An Integrated eource Management and Schedulng Sytem for Grd Data Streamng Applcaton Wen Zhang, Junwe Cao 2,3*, Yheng Zhong,3, Lanchen Lu,3, and Cheng Wu,3 Department of Automaton, Tnghua Unverty, Bejng 00084, Chna 2 eearch Inttute of Informaton Technology, Tnghua Unverty, Bejng 00084, Chna 3 Tnghua Natonal Laboratory for Informaton Scence and Technology, Bejng 00084, Chna *Correpondng emal: jcao@tnghua.edu.cn Abtract Grd data treamng applcaton are novel from other n that they requre real-tme data upply whle the proceng gong on, whch necetate harmonou collaboraton among proceor, bandwdth and torage. Tradtonal chedulng approache may not be uffcent for uch applcaton, for they uually focu on only one apect of reource, manly computatonal reource. A reource management and chedulng ytem for uch applcaton developed n th paper, whch reponble for enablng ther runnng baed on Globu toolkt. An ntegrated cheme propoed, ncludng admon control, applcaton electng, proceor agnng, allocaton of bandwdth and torage, wth correpondng algorthm elaborated. Evaluaton reult how excellent performance and calablty of th ytem.. Introducton Streamng applcaton are ganng ther popularty recently, and n mot cae data are puhed to the computatonal reource for dtrbuted proceng wth real-tme contrant, o the proceng rate mut match the data arrval rate. Nowaday, new knd of treamng applcaton are emergng wth dfferent requrement and charactertc. For example, LIGO (Laer Interferometer Gravtatonal-wave Obervatory) [] generatng TB centfc data per day and tryng to beneft from proceng capablte provded by the Open Scence Grd (OSG) [2]. Snce mot OSG te are CU-rch but torage-lmted wth no LIGO data avalable, data treamng upport are requred n order to utlze OSG CU reource. In uch a data treamng cenaro, data hould be pulled rather than puhed to the computatonal ytem n the form of tream of tuple, and proceng contnuouly executed over thee tream a f data were alway avalable from local torage. What more, data arrval rate mut be controlled to match the proceng peed to avod wate of computatonal capacty or data overflow. Meanwhle, proceed data have to be cleaned up to ave pace for the ubequently comng data. Such applcaton are novel n that () they are contnuou and long runnng n nature; (2) they requre effcent tranmon of data from/to dtrbuted ource/nk n an end-uer-pullng way; (3) t often not feable to tore all the data n entrety for later proceng becaue of lmted torage and hgh volume of data to be proceed; (4) they need to make effcent ue of hgh performance computng (HC) reource to carry out compute-ntenve tak n a tmely manner. Grd computng [3] pave a new way for uch knd of applcaton, gvng brth to the ocalled Grd Data Streamng applcaton. Such applcaton requre the combnaton of bandwdth uffcency, adequate torage and proceor to guarantee mooth and hgh-effcency proceng, makng them dfferent from other batchorented one. Mot chedulng nfratructure avalable n the fled of grd, uch a Legon [4], Nmrod/G [5] and Condor [6], are largely geared to upport batch-orented applcaton rather than the treamng one. Some cheduler are developed to upport data treamng applcaton, uch a E-Condor, GATES [7], and Streamlne [8], but they jut concern on computatonal reource allocaton, payng lttle attenton to torage and network bandwdth. egau [9] ha the mot mlar motvaton wth the work decrbed n th paper, but t handle data tranfer, job proceng and data cleanup n a workflow manner. EnLIGHTened computng [0] and G-lambda [] project, whch provde co-allocated computng and network reource wth advance reervaton, but they don t concern wth pecfc requrement of Grd data treamng applcaton. In th paper, an ntegrated reource management and chedulng ytem developed from vewpont of the reource, ncludng proceor, torage and bandwdth, to make effcent ue of them and accommodate a many treamng applcaton a poble to acheve hgh throughput. Th reource
