A mobility-based framework for adaptive clustering in wireless ad hoc networks

A mobility-based framework for adaptive clustering in wireless ad hoc networks
of 22
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
  1466 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 17, NO. 8, AUGUST 1999 A Mobility-Based Framework for AdaptiveClustering in Wireless Ad Hoc Networks A. Bruce McDonald,  Student Member, IEEE,  and Taieb F. Znati,  Associate Member, IEEE   Abstract— This paper presents a novel framework for dynam-ically organizing mobile nodes in wireless ad hoc networks intoclusters in which the probability of path availability can bebounded. The purpose of the         cluster is to help minimizethe far-reaching effects of topological changes while balancingthe need to support more optimal routing. A mobility model forad hoc networks is developed and is used to derive expressionsfor the probability of path availability as a function of time. It isshown how this model provides the basis for dynamically group-ing nodes into clusters using an efficient distributed clusteringalgorithm. Since the criteria for cluster organization dependsdirectly upon path availability, the structure of the cluster topol-ogy is adaptive with respect to node mobility. Consequently, thisframework supports an adaptive hybrid routing architecture thatcan be more responsive and effective when mobility rates are lowand more efficient when mobility rates are high.  Index Terms— Ad hoc networks, dynamic clustering, hier-archical routing, mobile computing, mobility models, routingalgorithms, wireless networks. I. I NTRODUCTION A DVANCES in wireless technology and portable com-puting along with demands for greater user mobilityhave provided a major impetus toward development of anemerging class of self-organizing, rapidly deployable network architectures referred to as ad hoc networks [2], [12]. Anad hoc network is comprised of wireless nodes and requiresno fixed infrastructure. Any device with a microprocessor,whether highly mobile or stationary, is a potential node inan ad hoc network. This includes mobile telephones, motorvehicles, roadside information stations, satellites, and desktopor hand-held computing devices. Unlike existing commercialwireless systems and fixed infrastructure networks, ad hocnetworks cannot rely on specialized routers for path discoveryand traffic routing. Consequently, mobile end systems in an adhoc network are expected to act cooperatively to route trafficand adapt the network to the highly dynamic state of its linksand its mobility patterns.Ad hoc networks evolved largely from the DARPA packet-radio (PR) network program [1], [16], [20]. They are expectedto play an important role in future commercial and military Manuscript received June 15, 1998; revised March 17, 1999.A. B. McDonald is with the Department of Information Science andTelecommunications, University of Pittsburgh, Pittsburgh, PA 15260 USAand the Department of Neurophysiology, Children’s Hospital of Pittsburgh,Pittsburgh, PA 15260 USA (e-mail: F. Znati is with the Department of Computer Science and the Departmentof Information Science and Telecommunications, University of Pittsburgh,Pittsburgh, PA 15260 USA (e-mail: Item Identifier S 0733-8716(99)04804-0. settings where mobile access to a wired network is eitherineffective or impossible. Potential applications for this classof network include instant network infrastructure to supportcollaborative computing in temporary or mobile environments,mobile patient monitoring for improved critical care, dis-tributed command and control systems, and mobile accessto the global Internet. Furthermore, ad hoc networks havethe potential to serve as a ubiquitous wireless infrastructurecapable of interconnecting many thousands of devices [31]with a wide range of capabilities and uses. In order to achievethis status, however, ad hoc networks must evolve to supportlarge numbers of heterogeneous systems with a wide range of application requirements [5], [21].Communication between arbitrary endpoints in an ad hocnetwork requires routing over multiple-hop wireless paths. Themain difficulty arises because without a fixed infrastructure,these paths consist of wireless links whose endpoints are likelyto be moving independently of one another. Consequently,node mobility causes the frequent failure and activation of links which leads to increased network congestion, while thenetwork’s routing algorithm reacts to the topology changes.Unlike fixed infrastructure networks where link failures arecomparatively rare events, the rate of link failure due to nodemobility is the primary obstacle to routing in ad hoc networks.The effectiveness of adaptive routing algorithms dependsupon the the timeliness and detail of the topology informationavailable to them. However, minimizing the exchange of information is crucial for efficient operation. In an ad hocnetwork, significant rates of topological change are expected;consequently, the distribution of up-to-date information caneasily saturate the network. Furthermore, information