/g. They have developed proprietary software that uses multiple radios (1 to 4 radios) to create a dynamic tree tropology to implement WMNs. They also have developed protocols for dynamic channel selection, enabling them to maintain high level of throughput over multiple hops. Their technology is radio agnostic and they provide a centrally managed software to monitor the network. Mesh Dynamics have deployments in the area of video surveillance, broadband access, public safety and campus networks. They have broadband access deployments in Fresno, California and Red River, New Mexico where the subscriber density is very low and wired networks are not economically feasible.

41 26 Microsoft Microsoft has actively been researching WMN with an application focus on community access networks. They have developed software called Mesh Connectivity Layer (MCL) that enables Windows based computers to use multiple radio cards to increase throughput and act as mesh routers. They have developed a routing protocol called Link Quality Source Routing (LQSR) which is based on Dynamic Source Routing (DSR). It has been designed to be transparent to higher and lower layers and therefore can be used with existing software and hardware. Currently, there are no commercial implementations or products based on MCL. Apart for these companies - Intel, Philips, Motorola, Cisco and many other companies are working towards the standardization of mesh networks in the form of IEEE s family of protocols. 2.4 Problems and Challenges in WMNs WMNs have been made possible due to significant advances in current technology and the maturing of protocols. While siginificant advances have been made to make WMNs possible, many problems that still remain, to fully realized their potential. There challenges are at different layers of a WMN, like capacity management, fairness, addressing and routing, mobility management, energy management, service levels, integration with the Internet, etc. Many of these aspects have been briefly discussed in [34] Phycial Layer Issues The radio models used in modeling a WMN fundamentally define how the protocols and algorithms would work to develop an efficient wireless network. Each of these models

42 has a unique set of properties that offer different design issues. Below are listed the most common radio models in use today and their charateristics Single Radio Single Channel: Most of today s typical data networks based on IEEE use a single half duplex radio and use a MAC protocol to share a single channel. In this environment, when one node transmits all other nodes have to listen and cannot transmit without causing a collision. Further, since single radio nodes are half duplex, i.e., they cannot transmit and receive a signal simultaneously, the bandwidth utilization is significantly reduced. In most of the current research, hardware and software augmentation is used to incrementally improve the performance of WMN based on this model. 2. Single Radio Multiple Channels (SR-MC): Single Radios can tune into several nonoverlapping channels. For example in b there are 3 non-overlapping or orthogonal channels but only one is used by a network at any given time a has 11 non-overlapping channels, thus brining forth the possibility of using channel diversity to reduce contention zones and increase the capacity and throughput existing single radio based wireless networks. 3. Multiple Radio Multiple Channels (MC-MR): In this model the wireless node has multiple radios and can use multiple non-overlapping channels at a time. This is known as MIMO (Multiple Input, Multiple Output). It is the most promising model, especially given the development of Software Defined Radios (SDR) and improved radio and antenna design. 4. Directional Antennas: Directional antennas provide higher gain and enable spatial

43 28 multiplexing to reduce interference. However designing network protocols that take advantage of these properties is a non-trivial problem due to additional deafness and exacerbated hidden node problem that is caused by the spatial sectoring of signals [19]. In our research we exploit the selective directional reception of signals in a directional antenna to conserve spectrum, reduce interference and signal to noise ratio (SNR). In [23] we made the use of directional tri-sectional antennas to exploit the geometric regularity of the network to present a fault recovery routing algorithm. To summarize, at the physical layer, the ability to reconfigure radios using technologies like software radios [35] and cognitive radios can increase the ability of the network to adapt and heal from failure. Additionally, the use of directional antenna has been used to decrease the contention space. However they also increase deafness and raise the number of hidden nodes in the network. However to capitalize on these innovations the MAC and routing layers need to better utilize the capability of dynamically controlling radio frequency and range Medium Access Layer Issues The ALOHA protocol [6] was the first foray into packet switched radio networks with shared wireless medium. Since then wireless communications have evolved significantly in efficiently, bandwidth and complexity. To improve the quality of packet switched network, Togabi and Klinerock suggested the Carrier Sense Multiple Access (CSMA) [64] similar to that used in their wired counterpart. In the same paper they also discuss the hidden-node problem, which severely limits the effectiveness of CSMA. The hidden node

44 29 problem had been addressed with an additional busy signaling channel in [64] and an improved version in [21]. Another approach which is adopted by the standards is to use the Request To Send and Clear to Sent (RTS/CTS) hand shake proposed in [37]. However, in the context of WMNs, improvements to the traditional contention based protocols are usually not sufficient to improve allocation efficiently and fairness. Traditional MAC protocols are limited when we need to take advantage of newer underlying models. For example, multiple channels and multiple radios bring new problems of channel assignment and medium access. While Multiple Input and Multiple Output (MIMO) radios been proposed to increase the capacity of WMNs to mitigate unfair access and under utilization, current MAC protocol cannot take advantage of this underlying techological improvement. To overcome this, network MAC protocols like [15] use prepartitioning and local pooling to improve throughput significantly. One of the earliest implementation of this has been [4] which uses simple greedy approach to allocate the radio and channel based on channel quality. Other approach has been to use cross-layer techniques for channel assignment [32] where the transport layer shares information with the MAC layer to improve efficiency. While these approaches are novel, to work in large scale WMNs with standard based protocols, there are still several issues that need to be addressed. At the Medium access layer improvements in random access MAC are only suitable for small scale WMNs like client only ad-hoc WMNs. To achieve a larger scale, enhancements to the CSMA/CA like adjusting contention window or back-off [7] are not sufficient. MAC protocols using Time Division Multiplexing (TDMA) and Code Division Multiplexing (CDMA) also need to be investigated. Wireless mesh networks, constructed with multiple

45 mesh routers, using multi-radio multi-channel wireless mesh network need new and novel 30 approaches for channel assignment [8]. We need to also improve protocols that use the newer multi-radio, multi-channel technologies [56] and use cross layering approaches between routing and medium access to improve network performance and scalability Transport Layer Issues At the transport layer WMNs have their own unique challenges. Ideally the transport protocols would efficiently utilize all available network resources and allocate them fairly. However, access fairness in wireless mesh networks is inherently complicated due to the interdependencies among neighboring wireless links [30]. There has been a lot of fostering research at the transport layer that study the performance of TCP in a multi-hops setup [29, 45]. Most studies show the limitations of such protocols in contexts of WMNs and propose improvements [60]. However there are many challenges to be overcome to make the transport protocols more effective in the WMNs environment Network Layer Issues At the network layer, WMNs need new and improved protocols that take advantage of the its distinct characteristics and traffic flow direction, which is highly skewed between the client and the gateway. Many routing in wireless multi-hop networks has been actively studied in recent years. In [53], the authors cover many approaches used in ad-hoc wireless networks from table driven to source initiated on-demand routing like Dynamic Source Routing (DSR). WMNs have similar challenges as ad-hoc networks but the two differ in terms of structure and scale [34]. As previously discussed, Roofnet [17] architecture is based on a multi-hop routing

46 31 protocol that allows packets to be wirelessly routed to a gateway based on a routing metric which is an expected transmission delay. Others provide similar approaches using different routing metrics, different technologies, or cross layer optimization. In [5,52] the authors present a cross layered heuristic to improve the overall good-put of a wireless mesh. They also exploit multiple channels and multiple radios (MR-MC). In [8], the authors present a theoretical function to optimize the capacity of such MC-MR based networks. In [27], the authors have proposed a new routing metric called Weighted Cumulative Expected Transmission Time (WCETT) to make better routing decisions. While, there is active research in this area, there are many issues yet to be investigated that are beyond the scope of this dissertation Topological and Deployment Issues Planning and deploying mesh network is another very challenging problem. Planning includes determining the number of gateways, optimal placement of gateways and relay nodes, maximizing coverage while minimizing the operational cost. In this research we will study issues related to planning and deployment in depth. There are typically two approaches to deployment - 1. Structured Deployment: When a telecommunication company wants to provide services in a new area, it has the flexibility of choosing the topology of the network. This flexibility may translate into improved network performance by capturing the regularity of the deployed mesh network. 2. Organic Deployment: In most cases the mesh network will be deployed organically or incrementally over existing infrastructure. In such a case, there are limited

47 options for the network architect to develop a network topology of choice. Most network protocols are designed for such deployments. 32 To study the effect of hop distance and distance from the wired gateway we conducted an experiment and simulations to understand the implication of the gateway placement and hop distance can have on WMNs Roofnet Experiment To study the impact of hop-count on throughput, we formed a small wireless mesh network consist of five nodes. We setup a Roofnet [17] based test bed at Kent State University, Computer Science department. We used five Wireless Media Routers, each equipped with a 200MHz MIPS32-like CPU on Linux kernel. We changed the firmware and re-flashed it with the open source version of the Roofnet software. The network essentially formed a g based single radio network. We then setup the network to look like a logical network shown in Figure 2.5. We then connected the first router ( A ) to one of the experimental server, via a wired 100 Mbps Ethernet. The server provided a FTP server with five files of different sizes. B mars.cs.kent.edu Gateway A C D E Figure 2.5: Roofnet testbed setup at Kent State University, Dept. of Computer Science With the Roofnet protocol installed on each router, we measured the effect of the

48 Throughput in kb/s 33 number of hops as well as the hop distance (in db) between each pair of relay routers. In the first experiment, we measured the effect of distance of the client from the router. Here, we used a laptop equipped with a standard card and placed it at varying distance from the gateway router ( A ). We then used FTP to download files of different sizes. Each trial was conducted five times at different times of the day and the results were averaged. The experiment was then repeated by increasing the hop distance. We then calculated and compared the average throughput of the network. In Figure 2.6, we observe the exponential drop in throughput as we move further away from the access point. The typical range of a network is -70 dbm and in our results we took measurements at -66 dbm and saw the throughput drop by almost 800% dbm 49 dbm 66 dbm Power at Receiver Small Medium Large Figure 2.6: Effect of distance on throughput, Roofnet testbed In the second experiment, we connected the same client laptop to the network to different routers. First at a hop distance of one at router A, then at a hop distance of two to router C, then to router D and finally router E (See Figure 2.5). In each trial we downloaded five files of varying sizes five times as before. Then we repeated the

49 Throughput in kb/s 34 same trials at increasing hop distances. The result of this experiment is shown in Figure 2.7 in which throughput performance decrease exponentially as hop distance increases. Several factors contribute to the throughput degradation including the single radios limitation of half duplex communication [3], increased blocking probability, reduced access fairness [30], and increase in the contention domain [38]. The latter is probably the most important factor influencing the throughput, mainly as the result of inherent exponential back-off implemented in hop 2 hop 3 hop 4 hop Hop Distance Figure 2.7: Effect of hops on throughput, Roofnet testbed The two major observation from Figure 2.7 are: (i) the transmission range between routers has to be factored into the clustering and deployment, and (ii) the throughput disparity for nodes distant away from the gateway has to be mitigated. These are discussed in Chapter 6

50 35 Hop Distance Simulation using Qualnet In addition to Roofnet experiment, we also used Qualnet [57] simulator to verify our experimental results. We setup the simulation to study the effect of hop distance on throughput and access fairness. We setup five wireless nodes as a linear multi-hop network, as shown in Figure 2.8, with the first node being the traffic source and the gateway node being the traffic destination. 3-hop flow 2-hop flow W 4-hop flow Figure 2.8: Simulation setup to study effect of hop distance We then increased the load, by increasing the traffic generated on the source node and observed the effect of load on the throughput. We repeated the simulation with traffic flows between source destination pairs. In this experiment, we limit the hop-counts to 3, as the throughput performance deteriorates significantly beyond 3 hops. Figure 2.9 shows the throughput performance varying based on the load for different hop distances. We observed that the throughput of the flows with fewer hop count is significantly better than the throughput of flows with higher hop count. As load increases beyond 40-50%, the throughput for all scenarios start decreasing, mainly due to the contention resolution and back-off algorithms provisioned in the protocol. This result is in line with our analysis and the Roofnet experiments. As can been seen from the experiments and simulations - the number of hops and the distance from a wired gateway have a significant impact on the performance of WMNs.