management and chedulng ytem tre to allocate proceor, torage and bandwdth ynchronouly to guarantee uch applcaton to execute moothly wth hgh effcency. Baed on Globu toolkt [2], th ytem able to dcover and manage reource geographcally dtrbuted and belongng to dfferent management doman n a tranparent and ecure way. Some key algorthm are propoed, ncludng admon control, applcaton electng, proceor agnng, bandwdth allocaton and torage allocaton. Evaluaton reult how excellent performance and calablty of th ytem. The ret of th paper organzed a followng: Secton 2 decrbe the overall archtecture and mechanm of th reource management and chedulng ytem, whoe core algorthm are elaborated n the next ecton; ome evaluaton reult are ncluded n Secton 4, and the followng ecton conclude th paper. 2. Sytem Archtecture The archtecture of our reource management and chedulng ytem hown n Fgure and t key component nclude but are not lmted to: Clent Tool Th tool an nterface for uer to ubmt ther applcaton wth ther requrement n XML format, ncludng the executable, proceor type and amount, mnmum bandwdth and torage, data ource, jut lke but more than what Condor ubmon doe. It alo capable of montorng the tatu of ubmtted applcaton and that of the reource n the whole grd. Nowaday, t carred out n command lne, and n the future a graphcal uer nterface (GUI) wll be avalable. Management Engne The management engne accept uer ubmon of applcaton and put them nto the queue, whch wll be acceed by the cheduler. It man functon to provde grd upport for treamng applcaton, uch a ecurty, reource dcovery and management. The component of Globu toolkt ued here nclude GAM (Globu eource Allocaton Manager), MDS (Meta-computng Drectory Servce), GSI (Globu Securty Infratructure), GASS (Global Acce to Secondary Storage), NWS (Network Weather Servce), GIS (Grd eource Informaton Servce), GIIS (Grd Index Informaton Servce) and o on. Scheduler Th the core component n the whole archtecture and t key algorthm wll be dcued n detal n Secton 3. It reponble to carry out admon control, applcaton electng, proceor agnment, and bandwdth and torage allocaton. It ntructon wll be executed by the dpatcher. Dpatcher The dpatcher n charge of endng executable wth ther decrpton fle to approprate proceor and nvokng a remote component,.e., applcaton wrapper. Th component wll nteract wth the ervce provded by grd mddleware, uch a GAM. Fgure. Sytem archtecture Applcaton Wrapper Th component wll pare the decrpton fle accordng to the XML chema, ntalze executon of executable, and tart data tranmon to pecfed torage wth allocated bandwdth. Alo, t wll end back the reult through dpatcher. Another functon to montor the uage of torage to determne data tranmon tatu, ee more detal n ubecton 3.5. Job ubmon Clent Tool Wrapper emote Data Source Management Engne Scheduler Dpatcher Clent Tool Wrapper Queue Management Securty eource Dcovery Grd Drectory Servce Other Mddleware Servce roceor Storage Bandwdth Scheduler MDS erver Grd eource eal-tme data upply Bandwdth Dpatcher GAM erver Fgure 2. Overall mechanm Storage Wrapper Uer proce The overvew of the runnng mechanm llutrated n Fgure 2. Bede allocaton of computatonal reource a mot tradtonal reource management and chedulng ytem do, t alo deal wth allocaton of bandwdth and torage to upport real-tme data upply, whch requred by data treamng applcaton. Management and chedulng of proceor, bandwdth and torage are carred out n an ntegrated way rather than ndependently. 3. Key Algorthm Th ecton jut elaborate on the key algorthm a the core of th reource management and chedulng ytem,.e., the cheduler. Note that although proceor agnment, allocaton cheme for torage and bandwdth are decrbed and evaluated
eparately, they are carred out ynchronouly a ntegraton. 3.. Admon control When a new job ubmtted, admon controller would decde to run t ntantly or jut keep t n the watng queue. Th decon made accordng to the uage tatu of reource and the requrement of the job. Each job can allege t mnmum requrement of reource, e.g., t need ome proceor, bandwdth and torage. An XML chema developed for the applcaton to expre ther requrement n the manner mlar to eource Decrpton Language (SL). For each applcaton, t can declare t mnmum requrement of reource lke = [ p b t ] where p tand for the number of proceor t requre, o p = for mple applcaton (.e., tandalone applcaton) and p > for compoed applcaton (uch a a ppelne); b and t tand for the requred mnmum bandwdth and torage repectvely. Th nformaton wll be ncluded n the ubmon fle n XML format. Suppoe the runnng applcaton n the computng pool form a et, denoted a S, and the total amount of proceor, bandwdth and torage be denoted a, B and S repectvely. Some applcaton have ther pecal requrement upon proceor, for example, applcaton compled on X86_64 cannot run on I386 proceor, o not every proceor utable for each applcaton. Suppoe thoe proceor elgble for applcaton form a et, called, and the number of free (not occuped or reerved) proceor n t when come denoted a. In any one of the followng three cae, a new applcaton, n, would jut be kept n the watng queue for there are no enough reource (utable and enough proceor, enough bandwdth and enough torage