Multicast in Wireless Mesh Networks

Transcription

1 Multicast in Wireless Mesh Networks JIN XU A thesis submitted to the Faculty of Graduate Studies in partial fulfilment of the requirements for the degree of Master of Science Graduate Programme in Computer Science and Engineering York University Toronto, Ontario October 2006

2 Abstract Wireless Mesh Networks (WMNs) form a new class of networks that has emerged recently. In a WMN, the wireless mesh routers are stationary and form the wireless mesh backbone, which provides multi-hop connectivity for mobile mesh hosts to communicate with either other mesh hosts or the Internet via access points. In such WMNs, multicast is an efficient way to distribute data to a group of receivers. In this thesis, we first propose a framework that supports multicast routing between the wired Internet and wireless mesh networks in a simple and efficient manner. We then propose a protocol named (Shortest Path Tree algorithm for wireless Mesh networks) for efficient and scalable multicast routing inside the mesh backbone of a WMN. The protocol builds source-based trees based on the network topology. It avoids flooding and employs an effective mechanism to prevent the implosion and exposure problems when a tree is constructed and when nodes join and leave. Our simulation results show iv

3 that the protocol outperforms existing protocols such as ODMRP (On- Demand Multicast Routing Protocol), MOSPF (Multicast Extensions to Open Shortest Path First), MST (Minimum Steiner Tree) and MNT (Minimum Number of Transmissions) in terms of packet delivery ratio, throughput, and end-to-end delay. Based on the protocol, we further propose an effective reliable multicast protocol for WMNs called Mesh Reliable Multicast (MRM). With the assistance of the underlying multicast routing protocol, MRM builds a recovery tree that dynamically coincides with the multicast routing tree. The cost of constructing the recovery tree is kept to a minimum, and packet losses are repaired locally using the recovery tree. Simulation results justify the correctness and effectiveness of the proposed MRM protocol in WMNs. v

4 Acknowledgements This thesis could not have been written without the encouragement, collaboration, and support of a tremendous number of people. To my supervisor, Professor Uyen Trang Nguyen, thank you for your patience in thesis supervision, encouragement, advice and guidance. Step by step, you have shown me how to discover the problems and solve them, which has resulted in this valuable research. To my committee members, Professors Amir Asif, Mariana Kant and Baoxin Hu, thank you for taking the time to read this thesis and giving me valuable comments. I am grateful to the administrative assistants and system administrators of the department for assisting me in many different ways. I would also like to thank my colleagues Qian Bai, Frank Xu, Xong Xing, Lan Nguyen, Van Ngo, David Xia, Pengfei Liu and Haoyuan Wang. vi

5 port. Finally, my special thanks to my family for their unconditional love and sup- vii

6 Table of Contents Abstract iv Acknowledgements vi Table of Contents viii List of Figures xii 1 Introduction Motivations Contributions The Multicast Framework The Mesh Multicast Routing Protocol The Mesh Reliable Multicast Protocol Thesis Outline viii

7 2 Literature Survey Multicast Routing Protocols Multicast Routing on the Internet Multicast Routing in MANETs Multicast Routing in WMNs Reliable Multicast Protocols Group-based Protocols Structure-based Protocols Simulators The Multicast Framework Network Model The Multicast Framework Summary Multicast Routing in WMNs Multicast Routing in WMNs: The Protocol Assigning APs to Mesh Routers Acquiring the Network Topology ix

8 4.1.3 Multicast Routing in the Mesh Backbone: the Protocol Integration with Multicast Routing in the Internet Experimental Results vs. ODMRP vs. MOSPF vs. MST/MNT Summary Reliable Multicast in WMNs Reliable Multicast in WMNs: the MRM protocol Overview Protocol Details Experimental Results Performance Metrics Simulation Setting Simulation Results Summary Conclusion 126 x

9 Bibliography 128 xi

10 List of Figures 1.1 Architecture of a WMN Deployment of a WMN in a city Deployment of a WMN in a rural area Tree-based protocols The sender is an Internet host; receivers are in the WMN (Case 1(b)) The sender and receivers are in the same WMN (Case 2(a)) The sender is a mesh host; receivers (R 3 and R 4 ) are Internet hosts (Case 2(b)) An join operation example between a sender S and its three receivers R 1, R 2 and R Case 1: The sender in Internet and some receivers in WMN xii

11 4.3 Case 2: The sender in WMN and some receivers in Internet vs. ODMRP: functions of traffic load, one sender, 10 receivers, network of 100 nodes vs. ODMRP: functions of traffic load, one sender, 30 receivers, network of 300 nodes vs. ODMRP: functions of number of receivers, one sender, traffic load = 40 pkts/s, network of 100 nodes vs. ODMRP: functions of number of receivers, one sender, traffic load = 40 pkts/s, network of 300 nodes vs. ODMRP: functions of number of senders, 40 receivers, traffic load = 10 pkts/s per sender, network of 100 nodes vs. ODMRP: functions of number of senders, 40 receivers, traffic load = 10 pkts/s per sender, network of 300 nodes The impact of multicast on unicast: vs. ODMRP, 10 common receivers/sender, network of 100 nodes The impact of multicast on unicast: vs. ODMRP, 10 common receivers/sender, network of 300 nodes xiii