arrivinglate due to latency can drive network routing into instability.Since the rate of link failure is directly related to node mobility,greater mobility increases both the volume of control trafficrequired to maintain routes and the congestion due to trafficbacklogs. Thus, a crucial algorithm design objective to achieverouting responsiveness and efficiency is the minimization of reaction to mobility [27].Existing schemes for routing in ad hoc networks can beclassified according to four broad categories, namely, proactiverouting, flooding, reactive routing, and dynamic cluster-basedrouting. Proactive routing protocols periodically distributerouting information throughout the network in order to pre-compute paths to all possible destinations. Although thisapproach can ensure higher quality routes in a static topology,it does not scale well to large highly dynamic networks.By contrast, flooding-based routing requires no knowledge of  0733–8716/99$10.00  󰂩  1999 IEEE  MCDONALD AND ZNATI: MOBILITY-BASED FRAMEWORK FOR ADAPTIVE CLUSTERING IN WIRELESS AD HOC NETWORKS 1467 network topology. Packets are broadcast to all destinationswith the expectation that they will eventually reach theirintended target. Under light traffic conditions flooding can bereasonably robust. However, it generates an excessive amountof traffic in large networks, and it is difficult to achieveflooding reliably [30] when the topology is highly dynamic.Consequently, it does not seem that a routing strategy basedexclusively on proactive routing or flooding can achieve theobjectives required for ad hoc routing.In a reactive routing strategy, the design objective is ac-complished by maintaining paths on a demand-basis using aquery–response mechanism. This limits the total number of destinations to which routing information must be maintained,and consequently, the volume of control traffic required toachieve routing. The shortcomings of this approach includethe possibility of significant delay at route setup time, thelarge volume of far-reaching control traffic required to supportthe route query mechanism, and lower path quality relative toproactive routing. Furthermore, despite the objective of main-taining only desired routes, the route query could propagate toevery node in a network during the initial path setup, causingeach node to establish paths even when they are only requiredby certain sources.In dynamic cluster-based routing, the network is dynami-cally organized into partitions called clusters, with the objec-tive of maintaining a relatively stable effective topology [21].The membership in each cluster changes over time in responseto node mobility and is determined by the criteria specifiedin the clustering algorithm. In order to limit far-reachingreactions to topology dynamics, complete routing informationis maintained only for intracluster routing. Intercluster routingis achieved by hiding the topology details within a cluster fromexternal nodes and using hierarchical aggregation, reactiverouting, or a combination of both techniques. The argumentmade against dynamic clustering is that the rearrangementof the clusters and the assignment of nodes to clusters mayrequire excessive processing and communications overhead,which outweigh its potential benefits. If the clustering al-gorithm is complex or cannot quantify a measure of clusterstability, these obstacles may be difficult to overcome.A desirable design objective for an architectural framework capable of supporting routing in large ad hoc networks subjectto high rates of node mobility incorporates the advantagesof cluster-based routing and balances the tradeoff betweenreactive and proactive routing while minimizing the shortcom-ings of each. Furthermore, the consequences of node mobilitysuggest the need to include a quantitative measure of mobilitydirectly 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 routingdecisions on this metric can help minimize the reaction totopological changes. Such a strategy can limit the propagationof far-reaching control information while supporting higherquality routing in highly mobile environments.The purpose of this paper is to present the clusterframework, which defines a strategy for dynamically organiz-ing the topology of an ad hoc network in order to adaptivelybalance the tradeoff between proactive and demand-basedrouting by clustering nodes according to node mobility. This isachieved by specifying a distributed asynchronous clusteringalgorithm that maintains clusters which satisfy thecriteria that there is a probabilistic bound on the mutualavailability of paths between all nodes in the cluster over aspecified interval of time . In order to evaluate thecriteria, a mobility model is proposed that characterizes themovement of nodes in large ad hoc networks. It is shownhow this model is used to determine the probability of pathavailability when links are subject to failure due to nodemobility.Based on the cluster framework, intracluster routingrequires a proactive strategy, whereas intercluster routingis demand-based. Consequently, the