A Load-balancng and Energy-aware Clusterng Algorthm n Wreless Ad-hoc Networks Wang Jn, Shu Le, Jnsung Cho, Young-Koo Lee, Sungyoung Lee, Yonl Zhong Department of Computer Engneerng Kyung Hee Unversty, Korea {wangjn, sl8132, sylee, zhungs}@oslab.khu.ac.kr {chojs,yklee}@khu.ac.kr Abstract. Wreless ad-hoc network s a collecton of wreless moble nodes dynamcally formng a temporary communcaton network wthout the use of any exstng nfrastructure or centralzed admnstraton. It s characterzed by both hghly dynamc network topology and lmted energy. So, the effcency of MANET depends not only on ts control protocol, but also on ts topology and energy management. Clusterng strategy can mprove the performance of flexblty and scalablty n the network. Wth the ad of graph theory, genetc algorthm and smulated annealng hybrd optmzaton algorthm, ths paper proposes a new clusterng strategy to perform topology management and energy conservaton. Performance comparson s made between the orgnal algorthms and our two new algorthms, namely an mproved weghtng clusterng algorthm and a novel Genetc Annealng based Clusterng Algorthm (GACA), n the aspects of average cluster number, topology stablty, load-balancng and network lfetme. The expermental results show that our clusterng algorthms have a better performance on average. 1 1 Introducton Wreless ad hoc wreless network s a collecton of wreless moble nodes that selfconfgure to form a network wthout the ad of any establshed nfrastructure [1]. It can be rapdly deployed and reconfgured where the communcaton nfrastructure s ether unavalable or destroyed. However, t s confronted wth many challenges too, such as the moblty of hosts, the dynamc topology, the mult-hop nature n transmsson, the lmted bandwdth and battery, etc. So, the study of MANET (Moble Adhoc NETwork) s a very demandng and challengng task. Up to now, there are many routng protocols based on varous strateges n MANET, and they can be classfed nto several knds as follows: (1) proactve and reactve; (2) flat and herarchcal; (3) GPS asssted and non-gps asssted, etc. These knds of protocols can be used solely or together. Here we manly dscuss the herarchcal routng protocols, whch are based on the clusterng algorthm [2, 3]. 1 Dr. Sungyoung Lee s the correspondng author.
The rest of the paper s organzed as follows. In secton 2, some relevant background and commonly used clusterng algorthms are presented. Based on whch, an mproved clusterng algorthm s proposed n secton3. In secton 4, another novel Genetc Annealng based Clusterng Algorthm (GACA) s gven so as to optmze the overall network performance. The smulaton results and comparson s made n the aspects of average cluster number, topology stablty, load-balancng and network lfetme n secton 5. Secton 6 concludes the paper. 2 Related Work Smlar to the cellular network, the MANET can be dvded nto several clusters. Each cluster s composed of one clusterhead and many normal nodes, and all the clusterheads form an entre domnant set. The clusterhead s n charge of collectng nformaton (sgnalng, message, etc.) and allocatng resources wthn ts cluster and communcatng wth other clusterheads. And the normal nodes communcate wth each other through ther clusterhead, no matter they are n the same cluster or not. Several orgnal clusterng algorthms have been proposed n MANET. These nclude: (1) Hghest-Degree Algorthm; (2) Lowest-ID Algorthm; (3) Node-weght Algorthm; (4) Weghted Clusterng Algorthm. (5) Others, lke RCC (Random Competton based Clusterng), LCC (Least Cluster Change), LEACH etc. We wll gve some of them a bref descrpton as follows. 2.1 Hghest-Degree Algorthm The Hghest-Degree Algorthm was orgnally proposed by Gerla. and Parekh [4,5]. A node x s consdered to be a neghbor of another node y f x les wthn the transmsson range of y. The node wth maxmum number of neghbors (.e., maxmum degree) s chosen as a clusterhead. Experments demonstrate that the system has a low rate of clusterhead change but the throughput s low under the Hghest-Degree Algorthm. As the number of nodes n a cluster ncreases, the throughput drops and hence a gradual degradaton n the system performance s caused. All these drawbacks occur because ths approach does not have any restrcton on the upper bound of node degree n a cluster. 2.2 Lowest-ID Algorthm Ths Lowest-ID Algorthm was orgnally proposed by Baker and Ephremdes [6]. It assgns a unque d to each node and chooses the node wth the mnmum d as a clusterhead. As for ths algorthm, the system performance s better compared wth the Hghest- Degree Algorthm n terms of throughput. But t does not attempt to balance the load unformly across all the nodes.
2.3 Node-weght Algorthm Basagn et al. [7] proposed two algorthms, namely dstrbuted clusterng algorthm (DCA) and dstrbuted moblty adaptve clusterng algorthm (DMAC). In these two approaches, each node s assgned a weght based on ts sutablty of beng a clusterhead. A node s chosen to be a clusterhead f ts weght s hgher than any of ts neghbor s weght; otherwse, t jons a neghborng clusterhead. Results show that the number of updates requred s smaller than the Hghest- Degree and Lowest-ID Algorthms. Snce node weghts vary n each smulaton cycle, computng the clusterheads becomes very expensve and there are no optmzatons on the system parameters such as throughput and power control. 2.4 Weghted Clusterng Algorthm The Weghted Clusterng Algorthm (WCA) was orgnally proposed by M. Chatterjee et al.[8]. It takes four factors nto consderaton and makes the selecton of clusterhead and mantenance of cluster more reasonable. As s shown n equaton (1), the four factors are node degree dfference, dstance summaton to all ts neghborng nodes, velocty and remanng battery power respectvely. And ther correspondng weghts are w 1 to w 4.Besdes, t converts the clusterng problem nto an optmzaton problem and an objectve functon s formulated. W w1 + w2d + w3v + w4 = E (1) However, only those nodes whose neghbor number s less than a fxed threshold value can be selected as a clusterhead n WCA. It s not very desrable n the practcal applcaton. For example, many well-connected nodes whose neghbor number s larger than the fxed threshold mght be a good canddate as well. Besdes, ts energy model s too smple. It treats the clusterhead and the normal nodes equally and ts remanng power s a lnear functon of tme, whch s also not very desrable. So, we proposed an mproved clusterng algorthm as follows. 3 The Improved Weghted Clusterng Algorthm From the dscusson mentoned above, we can see that most clusterng algorthms, except for the WCA, only take one of the followng factors nto consderaton, such as the node degree, ID, speed or remanng power. When the problem n one aspect s solved, some other problems are ntroduced smultaneously. Inspred by the basc dea of WCA, we proposed an mproved clusterng algorthm. On the one hand, WCA only chooses those nodes whose neghbor number s less than a fxed threshold as a clusterhead canddate. However, many well-connected nodes whose neghbor number s larger than the fxed threshold mght be a good
canddate as well. So we can also treat them as clusterheads canddates and select an affordable number of normal nodes from ther neghborng nodes. On the other hand, we establshed a more practcal energy-consumpton model whch we wll explan later. By solvng the optmzaton problem of mn ( W ), the clusterheads and ther afflated normal nodes are selected and a trade-off s made from four aspects. 3.1 Prncples of the Improved Weghted Clusterng Algorthm In order to determne the ftness value W of a node as a clusterhead, we need to consder from the followng four aspects. If the node degree s hgher, then the node s more stable as a clusterhead. Here we make a smple converson = N M, where N s the practcal degree of node and M s the maxmum degree. The smaller s, the better node wll be as a clusterhead. As for those nodes whose practcal degree s larger than the maxmum degree M, we also treat them as clusterhead canddates. Once they are chosen as clusterheads, we wll choose M nodes wth less W as ther normal nodes. It s a dstnctve dfference between the orgnal WCA and our mproved algorthm, and t can work very well under densely deployed ad hoc networks where the WCA becomes useless. If the node velocty V s lower, then the node wll be more stable as a clusterhead. If the dstance summaton of node to all ts neghbors D s smaller, t wll consume less transmsson power to communcaton wth the normal nodes wthn ts cluster. In other words, the cost wll be smaller. If the remanng battery power E s hgher, the longer t wll be for node to serve as a clusterhead. Here we make another converson and set an energy-consumng model. All the E s are set to zero ntatorly. If the node serves as a clusterhead, we assume that t consumes 0.1 unt of energy and f normal node, 0.02 unt of energy. Once some E s above 1 (normalzed), we beleve that ths node s out of energy and the network wll become useless rapdly due to the avalanche effect [9]. The energyconsumng relatonshp of 5:1 s commonly used among some papers. And t meets wth the mnmzaton problem very well. As for some specfc applcaton, one can nfer to the related techncal report, such as the Mca2 Motes [10].And the model s also applcable through mnor modfcaton. 3.2 Steps of the Proposed Algorthm Takng node as an example, we compute ts W accordng to the followng steps and then judge whether t s a clusterhead or a normal node.
Step 1: Compute ts practcal degree and then derve the equaton = N M. Step 2: Compute the dstance summaton D to ts neghborng nodes. Step 3: Set the velocty V accordng to the random waypont moblty model. Step 4: At frst, set E to zero and ncrease ther values accordng to the energyconsumng model. Our algorthm termnates once some E s above 1 (normalzed). Step 5: Compute W accordng to varous w under dfferent applcaton. Step 6: Takng the node wth mnmum W as the frst clusterhead and ts neghborng nodes as ts normal nodes wthn the same cluster. Then we go on wth ths process untl all nodes act as ether clusterheads or normal nodes. Step 7: All the nodes move randomly after some unt tme and t goes back to step 1 agan. And t termnates untl a maxmum number of tme s reached or some node s out of energy. 4 A Novel Genetc Annealng based Clusterng Algorthm (GACA) The selecton of clusterheads set, whch s also called domnant set n Graphc Theory, s a NP-hard problem. Therefore, t s very dffcult to fnd a global optmum. So, we can take a further step to use the computatonal ntellgence methods, such as Genetc Algorthm (GA) or Smulated Annealng (SA), to optmze the objectve functon. Consderng the length of our paper, we wll skp the prncples of GA and SA, and explan the steps of our new Genetc Annealng based Clusterng Algorthm (GACA) drectly. The steps of our GACA are as follows. And t usually takes 5 to 10 teratons to convergence. So, we can say that t converges very fast. Step 1: As for N nodes, randomly generate L nteger arrangements n the range of [1, N]. Step 2: Accordng to these random arrangements and the clusterng prncple of WCA, derve L sets of clusterheads and compute ther correspondng wold. Step 3: Accordng to the Roulette Wheel Selecton and Eltsm n GA, select L sets of clusterheads whch are better, and replace the orgnal ones. Step 4: As for each of the L sets of clusterheads, perform the crossover operator and derve the new L sets of clusterheads and ther wnew. Step 5: Accordng to the Metropols accept or reject crtera n SA, decde whether to take the one from L sets of clusterheads n wold or n wnew. And the new L sets of clusterheads n the next generaton are obtaned. Step 6: Repeat Step 3 to 5 untl t converges or a certan number of teraton s reached. And n our smulaton, t usually takes 5 to 10 teratons to converge.
Then the global optmal or sub-optmal soluton mn ( wnew ) (=1, 2 L) s obtaned and ther correspondng set of clusterheads s known. In Step 2, we make L random arrangements n order to reduce the randomness n the clusterng process, because there s much dfference n the set of clusterheads (or domnant set) for dfferent nodes arrangements. As for the Roulette Wheel Selecton n Step 3, we do not take the tradtonal selecton probablty P =, but w w e L P =. In that case, the set of L w ( w ) ( e ) = 1 clusterheads whose w s smaller wll have more chance to be selected. Besdes, to overcome the randomness n the probablty problem, we preserve the best set of w drectly to the old wnew n Eltsm. To further reduce the randomness and ncrease the probablty that the global soluton may occur, we perform M pars of crossovers as for each of L random arranged ntegers (.e. moble nodes). And the new L sets of clusterheads and ther w new (=1,2, L) are derved. In Step 5, we make the accept or reject decson accordng to the Metropols crtera. If wnew wold, then we accept wnew drectly. If w > w, we do not reject t drectly, but accept t wth some probablty. new old e wnew αt old In other words, f s larger than a randomly generated number n the range of (0,1), whch shows that wnew and wold may be very close to each other, we wll stll take t. Or else, we wll reject the one n wnew and take ts counterpart n wold. Besdes, we let T = α T (α s a constant between 0 and 1 and we normally take 0.9) after each teraton, so that wnew and wold must be closer f wnew s to be accepted. In ths way, our GACA wll not be trapped n the local optma and the premature effect can be avoded. In other words, the dversty of searchng space can be ensured and t s smlar to the mutaton operator n GA. = 1 5. Performance Evaluaton We set our smulaton envronment as follows. There are N nodes randomly placed wthn a range of 100 by 100 m 2, whose transmsson range vares from 15m to 50m. A Random Waypont moblty model s adopted here. And our GACA parameters are lsted n table 1.
Table 1. GACA parameters M L α ε 1 10 0.9 0.01 5.1 Analyss of Average Cluster Number As s shown n fgure 1, we smulate N nodes whose transmsson range vares from 15m to 50m. We can conclude that: (1) The average cluster number (ACN) decreases as the transmsson range ncreases. (2) As for a smaller transmsson range, the average number of cluster dffers greatly for varous N. But when the transmsson range s about 50m, one node can almost cover the entre network. So t only takes 3 to 5 clusters to cover all the N nodes. Fg.1. ACN under varous transmsson range Fg.2. ACN under varous maxmum veloctes Besdes, we do the same research under varous veloctes. Takng N=R=30 as an example, we can draw the concluson from fgure 2 that: the average number of cluster vares randomly between 5 and 7, and t s not related wth the velocty. In fact, t matches wth the practcal stuaton too. For example, when one node wth large velocty moves out of a cluster, t s hghly possble that some other node gets nto the same cluster. Or some of the nodes mght move toward the same drecton, whch results n a relatvely slow velocty and a stable cluster too. 5.2 Analyss of Topology Stablty As s mentoned before, the clusterhead and ther afflated normal nodes may change ther roles as they move. Here, we defne a cluster reafflaton factor (CRF) as follows: 1 (2) CRF= N 1 N2 2 here, s the average number of cluster, and N 1, N 2 are the degree of node at dfferent tmes. For example, we assume that clusterhead 1 and 2 have 6 and 5 neghbors at frst,.e. N11 6, N21 5 = =. As they move after one unt tme, ther