12 4.12 vs. MOSPF: functions of traffic load, one sender, 20 permanent receivers plus 20 transient receivers (join/leave interval=5s), network of 100 nodes vs. MOSPF: functions of traffic load, one sender, 20 permanent receivers plus 20 transient receivers (join/leave interval=5s), network of 300 nodes The impact of multicast on unicast: vs. MOSPF, 10 common permanent receivers/sender plus 20 common transient receivers (join/leave interval=5s), network of 100 nodes The impact of multicast on unicast: vs. MOSPF, 10 common permanent receivers/sender plus 20 common transient receivers (join/leave interval=5s), network of 300 nodes vs. MOSPF: functions of traffic load, one sender, 20 permanent receivers plus 20 transient receivers (join/leave interval=10s), network of 100 nodes vs. MST/MNT: functions of traffic load, one sender, 20 receivers, network of 100 nodes vs. MST/MNT: functions of traffic load, one sender, 40 receivers, network of 100 nodes xiv

13 4.19 vs. MST/MNT: functions of traffic load, one sender, 40 receivers, network of 300 nodes vs. MST/MNT: functions of traffic load, one sender, 80 receivers, network of 300 nodes Average path lengths of, MST and MNT, network of 100 nodes Average path lengths of, MST and MNT, network of 300 nodes The impact of multicast on unicast: vs. MST/MNT, 10 common receivers/sender, network of 100 nodes The impact of multicast on unicast: vs. MST/MNT, 10 common receivers/sender, network of 300 nodes vs. MST/MNT: functions of traffic load (20 unicast flows in background), one sender, 20 receivers, network of 100 nodes vs. MST/MNT: functions of traffic load (20 unicast flows in background), one sender, 40 receivers, network of 100 nodes vs. MST/MNT: functions of traffic load (20 unicast flows in background), one sender, 80 receivers, network of 100 nodes xv

14 4.28 vs. MST/MNT: functions of traffic load (20 unicast flows in background), one sender, 20 receivers, network of 300 nodes vs. MST/MNT: functions of traffic load (20 unicast flows in background), one sender, 40 receivers, network of 300 nodes vs. MST/MNT: functions of traffic load (20 unicast flows in background), one sender, 80 receivers, network of 300 nodes Example of DR selection The sender s operation A sender s send window A receiver s receive window and the associated bit vector nodes in a 2000m 2000m terrain nodes in a 3000m 3000m terrain xvi

15 Chapter 1 Introduction Wireless Mesh Networks (WMNs) [1, 2] form a new class of networks that has emerged recently. Major components of a WMN include wireless mesh routers, wireless mesh hosts, and access points (or gateways) that act both as Internet routers and wireless mesh routers (Figure 1.1). The mesh routers in a WMN are stationary; they form the wireless mesh backbone, which provides multi-hop connectivity from mesh hosts to either other mesh hosts or the Internet via access points. The mesh hosts can be stationary or mobile; they can form a wireless local area network (LAN) or a mobile ad-hoc network (MANET), and communicate with the outside world via connections to the mesh routers. A WMN is dynamically self-organized and self-configured with nodes in the network automatically 1

16 establishing and maintaining mesh connectivity among themselves. This feature brings many benefits to WMNs such as low installation cost, large-scale deployment, reliability, and self-management. Figure 1.2 illustrates a typical application of WMNs deployed in a new town. The wireless mesh routers are installed on the roofs of the buildings to cover an area of the town. It is also possible to implement a wired network in the area. However, a WMN incurs a much lower installation cost. Another example is a WMN deployed in a rural area, as shown in Figure 1.3. In this example, it is very hard to build a wired network due to the mountains and the forests. Deploying a WMN, on the other hand, is relatively easy. Figure 1.1: Architecture of a WMN 2

17 Figure 1.2: Deployment of a WMN in a city Figure 1.3: Deployment of a WMN in a rural area 3

18 1.1 Motivations WMNs are promising for providing Internet access to remote areas. They can provide large coverage area, reduce dead-zones in wireless coverage, and lower costs of backhaul connections for base stations. Their promise of rapid deployability and reconfigurability makes them suitable for important applications such as disaster recovery, homeland security, transient networks in convention centers, hard-to-wire buildings such as museums, unfriendly terrains, and rural areas with high costs of network deployment. Most of the existing work on WMNs concentrates on the issues of unicast routing and channel assignment when multiple channels are being used [3, 4, 5, 6], network architectures [7, 8], performance evaluation and analysis [9, 10], and network capacity analysis [11]. In this thesis, we focus on providing multicast services in WMNs. Multicast is a form of communication that delivers information from a source to a set of destinations simultaneously in an efficient manner; the messages are delivered over each link of the network only once (excluding retransmissions) and only duplicated at branch points, where the links to the destinations split. Important applications of multicast include distribution of stock quotes, billing records, software, and newspapers; audio/video conferencing; distance education; 4