framework specifies anadaptive-hybrid scheme whose balance is dynamically de-termined by node mobility. In networks with low rates of mobility, clustering provides an infrastructure that ismore proactive. This enables more optimal routing by in-creasing the distribution of topology information when therate of change is low. When mobility rates become veryhigh, cluster size will be diminished and reactive routingwill dominate. The cluster framework decouples therouting algorithm specification from the clustering algorithm,and thus, it is flexible enough to support evolving ad hocnetwork routing strategies [13], [15], [27], [29] in both theintra- and intercluster domains.The remainder of the paper is organized as follows:Section II presents a review of the significant contributionsin the area of dynamic clustering for ad hoc networks.The characterization of the cluster and the clusterrouting methodology is described in Section III. Details of the the cluster algorithm are presented in Section IV.The mobility model used to characterize link and pathavailability is developed in Section V, and simulationresults demonstrating the effectiveness of the clusterframework are presented in Section VI. Finally, conclusionsof this work are presented in Section VII.II. R ELATED  W ORK Several dynamic clustering strategies have been proposedin the literature [10], [21], [25], [31]. While these strategiesdiffer in the criteria used to organize the clusters and theimplementation of the distributed clustering algorithms, noneof the proposed schemes uses prediction of node mobility asa criteria for cluster organization. Clustering decisions in eachof these schemes are based on static views of the network at the time of each topology change. Consequently, they donot provide for a quantitative measure of cluster stability. Incontrast, the cluster strategy forms the cluster topologyusing criteria based directly on node mobility. According to[31], the ability to predict the future state of an ad hoc network comprised of highly mobile nodes is essential if the network control algorithms are expected to maintain any substantivequality-of-service (QoS) guarantees to real-time connections.The multimedia support for wireless network (MMWN)system proposed by Ramanathan and Steenstrup [31] is basedupon a hybrid architecture that includes the characteristics of   1468 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 17, NO. 8, AUGUST 1999 ad hoc and cellular networks. Their framework uses hierarchi-cal routing over dynamic clusters that are organized accordingto a set of system parameters that control the size of eachcluster and the number of hierarchical levels. Aggregation of routing information is used to achieve scalability and limit thepropagation of topological change information. A multilevelstrategy is used to repair virtual circuit (VC) connections thathave been disturbed due to node mobility. MMWN does notpredict node movement. Consequently, it is unable to providea quantitative bound on the stability of its cluster organization.Krishna  et al.  [25] proposed a scheme that dynamicallyorganizes the topology into clusters, where nodes in acluster are mutually reachable via -hop paths. The algorithmconsiders and reduces to finding cliques in the physicaltopology. Using a first-fit heuristic, the algorithm attemptsto find the largest cliques possible. Although the algorithmdoes not form optimal clusters, it still requires a three-passoperation each time a topology change occurs: one for findinga set of feasible clusters, a second for choosing the largestof the feasible clusters that are essential to maintain clusterconnectivity, and a third to eliminate any existing clusters thatare made superfluous by the new clusters.The objective of the scheme proposed by Lin and Gerla[21] differs significantly from the previous examples. Ratherthan using clustering to minimize the network’s reactionto topological changes, their scheme is intended to providecontrolled access to the bandwidth and scheduling of the nodesin each cluster in order to provide QoS support. Hierarchicalrouting and path maintenance were a secondary concern. Theproposed algorithm is very simple and uses node ID numbersto deterministically build clusters of nodes that are reachableby two-hop paths.The zone routing protocol (ZRP) proposed by Haas andPearlman [13] is a hybrid strategy that attempts to balance thetradeoff between proactive and reactive routing. The objectiveof ZRP is to maintain proactive routing within a zone andto use a query–response mechanism to achieve interzonerouting. In ZRP, each node maintains its own hop-countconstrained routing zone; consequently, zones do not reflecta quantitative measure of stability, and the zone topologyoverlaps arbitrarily. These characteristics differ fromclusters, which are determined by node mobility and do notoverlap. Both strategies assume a proactive routing protocolfor intrazone/cluster routing, and each organizes its topologybased upon information maintained by that protocol. ZRP alsodefines the query control scheme to