A Multi-Head Clustering Algorithm in Vehicular Ad Hoc Networks

Transcription

1 International Journal of Computer Theory and Engineering, Vol. 5, No. 2, April 213 A Multi-Head Clustering Algorithm in Vehiular Ad Ho Networks Shou-Chih Lo, Yi-Jen Lin, and Jhih-Siao Gao Abstrat Clustering is an important researh topi in wireless networks, beause luster strutures an failitate resoure reuse and inrease system apaity. In this paper, we present a new lustering algorithm that onsiders both node position and node mobility in vehiular ad ho environments. The proposed algorithm intends to reate stable lusters by reduing relustering overhead, prolonging luster lifetime, and shortening the average distane between luster heads and their luster members. Most important, this algorithm supports single and multiple luster heads. Simulation results show the superiority of our lustering algorithm over the other three well-known algorithms. Index Terms VANET, lustering, multi-head, head eletion, mobility. I. INTRODUCTION Clustering is a tehnique to group nodes into several lusters. Eah node in the luster struture plays one of three roles: Cluster Head (CH), Cluster Gateway (CG), and Cluster Member (CM). A CH is a leading node of a luster and is responsible to oordinate all CMs in its luster. A CG is a border node of a luster that an ommuniate nodes belonging to different lusters. In mobile wireless networks, lustering is a pratial skill to redue the omplexity of network management [1], [2], [3]. For example, CHs an alloate hannel resoure (time slots or frequeny spetrums) to their CMs to avoid any happenings of transmission ollisions and inrease resoure utilization within a luster. Moreover, a virtual bakbone network an be built by CHs and CGs to manage all routing jobs. This enables a salable wireless routing protool. Clustering has many appliation domains. However, developing effiient lustering tehniques in mobile environments is not an easy job. Node mobility will frequently destroy existing luster strutures. Relustering overhead beomes an important ost metri. In this paper, we onsider a data sharing appliation in a Vehiular Ad ho Network (VANET). VANET is a speialized Mobile Ad Ho Network (MANET) that onnets vehiles and roadside failities. The major funtion of VANET is to provide real-time servies and emergeny warnings for drivers and passengers. VANET provides both Manusript reeived September 14, 212; revised November 28, 212. Shou-Chih Lo and Jhih-Siao Gao are with the Department of Computer Siene and Information Engineering, National Dong Hwa University, Hualien, Taiwan ( slo@mail.ndhu.edu.tw, d97215@ems.ndhu.edu.tw). Yi-Jen Lin was with the Department of Computer Siene and Information Engineering, National Dong Hwa University, Hualien, Taiwan ( evelyn324@hotmail.om). DOI: /IJCTE.213.V Roadside-to-Vehile Communiation (RVC) and Inter-Vehile Communiation (IVC). Our data sharing appliation is motivated by a typial senario, where a passenger in a ar would like to download an interested multimedia file from neighboring ars via IVC. Here, we use the luster struture to failitate the finding, uploading, and downloading of multimedia files. Vehiles that are willing to share data are grouped into lusters. In a luster, CMs an upload their shared data and query interested data to the CH(s). CMs an also download interested data from the CH(s). To speed up data downloading, a bit-torrent downloading mehanism from multiple seed nodes (CHs) is reommended. Therefore, we need to onstrut a luster with multiple CHs. Several lustering algorithms [4]-[1] have been proposed for mobile networks. However, they have some weaknesses when applied to our data sharing appliation. First, these algorithms do not support an arbitrary number of CHs within a luster. Seond, most of them are designed for MANET but VANET. VANET has its own unique features suh as highly dynami topology, suffiient energy and storage, and geographial environment onstrains. A mobile node in MANET an move in arbitrary diretions but an only move along the street in VANET. Moreover, most vehiles are equipped with GPS (Global Positioning System) devies. The loation and mobility information about a vehile is available, whih failitates the