Sandp Kumar Goyal et al. / Internatonal Journal of Engneerng and Technology (IJET) Adaptve and Dynamc Load Balancng n Grd Usng Ant Colony Optmzaton Sandp Kumar Goyal 1, Manpreet Sngh 1 Department of Computer Engneerng, M.M. Unversty, Mullana, Ambala, Haryana (1337), Inda skgmmec@gmal.com Department of Computer Engneerng, M.M. Unversty, Mullana, Ambala, Haryana (1337), Inda drmanpreetsnghn@gmal.com Abstract Grd Computng nvolves coupled and coordnated use of geographcally dstrbuted resources for purposes such as large-scale computaton and dstrbuted data analyss. Wth the Grd becomng a vable hgh-performance alternatve to the tradtonal supercomputng envronment, a sutable and effcent load balancng algorthm s needed to equally spread the load on each computng node n the grd. Ths research work presents Ant based algorthm to solve the load balancng problem n computatonal grd. In proposed algorthm, the pheromone s assocated wth resources, rather than path. The ncrease or decrease of pheromone represent load and depends on task status at resources. The man objectve of proposed algorthm s to map tasks to resources n a way that balance out the load resultng n mproved utlzaton of resources. Keyword- Grd Computng, Load Balancng, Ant Colony Optmzaton, Resource Utlzaton. I. INTRODUCTION Grd nfrastructure ntegrates large computatonal and storage resources, data, servces and applcatons from dfferent dscplnes [1], []. Due to random task arrval patterns and heterogeneous nature of resources, resources n one grd ste may be over-loaded whle others n a dfferent grd ste may be under-loaded. It s therefore, requred to dspatch tasks to dle or under-loaded stes to obtan better resource utlzaton and reduce the average task response tme. Task schedulng [3] and load balancng [] are key grd servces, where ssues of load balancng represent a common concern for most grd nfrastructure developers. Load-balancng algorthms can be classfed nto statc and dynamc approaches. Statc load-balancng algorthms [5], [] assumes that a pror nformaton about all the characterstcs of the tasks, the computng nodes and the communcaton network are known and provded. In contrast, dynamc load-balancng algorthms [], [] attempt to use the runtme state nformaton to make more nformatve decsons n sharng the system load. It s now commonly agreed that despte the hgher runtme complexty, dynamc algorthms can potentally provde better performance than statc algorthms. Ant Colony Optmzaton (ACO) s a relatvely new computatonal and behavoural paradgm for solvng optmzaton and combnatory problems; t s based on the prncples that control the behavour of natural systems. In ths paper, Ant based algorthm for load balancng n Grd s proposed. The research work ams on mprovng the way ants search the best resources n terms of mnmzng the processng tme of each task and at the same tme balancng the load on avalable resources. II. ANT COLONY OPTIMIZATION Many aspects of the collectve actvtes of socal nsects, such as ants, are self-organzng. Ths means that complex group behavour emerges from the nteractons of ndvduals who exhbt smple behavours by themselves. Examples of these collectve actvtes among ants are fndng food and buldng nests [7]. The results of self-organzaton are global n nature, but come about from nteractons based entrely on local nformaton. To acheve ths, self-organzaton reles on several components: () postve feedback () negatve feedback () multple nteractons. The capabltes of a sngle ant are very lmted compared to those of a colony. In some speces, ants are mostly blnd and they communcate poorly. But collectvely, ants can establsh the shortest route between a source of food and ther nest and effcently move the food to ther home [9]. Ants communcate wth each other through the use of pheromones. As ants traverse a path, they depost pheromones. Pheromones are chemcal substances that attract other ants and are deposted by ants on the ground as they travel. Ants move randomly, but when they encounter a pheromone tral, they decde whether or not to follow t. If they do so, they lay down ther own pheromone on the tral as well, renforcng the pathway. The probablty that an ant chooses one path over other ncreases proportonal to the amount of pheromone present. The more ants that use a gven tral, the more attractve that tral becomes to subsequent ants. ISSN : 975- Vol No Aug-Sep 17
Sandp Kumar Goyal et al. / Internatonal Journal of Engneerng and Technology (IJET) Ant Colony Optmzaton (ACO) s a meta-heurstc usng artfcal ant to fnd desrable solutons to dffcult combnatoral optmzaton problems [7]. The behavour of artfcal ants s based on the trats of real ants as descrbed above, plus addtonal capabltes that make them more effectve, such as a memory of past actons. Each ant of the colony bulds a soluton to the problem under consderaton and uses nformaton collected on the problem characterstcs and ts own performance to change how other ants see the problem []. III. RELATED WORK In the past decade, a lot of research has been drected towards the task of effectve load balancng algorthms [] for dstrbuted computng systems. Generally, there are two methods for creatng load balancng n Grd envronment: statc load balancng and dynamc load balancng. In some algorthms the combnaton of these two methods are used whch are referred as combned algorthms. One of these combned algorthms s the load balancng algorthm based on the table of effectve nodes [11]. Another combned algorthm proposed for load balancng s Basc Hybrd algorthm [] whch used the combnaton of two methods Deferred and Random. The Deferred method uses the dynamc nformaton of the ste and the Random method uses the statc nformaton of the ste. Some of the algorthms have been developed by usng genetc approach through whch the selecton of the nodes are done by genetc operators whch nclude three operators of reproducton, exchange and mutaton [1]. There are also a number of algorthms that use the tree method for load balancng [17]. A recent approach s the use of ACO for schedulng jobs n grd [13]. ACO algorthm s used n grd computng because t s easly adapted to solve both statc and dynamc combnatoral optmzaton problems. In [1], ACO has been used as an effectve algorthm n solvng the load balancng problem n grd computng. A study n [15] proposed a new algorthm that s based on an echo ntellgent system, autonomous and cooperatve ants. In ths algorthm, the ants can procreate and also can commt sucde dependng on exstng condton. Ant level load balancng s proposed to mprove the performance of the mechansm. ACO algorthm for load balancng n dstrbuted systems through the use of multple ant colones s proposed n []. In ths algorthm, nformaton on resources s dynamcally updated at each ant movement. The study to mprove ant algorthm for job schedulng n grd computng s based on the basc dea of ACO was proposed n [1]. The pheromone update functon n ths research s performed by addng encouragement, punshment coeffcent and load balancng factor. Balanced job assgnment based on ant algorthm for computng grds called BACO was proposed n []. A game-theoretc-based soluton [1] to the grd load-balancng problem s proposed. The developed algorthm combnes the nherent effcency of the centralzed approach and the fault-tolerant nature of the decentralzed approach. [17] proposed a framework consstng of dstrbuted dynamc load balancng algorthm n perspectve to mnmze the average response tme of applcatons submtted to grd computng. In [7], authors has presented a Multple Ant Colony Optmzaton (MACO) approach for load balancng n crcut swtched networks. MACO uses multple ant colones to search for alternatves to an optmal path. Each group of moble agents corresponds to a colony of ants, and the routng table of each group corresponds to a pheromone table of each colony. Route [3] s a load balancng algorthm addressed to grd computng envronments where there s a large amount of resources, heterogenety, hgh communcaton latency, large number of users and dstrbuted locaton. An enhanced ant algorthm for load balancng n grd computng s proposed n [9]. The proposed algorthm wll determne the best resource to be allocated to the jobs based on job characterstcs and resource capacty, and at the same tme to balance the entre resources. In [19], dynamc grd schedulng algorthm based on adaptve ant colony algorthm was proposed. In ths algorthm, the evaporaton rate value was adaptvely changed and a mnmum value of zero was fxed. The local and global pheromone updates were used n order to control the pheromone value of each resource. IV. ANT BASED LOAD BALANCING ALGORITHM () In proposed ant algorthm, the pheromone s assocated wth resources, rather than path. The ncrease or decrease of pheromone represent load and depends on task status at resources. The notatons used n the descrpton of are llustrated n Table I. ISSN : 975- Vol No Aug-Sep 1
Sandp Kumar Goyal et al. / Internatonal Journal of Engneerng and Technology (IJET) TABLE I Notatons Used T R Number of Tasks th Avalable Resource τ () Intal Pheromone Assocated wth R τ (t) Current Pheromone Assocated wth R p (t) Probablty of Task Assgnment to R α Relatve Performance of Pheromone Tral Intensty β Relatve Importance of Intal Performance Attrbutes ρ Pheromone Decay Parameter wth Value Between and 1 Δ Pheromone Varance N Number of FT t RU The workng of proposed algorthm s as follow: Step 1. Intalze the value ofα, β, ρ, Δ, N, T, Fnsh Tme of Task t on Resource R Resource Utlzaton of R Step. Select the next task t. Step 3. Determne the transton probablty (load) of each resource p ( t) = j α [ τ ( t) ] * [ η ] r j β [ τ ( t) ] α * [ η ] β r j r RU and also set pheromone trals for each resource. R j as: Step. Fnd resource R wth hgh transton probablty among all resources: p ( t) = max p ( t) l N l.e. resource R s havng mnmum load. Step 5. Assgn task t to R. Step. Set T = T - 1. Step 7. Check whether any task completon or falure reported. If no, go to Step 11. Step. If (task completon at any resource R ) then Increase pheromone of R as: τ ( t ) = τ ( t ) + Δ reportng R as lghtly loaded. t Step 9. RU = RU + FT Step. If (task falure at any resource R ) then decrease pheromone of τ ( t) = τ ( t) Δ reportng R as heavly loaded. Step 11. If (T>) then go to Step. Step. For each resource R, 1 N RU Compute RU = N RU k = 1 Prnt resource utlzaton of R. k R as: ISSN : 975- Vol No Aug-Sep 19
Sandp Kumar Goyal et al. / Internatonal Journal of Engneerng and Technology (IJET) V. SIMULATION RESULTS & DISCUSSION In ths secton, some experments that have been carred out to test the effcency and effectveness of proposed algorthm are presented. The functonal code s mplemented usng GrdSm on an Intel core duo, GHz wndow based laptop. Table II specfes the smulaton envronment: TABLE II Smulaton Parameters Smulaton Runs No. of 3 No. of Tasks 5 Task Sze (MI) - Processng Power of (MIPS) 3 - In order to determne whether can search a near optmal schedule for a large number of tasks or resources, the smulaton was performed n two scenaros. A. Scenaro 1 (Effect of load n terms of tasks on average resource utlzaton) The number of tasks s vared from 5 to whle keepng resources as and the result on resource utlzaton s depcted n Fg.1, 3, and 5. The ndvdual assgnment of 5 tasks to resources under proposed algorthm and (Wthout Ant based Load Balancng Algorthm) s shown n Fg.. Resource Utlzaton(%) 1 1 15 9 3 1 3 5 7 9 Fg. 1. Effect of load varaton on average resource utlzaton wth Tasks = 5 and =. 1 3 5 7 9 111311511711935 Tasks Fg.. Assgnment of Tasks wth Tasks = 5 and =. ISSN : 975- Vol No Aug-Sep 17
Sandp Kumar Goyal et al. / Internatonal Journal of Engneerng and Technology (IJET) 5 Resource Utlzaton(%) 15 5 1 3 5 7 9 11 Fg. 3. Effect of load varaton on average resource utlzaton wth Tasks = 5 and =. Resource Utlzaton(%) 1 1 1 1 3 5 7 9 11 Fg.. Effect of load varaton on average resource utlzaton wth Tasks = 75 and =. Resource Utlzaton(%) 1 1 1 3 5 7 9 11 Fg. 5. Effect of load varaton on average resource utlzaton wth Tasks = and =. Ths mprovement s expected as s keepng track of the state of all resources at each pont n tme whch makes t able to take optmal decsons. B. Scenaro (Effect of scalablty on average resource utlzaton) ISSN : 975- Vol No Aug-Sep 171
Sandp Kumar Goyal et al. / Internatonal Journal of Engneerng and Technology (IJET) The number of resources s vared from to 3 whle keepng number of tasks as 75. Fg. and 7 shows the effect of scalablty on the performance of algorthms under observaton n terms of load balancng. Utlzaton(%) 1 3 5 7 9 11 13 1 15 1 17 1 19 1 Resource Fg.. Effect of scalablty wth Tasks = 75 and =. Resource Utlzaton (%) 1 3 5 7 9 111311511711935793313 Fg. 7. Effect of scalablty wth Tasks = 75 and = 3. Fg. shows that proposed algorthm s better than as the standard devaton for s not more than.77 and the standard devaton for ranges from.35 to 3.1. 3.7.35 Resource.77.57.71 3.1.5 1 1.5.5 3 3.5 Standard Devaton Fg.. Comparson n terms of standard devaton wth Tasks = 75. ISSN : 975- Vol No Aug-Sep 17
Sandp Kumar Goyal et al. / Internatonal Journal of Engneerng and Technology (IJET) VI. Concluson Load balancng s one of the man ssues n the grd envronment. Recent researches have proved that load balancng on computatonal grds s best solved by heurstc approach. Hence, an ant based load balancng algorthm s developed to allocate tasks to proper resources. In order to verfy the performance of proposed algorthm, the smulaton s performed. The results of the experments are also presented and the strength of the algorthm s nvestgated. The smulaton result concludes that the proposed algorthm enhances performance n terms of resource utlzaton. REFERENCES [1] I. Foster, and C. Kesselman, The Grd: Blueprnt for a new Computng Infrastructure, nd ed.: Morgan Kauffman publshers,. [] M. Bote-Lorenzo, Y. Dmtrads, and E. Gomez-Sanchez, Grd characterstcs and uses: a grd defnton, n Proc. 1 st European Across Grds Conference (ACG 3),, pp. 91-9. [3] A.Y.Zomaya, and Y.H.The, Observatons on usng genetc algorthms for dynamc load-balancng, IEEE Transactons on Parallel and Dstrbuted Systems, Vol., No. 9, pp. 99-9, 1. [] M. Sngh, and P.K. Sur, An effcent decentralzed load balancng algorthm for grd, n Proc. nd IEEE Inter. Conf. (IACC),, pp. -13. [5] A.N.Tantaw, and D. Towsley, Optmal statc load balancng n dstrbuted computer systems, Journal of Assocaton for Computng Machnery, Vol. 3, No., pp. 5-5, Aprl 195. [] J. Xu, and K. Hwang, Heurstc methods for dynamc load balancng n a message-passng multcomputer, Journal of Parallel and Dstrbuted Computng, Vol. 1, pp. 1-13, 1993. [7] A. D. Al, and M. A. Belal, Multple ant colones optmzaton for load balancng n dstrbuted systems, n Proc. Inter. Conf. (ICTA 7), 7. [] P. McMullen and P. Tarasewch, Usng ant technques to solve the assembly lne balancng problem, Insttute of Industral Engneers Trans., vol. 35, no. 7, pp. 5-17, 3. [9] H. J. A. Nasr, K. R. K. Mahamud, and A. M. Dn, Load balancng usng enhanced ant algorthm n grd computng, n Proc. nd Inter. Conf. Computatonal Intellgence, Modellng and Smulaton,, pp. 1-15. [] L. M. Khanl, and B. Ddevar, A new hybrd load balancng algorthm n grd computng systems, Internatonal Journal Computer Scence and Technology, vol., no. 5, 11. [11] K. Q. Yan, S. C. Wang, C.P Chang, and J. S. Ln, A hybrd load balancng polcy underlyng grd computng envronment, Computer Standards and Interfaces, vol. 9, no., pp. 11-173, 7. [] S. Zkos, and H. D. Karatza, Communcaton cost effectve schedulng polces of nonclarvoyant jobs wth load balancng n a grd, Journal of Systems and Software, vol., no., pp. 3-11, 9. [13] S. Fdanova, and M. Durchova, Ant algorthm for grd schedulng problem, n Proc. 5th Inter. Conf. Large Scale Scentfc Computng (LSSC), vol. 373,, pp. 5-. [1] Y. L, A Bo-nspred adaptve job schedulng mechansm on a computatonal grd, Internatonal Journal of Computer Scence and Network Securty (IJCSNS), vol., no. 3, pp. 1-7,. [15] M. Saleh, and H. Deldar, Grd load balancng usng an echo system of ntellgent ants, n Proc. th Inter. Conf. Parallel and Dstrbuted Computng and Networks,, pp. 7-5. [1] Y. L, Y. Yang, M. Ma, and L. Zhou, A hybrd load balancng strategy of sequental tasks for grd computng envronments, Future Generaton Computer Systems, vol. 5, no., pp. 19-, 9. [17] B. Yagoub, and Y. Slman, Task load balancng strategy for grd computng, Journal of Computer Scence, vol. 3, no. 3, pp. 1-19, 7. [1] R. Subrata, A. Y. Zomaya, and B. Landfeldt, Game theoretc approach for load balancng n computatonal grds, IEEE T.rans. Parallel and Dstrbuted Systems, vol. 19, no. 1, pp. -7,. [19] Lu., A. Wang, Grd task schedulng based on adaptve ant colony algorthm, n Proc. Inter. Conf. Management of e-commerce & e- Government,, pp. 15-1. [] A. Al, M. A. Belal, and M. B. Al-Zoub, Load balancng of dstrbuted system based on multple ant colones optmzaton, Journal of Appled Scences, vol. 7, no. 3, pp. 33-3,. [1] H. Yan, X. Shen, X. L, and M. Wu, An mproved ant algorthm for job schedulng n grd computng, n Proc. th Inter. Conf. Machne Learnng and Cybernetcs, 5, pp. 957-91. [] R. Chang, J. Chang, and P. Ln, Balanced job assgnment based on ant algorthm for grd computng, n Proc nd IEEE Asa-Pacfc Servce Computng Conference (APSCC), 7, pp. 91-95. [3] R. F. Mello, L. J. Senger, and L. T. Yang, A routng load balancng polcy for grd computng envronments, n Proc. th Inter. Conf. Advanced Informaton Networkng and Applcatons (AINA ),, pp. 153-15. [] S. K. Goyal, R. B. Patel, and M. Sngh, Adaptve and dynamc load balancng methodologes for dstrbuted envronment: a revew, Internatonal Journal of Engneerng Scence and Technology (IJEST), vol. 3, no. 3, pp. 135-1, 11. AUTHOR S BIOGRAPHY Sandp Kumar Goyal receved hs B.Tech, M.Tech. from Kurukshetra Unversty, Kurukshetra, Inda and s currently enrolled as a Ph.D. scholar n the department of Computer Scence and Engneerng at M.M.Unversty, Mullana, Ambala, Haryana, Inda. He s presently servng as Assoc. Professor n Computer Engneerng Department of M.M. Engneerng College, Mullana, Ambala. He s n teachng snce. He has publshed research papers n Internatonal and Natonal journals and conferences. Hs research area s Load balancng Methodologes n dstrbuted envronment. ISSN : 975- Vol No Aug-Sep 173
Sandp Kumar Goyal et al. / Internatonal Journal of Engneerng and Technology (IJET) Dr. Manpreet Sngh receved hs B. Tech., M.Tech. and Ph.D. from Kurukshetra Unversty, Kurukshetra, Inda. He s presently servng as Professor and Head, Computer Scence and Engneerng Department of M. M. Engneerng College, Mullana, Ambala. He has about 13 years of experence n teachng and research. He has publshed 3 research papers n Internatonal and Natonal Journals and Conferences. Hs s current research nterest ncludes Grd Computng, Cloud Computng and MANETs. ISSN : 975- Vol No Aug-Sep 17