19 and Internet games. Not much research on multicast in WMNs has been done. The only existing work is the routing protocol proposed by Ruiz et al. [12]. In this thesis we propose effective solutions to the problem of multicast in WMNs. 1.2 Contributions Specifically, we propose solutions to the following three problems. First, we propose a framework that supports multicast routing between the wired Internet and wireless mesh networks in a simple and efficient manner. Second, we design a protocol named (Shortest Path Tree algorithm for WMNs) for efficient and scalable multicast routing inside the mesh backbone of a wireless mesh network. Third, we propose an effective reliable multicast protocol named Mesh Reliable Multicast (MRM) based on the protocol The Multicast Framework Our proposed multicast framework offers the following advantageous features: 1. Transparency of hosts locations. Communications between an Internet host and a mesh host require assistance from access points. On the other hand, mesh hosts located in the same WMN may send data to each other via mesh 5

20 routers without involving the access points (APs). These two routing scenarios imply that we would need a mechanism to signify the locations of the communicating hosts (i.e., in the Internet, in the same WMN or in another WMN) in order to use the appropriate routing method (i.e., with or without assistance from APs). One such mechanism would be to create different IP address formats for hosts in the Internet and hosts in WMNs respectively. This approach would require efforts to define the address formats and to standardize them. An alternative would be to let a host advertise its location (in the Internet or in a WMN) to other hosts. This mechanism would create traffic overhead in both the Internet and WMNs. Considering the above disadvantages, we designed a multicast framework that does not require any knowledge of host locations. Yet it effectively supports communications between the Internet hosts and mesh hosts via access points, and multi-hop ad hoc communications between mesh hosts residing in the same WMN. 2. No changes to Internet multicast routing. The framework does not require any modifications to the current IP multicast model in the Internet, nor to the existing multicast routing protocols in the Internet. Multicast routing in the wired network can use any of the existing Internet routing protocols such as DVMRP [21], MOSPF [22], PIM [24], or CBT [25]. This enables the framework 6

21 to be deployed quickly and interact seamlessly with the IP multicast and the existing Internet multicast routing protocols The Mesh Multicast Routing Protocol The problem of efficient multicast routing in a wired network has been studied in depth. Several multicast routing protocols have been proposed for wired networks such as DVMRP (Distance Vector Multicast Routing Protocol) [21], MOSPF (Multicast Extensions to Open Shortest Path First) [22], PIM (Protocol Independent Multicast) [24], and CBT (Core-Based Tree) [25]. Although these protocols work well in wired networks, they are not suitable for wireless environments. In wireless networks, bandwidth is a scarce resource and wireless links are more error-prone than their wired counterparts. On the other hand, multicast routing protocols proposed for wireless mobile ad hoc networks such as MAODV (Multicast Ad hoc On-Demand Distance Vector Routing) [32], ODMRP (On-Demand Multicast Routing Protocol) [33], CAMP (Core-Assisted Mesh Protocol) [34], or AMRIS (Ad hoc Multicast Routing protocol utilizing Increasing id-numbers) [35] do not work efficiently in WMNs because they assume nodes are mobile while mesh routers are static. Mechanisms designed to deal with node mobility such as periodic route refreshment by flooding [33] or 7

22 beacon exchange [34] would create a lot of unnecessary overheads in WMNs. Recently, Ruiz et al. [12] argued that, in a wireless environment, the transmission of a multicast data packet from a node to any number of its neighbors can be done with a single data transmission. Thus in a multi-hop wireless mesh network the minimum-cost multicast tree is one which connects the source and receivers by issuing a minimum number of transmissions, rather than having a minimum edge cost. They also showed that the problem of minimizing the number of transmissions is NP-complete and proposed several heuristics to compute the approximation of such optimal trees, which we call Minimum Number of Transmissions Trees (MNT). Although MNT trees have a lower total edge cost and require less transmissions than Shortest Path Trees (SPTs), the mean source-to-destination path length in a MNT tree is longer than that in a shortest path tree. Since wireless links are error-prone, the longer the path, the higher the probability that a data packet gets lost or damaged, due to packet collision and/or congestion, resulting in a throughput reduction [54]. The proposed multicast routing protocol considers the broadcast nature of wireless mesh routers as well as their static placement/positions in a WMN so as to be efficient and scalable. Specifically, the proposed protocol builds a shortest-path multicast routing tree, which is rooted at the sender and spans 8

23 all receivers. It avoids flooding and employs effective mechanisms to prevent the feedback implosion and exposure problems during the tree construction process and when members join and leave. Moreover, the protocol can easily integrate with multicast routing in the Internet without any changes to the existing Internet multicast routing protocols. Our simulation results show that the protocol outperforms the existing protocols such as ODMRP [33], MOSPF [22], MST [14] and MNT [12] in terms of packet delivery ratio, throughput, and delay The Mesh Reliable Multicast Protocol Distribution of data using multicast in an unreliable network does not guarantee reliable delivery, which is the prime requirement for many important applications, such as distribution of software, stock quotes, electronic newspapers, billing records, and medical images. The second part of the thesis focuses on designing and implementing an effective reliable multicast protocol called Mesh Reliable Multicast (MRM) based on the aforementioned multicast routing protocol. Although many reliable multicast protocols have been proposed for wired networks, such as SRM (Scalable Reliable Multicast) [36], RMTP (Reliable Multicast Transport Protocol) [38], TMTP (Tree-based Multicast Transport Protocol) [39], 9