design of a more effiient lustering algorithm. We would first introdue our proposed single-head lustering algorithm by detailing the proess of CH eletion. Then, a multi-head lustering algorithm is illustrated. The remainder of this paper is organized as follows. A brief survey on lustering algorithms in mobile environments is given in Setion 2. Our proposed algorithm is presented in Setion 3. We evaluate the performane in Setion 4. Finally, we draw signifiant onlusions in Setion 5. II. RELATED WORK A lustering algorithm inludes two tehnial parts: luster establishment and luster maintenane. In luster establishment, we need to identify the role (CH, CG, or CM) of eah node in the network. The eletion of CHs is a ore tehnique. In luster maintenane, relustering (luster merge and split) needs to be performed to handle the effets of node mobility and node failure. An effiient lustering algorithm should take are of performane metris suh as luster lifetime and relustering overhead. We lassify existing lustering algorithms aording to two ategories: luster struture and eletion riterion. From

2 International Journal of Computer Theory and Engineering, Vol. 5, No. 2, April 213 the struture point of view, lustering algorithms either have different topologies (overlapping or non-overlapping) or CH numbers. Most of lustering algorithms selet one CH in a luster. Two CHs are seleted in [11]. A CG an belong to more than one luster in overlapping lusters while a CG belongs to only one luster in non-overlapping lusters. The distane between a CM and a CH is typially one-hop. Overlapping lusters (suh as [12], [13], and [14]) and non-overlapping lusters (suh as [15]) have the advantages of high robustness and low hannel ontention respetively. Different evaluation methods an be applied in CH eletion and an be lassified into five types. With an ID-based method, a node with the largest or smallest identifiation number (ID) in its neighborhood is eleted to be a CH. With a degree-based method, a node with the largest degree value (indiating the number of its diret neighbors) in its neighborhood is eleted as a CH. Loal nodes with a ompetition-based method ompete to be a CH. With a mobility-based method, a node that has stable relative mobility to its neighbors is eleted as a CH. In a multiple-metris-based method, several evaluation fators are ombined together by a weighted average funtion to qualify a node for a CH. In the following, we introdue some well-known lustering algorithms. (Lowest ID) [4] elets a node with the smallest ID as a CH. This approah is simple but may ause several relusterings and small-size lusters. HCC (High Connetivity Clustering) [4] elets a node with the largest degree value as a CH. HCC reates a less number of lusters than, so the maintenane ost is low. However, HCC is more unstable than to node mobility. MMDA (Max-Min D-hop lustering Algorithm) [5] and RCC (Random Competition based Clustering) [6] use a ompetition method to elet CHs. WCA (Weighted Clustering Algorithm) [7] and CEMCA (Connetivity, Energy and Mobility driven Clustering Algorithm) [1] onsider several fators suh as ID, node degree, power level, mobility, and loation in the CH eletion. (Least Cluster head Change) [8] improves and HCC on relustering overhead. Only two onditions ause a new CH eletion: 1. Two lusters merge into one luster. 2. A CM annot listen to any ativity of surrounding CHs. Instead of using IDs, (MOBIlity metris Clustering) [9] uses relative mobility in CH eletion to redue the effet of node mobility. A mobile node that is relatively less mobile with respet to its neighbors is eleted as a CH. The relative mobility is omputed by the distane hange of a node observed from the other node. measures signal strength to derive the distane between two nodes. This measurement is simple but inaurate. DPP (Diretional Propagation Protool) [11] extends into vehiular environments. Two CHs that are loated at the head and the tail of a luster are seleted. CBMAC (Cluster-Based Medium Aess Control) [16] uses three riteria in CH eletion: degree, position, and relative speed. III. CLUSTERING ALGORITHMS At first, the design of our single-head lustering algorithm is explained. Then, the multi-head version is