51 Load (Kbps) Applied Load 2 hop 3 hop 4 hop Figure 2.9: Load vs. throughput for varying hop distance. In the next few chapters we study this problem in depth and propose algorithms that help improve the performace of WMNs.

52 CHAPTER 3 Gateway Placement in WMNs 3.1 Introduction Wireless gateway placement is an important design factor for the deployment of mesh networks. If a wireless gateway is placed at the edge of the network, the average communication cost will intuitively be much higher than if the gateway was placed in the center of the network. It is thus very important to place the gateways so as to minimize the number of gateways to keep cost at a minimum. At the same time, their placement should meet the end user application performance metric of minimum bandwidth and an upper bound on the delay. Along with QoS constraints, other factors like minimizing interference, increasing fault tolerance and provisioning for demand growth are also important factors that influence the location of the wireless gateways. 3.2 Background and Related Work The problem of gateway placement in wireless mesh networks is similar to the facility location problem in which one or more distribution centers serve a spatiality distributed set of demand locations subject to one or more objective functions depending on the interaction between demand locations and distribution centers [40]. However, the facility location problem does not capture all aspects of the gateway placement problem in wireless mesh networks [18]. In particular, issues such as relay load, link capacity and other QoS constraints that we address here, are not within the scope of the facility location 37

53 38 problem. Many approximation algorithms have been developed for the facility location problem [55] and the k-median problem [11]. Most of these approaches are based on Euclidean distance and not on hop-count hence making it difficult to define QoS metrics like endto-end delay. Additionally, the capacity of the links that correspond to bandwidth is not accounted for in facility location problem, thus warranting a different solution for the gateway placement problem. Another similar problem is the placement of servers and web proxies on the Internet to optimize the delivery of content [39]. However, the gateway placement problem and proxy placement problem differ in terms of temporal locality and caching issues. In [13] a hierarchal placement scheme has been proposed for WMNs. The authors consider cluster size constraint and allow overlap with other clusters. However, they do not take additional constraints like relay load into consideration. The earliest work that directly addressed the placement of gateways in a wireless mesh network [14] describes the problem as a capacitate facility location problem with additional constraints. The author solves the placement problem by breaking it into two sub-problems. First, a polynomial time approximation algorithm that cuts the network into disjoint clusters using a shifting algorithm or a greedy dominating independent set algorithm. Once the initial clustering is completed, each cluster is evaluated to that ensure QoS constraints are met. If the QoS is violated, then the cluster is sub-divided into smaller clusters at the node where QoS is violated. However, for the solution to work, it is assumed that the underlying medium access protocol is TDMA(Time-Division Multiple Access). TDMA protocols require synchronization, which is hard to achieve in

54 39 large multi-hop wireless networks. Additionally, the proposed solution generates higher fragmented clusters as the clustering is addressed in two stages. To overcome these limitations, a more general scheme has been proposed in [18]. The authors provide an ILP formulation that takes into consideration the demand constrains of each node and channel interference by studying the problem under two different link models.then they propose a greedy placement algorithm that iteratively picks a gateway that satisfies the maximum traffic demand. However, their greedy approach can lead to an imbalanced load and does not necessarily support all the quality of QoS requirements. Similar to [14], in [10], the authors form a cluster and a spanning tree within each cluster to obtain a near optimal solution. The authors in [10] build the cluster and spanning tree in parallel instead of a sequential approach used in [14]. Thus they were able to use a lesser number of gateways than those used in [14] for certain QoS conditions. 3.3 System Model and Assumptions We model an n-node wireless mesh network as an undirected connected graph G(V, E), in which a vertex represents a wireless node and an edge represents a communication link between two nodes. E(G) is a set of all edges and V (G) is the set of all nodes in G. We assume that each node has a transmission radius of t r units. An edge e E between two nodes u, v V exists if and only if their Euclidean distance d(u, v) t r. For practical purposes, a node antenna is assumed to be omni-directional, thus enabling communication with any of its neighbors within transmission range t r. We also assume that the MAC layer can use multiple channels and radios, thus enabling full-duplex between communicating nodes. These assumptions are quite reasonable and practical,

55 given the state of the art in MIMO (Multiple Input, Multiple Output) and a vast amount of research highlighted in [3]. An n-node graph G(V, E), can be represented by its adjacency matrix A n n, where, 1 if d(v i, v j ) t r, 1 i, j n A[i, j] = (1) 0 if d(v i, v j ) > t r, v i, v j V. Definition 3.1 (Cluster) A cluster, C of G(V, E) is an acyclic subgraph of G, i.e., C(V, E ) G(V, E), such that, V V, E E. 40 However, the size and the node membership of a cluster can be constrained by other limiting factors, such as hop distance, node degree, relay load, etc. Definition 3.2 (Clustering) A clustering of graph G(V, E) is defined as a set of clusters Ψ, where, Ψ = {C 1, C 2,, C k }, 1 k n, (2) with the following properties. Property 1: Property 2: Property 3: k i,j=1c i C j =, i j V (Ψ) = V (G). E(Ψ) = k i=1 E(C i) E(G). (3) The set of all possible clusterings of G can be denoted by, A(G) = {Ψ 1, Ψ 2,, Ψ m }, m n. (4) Similarly, we define Γ, as a set of spanning trees, Γ = {T 1, T 2,, T k }, 1 k n. (5)

56 41 where T i is a spanning tree covering cluster C i, 1 i k, and its root c i is the cluster head. Our solution to the placement problem with some QoS constraints consists of two steps. Initially, we partition the network into a set of disjoint clusters based on some graph properties and QoS constraints such as hop count and relay load. Then, through iterative refinements that consists of Split, Merge and Shift operations, the algorithm tries to form a clustering Ψ A(G), such that, k = Ψ = min{ Ψ 1, Ψ 2,, Ψ m }, m n. (6) Finally, for each cluster C i Ψ, 1 i k, an optimal place for its cluster head c i V (C i ) on the spanning tree T i is found. The location of the cluster head c i (gateway) minimizes the cost function of the cluster in terms of hop count and relay load Clustering with QoS Constraints The formation of a cluster is subject to several QoS constraints that can be expressed in terms of an upper bound on maximum hop count (delay), maximum cluster size (capacity), and maximum relay node (throughput). These can be formulated as follows: Hop Distance (H q ): A significant portion of the end-to-end delay is associated to the number of hops a packet traversed when compared to the propagation delay. Everything else being the same, the delay constraint is directly proportional to the hop distance between two nodes. To ensure a bound on the maximum delay, we limit the maximum hop distance H q within a cluster, h(c i, v) H q, v C i and v c i (7)

57 42 where h(c i, v) is the shortest path, in terms of number of hops, between cluster head c i and node v. Cluster Size (S q ): To guarantee a minimum throughput for each cluster and to minimize the blocking probability between a node and its cluster head, a cap is put on the number nodes that can be supported by each cluster and the cluster head, i.e., C i S q C i Ψ (8) Relay Load (L q ): A part from delay and cluster size constraints, the feasibility of clustering is constrained by the relay(transit) load of the intermediate nodes and it can be expressed as, L(v) = l(v ) L q v v, (9) v T v where T v is a sub-tree of the spanning tree T i and rooted at v. This is illustrated in Figure Figure 3.10: Relay Load Flows. Definition 3.3 The conjuction matrix K n n is an asymmetric binary matrix with, 1 if v j C i (a) K[i, j] = (10) 0 otherwise (b)

58 43 K is simply an instance of clustering Ψ A(G), where, K[i, i] = 1, if v i = c i (cluster head) C i = n j=1 K[i, j] (b) (11) Definition 3.4 (Inter-cluster and intra-cluster links) The set of edges which have one end node in C i and the other end node in C j are called inter-cluster edges and denoted by: (a) E(C i, C j ) := {(u, v) E u C i, v C j }. Alternatively, the set of edges which have their both end nodes in one cluster are called intra-cluster edges and denoted by : E(C i ) := {(u, v) E u, v C i }. Definition 3.5 (Neighborhood Relationship( )) Two clusters are neighbors if they share one or more inter-cluster links, i.e., C i C j iff v i C i and v j C j A[v i, v j ] = 1 and K[v i, v j ] = 0 (12) where A is the adjacency matrix of G and K is the conjuction matrix corresponds to clustering Ψ. 3.4 Gateway Placement Algorithms Integer Linear Program Formulation We can formulate the clustering problem with the QoS constraints described in Equations (7), (8), and (9) as an optimization problem using Integer Linear Programming(ILP) as follows. n min K[i, i], (13) i=1

59 44 Subject to: h(i, j) H q 1 i, j n (a) n j=1 K[i, j] = 1 (b) (14) n i=1 K[i, j] S q (c) u T vi L(u) L q u v i (d) Condition 14(a) satisfies the maximum hop distance constraint. Condition 14(b) ensures that the clusters are mutually exclusive. Condition 14(c) satisfies the maximum cluster size and the maximum load on each gateway. Condition 14(d) limits the relay load on each member of a cluster. Solving the above ILP is N P-hard and it has been shown to be reducible to a set cover problem. However, we use the above notation to describe our near optimal algorithm in Section We propose a polynomial time heuristic called Split-Merge-Shift (SMS), that is fast and within provides results that are close to the optimal solution. The solution is close to the optimal as we have imperial simulation results that shows the results obtained by the SMS algorithm to be very close to the optimal solution. However, we were able to demostrate this only on smaller graphs as the combinatory algorithm used to obtain the optimal solution had a very high time complexity.

60 Split-Merge-Shift Algorithm (SMS) The SMS algorithm starts with an initial clustering graph and then it goes through a few refinement iterations of Split, Merge, and Shift (SMS) operations, as described below, to form the final clustering. A high-level description of the Algorithm 3.1 is illustrated in Figure 3.11 in which a random irregular mesh network has been partitioned into 5 clusters. Initialization Phase Algorithm 3.1 takes the adjacency matrix A that corresponds to network G(V, E) along with the three QoS constraints; H q, S q and L q given in Equation (14) as inputs, where H q is the maximum hop distance, S q is the maximum cluster size, and L q is the maximum relay load (traffic flow). The algorithm starts by exploiting node density (i.e., node degree) to select the initial cluster heads. Since we assume a uniform traffic model, we choose a node with the highest degree as an initial candidate for a cluster head. Subsequently, we include all nodes that are one hop away from the cluster head to join the cluster as long as the QoS constraints of Equation (14) are met. The greedy selection process continues until every node in G has been assigned as either a cluster head or a cluster member. This is illustrated in Figure 3.11(b) and on lines 2-7 of Algorithm 3.1. This gives us the initial clustering Ψ. Ψ is implemented as the conjunction matrix K, defined in Equation (10). It is possible that during the initial phase a singleton cluster is created. A singleton (or trivial ) cluster is a cluster with only one node (cluster head). After the initial phase the algorithm begins merging clusters.

