Advanced Wireless Networks - 4G Technologies

Figure 13.35 Message reduction using fisheye.

Table 13.2 Node density (nodes vs area)

is calculated as a linear function of the original packet bits. The process has its analogy in coding theory. The resulting packet (information and overhead) is fragmented into smaller blocks and distributed over the available paths. The probability of reconstructing the original information at the destination is increased as the number of used paths is increased.

A lot of research has been done in the area of multipath routing in wired networks. One of the initial approaches to this problem was the dispersity routing [44]. In order to achieve self-healing and fault tolerance in digital communication networks, diversity coding is suggested in Ayanoglu et al. [45]. In Krishnan and Silvester [46], a per-packet allocation granularity for multipath source routing schemes was shown to perform better than a per-connection allocation.

Figure 13.36 Control O/H vs number of nodes.

An exhaustive simulation of the various tradeoffs associated with dispersity routing is presented in Banerjea [47]. The inherent capability of this routing method to provide a large variety of services was pointed out. Owing to this fact, numerous schemes employing multipath routing have been proposed for wired networks in order to perform QoS routing [48–55]. All these protocols are based on proactive routing, since they maintain tables that reflect the state of the entire network. For this reason, owing to the unreliability of the wireless infrastructure and the nodal mobility, which can trigger an excessive amount of updates in the state tables, they cannot be successfully applied to mobile networks.

The application of multipath techniques in mobile ad hoc networks seems natural, as multipath routing allows reduction of the effect of unreliable wireless links and the constantly changing topology. The on-demand multipath routing scheme is presented in Nasipuri and Das [56] as a multipath extension of dynamic source routing (DSR) [57], described in Section 13.1. The alternate routes are maintained, so that they can be utilized when the primary one fails. TORA [58], routing on demand acyclic multipath (ROAM) [59] and ad hoc on-demand distance vector-backup routing (AODV-BR) [60], which is based on the ad hoc on-demand distance vector (AODV) protocol [61], are also examples of schemes that maintain multiple routes and utilize them only when the primary root fails. However, these protocols do not distribute the traffic into the multiple available paths.

Another extension of DSR, multiple source routing (MSR) [62], proposes a weighted round-robin heuristic-based scheduling strategy among multiple paths in order to distribute load, but provides no analytical modeling of its performance. The split multipath routing (SMR), proposed in Lee and Gerla [63], focuses on building and maintaining maximally disjoint paths; however, the load is distributed only in two routes per session. In Papadimitratos et al. [64], the authors propose a novel and nearly linear heuristic for constructing a highly reliable path set. In Pearlman et al. [65], the effect of alternate path routing (APR) on load balancing and end-to-end delay in mobile ad hoc networks has been explored. It was argued, however, that the network topology and channel characteristics (e.g. route coupling) can

500 AD HOC NETWORKS

severely limit the gain offered by APR strategies. In an interesting application [66], multipath path transport (MPT) is combined with multiple description coding (MDC) in order to send video and image information in a multihop mobile radio network. In this section , we discuss a multipath scheme for mobile ad hoc networks based on diversity coding [45]. Data load is distributed over multiple paths in order to minimize the packet drop rate and achieve load balancing in a constantly changing environment.

Suppose that nmax paths are available for the transmission of data packets from a source to a destination. Any of the multipath schemes mentioned in the introduction can be employed in order to acquire these paths. No paths have nodes in common (mutually disjoint).

Each path, indexed as i, i = 1, . . . nmax, is either down at the time that the source attempts to transmit with probability of failure pi or the information is received correctly with probability 1 − pi . Since there are no common nodes among the paths, they are considered independent in the sense that success or failure of one path cannot imply success or failure of another. It should be noted here that, in wireless ad hoc networks, nodes are sharing a single channel for transmission, so node disjointness does not guarantee the independence of the paths. Taking this into account, the paths are ideally considered independent as an approximation of a realistic ad hoc wireless network. For a more realistic modeling of the paths in a wireless network, one may refer to Tsirigos and Haas [67], where path correlation is included in the analysis. The failure probabilities of the available paths are organized in the probability vector p = { pi }, in such a way that pi ≤ pi+1. The vector of success probabilities is defined as q {qi } = 1 − p = {1 − pi }.