introdued. Two assumptions are made: Eah mobile node has a unique ID and is equipped with a GPS devie. A. Eletion Criterion In a data sharing appliation, it is fairer for CMs that a CH is nearer the enter of a luster, beause the hop ount of a data transmission path from a CM to the CH is similar. Moreover, a CH should have stable relative mobility to its CMs for reduing CH reeletion times. Based on these onepts of enter position and relative mobility, we measure the RPM (Relative Position and Mobility) of eah mobile node as the riterion of CH eletion. Eah mobile node is supposed to periodially broadast a HELLO paket to its one-hop neighbors. This HELLO paket arries information about the loation, motion vetor, and RPM of a mobile node. By listening to these HELLO pakets, a node maintains a neighbor table that reords the mobility-related data of all its urrent one-hop neighbors (inluding the node itself). Clusters with nodes having diverse moving diretions are unstable, so we restrit all nodes in a luster to have the same moving diretion. Therefore, eah node reords only those neighbors that drive with the same diretion as it in its neighbor table. If a node has m entries in its neighbor table, without loss of generality, we assume that the first entry reords its own mobility-related data. Three data fields (x i, y i, v i ) are reorded in the ith entry, whih indiate the urrent loation (x i, y i ) and the urrent moving speed (v i ) of the reorded node. The RPM of a node is omputed as the following steps: 1) Compute the enter position from these m entries in the neighbor table as (1). m m 1 P ( x, y ) ( xi, yi ) (1) m i 1 i 1 2) Compute the relative distane to this enter position as (2). Re l _ Dist P1 P ( x1 x ) ( y1 y ) (2) 3) Sort these m moving speeds (salar values) and find the median as (3). v Median Sort{ v, v,..., v }} (3) { 1 2 m 4) Compute the relative speed to the median speed as (4). Re l _ Speed (4) 1 v 1 v 5) Compute the RPM (between and 1) as (5). Rel _ Dist1 RPM Max{Re l _ Dis, j 1~ m} Rel _ Speed1 (1 ) Max{Re l _ Speed, j 1~ m} is a weighted fator whih is set as.6 in our experiments. A mobile node delares itself a CH when its RPM is the smallest one in its neighbor table. j j (5) 243

3 International Journal of Computer Theory and Engineering, Vol. 5, No. 2, April 213 B. Cluster Establishment In the luster establishment phase, three node roles are used: UN (Undeided Node), CH, and CM. A UN is a node that is not urrently belonging to any luster. A node will play one of these three roles and may transit from one role to another role if luster struture is hanged. The following notations are used in our latter disussion: BI (Broadast Interval): time interval for a node to broadast a HELLO paket. CI (Contention Interval): time duration after two enounter CHs start ompeting to be a CH. CD (Contention Distane): distane gap after two enounter CHs start ompeting to be a CH. TI (Timeout Interval): time duration after an unreahable neighbor is removed from a neighbor table. UN_NUM i : number of UNs in the neighbor table of node i. UN_BOUND: threshold value for a node to start the CH eletion. CM i : number of CMs joining to a CH node i. Four ontrol pakets are used in our lustering algorithm: HELLO: periodially broadast paket by eah node that arries the mobility-related data, RPM, node role, and CM values. JOIN_INVITE: broadast paket issued from a CH to invite any possible CMs. JOIN_REPLY: broadast paket issued from a CM to aknowledge a join invitation from a CH. CH_RESIGN: broadast paket issued from a CH when deiding to resign from a CH. Cluster onstrution steps are explained below: 1) Initialize eah node to be a UN. 2) Eah node broadasts a HELLO paket per BI. 3) A node i starts the CH eletion as UN_NUM i UN_BOUND. A node delares itself a CH when its RPM is the smallest one in its neighbor table. To break the tie, the smallest Rel_Dist, Rel_Speed, and ID are onsidered in sequene. 4) A new CH broadasts a JOIN_INVITE paket to its neighbors. 