repectvely) for t. p > b t n n > B b n S > S t n S If an applcaton mnmum requrement can be atfed accordng to the current tatu of computng pool, t wll called a potental elgble applcaton (EA), whch mean that t may be permtted nto computng pool. 3.2. Applcaton electon EA form a queue, wthn whch maybe everal one atfy the admon control polcy. A electng polcy mut be appled to chooe ome from the queue and agn approprate reource for them. Thoe elected one wll be called elgble applcaton (EA). EA have dfferent weght, and the hgher weght mean that they can be elected wth bgger prorte. EA wll be clafed nto everal group accordng to ther weght, and n each group, the electng prncple frt-come-frt-erve (FCFS). The electng wll be heurtc and teratve: the frt comng EA wth the hghet weght wll be elected, and then the next one tll the lat one n t group (f there are) wll be teted n ther arrvng order; then t turn for the group wth econd hghet weght, tll all the group are teted. Notce that the EA wth hgher weght wll not be elected pror to thoe wth lower weght necearly, for whenever a EA accepted, the reource tatu wll change and ome EA wll become nelgble. To ome extent, th algorthm reemble frt-ft (FIFT) wth backfllng mechanm. What more, to avod that ome EA tarve for a long tme, ome reervaton polcy wll be adopted. Some reource wll be labeled a reerved when they are executng other applcaton, and a oon a they are free, they wll be agned to the applcaton whch reerve them. Weght of each applcaton wll ncreae a tme goe by, to avod uch cae where applcaton wth lower weght wll be dle forever. The weght wll be a functon of tme, wth the orgnally et value a ther ntalzaton w t = f w t ( o ) ( wo, 0) wo, f = where w 0 the ntal weght of applcaton and f(t) an non-decreang functon about tme t. A functon n cae w t = f w, t = w + d * f t, T ( o ) o ( ) f ( t T ) = floor( t / T ), where d the ncreae coeffcent and d >0; T the ncreae perod and functon floor return the nearet nteger toward mnu nfnty for t dvded wth T. Then w wll ncreae by d once a perod T. Agnng approprate value for d and T, after ome tme of watng, the applcaton wth lower weght ntally wll be endowed a hgh enough weght to be elected from EA queue. Combnaton of reervaton polcy and ncreang weght over tme wll guarantee each applcaton wll be accepted by the computng pool n approprate tme.
In one word, the electng algorthm tre to make full ue of reource and keep farne among applcaton. 3.3. roceor agnment A oon a EA are elected, t tme to agn reource for them. Applcaton may have ther own tyle,.e., they may be executed more moothly on ome proceor than on other. So t neceary to agn approprate proceor for applcaton, and purely random agnment wll not work. On the other hand, the proceor can be clafed nto everal group accordng to ther charactertc, ther archtecture for ntance. One applcaton wll acheve mlar performance on the proceor of a group, o t not neceary to launch t on each proceor for tral, but a proceor can act a the repreentatve of t peer n the ame group. Matchmakng wll be carred out to fnd canddate proceor for applcaton, and applcaton wll be agned to proceor n the matched group to run a hort perod of tme to get t performance nformaton. The applcaton wth hgher weght wll have hgher prorte to fnd ther matched proceor, and the proceor producng the hghet proceng effcency wll be elected. 3.4. Storage allocaton When new EA arrve, the cheduler reponble for allocatng bandwdth and torage for them, together wth the extng applcaton n the computng pool. The overall prncple for torage allocaton to make full uage of torage to ncreae robutne whle gettng ready for new comng applcaton. If there are only a few applcaton runnng n the pool, the torage allocated for each applcaton can be et to a hgh value. Whle the applcaton ncreae, the allocated torage for each applcaton may be decreaed. There mut be ome margn of torage for potentally comng applcaton. An teratve allocaton algorthm of torage propoed a followng: ntalzaton: uppoe there are n applcaton n the pool, to generate n random number, r ( 0, ), =,2,,n. Calculate each quota, q a followng r q = n r j= 2 If q *TS t, reerve thee number for ntally allocated torage for applcaton ; ele, repeat tep untl all of thee nequaton hold true, where t the mnmal requred torage of applcaton a mentoned n ubecton 3. and TS the total torage avalable for applcaton. 3 dynamc adjutment: perodcally, montorng uage tatu of each allocated torage, and thoe wth j hgh occupaton percentage wll be ncreaed whle other wll be decreaed; 4 when a new EA comng, decreae the amount of the bgget partton of torage; 5 when an applcaton fnhed, t torage wll be dvded and allocated to the mnmal partton; 6 repeat 3, 4 and 5 untl all the applcaton are completed. 3.5. Bandwdth allocaton Bandwdth allocaton play an mportant role n the whole reource allocatng cheme, for approprate bandwdth ndpenable to guarantee data upply for applcaton to make them run contantly. To make a flexble allocaton cheme, o-called utlty functon are ntroduced and genetc algorthm [3] adopted to maxmze ther um. Dfferent from tradtonal bandwdth allocaton, our cheme torage aware,.e., data tranmon may be ntermttent rather than contnuou to avod data overflow, for allocated torage for each applcaton lmted. When the torage full of data, tranmon wll be halted for a whle untl ome data have been proceed and cleaned up o that ome torage releaed for more data. At any moment, the amount of data n torage for each applcaton affected by data upply and clean-up at the ame tme, where the former tend to ncreae the amount whle the latter wll decreae t. The computng pool connected to Internet through whch the data are treamed to the applcaton beng executed on the proceor, and the total nput bandwdth, denoted a I, lmted, whch hared by the data tream. The data tream, called eon, denoted a, form a et S. Each eon wll be agned a bandwdth x, where x X, X =[b,b ] and b >0, B <. b tand for the leat bandwdth requred for eon, whle B the hghet bandwdth avalable for from the correpondng data ource. Seon wll have a utlty U (x ) when t data uppled at a rate x, where U (x ) called utlty functon and aumed to be concave, contnuou, bounded and ncreang n the nterval [b,b ]. Note that t not neceary that all the eon adopt dentcal utlty functon. We try to maxmze the um of the utlte of all the eon, mantanng farne among them. The problem can be decrbed a follow...t. max U ( x ) ( ) S S x I ( 2) x ( 3) X
Due to the uage tatu of torage, there are two poble tate for each at any tme,.e., actve and nactve, whch ndcate a data tranmon on or off. All the actve eon form a et, called S A, and t obvou that th et varyng becaue the tate of eon are changng. We jut allocate bandwdth for actve tranmon, o the contrant (2) may be rewrtten a I 4 S A x An teratve optmzaton algorthm propoed n [4] and t convergence analyzed, but t requred to be aware of the congeton on the path, whch hard to be atfed n the wde Internet. Accordng to our tuaton, we make ome modfcaton upon t a followng. Whle S A [ x + α ku ] ' x f x ρi X ( k+ ) S A x = 5 [ βk x ] f x > X ρi S A Otherwe ( k+ ) x = 0, S 6 Here, x (k) the bandwdth for eon S at the k th tep. {α k } and {β k }are two potve equence. For the ake of convenence, α k and β k are uually ubttuted a a fxed value, denoted a α and β repectvely. [ ] X denote a projecton on the et X and can be calculated a y = mn B, max b y () [ ] ( ) X, ' U the ub gradent of and () = S A A U U x ( 7) U U () x = ' x If we defne U formula (7) can be modfed a U = U x ( x ) = 0, S A () S And ρ the o-called afety coeffcent to avod bandwdth exce, where ρ (0, ),.e., there ome margn from the full ue of total bandwdth for flexblty and robutne. Some heurtc algorthm, uch a genetc algorthm wll be appled to fnd optmal value for them. A popular utlty functon can be expreed a U ( x ) = μ ln( + x ), S where µ may tand for the eon coeffcent. Eentally, bandwdth allocaton here a knd of addtve ncreae multplcatve decreae (AIMD) algorthm whch uually ued n TC congeton avodance, but t ha ome pleaant new character. It torage-aware, for the tranmon wll be topped f allocated torage nearly full, o data overflow wll be avoded whle enough data are uppled; t proceng-aware, for the proceng capacty of proceor wll be nflected n the varyng occupaton of torage, although t may not aware of the prece value, and uch bandwdth allocaton on-demand; of coure, t congeton aware. A mentoned above, the cheduler hould make allocaton cheme n an evolvng way to keep pace wth latet tuaton,.e., when an applcaton ubmtted or fnhed, or reource ncreae or decreae dramatcally, t wll be nvoked to make new cheme to correpond to latet condton. 4. Evaluaton A campu computatonal grd beng etablhed n Tnghua Unverty (Bejng, Chna) whch hold a large amount of upercomputer, peronal computer and other pecal ntrument. Globu toolkt 4.0. beng deployed to provde common grd ervce and a mple Certfcate Authorty ha been etablhed to gn certfcate for hot and uer whch wll be ued to etablh a ecure and tranparent envronment for data treamng applcaton. Th campu grd connected to Internet wth lmted bandwdth, and network fle ytem (NFS) etablhed to whch all the data tream are drected. Applcaton are ubmtted at moment complyng wth negatve exponental dtrbuton law, and ther requrement of reource are alo explctly expreed. Experment are carred out for 0,000 unt of tme and ome reult are obtaned. eource chedulng carred out once per 200 unt of tme to correpond to updated tuaton. 