61 Initial Phase Split Iterative Split, Merge and Shift (a) (b) (c) (d) (e) (f) Figure 3.11: Algorithm overview.

62 Algorithm 3.1: Gateway Placement input : A, S q, L q, H q ; [QoS constraints] output : Ψ = {C 1, C 2,, C k }; [Set of clusters] A[i, j], 1 i n; 1 D n j=1 [Initial clustering] ; 2 while V (Ψ) V (G) do 3 c i max{d[i], 1 i n, v i / Ψ}; [Selects a node with highest degree, not V (Ψ)]; 4 C i c i v j 1 j n, where 5 A[c i, v j ] = 1, v j / Ψ, Ψ < S q ; 6 Ψ = Ψ C i [Makes cluster part of solution]; 7 end 8 Ψ p Ψ ; flag false; 9 while Ψ Ψ p do 10 Ψ p Ψ ; 11 if flag then 12 C i FindSplitCandidate(Ψ); 13 Ψ Split(C i ) ; 14 end 15 while Changed(Ψ) do 16 [C i, C j ] FindMergeCandidates(Ψ); 17 Ψ Merge(C i, C j ) ; 18 end 19 while Changed(Ψ) do 20 [C i ] FindShiftCandidate(Ψ); 21 Ψ Shift(C i ) 22 end 23 flag true 24 end 47 Merge Operation The merge operation tries to merge neighboring clusters, as a refinement process to form clusters with stronger bond without violating the QoS constraints of the resultant merged cluster. It first selects a list of smallest clusters as candidates to merge with other clusters, then it chooses clusters with strongest bond (inter-cluster links) to be merged. The merge process continues until no more merging is possible. The limits posed on a cluster by the QoS constraints. It is important to note that when we find two candidates

63 48 to merge we first merge them and then check for QoS constraints. If there is no violation, we commit the merge to Ψ using matrix K, otherwise we ignore the merge. This allows for more clusters to be merged and reduces false positives. This process is illustrated in Figure 3.11(c) in which clusters C 2 and C 3 from Figure 3.11(b) are merged with C 1. Lines of Algorithm 3.1 perform the merge operation on lines Note that the split operation (lines 10 14) is not executed in the first iteration, right after the initial phase, to allow cluster consolidation after the greedy selection. Split Operation When no more merging between neighboring clusters is possible, the split operation picks the smallest cluster with most neighbors and breaks it into singleton clusters. These singleton clusters have a better chance to merge with their neighboring clusters. If the split operation creates more clusters than the previous iteration, then the previous state of the clustering is restored and the algorithm stops as described in the stopping condition below. After the initial merging is completed, the algorithm tries to further merge the clusters that were too large to merge as a whole. This is done by splitting a good candidate cluster into smaller clusters and then merging the smaller clusters with their neighbors. The splitting routine begins with the search for a candidate cluster to split. It basically looks for the smallest cluster with the most neighbors. Once a candidate has been identified, it breaks up the cluster such that each node of the original cluster become a cluster in itself, i.e. a singleton cluster. It is important to note that the cluster that has been split

64 are excluded from merging with each other. Thus the next merging iteration will select them for merging, but will not merge them with each other. 49 Shift Operation After the split and merge operation, based on the topology there may be a few nodes left that cannot join their adjacent neighbors due to QoS constrains of their immediate neighbors. However there may exist other clusters that are one or more cluster-hops away that could accommodate these nodes. The shift operation tries to merge a singleton cluster that cannot merge with any of its neighboring clusters. The algorithm first constructs a cluster graph G c corresponding to clustering Ψ p as shown in Figure The algorithm then finds an exchange path rooted at the singleton cluster (cluster C 2 in Figure 3.12). The exchange path is used to move (shift) nodes from one cluster to another until the candidate singleton cluster is merged. In Figure 3.12, a node v 28 in C 3 is shifted to C 4 to make a room open for singleton C 2 to merge with C 3. Stopping Condition At the beginning of each split operation the previous iteration s clustering solution is saved as Ψ p. If after the completion of one complete iteration (split, merge and shift) the solution set is larger than the previous solution set then the algorithm stops and returns Ψ p as the final clustering.

65 Cluster Graph Cluster Neighborhood Tree Shift Path Figure 3.12: Shift operation overview Algorithm Illustration We illustrate the running of the algorithm using a 40-node network topology show in Figure 3.11(a). In this example we setup H q = 3, S q = 8 and L q = 15. L q is relaxed to simplify the illustration. The initial phase (lines 2 7 of Algorithm 3.1) forms single hop clusters around the high degree nodes and that would give the initial clustering Ψ, as shown in Figure 3.11(b). At this stage the algorithm finds clusters that can be merged. It iteratively merges clusters C 1, C 2 and C 3, and then merges cluster C 9 with C 10 using lines of Algorithm 3.1. This is illustrated in Figure 3.11(c). The merging stops when no more adjacent clusters can be merged without violating the QoS constraints. In this example, the Shift operation is not applied as there are no singleton clusters. The split operation, lines of Algorithm 3.1 finds C 2 of Figure 3.11(c) as a suitable cluster to split and merge

66 51 its members with other neighboring clusters. The split is shown in Figure 3.11(d). After the split, the merge phase absorbs C 3 into C 4. C 2 cannot merge with C 4 due to the size constraint S q = 8. The algorithm then starts the Shift Phase. From Figure 3.11(d), the shift operation identifies C 2 (singleton with most non-singleton neighbors) as a candidate to perform the shift operation. It builds the neighborhood tree representing neighbors with which exchange is possible. From this tree it identifies C 5 (Figure 3.11(d)) as a candidate that can accept a node and builds an exchange path. The shift operation moves node 28 from C 4 to C 5 and node 30 into C 3, eliminating C 2. The output after the merge and shift is shown in Figure 3.11(e) The algorithm continues with splitting C 5 in Figure 3.11(e) and then merging and shifting the nodes until the final configuration is reached. This is shown in Figure 3.11(f). The algorithm stops when the clusters formed after one complete split-merge-shift operation are more than that of the previous iteration Algorithm Analysis Algorithm 3.1 begins with the formation of the initial clusters. Since each node is checked for membership until all the clusters are formed, the worst case running time would be O( V 2 ), i.e. if each node becomes an individual cluster. After the greedy select, the merge routine runs in a while loop until all clusters are merged. The worst case run time of the merge operation is when Ψ = V and all the QoS conditions are relaxed such that they do not interfere with the formation of the clusters. Thus the clusters will merge until the whole graph become a single cluster. Therefore, the while loop has an upper bound of O( V ). For the merge operation, the

67 52 various functions are bound by O( V 2 ). So the upper bound of the merge operation with the outer while loop is O( V 3 ). The split operation simply finds the smallest node with the most clusters and then splits the selected cluster into singleton clusters. Thus the split routine operations have an upper bound of O( V ). Finally, in the shift operation, the identification of the candidate cluster takes O( V ), building of the conditional spanning tree O( V 2 ) and performing the actual shifting O( V ). So the overall upper bound on the algorithm is O( V 3 ). 3.5 Performance We evaluated the SMS algorithm based on i) the number of clusters (gateways) it produces, ii) the clusters size variation, and iii) the maximum and average relay loads. We compared its performance with algorithms discussed in [10], [14], and [18]. We ran all algorithms on ten different 50 node networks and the results presented is the average. We first vary the cluster size, while relaxing the cluster radius and relay load. By varying the maximum cluster size (S q ), as shown in Figure 3.13, our near-optimal algorithm outperforms other heuristic algorithms in terms of the number of gateways. In our simulations, our algorithm produced about 30% less gateways. Similar results can be see for relay load L q as shown in Figure Figure 3.15 shows the variation in cluster sizes (S q ). The SMS algorithm generally produces uniform clusters. As expected, the greedy algorithm creates imbalanced clusters as noted earlier.

68 53 Number of Gateways Iterative Recursive Split-Merge-Shift Greedy Maximum Cluster Size (S q ) Figure 3.13: Effect of S q on gateway count Number Of Gateways Iterative Recursive Split-Merge-Shift Greedy Maximum Relay Load (L q ) Figure 3.14: Effect of L q on gateway count 3.6 Summary In this chapter, we have investigated issues related to the optimal placement of wired gateways in wireless mesh networks with Quality of Service constraints. We have developed a polynomial time approximation algorithm that can find a solution that is very

69 54 Cluster Size Variation Iterative Recursive Split-Merge-Shift Greedy Maximum Cluster Size (S q ) Figure 3.15: Effect of S q on cluster size variation near the optimal solution. We then compared our algorithm with others presented in literature. Further investigation is needed to study this problem with additional constraints like channel allocation. Another interesting WMN deployment problem is related to relay node and link failure and its effect on gateway selection. We are currently investigating some of these problems.

70 CHAPTER 4 Fault Tolerance in Cellular WMNs Wireless mesh networks can broadly be classified as wireless cellular mesh networks (WCNs) and wireless mesh networks (WMNs). While both, from graph theory perspective, are considered mesh networks, they are different in terms of connectivity and operations. WCNs form a regular mesh, often in the form of hexagons, and the backbone of the base stations has triangular connectivity. WMNs, on the other hand, are considered irregular mesh, with non-uniform node degrees. In this chapter, first we present a fault-tolerant WCN based on a hexagonal topology with multi-radio and vectorized directional antennas. For this model, we introduce an addressing and routing scheme that simplifies the network operations. Later, we introduce a fault-tolerant routing algorithm that copes with one or multiple link failure. We exploit the regularity and multi path characteristics of an augmented tri-sectionized hexagonal system to route around link or node failures. 4.1 Background and Related Work As mentioned in Chapter 1, WMNs offer many advantages in terms of scalability and reliability. One of it s biggest advantages is that it offer multiple redundant communication paths throughout the network. However routing protocols need to be developed to exploit this property for WMNs. As discussed in Section 2.4.4, routing plays an important role in WMNs. However, 55

71 56 very few studies have exploited the topological regularity of a cluster of cellular systems to achieve fault tolerance. One of the contributions of this research is to use the honeycomb structure of a cellular system to implement better addressing, routing and fault tolerant solutions for WMNs. 4.2 An Augmented Tri-Sectorized System with Directional Antennas Regular wireless mesh networks can be implemented by arranging the nodes on a planar graph using squares, triangles or hexagons [58]. The term mesh is most commonly used in reference to a 2D grid where each node is connected to its four of its nearest neighbors. On the other hand an augmented trisectorized hexagonal network (honeycomb) is formed by tiling the plane with regular hexagons and pacing a degree-3 node at each vertex [20]. Based on the arrangement of the hexagons and the exterior shape formed by combining hexagons, they exhibit different properties [58]. In this section we consider hexagonal and rectangular honeycomb networks. While a 2D grid is a viable option for implementing a wireless mesh network, hexagonal structures have their own advantages. First, hexagonal WMNs closely resemble the current cellular infrastructures and simplify their deployment. Second, the degree of a hexagonal honeycomb network is three compared to the degree of 2D grid which is four, while the two networks have the same diameter, d = 2 n [49], where n is the number of nodes. Here, the node degree corresponds to the number of distinct radio channels needed and that directly relates to the hardware cost.

72 Sectorized cells directional antennas Cell sectoring with directional antennas have been proposed and used in wireless cellular systems to achieve better coverage of a small or populated area with less power requirements. In addition to power conservation, cell sectoring allows better frequency channel allocations, more flexible channel borrowing, less co channel interference, and higher spectrum efficiency (i.e., networks capacity). One approach in cell sectoring is to place directional transmitters at the corners of the hexagonal cell where three adjacent cells meet, as shown in Figure It may appear that the trisectionized arrangement 60 j i (a) (b) Figure 4.16: (a) Omni directional antennas, (b) Tri-sectorized antenna. in Figure 4.16(b) requires three times the transmission towers as compared to the omni directional arrangement in Figure reftri1(a), but as shown in Corollary 4.1, the number of transmitters remains approximately the same as the cluster size increases. Corollary 4.1 The number of omni directional transmitters in a hexagonal cellular system is N o = i 2 + ij + j 2, which is the cluster size [54], where i represents the number of cells to be traversed along direction i, starting from the center of a cell, and j represent the number of cells to be traversed along direction j, which is a 60 o angle from i direction in a hexagonal cellular system as shown in Figure 4.16(a).