24 LBRRM (Log-Based Receiver-Reliable Multicast) [37], and LSM (Large-Scale Multicast) [40], these protocols can not be applied directly to WMNs due to the inherent broadcast nature of wireless transmissions. For example, SRM is a group-based protocol that multicasts the packet-loss request and the repair packet globally to the whole multicast group. Although SRM uses some mechanisms to reduce the number of messages generated, it is still inefficient to recover local losses using global mechanisms. Because bandwidth is a scarce resource in wireless networks, SRM consumes too much bandwidth and results in poor performance. Unlike SRM, RMTP uses a tree-based approach in which a recovery tree is constructed at the transport layer; the parents and the children keep state about each other; and the recovery messages traverse only between a parent and its children. Although RMTP is efficient to do the loss recovery locally, the recovery tree is different from the underlying routing tree, which is not efficient in WMNs. TMTP employs distributed algorithms based on heuristics to estimate the topology of the routing tree and builds a dynamic recovery tree at the transport layer. However, it is hard to create a recovery tree that faithfully follows the underlying routing tree. Moreover, the recovery tree construction is expensive and not suitable for WMNs. With the assistance of the underlying multicast routing protocol, the pro- 10

25 posed reliable multicast protocol MRM builds a recovery tree that dynamically coincides with the underlying multicast routing tree. The cost of constructing the recovery tree is kept to a minimum and packet losses are repaired locally using the recovery tree. In addition, MRM can easily interact with the reliable multicast protocols used in the Internet, such as RMTP. Experimental results are presented to demonstrate the effectiveness, scalability and performance of the MRM protocol in WMNs. 1.3 Thesis Outline The remainder of the thesis is organized as follows. Chapter 2 provides a literature survey of the existing multicast routing protocols and reliable multicast protocols. The multicast framework is described in detail in Chapter 3. Chapter 4 presents the protocol for multicast routing in a mesh backbone and its integration with multicast routing in the Internet. Experimental results are included to quantify the performance of the protocol in WMNs. The MRM protocol is described in detail in Chapter 5. We also present experimental results to demonstrate the correctness and effectiveness of the MRM protocol in WMNs. Chapter 6 concludes the thesis and outlines some directions for future work. 11

26 Chapter 2 Literature Survey In this chapter, we present a literature survey of the existing multicast routing protocols and reliable multicast protocols. 2.1 Multicast Routing Protocols In this section, we review the existing multicast routing protocols for the Internet, wireless mobile ad hoc networks (MANETs), and wireless mesh networks (WMNs). Multicast routing algorithms for the Internet can be divided broadly into three categories: shortest path tree algorithms, minimum cost tree algorithms, and constrained tree algorithms. A shortest path tree algorithm computes a tree 12

27 rooted at the sender and spanning all the receivers such that the distance from the sender to each receiver is minimum. Unlike shortest path tree algorithms, a minimum cost tree algorithm tries to minimize the overall cost of the tree. A constrained tree algorithm tries to optimize on both constraints (i.e., the shortest paths and the minimum cost). However, the majority of multicast routing protocols used in the Internet today are based on shortest path trees because they are easy to implement and they provide minimum delay from the sender to each receiver, which is a desirable property for most of the real-life multicast applications. Multicast routing protocols proposed for wireless mobile ad hoc networks can be classified into two categories: tree-based and mesh-based. Studies have shown that in a mobile ad hoc network, where the network topology changes frequently, mesh-based protocols outperform tree-based protocols due to the fact that redundant routes in a mesh provide alternative paths for data delivery when one or more links fail. For wireless mesh networks, Ruiz et al. [12] studied the problem of building a multicast tree which requires a minimum number of transmissions. They showed that the problem of minimizing the number of transmissions is NP-complete and proposed heuristics to compute the approximation of such an optimal tree. 13

28 2.1.1 Multicast Routing on the Internet As stated earlier, the multicast routing algorithms for the Internet can be divided broadly into the following three categories: 1. Shortest path tree algorithms 2. Minimum cost tree algorithms 3. Constrained tree algorithms Shortest Path Tree Algorithms The goal of the shortest path tree (SPT) algorithms is to compute a tree rooted at the sender and spanning all the receivers such that the distance between the sender and each receiver along the tree is minimum. The two most well-known algorithms for computing shortest path trees are Bellman-Ford and Dijkstra. SPTs by definition are per sender. Thus for multiple source multicast, separate multicast trees need to be computed, one for each sender. The dynamic versions of these static algorithms as used in a network are distance vector algorithms [16] and link state algorithms [16]. In the distance vector algorithms, the router attached to the sender broadcasts to its neighboring routers that it can reach the sender at a cost of one. That is, its 14