5) A UN joins to a luster when it reeives a JOIN_INVITE paket from a CH that is driving with the same diretion as it. This UN replies a JOIN_REPLY paket to the CH. 2) A UN beomes a CM if reeiving a JOIN_INVITE paket from a CH that is driving with the same diretion as it. CH: 1) Two CHs i and j will enter a ompetition proess if the following onditions are true: (a) CHs i and j ontat (an ommuniate with eah other) for a time period greater than CI. (b) CHs i and j drive in the same diretion. () The distane between i and j is smaller than CD. 2) In the ompetition proess, the CH having the larger number of CMs wins the ompetition. That is, if CM i CM j, then CH j resigns from a CH by broadasting a CH_RESIGN paket to its neighbors. In this ase, CH j beomes a CM of CH i. 3) A CH removes a CM from its CM list if it does not reeive any HELLO paket from this CM over a time period of TI. If the CM list is empty, this CH returns as a UN. CM: 1) If a CM does not reeive any HELLO paket from its CH over a time period of TI, this CM returns as a UN. 2) A CM returns as a UN, if it reeives a CH_RESIGN paket from its CH. D. Multi-Head Clustering The multi-head version is extended from the single-head version. We first onstrut lusters using the single-head algorithm. The seleted CH in eah luster is alled a master CH (MCH). Then, we selet some CMs from a luster to be slave CHs (SCHs). Therefore, eah multi-head luster ontains one MCH and several SCHs. In a vehiular environment, a luster usually has a retangular shape that aptures a ertain road segment. Having the positions of all CMs, an MCH an ompute the bounded retangle that enloses all its CMs. Then, this retangle is divided into some setors (with the number being speified) as the example shown in Fig. 2. The MCH selets one SCH that has the loally smallest RPM in eah setor (exept for the setor the MCH is loated in). C. Cluster Maintenane We individually explain the transition of eah node role as follows. The role transition diagram is depited in Fig. 1. UN : CM : MCH : SCH Fig. 2. Multi-head eletion. CM Two lusters are merged Fig. 1. Role transition diagram in a single-head luster. UN: 1) A UN beomes a CH if winning the eletion CH A new ontrol paket SCH_ASSIGN is introdued, whih is used by an MCH to announe the SCH list. A CM transits to a SCH when reeiving a SCH_ASSIGN paket and finding that its ID is in the SCH list. The MCH periodially announes the SCH list by using the same HELLO paket. In the luster maintenane, the proedures of UN, CH (as MCH), and CM remain unhanged. An additional node role of SCH 244

4 Average Number of CMs Average Number of Clusters International Journal of Computer Theory and Engineering, Vol. 5, No. 2, April 213 is introdued in our role transition diagram as shown in Fig. 3. Here, an SCH returns to as a CH when its MCH is leaving or resigned. Aept an SCH assignment CM SCH Aept a join invitation The MCH is leaving or resigned All CMs are leaving UN MCH Fig. 3. Role transition diagram in a multi-head luster. IV. PERFORMANCE EVALUATION To evaluate the performane of our proposed algorithm, we arry out simulations using NS-2 [17]. We ompare our algorithm (abbreviated as, Center-Position and Mobility) with,, and. The single-head version of is onerned, sine the other ompared targets are all single-head approahes. The ost metris used in our experiments are listed below: Average number of lusters: average number of lusters by averaging system observations per ten seonds. Average number of CMs: average number of CMs in a luster by averaging system observations per ten seonds. Average luster lifetime: average lifetime of a luster. Average idle time: average time duration for a node remaining as a UN in the system. Average resident time: average time duration for a CM to stay in the same luster. A. Simulation Model We onsider a real street environment whih is imported from the TIGER (Topologially Integrated Geographi Enoding and Referening) database [18]. A ity street map of size 15m 15m is used as shown in Fig. 4. Eah road has two lanes (forward and reverse). Under this street model, 2 vehiles are generated and their moving patterns are ontrolled by VanetMobiSim [19]. Eah vehile moves with a speed from 1 m/se to 2 m/se. The parameter settings in our simulation are listed in Table I. Eah ost result is omputed through the average of ten simulation runs. Win the eletion TABLE I: PARAMETER SETTINGS IN EXPERIMENTS. Parameter Value Simulation time 3 