73 58 Figure 4.17 illustrates the approximation. Figure 4.17: Transforming a Regular Cellular structure to a Trisectional Directional Cellular System N Figure 4.18: Tri-sectored node with three directional antennas and one omni directional antenna. In our augmented trisectionized hexagonal system, a router is placed at the corner of each hexagonal cell, as shown in Figure In this model, each transmitter is equipped with four radios. Three of the radios are used by directional antennas to provide interconnection among the neighboring routers. The fourth radio is used by an omnidirectional antenna that provides interconnection between mesh clients and the mesh router within a

74 59 sector. This model would double the number of transmitters, while it reduces the number of radios per transmitter. It also reduces the reuse distance, and significantly increases the network capacity. The following arguments can be made to justify the additional hardware cost while improving the routing and fault tolerance. First, in a cellular network each base station is served by an elaborate wired infrastructure which is expensive to deploy and equally expensive to operate and maintain. However, in case of WMN, only a few nodes, called gateway nodes need to be connected to the wired infrastructure. This fact alone can significantly reduce the operating expenses as compared to a cellular network and also make them much faster to deploy and adapt. Second, the honeycomb topology provides alternate routes with a simple routing algorithm and provides fault tolerance with different fault scenarios. This is particularly important to offset the unreliability of the wireless links. Having more nodes closer to each other enables better coverage and the ability to provide higher bandwidth. Third, having more nodes closer to each other lowers path losses significantly. For example, if a direct link supports 6 Mbps waveform, then shorter links will use 18 Mbps due to 6dB less path loss Honeycomb Networks A honeycomb network can be represented by a pruned 2D planner network as show in Figure We represent the honeycomb network as a graph G(V, E), where each node v is identified by its (x, y) coordinates. The set of edges E can be defined as,

75 Honeycomb Brick Representation Figure 4.19: Brick network representation of a honeycomb network. E : (x, y) (x y ± 1) (x, y) (x + 1 y) (x, y) V If (x + y) is odd (15) (x, y) (x 1 y) If (x + y) is even Without loss of generality, the origin of the (x, y) coordinates is assumed to be (0, 0) throughout this paper. This formulation makes the addressing scheme and routing algorithm much simpler to implement. It also provides a framework to use the regularity of honeycomb topology to design a fault tolerant routing algorithm for the network. 4.3 Routing in Honeycomb Mesh In [58] the author presents a routing algorithm that can be applied directly to a honeycomb network. The same concept is expanded in [46] for mobile wireless application. However we use a modified and yet equivalent network to simplify the routing algorithm from a 3-coordinate to a 2-coordinate system. suggested in [49]. We first present this algorithm and then introduce a fault tolerant routing that can handle multiple link

76 61 failures QoS Routing As a packet traverses the network at each point the node calculates the offset from the current position in the network to the destination. Let the source and destination coordinates be represented by S = (S x, S y ) and D = (D x, D y ), respectively. We define the offset, = D S = ( x, y ) = (D x S x, D y S y ). Let P = {S,, U,, D} be a path chosen by the routing algorithm from source S to destination D, and let U be the current node traversed. We define δ = (δ x, δ y ) to be the sign( ) such that, δ x = 1 if x > 0 0 if x = 0 δ y = 1 if y > 0 0 if y = 0 (16) 1 if x < 0 1 if y < 0 Further, we define 1 if v is even e(v) = 0 if v is odd v {U x, U y }. Each node is connected to four links each with a quality w i where 0 w i 1, and i {left, right, up, down}. The link quality is associated with a weight on that link which is dynamically measured as a function of signal strength, error rate, and queuing delay. Less preferred links are quantified by smaller weights in the routing process. We also define a minimum links quality, qthresh, as a cut-off point whether to include the

77 62 link in the routing path or not. Hence, 0 if w i < qthresh w i = 1 if w i qthresh An ideal link has qthresh = 1. First, we consider the case where there is no link (or node) failure. In this case, a link either exists or not, based on a predefined threshold qthresh. In the honeycomb topology of Figure 4.19, the existence of the three out of four links is determined by Equation (15). In this case, Algorithm 4.1 is used by each node to forward packets based on their destination based on their destination address D. Algorithm 4.1: Route(S, D) 1 (1) Initialize : U S, D U (2) while ( x y) 0 do (3) U Move(U), D U (4) end (5) proc Move(U) // δ x, δ y from Eq. (16) (7) if δ x < 0 then p = δ x (e(u x ) e(u y )) (8) else p = δ x (e(u x ) e(u y )) (10) fi (11) q = δ y (1 p ) (12) Boundary(p, q) (13) U x = U x + p (14) U y = U y + q (16) end (17) proc Boundary(p, q) // boundary conditions (18) cases: (20) p = 1 U x = x max : q = p, p = 0 (21) p = 1 U x = x min : q = p, p = 0 (22) q = 1 U y = y max : q = 1 (23) q = 1 U y = y min : q = 1 (25) end (26) end

78 63 The algorithm, starting from the source, iteratively tries to reduced the distance between the current position (U) of the packet to that of the final destination (D). Procedure Move choose a direction along x-axis if possible, otherwise it moves the packet along y-axis. Procedure Boundary checks for the boundary condition of the packet traversing the network. The algorithm is similar to the routing in a 2D mesh mesh network. However since links are pruned on alternate nodes along x-axis, the routing algorithm first tries to reduce x and then y. The core of the algorithm is the calculation of of δ x and δ y and p and q on lines 7,8 and 11. First the algorithm tries to determine the direction in which the packet should be forwarded. If the x axis link is available then p get a value of ( 1, 1). If the packet move along the x axis, then q is set to 0, as the packet can move along only one axis. This is represented by line 11. This algorithm achieves the shortest hop count as shown in Theorem 4.1. This is due to the fact that latency is more sensitive to hop count than Euclidean distance due to the gap between transmission delay and queuing delay, so this algorithm is more suitable than those based on minimizing the Euclidean distance. Theorem 4.1 Algorithm 4.1 produces the shortest hop-distance between the source and the destination. Proof: To prove that the routing algorithm described in the Algorithm 4.1 uses the minimum number of hops to reach the destination we need to prove that after each iteration of the algorithm the packet is moved at least one hop closer to the destination. This can be proved by induction.

79 64 Let S U 1 U 2 U k 1 U k U k+1 D be a path from source S to destination D, and let d = D S = D x S x + D y S y = x + y be the cumulative length of the path along x-axis and y-axis. First, we show that U 1 is one hop closer to D than S along the path by showing, D x U 1 x + D y U 1 y? < D x S x + D y S y Steps 7, 8, and 11 of Move in Algorithm 4.1 computes the amount of move on x-axis (p) and y-axis (q), respectively, based on the value (direction) of δ x and δ y in Equation (16), where p, q { 1, 0, 1}. Note that p and x as well as q and y have the same signs. Hence, D x U 1 x) + D y U 1 y = D x (S x + p) + D y (S y + q) < D x S x + D y S y (17) For general case, if then But, D x U k x + D y U k y < D x U k 1 x + D y U k 1 y (18) D x U k+1 x + D y U k+1 y < D x U k x + D y U k y (19) D x (U k x + p) + D y (U k y + q) < D x (U k 1 x + p ) + D y U k 1 y + q ) (20) where p, q, p, q { 1, 0, 1}, q = δ y (1 p ) and q = δ y(1 p ). Since δ y and δ y have the same signs, both sides or the Equation (18) are reduced by the same amount (±1) in Equation (20).

80 65 The Algorithm 4.1 provides the shortest route, however it cannot handle link failures. In the next section we develop an algorithm that can handle such failures, that are common to wireless networks. 4.4 Fault Tolerant Routing The shared wireless nature of the medium in WMNs makes them more susceptible to failure. For any routing algorithm to be successful in such an environment, it is essential to provide certain resilience to failures. In this section we study single and multiple link failures. We develop a localized temporal source routing algorithm to circumvent the faults. We focus on link faults as apposed to node failure since a wireless medium fails far more frequently than the wireless hardware components. However, the technique we developed here can easily be generalized to encompass node failures as well Fault Models Link failure can be model in two ways. In the first model, the link is completely unavailable due to the radio malfunction or a problem with the directional antenna, e.g. the antenna directions between two nodes are not aligned. In the second model, the link has an error rate above a certain threshold and the link cannot satisfy the QoS requirements of the traffic passing through. In either case, the link is treated as having a fault. The routing Algorithm 4.1 is not fault tolerant by itself. If a link fails the routing will fails in most cases. In routing algorithms that use spanning trees like [52] and [5], fault tolerance is suggested by means of re-associating the sub-tree that has lost its link to the root, with another spanning tree in the forest. However this technique involves an

81 66 d' d' d'' u'' X u' u' X u'' d'' (a) (b) Figure 4.20: Detours, d and d, around faults. increased message complexity and looses its effectiveness during short and frequent link failures. Further, this approach is highly dependent on the fault distribution. Since the honeycomb topology, considered in this study, provides multiple redundant paths, we take advantage of re-routing the packets along alternate paths. We define detours, d and d based on the two adjacent cells that share the faulty link and can provide localized re-routing to forward the packet around the faulty link. For example consider Figure A packet needs to get from u to u but the link connecting the two is faulty. Based on the location of the fault it can get around the link failure using the detour d and/or d. In general, if we have a y-axis fault, like in Figure 4.20(a), at u = (i, j) we can define the detours d and d via the adjacent cells as follow: d = (i, j + 1) (i + 1, j + 1) (i + 1, j) (i + 1, j 1) (i, j 1) (21) d = (i 1, j) (i 1, j 1) (i 1, j 2) (i, j 2) (i, j 1) (22)

82 67 Similar to Equation (22) and 22, d and d can be easily be defined for x faults and all the boundary conditions. These detours are basically topologically aware temporal source routing which enable the packet to avoid the faulty links that it encounters along the source-destination path. In the context of distributed computing, [31] have taken to turn model address fault tolerance. In this approach and its adaptation [66], the authors use the concept of a turn to route around the fault. They restrict the definition of the turn such that the turns do not cause a routing loop and at the same time these turn maneuvers are able to route packets around the fault. In our brick topology, normal routing would occur if there are no faults as show in Figure 4.21(a). However, if a link failure occurs on the y-axis as show in Figure 4.21(b) or along the x-axis as shown in the Figure 4.21(c), we can easily route around them. Further, there can be multiple faults as shown in Figure 4.21 (d). In 4.21 (d), the first detour, d is applied. After one hop into the detour, d encounters another fault. As this point it tracks back to the initial node that initiated the detour. Once the packet, traces back to the orignial fault detection node, it applies the detour d. Applying the detours enables the routing to re-routed the packets along the adjacent cells that share the faulty link. The regular routing as per Algorithm 4.1 (indicated by dotted arrows) is done till a fault link is encountered. Once a fault is detected, via the Physical layer, the detour d is applied. If there is a node failure along the detour d as shown in Figure 4.21 (d), the packet tracks back to the faulty node and then applies the second detour, d. This approach can handle multiple faults as it addresses each fault in a localized

83 u' 6 5 u' X 2 u'' 2 u'' (a) Normal routing (b) Routing with single y-fault u' 5 u' X 3 X 2 1 u'' 2 1 X u'' (c) Routing with single x-fault (d) Routing with double fault Figure 4.21: Routing around a link failure.