achieve interzone routing.Although ZRP is not a clustering algorithm and thecluster framework is not a routing protocol, the comparisondemonstrates a close relationship that could be leveraged byincorporating the cluster into ZRP. The use of clusters in ZRP could achieve more efficient and adaptivehybrid routing without significantly increasing its complexity.III. C LUSTER  F RAMEWORK Hierarchical routing has been shown to be essential in orderto achieve at least adequate levels of performance in verylarge networks [17], [18]. In fixed infrastructure networks,hierarchical aggregation achieves the effect of making a largenetwork appear much smaller from the perspective of therouting algorithm. Cluster-based routing in ad hoc networkscan also make a large network appear smaller, but moreimportantly, it can make a highly dynamic topology appearmuch less dynamic. Unlike the cluster organization of afixed network, the organization of an ad hoc network cannotbe achieved offline. The assignment of mobile nodes toclusters must be a dynamic process wherein the nodes areself-organizing and adaptable with respect to node mobility.Consequently, it is necessary to design an algorithm thatdynamically implements the self-organizing procedures inaddition to defining the criteria for building clusters.The objective of the cluster framework is to maintainan effective topology that adapts to node mobility so thatrouting can be more responsive and optimal when mobilityrates are low and more efficient when they are high. Thisis accomplished by a simple distributed clustering algorithmusing a probability model for path availability as the basis forclustering decisions. The algorithm dynamically organizes thenodes of an ad hoc network into clusters where probabilisticbounds can be maintained on the availability of paths to clusterdestinations over a specified interval of time.The cluster framework can also be used as the basisfor the development of adaptive schemes for probabilistic QoSguarantees in ad hoc networks. Specifically, support for QoSin time-varying networks requires addressing: 1) connection-level issues related to path establishment and managementto ensure the existence of a connection between the sourceand the destination and 2) packet-level performance issues interms of delay bounds, throughput, and acceptable error rates.Ideally, it is desirable to guarantee that the QoS requirementsof ongoing connections are preserved for their entire duration.Unfortunately, this is not possible in a time-varying network environment as connections may fail randomly due to usermobility. A more realistic and practical approach is to pro-vide some form of probabilistic QoS guarantees by keepingconnection failures below a prespecified threshold value andby ensuring with high probability that a minimum level of bandwidth is always available to ongoing connections.Based upon the intracluster routing model proposed inSection III-B, and using the estimates of path availabilityand other link status metrics provided through the routingalgorithm, a connection admission control algorithm coulddetermine with high probability whether or not sufficientresources are available to support the requirements of anintracluster connection over a specific period of time. Inorder to achieve similar QoS guarantees across the interclusterdomain, the cluster framework could be extended tosupport a dynamic hierarchical architecture in which resourceinformation within each cluster is aggregated and a secondlevel of clustering algorithm maintains paths betweenclusters using virtual links. 1 Hierarchical QoS-based routingand admissions control schemes are not considered further inthis paper. 1 A virtual link represents the set of physical links that connect nodes inone cluster to nodes in another cluster.  MCDONALD AND ZNATI: MOBILITY-BASED FRAMEWORK FOR ADAPTIVE CLUSTERING IN WIRELESS AD HOC NETWORKS 1469 The remainder of this section is organized as follows:the cluster is formally characterized in Section III-A.The implementation of routing is discussed in Section III-B.Finally, a methodology for selecting the system parametersand is presented in Section III-C.  A. Cluster Characterization The basic idea of the cluster strategy is to partitionthe network into clusters of nodes that are mutually reachablealong cluster internal paths 2 that are expected to be availablefor a period of time with a probability of at least . Theunion of the clusters in a network must cover all the nodesin the network.  Definition 1:  Let indicate the status of path fromnode to node at time . if all the links inthe path are active at time , and if one or morelinks in the path are inactive at time . The path availabilitybetween two nodes and at time is givenby the following probability expression:Pr  Definition 2:  Let be the path availability of pathfrom node to node at time . Path is defined as anpath if and only if   Definition 3:  Node and node are available if they are mutually reachable over paths.  