29 distance from the sender is one hop. Based on this information, the neighboring routers compute their distance to the sender and choose the minimum among all possible alternatives. Each of these routers then broadcasts its distance to its neighboring routers and the process is repeated. The link-state approach is based on the Dijkstra s shortest path algorithm in which each router knows the complete topology of the network at any instant of time. In order to maintain an up-to-date network topology, the routers flood a change in the state of a directly connected link as soon as such a change occurs. For example, if a directly connected link goes down or comes up, a router immediately broadcasts the information to all its neighbors and so on until every router gets the information and updates its own view of the network topology. Once a router has the complete topology, it can use the Dijkstra s shortest path algorithm to compute the shortest path form the sender to each receiver. These paths collectively constitute the shortest path multicast tree Minimum Cost Tree Algorithms Unlike shortest path tree algorithms whose goal is to minimize the distance from the sender to each receiver, minimum cost tree algorithms try to minimize the overall cost of the tree. By assigning a cost to each edge of the graph, minimum 15

30 cost tree algorithms compute the tree which minimizes the sum of the cost of its edges. In practice, the cost of an edge (a link) can be the packet transmission delay on that link, or the distance between the two vertices (routers). The problem of finding a minimum cost multicast tree is known as the minimum Steiner tree problem, which is NP-complete and there exist several heuristics that compute the approximate Steiner trees. For instance, the MST algorithms [14, 15] provide a 2-approximation, while Zelikovsky [17] proposed an algorithm which obtains a 11/6-approximation. In practice, the MST heuristic [14] is used because it is simpler to implement. We also implement the MST heuristic to evaluate its performance in WMNs. Following is a brief description of this heuristic. MST Heuristic Given a connected undirected graph G = (V, E, d) and a set of Steiner points (multicast group members) S, where V is the set of vertices in G, E is the set of edges in G, d is a distance function which maps E into the set of nonnegative numbers and S is a subset of vertices in V, consider the complete undirected graph G 1 = (V 1, E 1, d 1 ) constructed from G and S in such a way that V 1 = S and, for every (v i, v j ) in E 1, d(v i, v j ) is set equal to the distance of the shortest path from v i to v j in G. Notice that, for each edge in G 1, there corresponds a 16

31 shortest path in G. Given any spanning tree in G 1, one can construct a subgraph of G by replacing each edge in the tree by its corresponding shortest path in G. The MST heuristic algorithm is as follows. 1. Construct the complete undirected graph G 1 = (V 1, E 1, d 1 ) from G and S. 2. Find the minimal spanning tree T 1 of G Construct the subgraph G s of G by replacing each edge in T 1 by its corresponding shortest path in G. 4. Find the minimal spanning tree T s of G s. 5. Construct a Steiner tree T h from T s by deleting edges in T s, if necessary, so that all the leaves in T h are Steiner points Constrained Tree Algorithms Shortest path tree algorithms try to minimize the distance between the sender and each receiver. Another way of interpreting this is to say that the shortest path tree algorithms try to minimize end-to-end delay. On the other hand, minimum Steiner tree algorithms aim at minimizing the overall cost of the tree. However, in a real-life scenario, it is important to minimize both. For example, 17

32 in a multimedia multicast case, minimizing the end-to-end delay is as important as minimizing the overall cost. References [18], [19], and [20] have studied this problem and suggested heuristics to optimize on both cost and delay Multicast Routing Protocols in Practice The majority of the multicast routing protocols used in the Internet today are based on shortest path trees. The reasons are that they are easy to implement and they provide minimum delay from the sender to each receiver, which is a desirable property for most of the real-life multicast applications. The multicast routing protocols used (or proposed to be used) in the Internet are DVMRP [21], MOSPF [22], PIM [24], and CBT [25]. Distance Vector Multicast Routing Protocol (DVMRP) DVMRP is a distance vector routing protocol. It uses flooding and pruning to build the multicast tree. The routers in the leaf subnets have group membership information. When a router receives a flooded packet, it knows whether that packet will be useful for its subnet or not. In case there is no group member on the subnet, the leaf router sends a prune message to its neighboring routers. In addition, a leaf router can send a prune message through all interfaces except 18

33 for the one on the reverse shortest path to the sender. When an intermediate router receives prune messages from all interfaces except for the reverse shortest path interface, it forwards the prune message upstream. This way, the unwanted branches of the spanning tree get pruned off. When a router sends a prune message, it maintains information about the (Source, Group) pair for which the prune message was sent. This state is used to prevent propagation of the data packets when they arrive at those routers. DVMRP is a soft-state protocol in the sense that the state in the routers times out, and hence the process of flooding and pruning needs to be repeated periodically. However, if a member wants to join a group before the next flooding takes place and there is no host on the subnet currently subscribed to the group, DVMRP allows the corresponding router to send a graft message. The graft message propagates upstream using the reverse path forwarding interface until it reaches a router that is part of the shortest path tree. In fact, a graft message cancels the prune state at the relevant routers. Multicast Extensions to Open Shortest Path First (MOSPF) MOSPF is an enhancement of the unicast routing protocol OSPF [23]. OSPF is a link-state routing protocol in which the routers advertise the state of their directly attached links, and based on these advertisements, each router builds up a link- 19

34 state database. The OSPF link-state database provides a complete picture of the topology of an Autonomous System (AS). In order to support multicast routing, a new type of link-state advertisement, referred to as the group-membership-lsa (Link State Advertisement), has been added to OSPF. These advertisements help to pinpoint the locations of all multicast group members in the database. The path of a multicast data packet can then be calculated by building a shortest-path tree rooted at the source of the data packet. Note that each router in the domain has the complete description of the topology and the membership information, and each one of them uses exactly the same algorithm to computer the shortestpath tree rooted at the same node. Thus every router ends up computing the same tree and creating the corresponding forwarding entries for each group. The shortest path trees are built on demand (that is, when the first packet arrives), and the results of this computation are cached for use by subsequent packets having the same source and destination. Protocol Independent Multicast (PIM) PIM is a multicast routing protocol that does not depend on a specific unicast routing protocol. PIM uses a set of special routers called Rendezvous Points (RPs) such that any receiver who wants to join the multicast group needs to 20

35 send an explicit join request to a unique RP determined based on the multicast group address. The receivers explicitly join the RP resulting in the formation of a unidirectional shared RP-tree. Senders, on the other hand, do not explicitly join the RP but send their data encapsulated in register messages directly to the RP for distribution using the shared RP-tree. A sender who wants to multicast to the group starts by sending encapsulated packets to the corresponding RP, which then forwards the packets to the attached receivers. If the sender s traffic increases beyond some threshold, the shortest path route is set up between the sender and the corresponding RP. In addition, the intermediate routers between the RP and the receivers switch from the RP-based shared tree to a source-based shortest path tree. PIM uses soft state mechanisms to maintain the tree. That is, control messages are sent periodically by the relevant routers to refresh the state information. No explicit teardown mechanism is needed to remove states when a group ceases to exist. Core-Based Tree (CBT) CBT is a multicast routing protocol which sets up a single shared bi-directional tree connecting the senders and the receivers. Every multicast session has a special router called the core, which is instrumental in setting up the multicast 21

36 tree. Every member of the multicast group sends an explicit Join message towards the core, resulting in the creation of a branch ending in either the core or an existing branch of the tree. All senders use the same tree for transmitting data, and the routers simply forward the data on all the interfaces for the CBT except for the incoming interface. The main advantage of using a single shared tree per group is the reduction of state information that needs to be maintained at each router. The main drawback of using a single shared tree in a network with a large degree of connectivity is traffic concentration. If every sender uses the same shared distribution tree, traffic is concentrated along the links of the shared tree. This is not the case if separate trees are used by the sources. Another drawback of a single shared tree is that the sender and the receivers are not necessarily connected by the shortest path. The delivery delay can thus be higher in shared trees than in source-based shortest path trees. The CBT routers in the network maintain hard state. Therefore, there is a need for explicit tear down of the multicast tree. Although the protocols described above work well in wired networks, they are not suitable for the wireless environment. In wireless networks, bandwidth is a scarce resource, and wireless links are more error-prone than their wired coun- 22

37 terparts. DVMRP requires a multicast sender to periodically flood the network in order to prune or graft branches to keep the multicast tree up-to-date. In MOSPF, routers maintain multicast group membership information in addition to the topological information. Based on these two pieces of information, MO- SPF routers compute the shortest paths from the sender to the group members using the Dijkstra s shortest path algorithm. When a member joins or leaves the group, the group membership advertisement is flooded to the entire MOSPF routing domain to inform the routers in the domain. Flooding is expensive in a wireless environment because it consumes bandwidth and causes a high number of packet collisions and thus packet losses. PIM introduces the notion of a rendezvous point (RP), which acts as a meeting place for the receivers and the sender. The receivers explicitly join the RP resulting in the formation of an unidirectional shared RP-tree, and then the sender sends the data to the RP for distribution using the shared RP-tree. CBT sets up a single shared bidirectional tree connecting the senders and the receivers. All senders use the same tree for transmitting the data and the routers simply forward the data on all the interfaces (except for the incoming interface of the data) for the tree. In PIM and CBT, the routing paths are sub-optimal and longer than necessary. The data packets thus need to traverse more hops to reach the 23

38 receivers. Since the wireless links are more error-prone, the longer the path, the more likely it is for a data packet to be lost or damaged. In WMNs, we can take advantage of the knowledge of the network topology to build the tree in a more efficient manner Multicast Routing in MANETs Multicast routing protocols proposed for MANETs can be classified broadly into two categories: tree-based and mesh-based. Typical tree-based protocols are AM- RIS (Ad hoc Multicast Routing protocol utilizing Increasing id-numbers) [35] and MAODV (Multicast Ad hoc On-Demand Distance Vector Routing) [32]. Example mesh-based protocols are ODMRP (On-Demand Multicast Routing Protocol) [33] and CAMP (Core-Assisted Mesh Protocol) [34]. Existing studies [13] show that in a mobile ad hoc network, where the network topology changes frequently, mesh-based protocols outperform tree-based protocols in terms of packet delivery ratio and throughput due to the fact that redundant routes in the mesh provide alternative paths for data delivery when a path breaks. Among the multicast routing protocols proposed for MANETs, ODMRP is the most popular because it is simple and gives high packet delivery ratio. Following is a review of ODMRP. 24

39 On-demand Multicast Routing Protocol (ODMRP) ODMRP creates a mesh of nodes (the forwarding group), which forwards multicast packets via flooding within the mesh, thus providing path redundancy. In ODMRP, group membership and multicast routes are established and updated by the source on demand. When multicast sources have data to send, they flood JOIN QUERY packets over the network. When a node receives a non-duplicate JOIN QUERY, it stores the upstream node ID (i.e., backward learning) and rebroadcasts the packet. (The node will discard duplicate JOIN QUERY packets.) When a JOIN QUERY packet reaches a multicast receiver, the receiver creates a JOIN REPLY and broadcasts the JOIN REPLY to the neighbors. The JOIN REPLY packet includes the next hop ID and the sender IDs. When a node receives a JOIN REPLY, it checks if the next hop ID of one of the entries in the JOIN REPLY matches its own ID. If it does, the node realizes that it is on the path to the source and is thus a forwarding node. To reduce the control overhead, a forwarding node consolidates all the JOIN REPLY packets it receives during a period of ReplyConsolidationInterval, then broadcasts its own JOIN REPLY built upon matched entries. The JOIN REPLY is thus propagated by each forwarding node until it reaches the multicast source via the shortest delay path. This process constructs (or updates) the routes from the sources to the receivers, 25

40 and builds a mesh of nodes, the forwarding group. Multicast senders refresh the routes by sending JOIN QUERY periodically. ODMRP works independently of the unicast routing protocols and can itself be used as one Multicast Routing in WMNs Recently, Ruiz et al. [12] argued that in a wireless environment, the transmission of a multicast data packet from a node to any number of its neighbors can be done with a single data transmission. Therefore, in a multi-hop wireless mesh network, the minimum-cost multicast tree is one that connects the source and receivers by issuing a minimum number of transmissions, rather than having a minimal edge cost. They also showed that the problem of minimizing the number of transmissions is NP-complete and proposed MNT heuristics [12] to compute the approximation of such optimal trees. Although such minimum-cost trees save bandwidth, the paths from the source to the receivers in such a minimum-cost tree are longer than the paths produced by a shortest path algorithm. Since wireless links are more error-prone, the longer the path, the more likely that a data packet will be lost or damaged. Another problem of such minimum-cost trees is that when nodes join or leave the multicast group, the tree is no longer optimal and it must be rebuilt to maintain the optimality. This problem cannot 26

41 be solved without incurring traffic and computation overheads. The worst case occurs when nodes keep joining and leaving the group. Following is the description of one of the MNT heuristics proposed by Ruiz et al. [12], which we implemented to compare with our proposed protocol. MNT heuristic Algorithm MNT MF φ V V {s} aux R Cov(s) + {s} repeat v argmax v V ( Cov(v) )s.t.cov(v) 2 aux aux Cov(v) + {v} V V {v} MF MF + {v} until aux = φ or v = null if aux φ then Build Steiner tree among nodes in aux. end if In the above algorithm, s is the sender; V is the set of all nodes; R is the set of receivers; MF is the set of forwarding nodes; aux is the set of nodes to cover; Cov(v) is the set of nodes that node v covers. This algorithm systematically builds different cost-effective subtrees. Costeffectiveness refers to the fact that a node v is selected to be a forwarding node only if it covers two or more nodes. The algorithm starts by initializing the nodes- 27

42 to-cover set (aux) to all the receivers, except for those which are already covered by the source s. Initially, the set of forwarding nodes (MF) is empty. After the initialization, the algorithm repeats the process of building a cost-effective tree, starting with node v which covers more nodes in aux. Then, v is inserted into the set of forwarding nodes (MF) and it becomes a node to cover. In addition, the receivers covered by v (Cov(v)) are removed from the list of nodes-to-cover (aux). This process is repeated until all the nodes are covered, or it is not possible to find more Steiner nodes, guaranteeing the cost-effectiveness. In the latter case, the different subtrees are connected by a Steiner tree among their roots. 2.2 Reliable Multicast Protocols IP multicast provides only best-effort services. It does not guarantee message delivery, ordering, throughput, or end-to-end delay. Many multicast applications, however, have requirements beyond best-effort services. For example, bulk-data transfers such as file distribution require error-free delivery of data (but they can usually tolerate relatively large delay and delay jitter). Some applications, such as distributed whiteboards [36] and distributed games, require both errorfree and real-time delivery. Therefore, various multicast transport protocols are proposed to provide reliable delivery and/or quality-of-service guarantees on top 28

43 of the multicast routing protocols. An end-to-end reliable multicast service ensures that all packets from a sender are delivered to all receivers of the multicast session within a finite amount of time and free of errors. Lost or damaged packets are recovered by Automatic Repeat Request (ARQ) and/or Forward Error Correction (FEC). ARQ is a retransmission on demand mechanism based on retransmissions of lost packets. ARQ provides total reliability but retransmissions increase packet delivery latency. FEC [41] allows error correction codes (also referred to as parity packets) to be transmitted along with the original data from the sender to the receiver. Depending on how many error correction codes are sent and how many packets are lost at the receiver, it is possible to recover missing packets based on the received packets. The FEC-based approach thus reduces packet delivery latency at the cost of additional bandwidth needed for the parity packets. However, the optimal amount of error correction code in FEC is usually difficult to set due to heterogeneous loss probabilities in a multicast group and the burst of losses. Moreover, FEC alone is not enough to guarantee complete reliability in practice. Thus we chose ARQ as the error recovery mechanism for our reliable multicast protocol. 29

44 A reliable multicast protocol needs to address the following issues: Implosion problem: A problem that occurs when the loss of a packet triggers simultaneous messages, either requests or replies, from a large number of receivers. Exposure problem: A problem that occurs when recovery-related messages reach receivers which have not experienced loss. Recovery latency: The latency experienced by a member from the instant a loss is detected until the repair packet is received. A reliable multicast is scalable if it can support a very large number of receivers with low recovery latency, and minimal implosion and exposure. Several ARQ-based reliable multicast protocols have been proposed in the literature [36, 38, 39, 43]. These protocols can be classified into two broad categories: group-based, and structure-based. RBP (Reliable Broadcast Protocol) [42], RMP (Reliable Multicast Protocol) [43] and SRM (Scalable Reliable Multicast) [36] are typical group-based protocols. Example structure-based protocols are RMTP (Reliable Multicast Transport Protocol) [38] and TMTP (Tree-based Multicast Transport Protocol) [39]. 30

45 2.2.1 Group-based Protocols Group-based protocols require no structure among the members of a multicast group. Levine et al. [44] classify them into sender-initiated and receiver-initiated protocols Sender-initiated Protocols The sender-initiated approach places the responsibility of error recovery on the sender, which maintains the state information of all of its receivers. This is accomplished by having receivers return positive acknowledgments (ACKs) for the packets correctly received, and having the sender detect and retransmit lost packets. As the number of receivers increases, the sender is overwhelmed by a large amount of ACKs from the receivers and the task of retransmission. This scheme is thus not scalable, and the recovery latency may be high Receiver-initiated Protocols In contrast, receiver-initiated protocols place the responsibility of loss recovery on both the receivers and the sender. A receiver multicasts negative acknowledgments (NACKs) to other receivers and the sender when packets are lost. The sender or any receiver having the lost packet can multicast the packet. Pin- 31

46 gali et al. [45] showed that receiver-initiated protocols are far more scalable than sender-initiated protocols because the sender of a receiver-initiated protocol does not need to maintain the state of all the receivers, and the task of retransmission is distributed among all group members. However, receiver-initiated protocols may experience the problems of request implosion and repair implosion. In other words, several receivers may request for the same data, and several members may retransmit the same data. A suppression mechanism must be used to reduce the number of duplicate NACKs or duplicate repair packets. The suppression mechanism can be timer-based (as in SRM (Scalable Reliable Multicast) [36]), or probabilistic (as in TRM (Transport Protocol for Reliable Multicast) [46]). Exposure is also a potential problem with this approach since retransmission is done via multicast Structure-based Protocols This class of protocols imposes a logical structure among group members. The two common structures are rings and trees. 32

47 Ring-based Protocols Ring-based protocols arrange the receivers in the form of a ring. Data are transmitted in one direction around the ring. One of the receivers at any instance of time is designated as a token site. The token site is responsible for sending ACKs back to the sender. The sender retransmits packets if it does not receive an ACK from the token site within a timeout period. Receivers send NACKs to the token site when packets are lost. The token site then retransmits the lost packets. The token is passed to the next receiver in the ring when that receiver has correctly received all packets that the former site has received. One of the first ring-based reliable multicast protocols is the Token Ring Protocol (TRP) [42]. The Reliable Multicast Protocol (RMP) [43] is an updated version of TRP and is designed for wide area networks (WANs) Tree-based Protocols The basic principle of tree-based protocols is to distribute the burden of acknowledgment handling and retransmission among members of a multicast group in order to achieve better scalability of error recovery. These protocols require the receivers of a multicast group to be organized into local regions, each with a representative called a Designated Receiver (DR). The local regions are them- 33

48 S DR DR DR DR R R R R R R R R R R DR R R R Figure 2.1: Tree-based protocols selves organized into a hierarchical tree structure. The receivers in a local region request lost packets from their parent DR. The DR retransmits repair packets to the receivers in its local region. If the DR does not have a requested packet, it forwards the request to the next higher level DR, and so on, until the request reaches a DR that has the packet (see Figure 2.1). This local recovery leads to shorter recovery latency and reduces the number of data units flowing through a global network. The protocols that belong to this category are RMTP (Reliable Multicast Transport Protocol) [38], TMTP (Tree-based Multicast Transport Protocol) [39], LBRRM (Log-Based Receiver-Reliable Multicast) [37], and Lorax [48]. 34