neghbors (degrees) become 5 and 6,.e. N 12 = 5, N 22 = 6. We can derve that CRF s equal to 1. So, we beleve equvalently that one node n cluster 1 moves nto cluster 2 and one reafflaton s made. (a) Hghest-Degree Algorthm (b) WCA (c) GACA Fg.3. CRF under varous clusterng algorthms Under the maxmum velocty of 10 m/s, we compared the CRF performance of Hghest-Degree Algorthm, WCA and our GACA. From fgure 3, we can see that GACA has the lowest CRF, whch shows that t s the stablest clusterng strategy among three of them. And WCA has the hghest CRF value. The average CRF values of them are 1.56, 0.77 and 0.17 respectvely. Besdes, we dd some other experments about CRF. We got the concluson that the CRF ncreases as the velocty ncreases. 5.3 Analyss of Clusterhead Load-balancng We take the same defnton of load-balancng factor (LBF) as s defned n [8]: nc ( N nc LBF=, µ = ) 2 ( x µ ) n c where, n c s the average cluster number, N s the number of all nodes, and x s the practcal degree of node. The larger LBF s, the better the load s balanced. Takng N=20, M=4 as an example. The deal case s that there are 4 clusters and each clusterhead has a degree of 4,.e. 4 µ = ( 20 4) / 4 =. So LBF s n x =. Then, 4 c = nfnte, whch shows that the load s perfectly balanced. For smplcty, we do not consder the factor of network lfetme here (we wll dscuss t later n next secton). So we set the smulaton parameters as follows. (X,Y)=[100,100], N=20, R=30, M=4, maxmum velocty V max = 5 and w1 = 0.7, w2 = 0.2, w3 = 0.1, w4 = 0. It should be noted that we make N as our ( w = 0.7), because t represents the matchng degree of prmary focus of attenton 1 the practcal case and deal case drectly. Fgure 4 shows the LBF dstrbuton under Hghest-Degree Algorthm, WCA, our mproved weghted clusterng algorthm and
GACA. From fgure 4 we can see that: the Hghest-Degree Algorthm has the worst performance, WCA s secondary to t, and our two mproved clusterng algorthms are better. Besdes, the WCA wll become useless under densely deployed ad hoc networks whle our algorthm stll works well. And ther average values are 0.09, 0.38, 1.19 and 1.86 respectvely. (a) LBF under Hghest-Degree Algorthm (b) LBF under WCA (c) LBF under Our Improved Algorthm (d) LBF under our GACA Fg.4. LBF under varous clusterng algorthms 5.4 Analyss of Network Lfetme network lfetme 35 30 25 20 15 10 5 0 80 100 120 140 (X,Y) range GACA Our Improved Algorthm WCA Hghest-Degree Algorthm Fg.5. Network lfetme under varous clusterng algorthms
Fnally, we made a comparson between the aforementoned four clusterng algorthms n the aspect of network lfetme, as s shown n fgure 5. From whch, we can see that GACA acheves the best performance, our mproved weghted clusterng algorthm s second to t, the WCA and the Hghest-Degree algorthm are worse. 6 Concluson We proposed an mproved weghted clusterng algorthm based on the WCA and another novel Genetc Annealng based Clusterng Algorthm (GACA) n ths paper. Some performance comparson s made n the aspect of average cluster number, topology stablty, load-balancng and network lfetme. The smulaton results show that our two clusterng algorthms have a better performance on average. Acknowledgement Ths work was supported by grant No. R01-2005-000-10267-0 from Korea Scence and Engneerng Foundaton n Mnstry of Scence and Technology. References 1. Internet Engneerng Task Force MANET Workng Group. Moble Ad Hoc Network (MANET) Charter [EB/OL]. Avalable at http://www.etf.org/html.charters/manetcharter.html. 2. C.C. Chang. Routng n Clustered Multhop, Moble Wreless Networks wth Fadng Channel [C]. Proceedngs of IEEE SICON'97, Aprl1997, pp.197-211. 3. Mnglang Jang, Jnyang L, Y.C. Tay. Cluster Based Routng Protocol [EB/OL]. August, 1999 IETF Draft. 4. A.K. Parekh. Selectng routers n ad-hoc wreless networks [C]. Proceedngs of the SBT/IEEE Internatonal Telecommuncatons Symposum, August 1994. 5. M. Gerla and J.T.C. Tsa. Multcluster, moble, multmeda rado network [J]. ACM/Baltzer Wreless Networks, 1(3),1995, pp. 255-265. 6. D.J. Baker and A. Ephremdes. The archtectural organzaton of a moble rado network va a dstrbuted algorthm [J]. IEEE Transactons on Commcatonuns COM-29 11(1981) pp. 1694 1701. 7. S. Basagn. Dstrbuted clusterng for ad hoc networks [C]. Proceedngs of Internatonal Symposum on Parallel Archtectures, Algorthms and Networks, June 1999, pp. 310 315. 8. M. Chatterjee, S.K. Das and D. Turgut. An On-Demand Weghted Clusterng Algorthm (WCA) for Ad hoc Networks [C]. Proceedngs of IEEE GLOBECOM 2000, San Francsco, November 2000, pp.1697-1701. 9. Nshant Gupta, Samr R. Das. Energy-aware On-demand Routng for Moble Ad Hoc Network. [EB/OL], Avalable form http://crewman.uta.edu/~cho/energy.pdf. 10. MICA2 Mote Datasheet, http://www.xbow.com/products/product_pdf_fles/ Wrelesspdf/ 6020-0042-01_A_MICA2.pdf, 2004.