Definition 4:  An cluster is a set of avail-able nodes. Definition 4 states that every node in ancluster has a path to every other node in the cluster thatwill be available at time with a probability .The cluster characterization, as previously defined, requires amodel which quantifies the path availability as givenin Definition 1. Path availability is a random process thatdepends upon the mobility of the nodes which lie along agiven path. Consequently, the mobility characteristics of thenodes play an important role in the characterization of thisprocess. In Section V, a mobility model for large ad hocnetworks is proposed, and the probability distributions forthe aggregate distance and trajectory covered by a node overtime are derived. These distributions provide the basis fordeveloping analytical models for link availability. It is alsoshown how this model can be used to derive expressions forpath availability which can be efficiently evaluated by thecluster algorithm.  B. Cluster Routing Methodology The logical relationship between the cluster algo-rithm, the routing algorithm, and the other network-layerentities is depicted in Fig. 1. The cluster algorithm resideslogically between the routing-layer and the Internet MANET 3 2 A cluster internal path consists exclusively of nodes that are members of the cluster. 3 A MANET is a mobile ad hoc network.Fig. 1. Logical relationships among MANET network-layer entities. encapsulation protocol (IMEP) 4 [6]. As such, the cluster al-gorithm presents a logical topology to the routing algorithm,and it accepts feedback from the routing algorithm in order toadjust that logical topology and make clustering decisions. Tosupport the cluster framework, IMEP or an equivalentprotocol must identify a node’s cluster identifier number (CID)to neighboring nodes and include the CID in the encapsulationof the routing information packets. A protocol that provides thefunctionality of IMEP along with these enhancements will bereferred to in this paper as a network-interface layer protocol.A two-level routing algorithm adaptively subdivides the task of establishing and maintaining routes to mobile destinations.Intracluster routing uses a proactive strategy, whereby eachnode in a cluster maintains topology information and routes toevery cluster destination for the duration of the time that thenode remains in a given cluster. Routes to destinations outsideof a node’s cluster are established on a demand basis only.Consequently, a reactive routing strategy must be implementedto setup intercluster routes. 1) Intracluster Routing:  Intracluster routing can be imple-mented with any distributed routing algorithm that can proac-tively maintain routes to a set of mobile destinations. Similarto ZRP, which uses hop-count [11], the cluster uses apath availability based membership to limit the propagationof routing updates. Those MANET protocols that have beendesigned specifically to operate as reactive protocols can stillfunction as intracluster protocols in which the demand for aroute is produced by the cluster algorithm. However, becausethe criteria establishes a lower bound on the availabilityof cluster paths, preference is given to those algorithms capableof incorporating link and path availability information asa routing metric and of using maximum availability as anoptimization criteria when establishing paths. Arguments havebeen made against path optimization in ad hoc networks [7],[27]; however, these arguments are based upon the assumptionof a monolithic network without clustering. clustersgradually adapt the cluster topology to maintain a consistent 4 The IMEP layer is designed to provide services to upper-layer network entities such as link status sensing, neighbor discovery, one-hop neighborbroadcast, control packet aggregation, and address resolution.  1470 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 17, NO. 8, AUGUST 1999 level of path stability such that path optimization becomeseffective.In Section V-C, it is shown how the path availability canbe calculated from the individual link availabilities alongthe path. This can be accomplished based on aggregatedpath information using a modified Bellman–Ford algorithmto find the maximum availability path or using Dijkstra’salgorithm if complete link status information is available.Because path availability is a time-varying quantity that de-pends on the individual link properties, it is more efficientif the characteristics of the links are known at each node.To overcome the shortcomings of link-state protocols, severalalternatives that provide complete link status information alongselected paths have been proposed for use in highly dynamicenvironments. Examples that are well-adapted for intraclusterrouting include the link vector algorithm (LVA) [9] and thewireless routing protocol (WRP) [24]. Based on completelink status information, it is possible to estimate the currentlink availability anywhere in the cluster without periodicrouting updates. If aggregated routing information is used,the path availability information will become outdated withoutperiodically updating the path status.A node’s membership in a cluster is implied by the contentsof the routing tables distributed among the active nodes inthat cluster. No distinct cluster table is required since it isthe routing algorithm as modulated by the cluster algorithmwhich determines the set of nodes that are in the cluster. Con-sequently, a node’s routing table gives a complete picture of itscurrent view of the cluster. Accordingly, cluster convergenceis simply a matter of the convergence of the routing tables inthe cluster. 