84 69 manner. It is important to note that if two or more faults occur on non-adjacent cells (i.e. two cells apart) along the path, they are treated as separate single faults as the algorithm uses temporal source routing. Thus addressing the individual fault with local topological information, without the need of being aware of global state of the network. To implement this detour technique, we propose an algorithm that uses a two phased approach to circumvent the faulty links - fault detection and fault avoidance. 1. Fault detection: In this phase the routing is carried out normally as discussed in presented in Algorithm 4.1. Typically a node is aware of its links. Thus if any of the link fails, the routing algorithm is made aware of this. The fault detection is mostly the responsibility of the physical layer or a function of the Medium Access Layer. 2. Fault avoidance: Once a fault has been detected, the algorithm goes into recovery mode. During recovery, based on the direction in which the packet needs to travel from the current node to the destination, the recovery algorithm using the predefined detours to route around the fault(s). We next look at different fault cases this algorithm can handle. 1. Single Fault: This approach can handle all single link faults as long as the fault does not disconnect the faulty node from the network. We can use either of the adjacent cells to apply a detour d or d to circumvent any single fault. 2. Double Fault: This approach can also handle all double localized faults as long as the faults do not disconnect the faulty node from the network. This is true because

85 70 Algorithm 4.2: RouteUnderFault(S, D)) 1 (1) Initialize : U S, D U (2) while ( x y) 0 do (3) Get from Algorithm 4.1 (5) Check for faulty link to next hop (7) if fault detected go to recovery (9) Apply detour d from Eqn. (22) (11) if d fails (12) Track back to initial fault node (13) Apply detour d from Eqn. (22) (15) if d fails (16) Drop Packet (17) end (18) end (19) end (20) Resume normal routing (21) end the topology provides two or more alternate paths from u to u. If there is a fault along the detour d, then the packet can trace back to the faulty node (u ) and try d. 3. Three or more faults: This approach can handle three or more faults as long as all the three faults are not distributed evenly across the fault and the two potential detours. As long as this criteria is met, the fault tolerant Algorithm 4.2 will be able to route around any number of faults from the source to the destination. The localized temporal routing begin with the recovery mode. As soon as the fault is detected, the packet is forwarded along one of the alternate detours, d or d. During this phase the packet is modified to add the detour path into the packet, so that it can track back to the faulty node in case the chosen detour has a fault along its path too. We implemented this algorithm and verified its working by trying different networks

86 71 dimentions with every possible combination of source-destination pairs and possible fault locations. A more in-depth comparative analysis of this approach with that of others similar routing techinques can be an interesting future study. 4.5 Comparision of Routing Schemes The fault tolerant routing approach presented in the previous section increases the deliverability of packets to the final destination. However at the same time this scheme increases the packet delay caused by the detour, due to the additional hops needed to route around the faulty link. While comparing to the shortest path, the fault tolerant routing add 4 additional hops per fault in the best case and 10 hops (back tracking and re-routing) in the worst case, i.e. if there are two local faults on adjacent cells. Depending on the type of traffic the fault tolerant routing can provide improved service quality and reliability. In this chapter we have presented an efficient and deterministic fault-tolerant routing algorithm that can handle multiple link failures, which are common to WMNs.

87 CHAPTER 5 Fault Tolerance in WMNs There are several approaches one can take to introduce fault tolerance into a wireless mesh network that is broken into clusters that have wired gateways as their cluster head. One of the goals of the study is to find a technique that can achieve the most resilience to fault with the least amount of over provisioning. 5.1 Background and Related Work Faults in WMNs can be classified as node (gateway, relay) failure or link (radio) failure. The latter could be associated with power, channel interference, fading or other shared wireless medium phenomenon. In wireless network the rate of link failure is several magnitudes higher than node failure or the failure of other elements [16]. In this chapter we primarily focus on how to handle a single or multiple link failure. From graph theory perspective, a node failure would result in multiple link failure. A node (relay or router) failure has also been investigated as long as the resulting graph remains connected. Link failure in a wireless network is commonly caused due to interference in the medium or traffic congestion and on rarer occasion due to the radio malfunction. Link failure does not necessirly mean link disconnectivity, but also includes the failure to satisfy the the QoS requirements such as delay and packet drop rates. Fault tolerance has been getting a lot of attention in the area of sensory networks [44], due to the higher node failure rate, large scale of the network and the desire to increase 72

88 73 automation. Fault recovery in such networks has been addressed in terms of routing [9], topology control [59] and power assignment [43]. Fault tolerance in wireless mesh networks, have been studied at the networking layer using routing protocols that find alternate paths to route packets from the source to destination, if the preferred path fails. While redundancy in terms of alternate routes(path) is necessary to tolerate some faults, it is not sufficient since re-routing or moving nodes affected by a fault to another cluster may result in QoS degradation In [18] the authors discuss fault tolerance with respect to gateway placements. To address node and link failures they modify the gateway placement LP (Linear Programming) formulation and add a fault tolerance constraint to ensure over-provisioning via multiple independent paths. They propose a greedy heuristic to address gateway placement that iteratively picks up nodes that increasingly satisfy the traffic demand without necessarily selecting a node that satisfies the most demand. Optimal gateway placement of gateways in a WMN with QoS constraints is known to be a N P-hard problem [55]. Several heuristic are proposed in [14], [18], [10] and [24] to place gateways efficiently in a given network. However limited work has been done to make these solutions fault-tolerant. 5.2 System Model Network and QoS Model We model an n-node wireless mesh network as an undirected connected graph G(V, E) as presented in Section 3.3. The vertex represents a wireless node and an edge represents a communication link between two nodes. E(G) is a set of all edges and V (G) is the set

89 74 of all nodes in G. The formation of a cluster is subject to several QoS constraints that can be expressed in terms of an upper bound on maximum hop count (delay), maximum cluster size (capacity), and maximum relay node (throughput). These have been defined in Section 3.3. Assuming the WMN has already had the gateways placed and routers clustered, a mesh network with the clustered graph is show in Figure Clustered graph Figure 5.22: A clustered wireless mesh with the overlay cluster graph Fault Model Transmission link failures are mainly caused by outdoor noise, interferences and multipath fading. In addition to transmission link failure, traffic congestion and protocol limitations can also contribute to communication failures. The fault model that is described

90 75 in this section focuses only on link failures, f(u, v), as node failures can be treated as several adjacent link failures. In this model if the link does not meet the QoS requirements in terms of delay, jitter or error rate, it will be purged from the overlay spanning tree. Clearly, network connectivity, in terms of maximum and minimum node degrees, fault location, and flexibility of QoS parameters directly affect the number of faults that can be tolerated in the network. Definition 5.1 (Fault Set) The fault set, F, f(u, v) V is the collection of all active faults in the network at any given point of time. The set is transient and changes with time. The detection of faults is typically assumed to be a function of lower layers, especially PHY and MAC. Each gateway is responsible for its members. To make the network fault-tolerant, some spare capacity in terms of constraining the maximum cluster size, (S q ), the maximum relay load (L q ), and the maximum hop distance (H q ) are provisioned in the original clustering algorithm or the network overlay design in order to accommodate certain number of faults. Definition 5.2 (Fault Provisioning Factor) The fault provisioning factor (δ), is a factor over which the basic QoS constraints can be extended to accommodate network faults. δ H, δ S and δ L represent the additive increment in H q, S q and L q. When a fault occurs, the recovery algorithm ensure that it does not violate these constraints plus the corresponding δ. We assume the gateway that is associated with an overlay cluster has a full knowledge of all its wireless nodes (relay nodes or clients)

91 76 in its sub-tree. When a link fails, the gateway can initiate an overlay reconfiguration process. This may result in finding a new spanning tree within the cluster while keeping the QoS constraints intact, or it may require the partially disconnected sub-tree to join the neighboring clusters and that is discussed in Section Fault Recovery Fault recovery algorithms can be classified broadly with respect to the amount of resource provisioning they need to accommodate different groups of faults. They can be classified into proactive or reactive provisioning algorithms. In proactive provisioning algorithms, during the deployment of the wireless mesh, a minimum amount of redundancy, in terms of spare links (or gateways) are provided to keep the network fully connected at all time. One way to achieve this is allow each node to be on associated with more than one wired gateways, thus forming overlapping clusters [67]. While this approach provides redundancy, it will not always guarantee fault-tolerance due to topological constraints of the network. In a more selective approach, the QoS constraints such as cluster size (S q ), hop distance (H q ), or relay load (L q ) are used to provide spare resources to accommodate faults while maintaining the QoS constraints. This provides some flexibility for network nodes (mobile or relay nodes) under fault to join other clusters. Relaxing the above QoS parameters will require more gateways during deployment and this comes at an increased cost. The main disadvantage of over provision solutions is the absence of sub-optimal operation of the network when there are no failures.

92 77 In reactive topology control algorithms, the initial QoS parameters, without over provisioning, are used for deployment. When a fault occurs, a new minimum spanning tree that avoids the faulty link is configured within the faulty cluster, or re-clustering is performed among the neighboring clusters. The advantage of this approach is that the network operates optimally at all time with or without faults. The cost of spare resources has now been shifted to the computational cost of dynamic clustering under faults. The fault recovery discussed in this section focuses on dynamic topology control. Any additional provisioning increases the probability of fault recovery Topology Control Fault Recovery Algorithm In a WMN, when a fault occurs, assuming that the node does not loose connectivity to the network, there are two possible ways the node or a disconnected sub-tree can recover from the fault 1. It can rejoin the same cluster using an alternate path if there exists one. 2. It can join one of the neighboring clusters. In each case the recovered topology should not violate the QoS requirements provisioned in the original deployment. The fault recovery algorithm takes the the original mesh adjacency matrix (A), the clustered mesh network (Ψ), the fault set (F ), the fault provisioning factor (δ), and QoS constraints (S q, L q, H q ). The algorithm assumes that no new gateway can be created as the system has already been deployed.

93 78 Since each cluster forms a spanning tree, depending on the location of the fault, the disconnected sub-tree can be too large, in terms of QoS constraints, to be accommodated into an adjacent cluster. In otherwordrs, the merger of the cluster with the disconnected sub-tree (T v ) should not violate the QoS beyond the fault-provisioning factor. C i T v S q + δ S (23) In addition to the size, the topological properties of T v may violate the hop distance constraint, H q + δ H, or the relay load constraint,l q + δ L, not making it possible to join a neighboring cluster. To overcome these challenges, the rearrangement algorithm basically breaks each of the nodes on the disconnected tree to form a single node pseudo cluster (singleton). Each of these then seek membership in one of their neighboring clusters. The recovery algorithm first tires to resolve the fault locally by finding an alternate path that can reconnect the disconnected tree as shown in line 3-7 in the Algorithm 4. The function CanJoinV ianeighbors(t v, C f ) tries to find alternate paths from the root of the disconnected sub-tree (v C f ) to a neighboring node that connects it to the gateway without violating the QoS parameter. It is important to note that when we run the recovery algorithm the fault provisioning factor (δ) is used to augment the QoS constraints. We study the effect of the fault provisioning factor in Section 5.4. If local recovery of the disconnected tree is not possible, the algorithm breaks up the disconnected tree T v and forms pseudo singleton clusters. From the singleton set, the most connected pseudo singleton cluster C i is chosen to join one of neighboring clusters. The algorithm finds a suitable neighboring cluster C c which can accept C i without violating

94 Algorithm 5.1: Fault Recovery Algorithm input : A, Ψ, F, S q, L q, H q, δ; [mesh, faults, QoS ] output : Ψ = {C 1, C 2,, C k }; [set of clusters] 1 count 0 ; 2 while F f aultcount do 3 C f C i, (u, v) C i, C i Ψ [the faulty cluster] T v SubT ree(v) [disconnected subtree] 4 if CanJoinV ianeighbors(t v, C f ) then 5 Join(T v, C f ) ; 6 count count + 1 ; 7 continue ; [end iteration] 8 end 9 Φ Split(T v ) ; [form singletons] 10 while Φ 0 do 11 C i P ickmostconnected(φ) ; [pick singleton] 12 Φ Φ C i ; 13 C c GetCandidateCluster(C i, δ) ; 14 if C c > 0 then 15 Ψ Shift(C c, C i ) ; [Join neighbor] 16 end 17 end 18 count count + 1 ; 19 end 79 the QoS parameters. Lines 9-17 of Algorithm 4 accomplishes this. As can be seen in Figure 5.23(a), in the simple, the fault can be recovered within the cluster without reconfiguring the clusters. In Figure 5.23(b), since the leaf node of a cluster is disconnected, it can join the adjacent cluster without violating QoS parameters. In Figure 5.23(c), which is a little more complex since a complete branch is disconnected. Here, the algorithm splits the disconnected branch into two single node clusters (singletons) and then join a neighboring cluster that can accept them. The recovery algorithm can be run partially distributed by each gateway node as a middleware application.

95 (b) (c) (a) 4 x Fault Change in Spanning Tree Figure 5.23: Fault recovery algorithm working with three cases 5.4 Algorithm Evaluation It is important to make the recovery algorithm scalable. The fault recovery algorithm complexity should be limited to the size of the perturbation and not the whole network. When small faults occur in very large networks, the recovery algorithm can quickly recover from them. In our algorithm, we try to make a local correction first by trying to reconnect with one of the neighbors and if that is not possible, it then tries to break up the disconnected spanning tree and join the neighboring clusters. Most of the algorithm works towards localized fault correction. The complexity of the algorithm is linearly

96 81 proportional to the number of faults in the system. The recovery algorithm is probabilistic in nature as its ultimate performance, in terms of recovered failures depends on the topology, location and QoS constraints. We study some of these aspect in the Section Simulation Results We ran our algorithm on over 50 different topologies representing different network sizes (n) ranging from 50 to 100 nodes. The nodes were randomly placed according to a uniform distribution on a square area. For each topology, the transmission range of each node (t r ) was varied to change the network degree and connectivity properties. We also use a several typical topologies and network traffic applications and implemented them in Qualnet [57] to study the affect of faults on network traffic. In Figure 5.25, we can see the effect of fault location on the probability of fault recovery. As the faults get closer to the gateway, i.e. the disconnected sub-tree becomes larger, the fault recovery probability reduces. As can be seen in the graph, the fault recover algorithm is very effective and has a high recovery probability. In Figure 5.25, we study the effect of fault on throughput. Here, we compare nonclustered mesh with a clustered mesh. As seen, clustered mesh gives throughput performance as load increases in comparison to non-overlay meshes. This is despite the fact that clustering reduces the number of alternate paths. In Figure 5.26 we can see similar improvements in delay performance. Clustered mesh handle faults better as it can contain the delay much better even at higher loads. Same is the case with jitter as shown in Figure 5.27

97 Normalzied Throughput Recovery Probability 82 1 Location Vs Recovery Probability Relay Load (Location Depth) Faults = 1 Faults = 2 Faults = 3 Faults = 4 Faults = 5 Figure 5.24: Effect of fault location on recovery probability (50 node network) 1.2 Load Vs Throughput Normalized Load Non-Overlay Mesh Overlay Mesh Non-overlay With Fault Overlay with Fault Figure 5.25: Load Vs Throughput with two faults and CBR traffic 5.6 Summary In this chapter we have discussed the importance of fault tolerance in wireless mesh networks. We have developed and simulated a reactive fault recovery algorithm that can be used to probabilistically recover from one or more faults. We studied the effect of the fault location on the probability of fault recovery. We have also studied the effect of faults on actual network traffic in terms of throughput, delay and jitter.

98 Jitter (ms) Delay (ms) 83 4 Load Vs Delay Normalized Load Non-Overlay Mesh Overlay Mesh Non-overlay With Fault Overlay with Fault Figure 5.26: Load Vs Delay 0.12 Load Vs Jitter Normalized Load Non-Overlay Mesh Overlay Mesh Non-overlay With Fault Overlay with Fault Figure 5.27: Load Vs Jitter

99 CHAPTER 6 Falut-Tolerance Provsioning Through Cluster Overlapping in General WMNs As we have discussed thus far, wide spread deployment of wireless mesh networks for broadband access requires careful deployment and planning in terms of laying down the network infrastructure. Deploying such networks comes with some major inter-related issues including capacity planning, scalability and access reliability. In this chapter we will look at another planning approach that aims at increasing access fairness and fault tolerance using an overlapping clustering technique. It provides alternate paths for nodes residing at the edge of the clusters and mitigates upstream blocking towards the gateways to control delay, congestion and loss rate. 6.1 Introduction The shared wireless nature of the medium makes them more susceptible to failure. Transmission link failure, in which a wireless link experience an excessive loss rate or prolonged delays, is a common case of failure. Wireless mesh networks may also be subject to a variety of other faults including faults in network elements and protocol faults [44]. For a successful deployment of WMNs in such an environment, it is essential to provide certain resilience to the network connectivity during planning and deployment to avoid potential failures [34]. WMNs planning involves several inter-dependent factors that include network topology, network coverage, traffic demand, and capacity assignment. 84

100 85 The optimal number of gateways and their locations have to be determined in advance and before deployment. In this chapter, we discuss how optimal layout cab be integrated to the planning phase of a WMN deployment to achieve more reliability. It will allow diverse routing and fault-tolerant provisioning, particularly for links that face higher blocking probability to access a gateway. Generally, an edge node of a cluster faces more blocking probability and hence higher delay and loss rate than nodes closer to the cluster head. A new clustering technique is introduced by which the gateway placement algorithm allows redundant cluster membership to improve access reliability while keeping the optimality intact. The rest of the chapter we summarizes related work with respect to the gateway placement in Section 6.2 and cover basic preliminaries and definitions in Section 6.3. Section 6.4 presents a new network clustering techniques based on maximal independent sets. Section 6.5 extends the clustering algorithm of Section 6.4 for joint cluster membership for disadvantaged nodes. 6.2 Research Issues As discussed in Chapter 3 gateway placement has a significant impact on the overall network performance including its financial viability and access reliability. In WMNs, traffic congestion is mostly due to up-stream aggregate traffic heading towards a gateway and that can be controlled by proper placement of gateways. While minimizing the number of gateways will reduce the deployment cost, fewer gateways will increase the average hop distance and consequently increases the average delay and the average relay load of the intermediate routers.

101 86 Given a significant portion of delay a packet suffers is associated with the hop count the packet travels, it is important to put a limit on the hop count towards a gateway and then optimize the other objectives. In multi-hop networks, the throughput performance of a connection decays exponentially with an increase in hop count. This is illustrated in Figure 6.28 for a simple network that is illustrated in Figure 6.29 in which a packet hops towards a gateway. A packet may runs into successive contentions and that results in higher blocking probability on each hop along the path towards its intended gateway. Each link carries a local traffic load (ρ) and relays up-stream traffic from previous nodes. Figure 6.28 illustrates the theoretical end-to-end throughput as a function Erlang blocking probability for different traffic load (ρ). An Erlang-B blocking probability with different traffic loads has been used to calculate the end-to-end throughput in Figure The blocking probability exasperate further when horizontal flows(traffic) emerge to the upstream vertical path towards the gateway. Therefore, explicit redundancy in terms of alternate paths towards a gateway will improve an individual and the overall throughput significantly. While Erlang blocking probability has been studied extensively in the content of switching networks and telephony, it is applicable to applications such as VoIP in packet switching networks or wireless cellular systems. Similar results were obtained from several experiments conducted with the Roofnet [17] deployment as discussed in Chapter 2 Section While ad hoc routing algorithms such as AODV (Ad hoc On-demand Distance Vector) and some of its variations can be used to route packets in multi-hop wireless mesh networks, generally they face a few shortcomings when directly applied to WMNs. First,

102 87 Figure 6.28: Maximum throughput performance across different traffic loads. ρ 1 ρ ρ ρ ρ... G Figure 6.29: A linear multi-hop network. their throughput performance does not typically scale to meet the expectation, particularly for real-time applications that are delay-sensitive or even loss-sensitive for data transmission. Their effective performance in terms of QoS requirements such delay, loss and jitter depends strongly on the underlying topology and the transmission range. Second, unlike ad hoc networks, where the traffic flows between arbitrary nodes, WMN traffic is either to or from a designated gateway (similar to a cellular system). A WMN routing algorithm must exploit this property to gain efficiency, which is the intention of this paper. Third, ad hoc routing algorithms are designed to deal with the possibility of highly

103 88 mobile nodes and that requires a significant amount of overhead for route discovery, mobility and maintenance. On the other hand, WMNs routers have minimal mobility. This is yet another characteristic that can be exploited for efficiency. Finally, in terms of planning, ad hoc network planning is mostly done manually without any systematic approach, and often without paying attention to the overall system cost. Because of their relatively fix position (or change of position is limited within a certain range) of WMN nodes, the implication is that the routing paths can be created that are likely to be stable. This will substantially reduce the routing overhead. The most commonly used topology for WMNs is a grid layout which is due to the layout of building and blocks. The relatively stationary topology of WMNs suggests that we can develop a more simplified routing algorithm along with a systematic approach to the planning and deployment. All these necessitate a different approach to the planning, deployment, and routing in WMNs which is the focus of this paper. Link failure in a wireless network is commonly caused due to interference in the medium or traffic congestion and on rarer occasion due to the radio malfunction. Fault tolerance has been getting a lot of attention in the area of sensory networks [44], due to the higher node failure rate, large scale of the network and the desire to increase automation. Fault recovery in such networks has been addressed in terms of routing [53], topology control [42], power assignment [28] and channel assignment [41, 56]. On the other hand, fault tolerance in wireless mesh networks, which are more stable than sensory networks, has been studied in the context of networking layer using routing protocols [27]. The routing protocols finds an alternate path to route a packet from a source to a destination if the primary path fails. However, all routing algorithms assume

104 89 some route redundancy in the underlying network topology, which is more apparent in WMNs than in sensor networks. In [18] the authors discuss fault tolerance with respect to gateway placements. To address node and link failures they modify the gateway placement LP formulation and add a fault tolerance constraint to ensure over-provisioning via multiple independent paths. They propose a greedy heuristic to address gateway placement that iteratively picks up nodes that increasingly satisfy the traffic demand without necessarily selecting a node that satisfies the most demand. Therefore, in this work we focus on building wireless network that are fault tolerant at the network topological level. In the following subsection, we give an overview of the gateway placement problem and provide the most common approaches proposed in the context of wireless mesh networks. 6.3 Preliminaries In planning, deployment or updating a wireless network, it is often necessary to determine the transmission range with an acceptable throughput. While there are many factors that affect the tranmission range, the theoretical transmission distance can be obtained from a few key specifications. Definition 6.1 (Transmission Range) Given the transmission power P t, the receiving power P r, the transmission range d can be calculated as, d = λ Pt G t G r (24) 4π P r F t

105 90 where G t, G r are the transmitting and receiving gains with an acceptable loss factor F t and λ is the wavelength of the communication channel. Definition 6.2 (Transmittance Matrix) We define the binary transmittance matrix T = [t ij ] as 1 if d ij t r, i j t ij = 0 otherwise. 1 i, j N (25) where d ij be the Euclidian distance between node i and node j obtained from Equation 24. Definition 6.3 (Reachability Matrix) The h-hop binary reachability matrix R h = [r ij ] is defined as R h = T 1 T 2 T h = h T k, (26) where is the binary OR operation, and 1 if node i is at most h hops away fromnode j, i j r ij = 0 otherwise. Definition 6.4 (Hop Count Matrix) The entries of the hop count matrix H = [h ij ] k=1 (27) give the hop distance between nodes within the reachability range such that, k if node j is within k h hops from node i h ij = 0 otherwise. (28) where h = max{h ij 1 i, j N}. Corollary 6.1 The reachability matrix R h = [r ij ] can be obtained from the hop-count matrix H as, 1 if h ij > 0 r ij = 0 otherwise. (29)

106 We have already defined Clustering in Section 3.3. We are redefining them below as we will modify the clustering constraints in the context of the proposed solution. 91 Definition 6.5 (Cluster) A cluster C(V, E ) G(V, E) is an acyclic subgraph of G such that, V V and E E. Definition 6.6 (Clustering) A clustering is a way of partitioning graph G(V, E) and can be formally defined as a set of clusters Ω, where, Ω = {C 1, C 2,, C k }, 1 k N, (30) with the following properties: P 1 : i, j = 1 k C i C j =, i j P 2 : k i=1 C i = G (31) P 3 : V (Ω) = V (G) P 4 : E(Ω) = k i=1 E(C i) E(G). Property P 1 guarantees that clusters are independent with no nodes in common. Relaxing this property allows overlapping clusters. Ω can be represented by an N N asymmetric binary matrix with k non-zero rows, each representing a cluster with exactly a 1 on each column, characterizing each node to belong only to one cluster. Formally, Ω N N = [ω ij ] {0, 1} with the following constraints, (a) : (b) : N w ij = 1, i=1 N w ij > 0, j=1 1 j N for some i (32)