s MAC protool 82.11p Transmission range (TR) 1 m 25 m Broadast interval (BI).5 s Contention interval (CI) 4 BI Timeout interval (TI) 2 BI Contention distane (CD) TR/2 UN_BOUND 1 B. Simulation Results The experimental results are depited in Figs. 5~9. and are ID-based lustering algorithms. They will group vehiles on both forward and reverse lanes, so their onstruted lusters tend to be large and hene the average number of lusters is small. These results are more signifiant if larger transmission ranges are used. However, under high node mobility, these two algorithms are not stable. The average luster lifetime is short. relaxes ertain riteria on relustering as ompared with. The luster lifetime of is a little bit higher than that of. Sine the CH eletion is simple for ID-based approahes, a UN an easily join to a surrounding luster. Therefore, the average idle time is short. However, the luster is not stable and tends to be reonstruted. A CM frequently joins to a different luster, so the average resident time is short. This redues the opportunity of a CM to suessfully download data from CHs. As a onlusion, these ID-based lustering algorithms are not suitable for our data sharing appliation in vehiular environments Fig. 5. Average number of lusters vs. TR Fig. 4. Simulated street environment. Fig. 6. Average number of CMs vs. TR. 245

5 Average Idle Time (se) Average Cluster Lifetime (se) Average Resident Time (se) International Journal of Computer Theory and Engineering, Vol. 5, No. 2, April 213 Next, we ompare the performanes of and. is a mobility-based lustering algorithm while is a multiple-metris-based one. As an been seen in the figures, generates smaller lusters than. Also, vehiles on forward and reverse lanes are not separated in, so its luster lifetime is shorter than. Those vehiles that oasionally join to a CH with a different moving diretion as them will frequently beome UNs as the CH is leaving. Moreover, the CH eletion in takes the most time among all. To ompute the relative mobility of one node against the other node, two suessive HELLO pakets are required. Therefore, the average idle time in is long. However, those vehiles that drive the same diretion with the CH will keep joining the same luster, so the average resident time of is long. As a onlusion, is still not suitable for our data sharing appliation, beause small-size lusters and long-idle periods prevent the searhing of a verity of shared data. Compared with other algorithms, provides stable lusters with long lifetime. Moreover, the short idle time and the long resident time qualify for a good lustering algorithm in data sharing. The idle time of is higher than and, beause we disallow any lusters with size of 1. There may have single UN in our system and these single UNs are useless to data sharing Fig. 7. Average luster lifetime vs. TR Fig. 9. Average resident time vs. TR. V. CONCLUSION We have introdued the importane of lustering to network and appliation designs. For example, grouping nodes into luster strutures make a network topology easy to maintain, and grouping users of the same interests failitates data sharing. In this paper, we onsider a data sharing appliation in a vehiular environment. Vehiles nearby are grouped into a luster in whih some nodes are seleted as luster heads. These luster heads serve as loal file servers that enable surrounding nodes to upload and download shared data. We therefore propose a multi-head lustering algorithm. Our proposed algorithm an selet a given number of luster heads with a uniform distribution in spae. Sine node mobility has a great influene on luster maintenane, both node position and node mobility are onsidered in our luster head eletion. Different node roles are identified with a lear role transition diagram. Compared with traditional lustering algorithms for mobile environments, we additionally onsider the unique feature in vehiular environments that vehiles drive along the street either with the same or opposite driving diretion. Our experimental results reveal that the proposed algorithm generates stable lusters with long lifetime. In the future, we will onsider more appliations for our multi-head lustering tehnique. The ooperation behavior among multiple luster heads need to be further disussed and evaluated ACKNOWLEDGMENT The authors would like to thank the anonymous referees for their helpful suggestions. This researh was partially supported by National Siene Counil of the Republi of China under Contrat No. NSC E Fig. 8. Average idle time vs. TR. REFERENCES [1] A. B. MDonald and T. F. Znati, A mobility-based framework for adaptive lustering in wireless ad ho networks, IEEE Journal on Seleted Areas in Communiations, vol. 17, no. 8, pp , August [2] C. R. Lin and M. Gerla, Adaptive lustering for mobile wireless networks, IEEE Journal on Seleted Areas in Communiations, vol. 15, no. 7, pp , September

6 International Journal of Computer Theory and Engineering, Vol. 5, No. 2, April 213 [3] T. C. Hou and T. J. Tsai, A aess-based lustering protool for multihop wireless ad ho networks, IEEE Journal on Seleted Areas in Communiations, vol. 19, no. 7, pp , July 21. [4] M. Gerla and J. T. Tsai, Multiuser, mobile, multimedia radio network, Wireless Networks, vol. 1, pp , Otober [5] A. D. Amis, R. Prakash, T. H. P. Vuong, and D. T. Huynh, Max-min d-luster formation in wireless ad ho networks, in Pro. IEEE Infoom Conf., pp , Marh 2. [6] K. Xu and M. Gerla, A heterogeneous routing protool based on a new stable lustering sheme, in Pro. Milom Conf., pp vol. 2, Otober 22. [7] M. Chatterjee, S. K. Sas, and D. Turgut, An on-demand weighted lustering algorithm (WCA) for ad ho networks, in Pro. IEEE Globeom Conf., pp , November 2. [8] C. C. Chiang, H. K. Wu, W. Liu, and M. Gerla, Routing in lustered multihop, mobile wireless networks with fading hannel, in Pro. IEEE Sion Conf., pp , April [9] P. Basu, N. Khan and T. D. C. Little, A mobility based metri for lustering in mobile ad ho networks, in Pro. Int. Conf. Distributed Computing Systems Workshop, pp , April 21. [1] F. D. Tolba, D. Magoni, and P. Lorenz, Connetivity, energy and mobility driven lustering algorithm for mobile ad ho networks, in Pro. IEEE Global Teleommuniations Conf., pp , November 27. [11] T. D. C. Little and A. Agarwal, An information propagation sheme for VANETs, in Pro. IEEE Intelligent Transportation Systems Conf., pp , September 25. [12] X. Ji and H. Zha, Sensor positioning in wireless ad-ho sensor networks using multidimensional saling, in Pro. IEEE Infoom Conf., pp , Marh 24. [13] P. Krishna, N. H. Vaidya, M. Chatterjee and D. K. Pradhan, A luster-based approah for routing in dynami networks, ACM Sigomm Computer Communiation Review, vol. 27, no. 2, pp , April [14] T. Wu and S. Biswas, A self-reorganizing slot alloation protool for multi-luster sensor networks, in Pro. Int. Symp. Information Proessing in Sensor Networks, pp , April 25. [15] J. Y. Yu and P. H. J. Chong, 3hBAC (3-hop between adjaent lusterheads): a novel non-overlapping lustering algorithm for mobile ad ho networks, in Pro. IEEE PACRIM Conf., pp , August 23. [16] Y. Gunter, B. Wiegel, and H. P. Grossmann, Cluster-based medium aess sheme for VANETs, in Pro. IEEE Intelligent Transportation, pp , September 27. [17] The network simulator. [Online]. Available: [18] Topologially integrated geographi enoding and referening. [Online]. Available: [19] Vanet Mobi Sim. [Online]. Available: Shou-Chih Lo reeived the B.S. degree in omputer siene from National Chiao Tung University, Taiwan, in 1993, and the Ph.D. degree in omputer siene from National Tsing Hua University, Taiwan, in 2. Sine 24, he has been with the Department of Computer Siene and Information Engineering of the National Dong Hwa University in Taiwan, where he is urrently an Assoiate Professor. His urrent researh interests are in the areas of mobile and wireless networks and delay-tolerant networks. I-Cheng Lin reeived the B.S. degree in Computer Siene from National Dong Hwa University, Taiwan, in 29. Sine 29, she has been with the ompany of the Trend Miro in Taiwan, where she is urrently a software engineer. Jhih-Siao Gao reeived his B.S. and M.S. degree in Computer Siene Information Engineering from Shu-Te University, Taiwan, in 26 and 28. Currently he is working towards the Ph.D. degree in Computer Siene and Information Engineering at National Dong Hwa University, Taiwan. His researh interests fous on wireless ommuniations network. 247