2) Intercluster Routing:  Intercluster routing is achieved us-ing a demand-based protocol that establishes paths by execut-ing a path-search query and response algorithm. With respectto this process, each node can be considered as a route cachefor the set of nodes in its cluster. In the worst case, theresponse phase will begin as soon as the first node in thetarget destination’s cluster receives the route query. Querieswill never be propagated further into the cluster in which thetarget destination resides.One methodology for maintenance of end-to-end routingbetween destinations in different clusters is direct implemen-tation of a flat-routed reactive routing protocol, such as thetemporally ordered routing algorithm (TORA) [26], [27] orthe ad hoc on-demand distance vector algorithm (AODV)[28]. Specifically, each node requiring a route first searchesfor the desired destination in its cluster routing table that isproactively maintained by the intracluster routing protocol. If the destination is not found, the node initiates a route discoveryprocess if it is the source node, or it propagates the request if itis processing another node’s route query. As such, every nodewill participate in two routing protocols: one within a clusterand one for noncluster destinations. Consequently, each nodewill be able to maintain routes to any connected destination.The problem with the flat-routed reactive approach is that itfails to take advantage of the cluster topology, which could beused to more efficiently manage the route discovery and main-tenance processes. To address this shortcoming, an improvedmethodology for cluster interconnection is proposed thatleverages the cluster topology in order to more efficientlydiscover and maintain end-to-end routing between nodes indifferent destinations by adapting the ZRP interzone routingprotocol (IERP). IERP assumes a topology comprised of asequence of overlapping zones and specifies a bordercastingtechnique that is used to efficiently construct routes acrossmultiple zones. However, IERP as defined for ZRP cannotbe directly applied to the interconnection of clusterssince these clusters are designed not to overlap. To bridgethe differences between the requirements of IERP and theproperties of the clusters, the following adaptations arerequired.1) The border nodes of an cluster consist of theset of nodes which are adjacent to nodes that are notmembers of the same cluster. Each border node treatsadjacent clusters 5 as supernodes and advertises reach-ability to those supernodes within their cluster.Consequently, each node in a cluster has knowledge of,and proactively maintains routes to, the set of adjacentclusters identified by the border nodes.2) The bordercasting process defined by IERP must bemodified to allow the exchange of the route query fromthe egress border node of one cluster to the ingressborder node of its adjacent clusters.3) Each cluster egress border node processing anIERP route query will append its CID to the routequery and forward one copy of the query to thoseneighboring nodes which are ingress border nodes inadjacent clusters. Consequently, a sequence of CID’s isaccumulated, which represents an intercluster route tothe desired destination.4) Each cluster ingress border node searches itsrouting table for the destination. If the destination isfound, the node appends its CID to the accumulatedsequence of CID’s in the route query and returns itin the response message that is sent back along theaccumulated sequence of clusters in reverse order.Unlike ZRP, this modified scheme builds a route as a se-quence of clusters rather than nodes. The specific paths acrosseach cluster are determined dynamically by the intraclusterrouting algorithm. Since each node in a cluster maintainsroutes to the set of adjacent clusters, this methodology providesa strategy which is robust, such that routes will remain viableso long as the cluster adjacencies remain intact—even if thespecific border nodes change. Thus, it is highly adaptivewith respect to node mobility and requires less reactive routemaintenance. C. Cluster Parameters Evaluation of paths requires specification of twosystem parameters, and . The effects of these parametersare tightly coupled, making it difficult to select optimal values.Large values for seem desirable, as they imply more clusterstability and reduce the computational requirements of cluster 5 An adjacent cluster to a node    is one with a border node that is adjacentto a border node in the node    ’s cluster.
Similar documents
View more...
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks