ISSN : 2347-8446 (Onlne) Inernaonal Journal of Advanced Research n Genec Algorhm wh Range Selecon Mechansm for Dynamc Mulservce Load Balancng n Cloud-Based Mulmeda Sysem I Mchael Sadgun Rao Kona, II K.Purushoama Rao I,II Dep. of Informaon Technology, LBRCE, JNTUK Unversy, Andhra Pradesh, Inda Absrac Consder a cenralzed herarchcal cloud-based mulmeda sysem (CMS) whch has a resource manager, cluser heads, and server clusers. Resource manager acceps clens requess for mulmeda servce asks and allocaes server clusers based on he ask characerscs. Cluser head dsrbues he ask o he servers whn s server cluser. An effecve genec algorhm wh an mmgran scheme s proposed for load balancng ha spreads he mulmeda servce ask load on servers wh he mnmal cos for ransmng mulmeda daa beween server clusers and clens, by no volang he maxmal load lm of each server cluser. Dynamc mulservce scenaro s consdered n whch each server cluser only handles a specfc ype of mulmeda operaon, and every clen requess a dfferen ype of mulmeda servce a a dfferen me. Keywords Cloud compung, load balancng, genec algorhm and mulmeda sysem. I. Inroducon Cloud-based mulmeda sysem (CMS) [3] ganed momenum because here s huge number of users demands for varous mulmeda compung and sorage servces hrough he Inerne a he same me [4], [5]. I provdes nfrasrucure, plaforms, and sofware as servce o large number of clens a he same me n order o sore and process her mulmeda applcaon daa n a dsrbued manner by meeng dfferen mulmeda QoS requremens hrough he Inerne. Cloud compung s srongly requred n mos mulmeda applcaons wh huge compuaon and for moble devces whch are power consraned. In general, a cloud servce provder offers requred facles o clens whch ake less cos o reques, process and o compue mulmeda servces. So, mulmeda applcaons are processed on powerful cloud servers wh less cos because clen should only pay for he ulzed resources by he me. We are consderng a cenralzed herarchcal CMS from [3] composed of a resource manager and a number of server clusers, each of whch s coordnaed by a cluser head, and we assume he servers n dfferen server clusers o provde dfferen servces. Operaon of CMS Resource manager assgns clen requesed mulmeda servce ask o dfferen server clusers accordng o he characerscs of he requesed asks. Cluser head wll dsrbue he assgned ask o some server whn he server cluser. The load of each server cluser affecs he performance of he whole CMS. Resource manager of he CMS s requred o dsrbue he ask load across server clusers, and hence, s mporan o be able o cope wh load balancng n he CMS. Servces offered by CMS are generang, edng, processng, and searchng a varey of mulmeda daa, e.g., hyperex, mages, vdeo, audo, graphcs, and so on wh varous requremens for he funcons provded by he CMS (sorage, cenral processng un, and graphcs processng un clusers), e.g., he requremen for QoS of hml webpage servces s looser han ha of vdeo sreamng servces. Fg. 1: Cenralzed Herarchcal cloud based Mulmeda Sysem [1] We assume n CMS, each server cluser can only handle a specfc ype of mulmeda servce ask, and each clen requess a dfferen ype of mulmeda servce a dfferen me. In hs paper, we propose a genec algorhm (GA) for he concerned dynamc load balancng problem for CMSs. In our seng of GA, ele mmgrans and random mmgrans are added o new populaon, because hey are suable for solvng he problems n dynamc envronmens [6]. The expermenal resuls show ha o a ceran exen, our approach s capable of dynamcally spreadng he mulmeda ask load evenly. II. Problem Descrpon Acually hs load balancng problem for he CMS s from [1], and we are exendng her model wh dfferen selecon mehod n GA. A. Sysem Overvew CMSs can be classfed no wo caegores: cenralzed and decenralzed. We are consderng a cenralzed CMS as llusraed n Fg. 1, whch consss of a resource manager and a number of server clusers each of whch s coordnaed by a cluser head. Dfferen from he decenralzed CMS, whenever receves requess from clen for mulmeda servce asks, he resource www.arcs.com 217 All Rghs Reserved, IJARCST 2014
Inernaonal Journal of Advanced Research n ISSN : 2347-8446 (Onlne) manager of he cenralzed CMS sores he global servce ask load nformaon colleced from server clusers, and decdes he amoun of clen s requess assgned o each server cluser so ha he load of each server cluser s dsrbued as balanced as possble n erms of he cos of ransmng mulmeda daa beween server clusers and clens. The decson of assgnmen s based upon he characerscs of dfferen servce requess and he nformaon colleced from server clusers. B. Problem Formulaon We are usng he same CMS problem formulaon as n [1], n whch me s dvded no dfferen me seps. A he h me sep, he CMS s modeled as a complee weghed bpare graph G = (U, V, E, φ, ψ, q, r, w ) n whch U s he se of verces for server clusers, V s he se of verces for clens, E s he se of edges beween U and V, n whch each edge e E represens he lnk beween server cluser U and clen j V. φ s a funcon used o represen ha server cluser can only cope wh mulmeda asks of ype φ, ψ s a funcon used o represen ha clen j requess he mulmeda servce of ype ψ j a he h me sep. q s a funcon used o represen ha server cluser can provde he mulmeda servce of QoS q, r s a funcon used o represen ha clen j requess he mulmeda servce of QoS requremen r a j he h me sep and fnally w s he wegh funcon assocaed wh edges, n whch w denoes he w value ha represens he cos for ransmng mulmeda daa beween server cluser and clen j a he h me sep, whch s defned as follows: w =, f d or φ ψ j w = d. l, oherwse (1) where d s he nework proxmy beween server cluser and clen j; l s he raffc load of he lnk beween server cluser and clen j ha s defned as follows: l = k K u C (2) kj k where K s he servers se n server cluser ; u s he server kj ulzaon rao of server k n server cluser due o clen j, and C k s s capacy. Noe ha he proxmy d beween server cluser and clen j n (1) s requred o be measured a every me sep due o dynamc change of nework opology. Proxmy beween nodes s calculaed wh landmark order and Bnnng scheme. Ths paper uses he same lnear programmng formulaon used n [1] for every h me sep wh consrans: Each clen only allows a mos one lnk o be assgned (Consran (4)), he ulzed capacy of each server cluser canno exceed s capacy a he h me sep (Consran (5)), mulmeda servce ype requesed by each clen j s conssen wh ha provded by server cluser (Consran (6)) and each clen j requess he mulmeda server of he QoS no more han ha offered by server cluser (Consran (7)). Dynamc Mulservce Load Balancng In Cms (Cms-Dynmlb): G =(U, V, E, φ, ψ, q, r,w ) s he bpare graph for CMS wh m servers and n clens for each me sep =1,2,.Ths paper uses Range selecon mehod for genec algorhm (GA) wh mmgran scheme [1] for beer resuls n solvng he problem. Algorhm 1 Dynamc Load Balancng Algorhm for = 1, 2, do creae complee weghed bpare graph G remove he edges n G volang (6) and (7) calculae {l } load and wegh {w } on server cluser by clen j by callng Algorhm2 assgn {x } by callng Algorhm 3 Algorhm 2 Calculae Weghs for each clen j V do measure he laency from clen j o each landmark compue he landmark order j of clen j oban he se of avalable server clusers U j for each U j do measure he laency from server cluser o each landmark compue he landmark order of server cluser f l = l j hen measure he nework proxmy d beween server cluser and clen j measure server ulzaon raos u for all k K kj respecvely w = l = calculae l and w by Equaons (2) and (1), Algorhm 3 Genec Algorhm (graph G, me sep ) f = 1 hen generae and evaluae he nal populaon P 0 of sze η n whch each chromosome has o sasfy (4) and (5). P 0 P 1 τ 0 whle < τ or he convergen condon s no acheved do selec he parenal pool Q from P by usng Range Selecon mechansm reproduce a new populaon P of sze η by performng crossover procedure on pars of chromosomes n Q wh probably p c perform muaon procedure on chromosome n P wh probably p m repar each nfeasble chromosome n P evaluae P f r e > 0 hen a number of he bes chromosome E 1, called ele, are seleced from he prevous generaon P 1 generae and evaluae r e η ele mmgrans 2014, IJARCST All Rghs Reserved 218 www.arcs.com
ISSN : 2347-8446 (Onlne) Inernaonal Journal of Advanced Research n f r r > 0 hen generae and evaluae r r η random mmgrans replace he wors chromosomes n P wh he above ele and random mmgrans P P +1 +1 end whle oupu he bes found chromosome as he soluon a he h me sep Algorhm 4 Inalze for j = 1 o n do arg(u j ) se of negers whch sore he ndces of he avalable cluser servers U j ha are lnked wh clen j n he reduced bpare graph G (where he lnks volang Consrans (6) and (7) have been removed n Algorhm 1) for p = 1 o η do consruc he p h chromosome c p = σ 1,, σ j,,σ n, n whch σ j s a number chosen arbrarly from arg(u j ) for each j {1,, n} le o be he se of he posons of 0 s n c p ha do no volae our problem consrans for = 1 o m do scan chromosome cp o compue he j U x l value for server cluser g j U x l k K C k le o be he se of o assocaed wh cluser server whle g > 0 do fnd he ndex x such ha σ x = n cp, l x > g,and he gap beween l x and g s he smalles f here s a poson o y n o ha can accommodae l x and sasfy all he problem consrans hen swap he values of o y and σ x remove o y n o g g l x σ x 0 end whle III. Our Genec Algorhm A. Our Algorhm GA s based on mang he evoluonal behavor of a populaon of chromosomes (each represens a canddae soluon) o fnd he soluon close o he global opmal soluon by usng hree basc evoluonal operaons: selecon, crossover, and muaon. The nal sep of he algorhm s o manan a populaon (generaon) of chromosomes ha are nalzed randomly or n some way. Nex, a number of he chromosomes n he populaon are seleced as he parenal pool, n whch each par of chromosomes are crossovered o produce chld chromosomes. Nex, pars of he orgnal chromosomes and he chld chromosomes consue he nex generaon. Afer evolvng a maxmal number of generaons or achevng a convergen condon, he soluon represened by he chromosome wh he bes fness value n he las generaon s oupued as he fnal soluon. A each me sep, Algorhm 1 eraes on o reallocae he nework load assgnmens so as o adap o he me change. Frs, Lne 2 consrucs a complee weghed bpare graph G. Lne 3 removes he lnks n G volang (6) and (7). Before usng our GA o calculae soluons, he nformaon of {l } and {w } s requred, so Lne 4 of Algorhm 1 calls Algorhm 2 o oban hose nformaon. Afer ha, Lne 5 of Algorhm 1 calls Algorhm 3 o compue our fnal load assgnmen soluon {x }. In algorhm 2, Lne 1 consders each clen j V o compue s wegh w wh each server cluser. We are usng dsrbued bnnng scheme o calculang he proxmy, so, Lnes 2 3 calculae he landmark order lj of clen j, and hen, for each avalable server cluser n he se Uj ha ncludes he server clusers conneced o clen j, Lnes 6 and 7 calculae he landmark dsance l of server cluser. In Lnes 8 14, f lj = l (.e., clen j and server cluser belong o he same landmark bn), we acually measure he nework proxmy beween hem, and hen compue her l and w values; oherwse, we drecly le he l and w values be. Wh he values of { l } and { w }, we are ready o apply he GA o compung he opmal load assgnmen n Algorhm 3. The GA runs dfferenly based on wo parameers: he npu bpare graph G and he npu me sep. Recall ha G may no be a complee graph any longer, because he lnks ha volae some problem consrans have been removed n Lne 3of Algorhm 1. Le η and τ denoe he sze of a populaon and he number of generaons, respecvely. In Lnes 1 5, he nal populaon of η chromosomes s generaed n wo cases. If = 1 (.e., hs s he frs me o run he GA), hen we randomly generae he nal populaon ha sasfes he remanng wo consrans ha we dd no consder ye (.e., (4) and (5)). Oherwse (.e., 1, whch mples ha here exsed he fnal populaon P 1 of he ( 1) τ h me sep), Lne 4 uses P 1 as he nal populaon a he h τ me sep Subsequenly, he whle loop n Lnes 7 23 repeas a mos τ eraons, each of whch produces a populaon P of +1 chromosomes,.e., he nex populaon a he h me sep. Lne 8 selecs a number of chromosomes from he populaon P as he parenal pool Q. Lnes 9 and 10 perform crossover and muaon operaors o he populaon P wh probables pc and pm, respecvely. In addon, snce our concerned problem consders dynamc scenaros a dfferen me seps, we add ele mmgrans and random mmgrans o adap o dynamc changes, as he mmgrans are used o solve dynamc problems convenonally [6]. Lnes 13 16 add ele mmgrans for ncreasng effcency of convergence, whle Lnes 17 19 add random mmgrans for ncreasng he populaon dversy. Lne 20 replaces he wors chromosomes n P wh he ele and random mmgrans. Afer fnshng he whle loop, he bes found chromosome s oupued as he soluon a he h me sep. Noe ha a mulmeda servce ask may no be able o be fnshed whn a sngle me sep,.e., akes a number of me seps o be fnshed. Hence, s no necessary o manan a specfc lnk for he same servce for some me seps because he nework s packe-based. B. Basc Elemens of Our GA To use GA n order o solve he CMS-dynMLB problem, we frs defne he basc elemens of GA (.e., populaon, chromosome, fness funcon) for he problem as follows. 1) Populaon: A populaon consss of a number of chromosomes, and he number of chromosomes depends on he gven nal populaon sze. www.arcs.com 219 All Rghs Reserved, IJARCST 2014
Inernaonal Journal of Advanced Research n ISSN : 2347-8446 (Onlne) 2) Chromosome: A soluon for he CMS-dynMLB problem consss of all he ndcaor varables { x U, j V}.The soluon for ndcaor varables { x } s encoded as a sequence of decmal numbers of lengh n:{σ 1,, σ j,, σ n } where σ j {0, 1, 2,,m} represens he lnk assgnmen of clen j wh he followng wo cases. 1) If σ j = 0, hen each x = 0 for any U,.e., clen j s no lnked a he h me sep. 2) Oherwse, σ j = k 0, meanng ha x = 1 and kj x = 0 for any k,.e., clen j s lnked o server cluser k unquely. 3) Fness Funcon: Fness funcon s he measure for deermnng whch chromosomes are beer or worse. We le he objecve (3) of our concerned problem as our fness funcon as follows: f (X()) = λ U j V x w / w + (1 λ) (1 j V max j V U x / V ) where X() s he profle of { x }. C. Man Componens of Our GA 1) Inalzaon: In Lnes 1 5 of Algorhm 3, we use wo cases o ge nal populaon a each me sep. If >= 1, hen jus ake he fnal populaon a he prevous me sep as he nal populaon a he curren me sep. For he oher case (.e., = 1), we randomly produce he nal populaon by usng Algorhm 4, whch s explaned as follows. Lnes 1 3 fnd he se arg(u j ) ha collecs he ndces of he avalable server clusers for each clen j. Nex, he se arg(u j ) s used o consruc a populaon of η chromosomes n Lne 5, and hen we repar each chromosome o be feasble n Lnes 6 21. The dea of he reparng operaon s ha f (5) s volaed, hen we fnd anoher avalable server cluser o serve he volaed load; oherwse, we drop he load. 2) Selecon, Crossover: The range selecon procedure s used. Each ndvdual n he populaon s assgned a numercal rank based on her fness, and selecon s based on hese rankng raher han absolue dfferences n fness. The advanage of hs mehod s ha can preven very f ndvduals o gan domnance a frs a he expense of he less f, whch would reduce he genec dversy of he populaon and could hamper he search for an accepable soluon [7]. One-pon crossover operaon s used n whch wo seleced chromosomes are crossovered for generang wo new chld chromosomes n order o keep some characerscs of s paren chromosomes. 3) Muaon: Changng some genes on some chromosomes. le pm be gven probably o muae. In he muaed chromosome, a nonzero gene σ x s chosen randomly, and hen we randomly fnd a zero gene σ y ha can accommodae σ x and do no volae any problem consran. Then, we swap he values of σ x and σ y. 4) Repar: The modfed chromosomes may become nfeasble, and hence repar s requred. The reparng operaon s referred o Lnes 6 21 n Algorhm 4, whch have been used n nalzng chromosomes. 5) Termnaon: If he dfference of average fness values beween successve generaons n he laes en generaons s no greaer han 1% of he average of he average fness values of hese en generaons, or he maxmum generaons are acheved, hen our GA sops. Afer ermnaon, he bes chromosome from he laes populaon s chosen, and s correspondng load assgnmen s oupued as he fnal soluon. IV. Implemenaon and Expermenal Resuls A. Implemenaon CloudSm 3.0.3[2] s used for smulaon wh java as programmng language. CloudSm has suppor for modelng and smulaon of large scale Cloud compung envronmens, ncludng daa ceners, on a sngle physcal compung node. I has necessary classes n java for smulang every eny ha nvolved n our projec lke Daacener, Hos, VM, Cloudle, DaacenerBroker ec. The CloudSm oolk suppors boh sysem and behavour modellng of Cloud sysem componens. I has avalably of a vrualzaon engne ha ads n creaon and managemen of mulple, ndependen, and co-hosed vrualzed servces on a daa cener node. Our smulaon was esed on an Inel Core 3-2328M CPU a 2.20 GHz wh 4-GB memory. B. Resuls Wh Range selecon mechansm mplemenaon n exsng genec algorhm for balancng dynamc mulmeda load n cloud based CMS here s a lle beermen. V. Concluson A genec algorhm approach for opmzng he CMSdynMLB was proposed and mplemened. The man dfference n our model from prevous models s ha we consdered a Range selecon mechansm n genec algorhm along wh praccal mulservce dynamc scenaro n whch a dfferen me seps, clens can change her locaons. References [1] Chun-Cheng Ln, Member, IEEE, Hu-Hsn Chn, Suden Member, IEEE, and Der-Junn Deng, Member, IEEE, Dynamc Mulservce Load Balancng n Cloud-Based Mulmeda Sysem n Sysems Journal, IEEE (Volume:8, Issue: 1),pp.225-234, March 2014. [2] Rodrgo N. Calheros, Rajv Ranjan, Anon Beloglazov, Cesar A. F. De Rose, and Rajkumar Buyya, CloudSm: A Toolk for Modelng and Smulaon of Cloud Compung Envronmens and Evaluaon of Resource Provsonng Algorhms, Sofware: Pracce and Experence (SPE), Volume 41, Number 1, Pages: 23-50, ISSN: 0038-0644, Wley Press, New York, USA, January, 2011. [3] W. Zhu, C. Luo, J. Wang, and S. L, Mulmeda cloud compung: An emergng echnology for provdng mulmeda servces and applcaons, IEEE Sgnal Process. Mag., vol. 28, no. 3, pp. 59 69, May 2011. [4] C.-F. La, Y.-M. Huang, and H.-C. Chao, DLNA-based mulmeda sharng sysem over OSGI framework wh exenson o P2P nework, IEEE Sys. J., vol. 4, no. 2, pp. 262 270, Jun. 2010. 2014, IJARCST All Rghs Reserved 220 www.arcs.com
ISSN : 2347-8446 (Onlne) Inernaonal Journal of Advanced Research n [5] W. Hu, H. Zhao, C. Ln, and Y. Yang, Effecve load balancng for cloud-based mulmeda sysem, n Proc. In. Conf. Elecron. Mech. Eng. Inform. Technol., 2011, pp. 165 168. [6] H. Cheng and S. Yang, Genec algorhms wh mmgran schemes for dynamc mulcas problems n moble ad hoc neworks, Eng. ppl.arf. Inell., vol. 23, no. 5, pp. 806 819, 2010. [7] R.Svaraj e al., Dr.T.Ravchandran, A Revew of selecon mehods n genec algorhm n IJEST, Vol. 3 No. 5, pp.3792-3797, May 2011. Auhor s Profle Mchael Sadgun Rao Kona pursung M.Tech Degree n Sofware Engneerng from Lakreddy Balreddy College of Engneerng, Mylavaram, AP, INDIA. Hs man research neres n sofware engneerng, Cloud Compung. K.Purushoam Rao s presenly workng as Asssan Professor n Dep. of Informaon Technology, Lakreddy Balreddy College of Engneerng, Mylavaram. www.arcs.com 221 All Rghs Reserved, IJARCST 2014