Let us now suppose that we have to send a packet of D data bits utilizing the set of available independent paths in such a way as to maximize the probability that these bits are successfully communicated to the destination. This probability is denoted as P. In order to achieve this goal, we employ a coding scheme in which C extra bits are added as overhead. The resulting B = D + C bits are treated as one network-layer packet. The extra bits are calculated as a function of the information bits in such a way that, when splitting the B-bit packet into multiple equal-size nonoverlapping blocks, the initial D-bit packet can be reconstructed, given any subset of these blocks with a total size of D or more bits. First, we define the overhead factor r = B/D = b/d where b and d take integer values and the fraction b/d cannot be further simplified. One should note that 1/r would be equivalent to coding gain in channel coding theory.

Next we define the vector v = {vi }, where vi is the number of equal-size blocks that is allocated to path i. Some of the paths may demonstrate such a poor performance that there is no point in using them at all. This means that we might require using only some of the available paths. If n is the number of the paths we have to use in order to maximize P, it would be preferable to define the block allocation vector v = {vi } as a vector of a variable size n, instead of fixing its size to the number of available paths nmax.

Given the fact that the probability failure vector is ordered from the best path to the worst one, a decision to use n paths implies that these paths will be the first n ones. Based on these observations, the allocation vector v = {vi } has the following form: v =

it follows that vi ≥ vi+1, because a path with higher failure probability cannot be assigned fewer blocks than a path with a lower failure probability. The original D-bit packet is fragmented into N w-size blocks, d1, d2, d3, . . . , dN , and the C-bit overhead packet

into M w-size blocks, c1, c2, c3, . . . , cM . Based on this we have N = D/w = a/r and M = C/w = (r − 1)N = (r − 1)a/r. Path 1 will be assigned the first v1 blocks of the B-bit sequence, path two will receive the next v2 blocks, and so on. Thus, path i will be assigned vi blocks, each block of size w. Like parity check bits in error correcting (N + M,M) block coding, the overhead symbols are generated as linear combination of the original packets as

where multiplication and summation are performed in Galois Fields G F(2m ). The relations between probability of successful packet transmission P, parameters N and M and link failure probabilities are available from coding theory and will not be repeated here. One of the important results from that theory is that the block size has to satisfy the following inequality, so that the original information can be recovered [45]:

By incorporating the previous definitions in Equation (13.2), we obtain an inequality for the number of blocks, into which we can split the B-bit packet

13.5 CLUSTERING PROTOCOLS

13.5.1 Introduction

In dynamic cluster-based routing, described so far in this chapter, the network is dynamically organized into partitions called clusters, with the objective of maintaining a relatively stable effective topology [70]. The membership in each cluster changes over time in response to node mobility and is determined by the criteria specified in the clustering algorithm. In order to limit far-reaching reactions to topology dynamics, complete routing information is maintained only for intracluster routing. Intercluster routing is achieved by hiding the topology details within a cluster from external nodes and using hierarchical aggregation, reactive routing or a combination of both techniques. The argument made against dynamic clustering is that the rearrangement of the clusters and the assignment of nodes to clusters may require excessive processing and communications overhead, which outweigh its potential benefits. If the clustering algorithm is complex or cannot quantify a measure of cluster stability, these obstacles may be difficult to overcome. A desirable design objective for an architectural framework capable of supporting routing in large ad hoc networks subject to high rates of node mobility incorporates the advantages of cluster-based routing and balances the tradeoff between reactive and proactive routing while minimizing the shortcomings of each. Furthermore, the consequences of node mobility suggest the need to include a quantitative measure of mobility directly in the network organization or path selection process.

Specifically, a strategy capable of evaluating the probability of path availability over time and of basing clustering or routing decisions on this metric can help minimize the reaction

502 AD HOC NETWORKS

to topological changes. Such a strategy can limit the propagation of far-reaching control information while supporting higher quality routing in highly mobile environments.

In this section we present the (c,t) cluster framework, which defines a strategy for dynamically organizing the topology of an ad hoc network in order to adaptively balance the tradeoff between proactive and on demand-based routing by clustering nodes according to node mobility. This is achieved by specifying a distributed asynchronous clustering algorithm that maintains clusters which satisfy the (c,t) criteria that there is a probabilistic bound c on the mutual availability of paths between all nodes in the cluster over a specified interval of time t. In order to evaluate the (c,t) criteria, a mobility model is used that characterizes the movement of nodes in large ad hoc networks. It is shown how this model is used to determine the probability of path availability when links are subject to failure due to node mobility.

Based on the (c,t) cluster framework, intracluster routing requires a proactive strategy, whereas intercluster routing is demand-based. Consequently, the framework specifies an adaptive-hybrid scheme whose balance is dynamically determined by node mobility. In networks with low rates of mobility, (c,t) clustering provides an infrastructure that is more proactive. This enables more optimal routing by increasing the distribution of topology information when the rate of change is low. When mobility rates become very high, cluster size will be diminished and reactive routing will dominate. The (c,t) cluster framework decouples the routing algorithm specification from the clustering algorithm, and thus, it is flexible enough to support evolving ad hoc network routing strategies described so far in both the intraand intercluster domains.

Several dynamic clustering strategies have been proposed in the literature [70–73]. These strategies differ in the criteria used to organize the clusters and the implementation of the distributed clustering algorithms. McDonald and Znati [74] use prediction of node mobility as a criteria for cluster organization. Clustering decisions in [70–73] are based on static views of the network at the time of each topology change. Consequently, they do not provide for a quantitative measure of cluster stability. In contrast, the (c, t) cluster strategy [74] forms the cluster topology using criteria based directly on node mobility. According to Ramanathan and Steenstrup [73], the ability to predict the future state of an ad hoc network comprising highly mobile nodes is essential if the network control algorithms are expected to maintain any substantive QoS guarantees to real-time connections. The multimedia support for wireless network (MMWN) system proposed by Ramanathan and Steenstrup [73] is based upon a hybrid architecture that includes the characteristics of ad hoc and cellular networks. Their framework uses hierarchical routing over dynamic clusters that are organized according to a set of system parameters that control the size of each cluster and the number of hierarchical levels. Aggregation of routing information is used to achieve scalability and limit the propagation of topological change information. A multilevel strategy is used to repair virtual circuit (VC) connections that have been disturbed due to node mobility. MMWN does not predict node movement. Consequently, it is unable to provide a quantitative bound on the stability of its cluster organization.

Vaidya et al. [72] proposed a scheme that dynamically organizes the topology into k clusters, where nodes in a cluster are mutually reachable via k-hop paths. The algorithm considers k = 1 and reduces to finding cliques in the physical topology. Using a first-fit heuristic, the algorithm attempts to find the largest cliques possible. Although the algorithm does not form optimal clusters, it still requires a three-pass operation each time a topology change occurs: one for finding a set of feasible clusters, a second for choosing the largest

of the feasible clusters that are essential to maintain cluster connectivity, and a third to eliminate any existing clusters that are made superfluous by the new clusters.

The objective of the scheme proposed by Lin and Gerla [70] differs significantly from the previous examples. Rather than using clustering to minimize the network’s reaction to topological changes, their scheme is intended to provide controlled access to the bandwidth and scheduling of the nodes in each cluster in order to provide QoS support. Hierarchical routing and path maintenance were a secondary concern. The proposed algorithm is very simple and uses node ID numbers to deterministically build clusters of nodes that are reachable by two-hop paths.

The zone routing protocol (ZRP), described in Section 13.2 is a hybrid strategy that attempts to balance the tradeoff between proactive and reactive routing. The objective of ZRP is to maintain proactive routing within a zone and to use a query–response mechanism to achieve interzone routing. In ZRP, each node maintains its own hop-count constrained routing zone; consequently, zones do not reflect a quantitative measure of stability, and the zone topology overlaps arbitrarily. These characteristics differ from (c,t) clusters, which are determined by node mobility and do not overlap. Both strategies assume a proactive routing protocol for intrazone/cluster routing, and each organizes its topology based upon information maintained by that protocol. ZRP also defines the query control scheme to achieve interzone routing. Although ZRP is not a clustering algorithm and the (c,t) cluster framework is not a routing protocol, the comparison demonstrates a close relationship that could be leveraged by incorporating the (c,t) cluster into ZRP. The use of (c,t) clusters in ZRP could achieve more efficient and adaptive hybrid routing without significantly increasing its complexity.

13.5.2 Clustering algorithm

The objective of the clustering algorithm is to partition the network into several clusters. Optimal cluster size is dictated by the tradeoff between spatial reuse of the channel (which drives toward small sizes) and delay minimization (which drives toward large sizes). Other constraints also apply, such as power consumption and geographical layout. Cluster size is controlled through the radio transmission power. For the cluster algorithm, we assume that transmission power is fixed and is uniform across the network. Within each cluster, nodes can communicate with each other in at most two hops. The clusters can be constructed based on node ID.

The following algorithm partitions the multihop network into some nonoverlapping clusters. The following operational assumptions underlying the construction of the algorithm in a radio network are made. These assumptions are common to most radio data link protocols [75–78]:

(A1) Every node has a unique ID and knows the IDs of its one-hop neighbors. This can be provided by a physical layer for mutual location and identification of radio nodes.

(A2) A message sent by a node is received correctly within a finite time by all of its one-hop neighbors.

(A3) Network topology does not change during the algorithm execution.

The distributed clustering algorithm is shown in Figure 13.37

506 AD HOC NETWORKS

then routing will be deterministic during this interval, and standard assumptions permit the ad hoc network to be modeled as a network of Jackson queues. Assuming that path availability is an ergodic process, c represents the average proportion of time a (c,t) path is available to carry data. Consequently, c places a lower bound on the effective capacity of the path over an interval of length t.

Let the link capacity be C b/s and the mean packet length 1/μ b. The effective packet service rate μeff over the interval t can be determined based upon the path availability according to Equation (13.4). Based on the Jackson model, each node can be treated as an independent M/M/1 queue. Using knowledge of the current aggregate arrival rate λ and the effective service rate μeff, the M/M/1 results can be applied to find the mean total packet delay T . Since this delay must be less than t, this approach establishes a lower bound on the path availability, as shown in Equation (13.7)

An effective adaptive strategy for determining the value of c controls the minimum level of cluster stability required to support the traffic load and QoS requirements of established connections. The choice of the parameter t is a system design decision that determines the maximum cluster size achievable for different rates of mobility when no traffic exists in the network.

13.5.3.2 (c,t) Cluster algorithm

There are five events which drive the (c,t) cluster algorithm, namely, node activation, link activation, link failure, expiration of the timer and node deactivation.

Node activation

The primary objective of an activating node is to discover an adjacent node and join its cluster. In order to accomplish this, it must be able to obtain topology information for the cluster from its neighbor and execute its routing algorithm to determine the (c,t) availability of all the destination nodes in that cluster. The source node can join a cluster if and only if all the destinations are reachable via (c,t) paths. The first step upon node activation is the initialization of the source node’s CID (cluster ID) to a predefined value that indicates its unclustered status. The network-interface layer protocol is required to advertise the node’s CID as part of the neighbor greeting protocol [79] and in the header of the encapsulation protocol. This enables nodes to easily identify the cluster status and membership of neighboring nodes and of the source of the routing updates – a necessary function to control the dissemination of routing information. When its network-interface layer protocol identifies one or more neighboring nodes, the source node performs the following actions. First, the source node identifies the CIDs associated with each neighbor. Next, it evaluates the link

availability associated with each neighbor according to either a system default mobility profile or mobility information obtained through the network-interface layer protocol or physical-layer sensing. The precise methodology and the information required for the evaluation of link availability is described later in this section.

Finally, the neighbors, having discovered the unclustered status of the source node, automatically generate and transmit complete cluster topology information, which they have stored locally as a result of participating in the cluster’s intracluster routing protocol. This topology synchronization function is a standard feature of typical proactive routing protocols when a router discovers the activation of a link to a new router. The source node does not immediately send its topology information to any of the neighbors.

Link activation

A link activation detected by a clustered node that is not an orphan is treated as an intracluster routing event. Hence, the topology update will be disseminated throughout the cluster. Unlike reactive routing that responds after path failure, the dissemination of link activation updates is a key factor to an (c,t) cluster node’s ability to find new (c,t) paths in anticipation of future link failures or the expiration of the timer.

Link failure

The objective of a node detecting a link failure is to determine if the link failure has caused the loss of any (c,t) paths to destinations in the cluster. A node’s response to a link failure event is twofold. First, each node must update its view of the cluster topology and re-evaluate the path availability to each of the cluster destinations remaining in the node’s routing table. Second, each node forwards information regarding the link failure to the remaining cluster destinations.

Expiration of c timer

The c timer controls cluster maintenance through periodic execution of the intracluster routing algorithm at each node in a cluster. Using the topology information available at each node, the current link availability information is estimated and maximum availability paths are calculated to each destination node in the cluster. If any of the paths are not (c,t) paths, then the node leaves the cluster.

Node deactivation

The event of node deactivation encompasses four related events, namely, graceful deactivation, sudden failure, cluster disconnection and voluntary departure from the cluster. In general, each of these events triggers a response by the routing protocol. As a result, nodes determine that the node that has deactivated is no longer reachable.

13.5.3.3 Ad hoc mobility model

The random ad hoc mobility model used in this section is a continuous-time stochastic process, which characterizes the movement of nodes in a two-dimensional space. Based on