107 92 Constraint (a) guarantees that each node belongs only to one cluster, and constraint (b) makes node i as a cluster head with its member nodes j, where ω ij = 1 (1 j N). Later, in Section 6.5, we relax the property P 1 in Equation 31 and its corresponding constraint (a) in Equation 32 to allow a node to participate in more than one cluster. 6.4 Planning with Disjoint Clustering By strictly applying property P 1 in Equation 31 along with constraint (a) in Equation 32, a clustering matrix in the form of Equation 25 can be formulated to represent an optimal set non-overlapping clusters covering the mesh network. One of the important QoS requirements in WMNs is to determine the maximum number of hops a packet can travel before reaching its intended destination (gateway). For that, we form the h-hop reachability matrix R h from Equation 26 that identifies the reachability set for each node on its rows. This can be viewed as an initial clustering (trivial clusters) in which every node is considered to be a cluster head with all its members within h-hop distance. Clearly, this will create the maximum possible number of clusters (N) with maximum overlap amongst them. However, condition P 1 in Equation 31 is not satisfied for non-overlapping clusters. To satisfy property P 1, we introduce a cluster graph in which cluster C i is connected to cluster C j if C i C j =, 1 i, j N, i j. We further define the corresponding clustering overlap matrix as follows. Definition 6.7 (Clustering Overlap Matrix) The entries of the clustering overlap

108 93 Matrix, O = [o ij ] is defines as, N r ik r jk i j o ij = k=1 0 i = j 1 i, j N, (33) where is the binary operation AND, and o ij is the inner product of row i and row j of R h. In effect o ij gives the number of common nodes in two adjacent clusters headed by nodes i and j are considered two cluster heads. We define the adjacency clustering matrix A = [a ij ] that describes relationships between clusters as follows. Definition 6.8 (Clustering Adjacency Matrix) The clustering adjacency matrix A = [a ij ] is defined as, 1 if o ij I c 0 a ij = 0 otherwise. (34) For disjoint clustering, first, we consider the case where I c = 1. We start with the transmission matrix T in Equation 25 and a maximum clustering radius of h. We compute the reachability matrix within h-hop distance for each node according to Equation 26. The clustering adjacency matrix A identifies the relationship between potential clusters in terms of node sharing. We define matrix A as the complement of A where, 1 if a i,j = 0 a ij = 0 if a i,j = 1 (35) A identifies all pair-wise disjoint clusters. Definition 6.9 (Inter Cluster Distance) Inter cluster distance D h is defined as the maximum number of hops between any two clusters.

109 94 To find the optimal location of cluster heads with maximum coverage, one has to find the maximum clique (maximum complete subgraph) of the graph associated with the adjacency matric A. We use the Algorithm original developed by [65] to find the largest clique (complete subgraph). The current implementation of the algorithm searches for maximal independent vertex sets in the complementer graph. Given we have applied the constraint of hop-count h on each cluster, depending on the network topology, the algorithm does not necessarily cover all the nodes in the clustering. Consider the 100-node mesh network of Figure The initial clustering is shown in Figure 6.30: A 100-node mesh network. Figure 6.31 in which 8 gateways optimally cover the network with maximum hop count

110 95 h = 2. The initial clustering does not cover nodes all the nodes mainly due to the intercluster constraints applies. For example, nodes {41, 45, 48, 52, 55, 57, 60, 64, 66, 73, 79, 81, 82} have not been assigned to any of the clusters due to: (i) the maximum 2-hop coverage (h = 2) by the cluster heads, and (ii) the inter-cluster distance I c = 1, i.e., neighboring clusters are at least one hop away from each other. However, uncovered nodes nodes Figure 6.31: Initial clustering with h = 2 and I c =1. are at most h hops away from a nearby cluster. For that we identify the inter-cluster distance matrix for the above clustering algorithm. Theorem 6.1 A node is either a cluster head or at most 2h hops away from a cluster head.

111 Proof: Given, the clustering algorithm forms only disjoint clusters, there are two cases. If k i=1 C i = G, then the clustering algorithm covers all nodes in the network and Property P 2 holds. Every node is within h hops from a cluster head and no nodes lies between two adjacent clusters, and hence D h = 1. If k i=1 C i G, then there is at least one node that does not belong to any of the clusters. Let v G but v / k i=1 C i be such a node. Let the closest cluster to v be C i with its cluster head node u. The h-hop reachability set of v is either disjoint or it has some nodes in common with the h-hop reachability set of u. Let R h (v) and R h (u) be the h-hop reachability sets for node v and u, respectively. 96 Case 1: R h (v) R h (u). Let w be a common node in both reachability sets. Then the hop distance H(v, w) h and the hop distance H(u, w) h. Hence H(u, v) 2h. Case 2: R h (v) R h (u) =. Then v by itself constitutes an independent reachability set within its h radius and forms an independent cluster. R h (v) R h (u) R h (v) R h (u) v w h u v h h u Case 1 Case 2 Figure 6.32: Inter-cluster distance. From Theorem 6.1, we can conclude the following corollaries. Corollary 6.2 The maximum inter-cluster distance D h = h.

112 Corollary 6.3 A node that has not been assigned to any clusters is at most h hops away from a neighboring cluster. 97 This is shown in Figure 6.31 in which nodes {41, 45, 48, 52, 55, 57, 60, 64, 66, 73, 79, 81, 82} are either one hop or two hops away from a neighboring cluster, where h = 2. After the initial clustering, we will find the nearest cluster for the remaining nodes them to join. This is shown in Figure Therefor, the radius of the final clustering is at most 2h = 4 Figure 6.33: Final disjoint clustering, 2 h 4, I c = 1. hops. While the inter-cluster distance I c = 1 is one hop among the neighboring clusters, due to the network topology some clusters are affected by the residual nodes left out from the constraint h in algorithm 7.

113 98 Algorithm 6.1: Disjoint Clustering Input : Transmittance Matrix T, h, I c = 1 Output: Array C of cluster heads 1 Calculate R h = h k=1 T k Eqn Calculate o ij = N k=1 r ik r jk Eqn i, j N 3 Calculate A from O for I c = 1 Eqn Calculate A from A 5 Use the maximal independent set [65] to identify the cluster heads. 6 Form clusters by incorporating the reachability set (from R h ) for each cluster head. 7 Assign nodes outsides clusters to the closest cluster Analysis The processing time involved in Steps 1-4 in Algorithm 7 are all based on twodimensional matrices (mostly sparse matrices) and bounded by O(N 2 ). The processing time and memory space in step 5 are bounded by O(N + m) and O(Nmδ), respectively, where N is the number of nodes, m is the number of edges and δ is the maximal independent sets of the graph [65]. 6.5 Planning with Joint Clustering By relaxing property P 1 in Equation 31 and constraint (a) in Equation 32, we can obtain clusters that can share available bandwidth at the edge of clusters. Nodes at the edge of clusters belong to more than on cluster simply because they are in disadvantage positions as far as gateway access is concerned. They can dynamically switch their cluster membership due to a weak or bad connection at the edge of each cluster. This can be achieved in two ways; i) making D h = 1 and allow the inter-cluster links be shared by the neighboring clusters, or ii) make clusters overlap by one or more hops. Figure 6.34 shows how inter-cluster links can shared by two neighboring clusters. Note that the objective

114 of this paper is to compensate access disparity with access redundancy for those nodes further away from a gateway to improve their throughput. 99 Figure 6.34: Joint clustering with R = 2, I c = 1. In this clustering scheme, nodes that are h hops away from a gateway have memberships in more than one cluster. The joint clustering algorithm is simply an extention of disjoint clustering algorithm with inter-cluster nodes having at least dual membership in neighboring clusters. This is shown in Algorithm 7. The difference between Algorithms 7 and 7 are in steps 3 and 7. Figures 6.35 and 6.36 show one hop (h = 1) clustering with I c = 1 and I c = 2, respectively. Similarly, Figure 6.37 for h = 2 and I c = 3 The choice for h and I c depends on the planning. Clearly, increasing I c reduces the number

115 100 Algorithm 6.2: Joint Clustering Input : Transmittance Matrix T, h, I c 1 Output: Array C of cluster heads 1 Calculate R h = h k=1 T k Eqn Calculate o ij = N k=1 r ik r jk Eqn i, j N 3 Calculate A from O for a given I c Eqn Calculate A from A 5 Use the maximal independent set [65] to identify the cluster heads. 6 Form clusters by incorporating the reachability set (from R h ) for each cluster head. 7 Assign nodes outsides clusters to adjacent clusters. Figure 6.35: h = 1, I c = 1 of clusters and hence the number of gateways and higher fault-tolerance. The drawback is the amount of delay.

116 101 Figure 6.36: h = 1, I c = Simulation Results We used Qualnet [57] simulator to verify our analysis. In the experiment as shown in figure 6.38, we allowed overlapping of clusters in the planning phase, thus enabling node S to be part of cluster C1 and C2. We created a traffic flow from node S to node D. We then studied the effect of load on the throughput with and without the overlapping clustering. Figure 6.39 illustrates the effect of load on the throughput with a node belongs to one or two clusters. The throughput for a node is significantly higher if it belong to two clusters.

117 102 Figure 6.37: h = 2, I c = Summary In this chapter we developed a new clustering technique that improves fault-tolerance in wireless mesh networks. It also mitigate throughput disparity for nodes distant way from a gateway by allowing them join multi-cluster membership.

118 Throughput (Kbps) 103 C1 C2 G S G G D C3 Figure 6.38: Overlapping clusters simulation setup Non-Overlapping Overlapping Load Figure 6.39: Load vs. throughput with and without overlapping

119 CHAPTER 7 Conclusion and Future Work In this dissertation we have discussed Wireless Mesh Networks as a promising architecture in the context of last mile and wide area network deployments. There are several advantages that WMNs offer that have been summarized in Section 1.4. However there are also several challenge to fully exploit their benefits as discussed in Section 2.4. In this research we focused on the placement of wired gateways subject to quality of service (QoS) constraints and fault recovery algorithms. The placement of wired gateways in multi-hop wireless mesh networks (WMNs) has a significant impact on network throughput performance, cost and capacity to satisfying the QoS requirements as well as fault tolerance. In the context of gateway placement, the QoS is influenced by the number of gateways, the number of nodes served by each gateway, the location of the gateways, and the relay load on each wireless router. We have developed a polynomial time approximation algorithm [24] that can find a solution that is very near the optimal solution. We then compared our algorithm with others presented in literature. We have discussed the importance of fault tolerance, especially in wireless networks due to the susceptible and shared nature of the wireless medium. Mesh networks, due to their multi-point to multi-point architecture, inherently lend themselves to being more resilient to faults. However protocols and deployment strategies are needed to exploit these properties of WMNs. 104

120 105 We first studied cellular mesh networks (WCNs) that make use of directional antenna and have a hexagonal topology. We have presented an efficient and deterministic faulttolerant routing algorithm that can handle multiple link failures, which are common to WCNs [23]. While it does improve QoS for some application, when it comes to faults, this scheme increases the packet delay caused by the detour, due to the additional hops needed to route around the faulty link. WCN networks offer lot of advantages in terms of fault tolerance against link failures, they come with their own set of challenges and limitations. One deployment challenge is that the directional antennas need to be in line of sight of each other. While this is possible in many cases, even in a three dimensional environment, it is impossible to achieve this in every environment. Therefore they cannot be deployed in every environment. For general WMNs, we developed a reactive fault recovery algorithm that can be used to probabilistically recover from one or more faults [25]. We studied the effect of the fault location on the probability of fault recovery using simulation. Depending on the type of traffic the fault tolerant routing can provide improved service quality and reliability. Finally we developed a new clustering technique that improves fault-tolerance in wireless mesh networks. It also mitigate throughput disparity for nodes distant way from a gateway by allowing them join multi-cluster membership [26]. We also plan to implement a distributed protocol that will enable overlay mesh networks to better handle faults using the technique. While we addressed few problem in WMNs, there are many interesting challenges that still remain. An interesting application to which we can extend the current work is load balancing. While a congested link may not be a fault, a similar re-routing approach

121 can be used to uniformly distribute the traffic load across the network. 106

122 BIBLIOGRAPHY [1] IEEE b WG. Wireless lan medium access control(mac) and physicallayer(phy) specifications - high-speed physical layer (phy) in the 2.4 ghz band. Standard, [2] s Task Group. update.htm. [3] J. Padhye A. Wolman A. Adya, P. Bahl. Reconsidering wireless systems with multiple radio. In ACM SIGCOMM Computer Communications Reviews, volume 35, [4] J. Padhye A. Wolman A. Adya, P. Bahl and L. Zhou. A multi-radio unification protocol for ieee wireless networks. In International Conferences on Broadband Networks (BroadNets), 2004, [5] Prof Chiueh A Raniwala. Architecture and algorithms for an ieee based multi-channel wireless mesh network. In Proc. of IEEE INFOCOM, 2005, [6] N. Abramson. The aloha system - another alternative for computer communications Fall Joint Computing Conf, AFIPS Confernece Proceedings, 37: , [7] S. M. Alamouti. A simple transmit diversity technique for wireless communications. Selected Areas in Communications, IEEE Journal on, 16(8): , [8] Mansoor Alicherry, Randeep Bhatia, and Li (Erran) Li. Joint channel assignment and routing for throughput optimization in multi-radio wireless mesh networks. In MobiCom 05: Proceedings of the 11th annual international conference on Mobile computing and networking, pages 58 72, New York, NY, USA, ACM Press. [9] Member-Datta Amitava. A fault-tolerant protocol for energy-efficient permutation routing in wireless networks. IEEE Trans. Comput., 54(11): , [10] Bassam Aoun, Raouf Boutaba, Youssef Iraqi, and Gary Kenward. Gateway placement optimization in wireless mesh networks with QoS constraints. IEEE Journal on Selected Areas in Communications, 24(11), Nov [11] Sanjeev Arora, Prabhakar Raghavan, and Satish Rao. Approximation schemes for euclidean k-medians and related problems. In STOC 98: Proceedings of the thirtieth annual ACM symposium on Theory of computing, pages , New York, NY, USA, ACM Press. [12] P. Bahl. Opportunities and challenges of community mesh networking, May [13] Suman Banerjee and Samir Khuller. A clustering scheme for hierarchical control in multi-hop wireless networks. In Proceedings of IEEE INFOCOM, pages ,

123 [14] Yigal Bejerano. Efficient integration of multihop wireless and wired networks with QoS constraints. IEEE/ACM Trans. Netw., 12(6): , [15] Andrew Brzezinski, Gil Zussman, and Eytan Modiano. Distributed throughput maximization in wireless mesh networks via pre-partitioning. IEEE/ACM Trans. Netw., 16(6): , [16] Alberto Cerpa, Jennifer L Wong, Louane Kuang, Miodrag Potkonjak, and Deborah Estrin. Statistical model of lossy links in wireless sensor networks. IPSN 05: Proceedings of the 4th international symposium on Information processing in sensor networks, [17] Benjamin A. Chambers. The grid roofnet: a rooftop ad hoc wireless network. MIT Masters Thesis, [18] Ranveer Chandra, Lili Qiu, Kamal Jain, and Mohammad Mahdian. Optimizing the placement of integration points in multi-hop wireless networks. Proceedings of IEEE ICNP, [19] Romit Roy Choudhury, Xue Yang, Ram Ramanathan, and Nitin H. Vaidya. On designing mac protocols for wireless networks using directional antennas. IEEE Transactions on Mobile Computing, 05(5): , [20] V. Milutinovic D. Milutinovic. The honeycomb architecture. IEEE Computer, 20:81 83, Apr [21] J. Deng and Z. Haas. Dual busy tone multiple access (dbtma): A new medium access control for packet radio networks. In IEEE ICUPC 98, Florence, Italy, October [22] Y Drabu and H. Peyravi. An adaptive bandwidth control algorithm for IP routers. In Communications in Computing, pages , Las Vegas, USA, June CSREA Press. [23] Y Drabu and H Peyravi. Fault tolerant routing in tri-sector wireless cellular mesh networks. Proceedings Parallel and Distributed Computer Systems, pages , [24] Yasir Drabu and Hassan Peyravi. Gateway placement in wireless mesh networks with qos constraints. Seventh International Conference on Networking (ICN 2008), Nov [25] Yasir Drabu and Hassan Peyravi. Fault recovery with QoS provisioning in wireless mesh networks. In Proceedings of the International Conference on Wireless Networks, pages , July [26] Yasir Drabu and Hassan Peyravi. Planning with joint clustering in multi-hop wireless mesh and sensor networks. In Proceedings of the International Conference Networking, January

124 [27] Richard Draves, Jitendra Padhye, and Brian Zill. Routing in multi-radio, multihop wireless mesh networks. In MobiCom 04: Proceedings of the 10th annual international conference on Mobile computing and networking, pages , New York, NY, USA, ACM Press. [28] S. Agrawal et al. Distributed power control in ad hoc wireless networks. PIMRC, [29] Zhenghua Fu, Petros Zerfos, Haiyun Luo, Songwu Lu, Lixia Zhang, and Mario Gerla. The impact of multihop wireless channel on tcp throughput and loss. In Proceedings of IEEE INFOCOM, pages , [30] Violeta Gambiroza, Bahareh Sadeghi, and Edward W. Knightly. End-to-end performance and fairness in multihop wireless backhaul networks. In In Proceedings of ACM MOBICOM, pages , [31] Christopher J. Glass and Lionel M. Ni. The turn model for adaptive routing. In 25 Years ISCA: Retrospectives and Reprints, pages , [32] Michelle X. Gong, Scott F. Midkiff, and Shiwen Mao. A cross-layer approach to channel assignment in wireless ad hoc networks. Mob. Netw. Appl., 12(1):43 56, [33] Bay Area Wireless Users Group. [34] Weilin Wang Ian F. Akyildiz, Xudong Wang. Wireless mesh networks: a survey. Computer Networks, 47(5): , March [35] J. Mitola III. Software Radio Architecture: Object- Oriented Approaches to Wireless System Engineering. Wiley Inter-Science, [36] Johnathan Ishmael, Sara Bury, Dimitrios Pezaros, and Nicholas Race. Deploying rural community wireless mesh networks. IEEE Internet Computing, 12(4):22 29, [37] P Karn. Maca - a new channel access method for packet radio. In ARRL/CRRL Amateure Radio 9th Computer Networking Conference, September [38] Vinay Kolar, Karthik Bharath, Nael B. Abu-Ghazaleh, and Janne Riihijarvi. Contention in multi-hop wireless networks: model and fairness analysis. In MSWiM 09: Proceedings of the 12th ACM international conference on Modeling, analysis and simulation of wireless and mobile systems, pages 21 29, New York, NY, USA, ACM. [39] P. Krishnan, Danny Raz, and Yuval Shavitt. The cache location problem. IEEE/ACM Trans. Netw., 8(5): , [40] Alfred A. Kuehn and Michael J. Hamburger. A heuristic program for locating warehouses. Management Science, 9(4): ,

125 [41] K. L. Eddie Law and Adam Kohn. Topology designs with controlled interference for multi-radio wireless mesh networks. In Mobility 08: Proceedings of the International Conference on Mobile Technology, Applications, and Systems, pages 1 6, New York, NY, USA, ACM. [42] Xiang-Yang Li, Peng-Jun Wan, Yu Wang, and Chih-Wei Yi. Fault tolerant deployment and topology control in wireless ad hoc networks: Research articles. Wirel. Commun. Mob. Comput., 4(1): , [43] Cao Liang, Xinming Huang, and Jing Deng. A fault tolerant and energy efficient routing protocol for urban sensor networks. In InfoScale 07: Proceedings of the 2nd international conference on Scalable information systems, pages 1 8, ICST, Brussels, Belgium, Belgium, ICST (Institute for Computer Sciences, Social-Informatics and Telecommunications Engineering). [44] Qi Han Lilia Paradis. A survey of fault management in wireless sensor networks. Journal of Network and Systems Management, 15(2): , June [45] Kitae Nahm, Ahmed Helmy, and C.-C. Jay Kuo. Tcp over multihop networks: issues and performance enhancement. In MobiHoc 05: Proceedings of the 6th ACM international symposium on Mobile ad hoc networking and computing, pages , New York, NY, USA, ACM. [46] Fabian Garcia Nocetti, Ivan Stojmenovic, and Jingyuan Zhang. Addressing and routing in hexagonal networks with applications for tracking mobile users and connection rerouting in cellular networks. IEEE Trans. Parallel Distrib. Syst., 13(9): , [47] US Department of Commerce. 21st century america s progress toward universal broadband internet access. In internet use report Feb2010.pdf, page 12, [48] Kouji Ohboshi. Mobile, broadband, ubiquitous, and the information renaissance. In ISSS 02: Proceedings of the 15th international symposium on System Synthesis, pages 1 1, New York, NY, USA, ACM. [49] Behrooz Parhami and Ding-Ming Kwai. A unified formulation of honeycomb and diamond networks. IEEE Trans. Parallel and Distributed Systems, 12:74 80, Jan [50] Charles E. Perkins and Elizabeth M. Royer. Ad-hoc on-demand distance vector routing. In IN PROCEEDINGS OF THE 2ND IEEE WORKSHOP ON MOBILE COMPUTING SYSTEMS AND APPLICATIONS, pages , [51] Community Wireless Network project in Seattle. [52] Ashish Raniwala, Kartik Gopalan, and Tzi cker Chiueh. Centralized channel assignment and routing algorithms for multi-channel wireless mesh networks. SIGMOBILE Mob. Comput. Commun. Rev., 8(2):50 65,

126 [53] E. Royer and C. Toh. A review of current routing protocols for ad-hoc mobile wireless networks, Apr [54] Mischa Schwartz. Mobile Wireless Communications. Cambridge Press, [55] David B. Shmoys, E. Tardos, and Karen Aardal. Approximation algorithms for facility location problems. In STOC 97: Proceedings of the twenty-ninth annual ACM symposium on Theory of computing, pages , New York, NY, USA, ACM Press. [56] Weisheng Si, Selvadurai Selvakennedy, and Albert Y. Zomaya. An overview of channel assignment methods for multi-radio multi-channel wireless mesh networks. J. Parallel Distrib. Comput., 70(5): , [57] Qualnet Network Simulator. [58] I. Stojmenovic. Honeycomb networks: Topological properties and communication algorithms. IEEE Trans. Parallel and Distributed Systems, 8: , Oct [59] Ning Student Member-Li and Jennifer C. Senior Member-Hou. Localized faulttolerant topology control in wireless ad hoc networks. IEEE Trans. Parallel Distrib. Syst., 17(4): , [60] Karthikeyan Sundaresan, Vaidyanathan Anantharaman, Hung-Yun Hsieh, and Raghupathy Sivakumar. Atp: a reliable transport protocol for ad-hoc networks. In MobiHoc 03: Proceedings of the 4th ACM international symposium on Mobile ad hoc networking & computing, pages 64 75, New York, NY, USA, ACM. [61] Csaba Szabó, Károly Farkas, and Zoltán Horváth. Motivations, design and business models of wireless community networks. Mob. Netw. Appl., 13(1-2): , [62] Rice University Technology For All (TFA). [63] Emulab Network Emulation Testbed. [64] F. A. Tobagi and L. Kleinrock. Packet switching in radio channels: Part 1: Carrier sense multiple-access models and their throughput-delay characteristics. IEEE Trans. Communication., 23(12): , [65] Shuji Tsukiyama, Mikio Ide, Hiromu Ariyoshi, and Isao Shirakawa. A new algorithm for generating all the maximal independent sets. SIAM J. Comput., 6(3): , [66] Jie Wu. A fault-tolerant and deadlock-free routing protocol in 2d meshes based on odd-even turn model. Computers, IEEE Transactions on, 52(9): , Sep [67] Adel Youssef, Mohamed F. Younis, Moustafa Youssef, and Ashok Agrawala. Establishing overlapped multihop clusters in wireless sensor networks. Int. J. Sen. Netw., 2(1/2): ,