4.. Admon control A the reource n the computng grd are lmted and each treamng applcaton hold t own requrement, t can be nferred that too many applcaton accepted by the computng grd wll lead to low proceng effcency f no admon control carred out, and ome expermental reult verfy t, a hown n Fgure 3. Hgher throughput acheved n the cenaro wth admon control than that n the cenaro wthout admon control. Note that n the latter cenaro, one proceor may have to deal wth more than one
applcaton at the ame tme, whch offend the aumpton made before that one applcaton wll occupy a proceor excluvely. Inadequate data upply for each applcaton and dcount of computatonal capacty due to competton among applcaton on one ngle proceor lead to a lower data proceng effcency a a whole. roceed Data (MB) 8 x 04 6 4 2 0 8 6 4 2 Wth admon control Wthout admon control 0 Fgure 3. Throughput wth/wthout admon control What more, admon control can make applcaton fnhed ooner than wthout, a demontrated n (a) and (b) n Fgure 4, where the red bar wth character tand for the pendng tatu and pnk bar wth character mean for the runnng tatu for each applcaton. The number of completed applcaton are 29 and 25 repectvely. More mportantly, mot of the makepan wthout admon control are longer than ther counterpart, whch advere for the requrement of qualty of ervce. 4.2. Bandwdth allocaton Bandwdth allocated to each runnng applcaton to guarantee ther data upply. arameter n bandwdth allocaton are obtaned wth genetc algorthm and are appled n each perod. Th bandwdth allocaton adaptve to the total avalable bandwdth and requrement of runnng applcaton. app 30 app 29 app 28 Applcaton Statu app 3 app 2 app (a) Wth admon control app 30 app 29 app 28 Applcaton Statu app 3 app 2 app (b) Wthout admon control Fgure 4. Statu of applcaton n teratve bandwdth allocaton To jutfy our bandwdth allocaton algorthm (named teratve allocaton), we compare t wth the even bandwdth allocaton, where the bandwdth allocated to the runnng applcaton equably a hown n Fgure 5. In (a), the total avalable bandwdth relatvely mall whch equal wth that n cae of Fgure 3, and 25 applcaton are fnhed; n (b), the avalable bandwdth relatve bg, o each applcaton can get enough data upply wth the even allocaton cheme, and the reult reemble that n (a) of Fgure 4. app 30 app 29 app 28 Applcaton Statu app 3 app 2 app app 30 app 29 app 28 Applcaton Statu app 3 app 2 app (a) Low bandwdth (b) Hgh bandwdth Fgure 5. Statu of applcaton n even bandwdth allocaton
roceed Data (MB) 8 x 04 6 4 2 0 8 6 4 2 Iteratve bandwdth allocaton/30 Even bandwdth allocaton/40 Even bandwdth allocaton/30 0 Fgure 6. Throughput for bandwdth allocaton cheme Throughput n the cae of teratve allocaton and even allocaton are hown n Fgure 6, where the number at the end of legend tand for the total avalable bandwdth. Then our allocaton cheme jutfed that t can acheve hgh throughput wth relatvely mall avalable bandwdth for t proceng-aware whle the even allocaton cheme not, o n that cae ome applcaton may tarve whle other may be allocated redundant bandwdth. 4.3. Storage uage Data upply n our cheme torage-aware,.e., data upply controlled by the uage of allocated torage, rather than pontaneouly. The prncple here jut enough data ok, not the more data the better. Sometme the data tranmon ntermttent, not alway contnuou. In th way hgh volume of data can be proceed wth jut reaonable torage, a hown n (a) of Fgure 7, where the ued torage jut vare n a lmted cope. If data upply contnuou and avalable torage bg enough, the occuped torage wll be of hgh volume, whch can be oberved n (b) of Fgure 7. Actually, mall torage can acheve hgh throughput n the treamng applcaton wth wellmade data upply and proceng cheme, whch the promnent charactertc of uch cenaro. elatve bg torage not neceary but rather derable, for more data can be tored before proceed to urvve network collape when no more data can be uppled. 4.4. roceor agnment roceor are agned to applcaton, one for a ngle applcaton excluvely. Here, the allocaton reemble job hop problem chedulng, a hown n Fgure 8, where the number n the horzontal bar tand for the correpondng applcaton executed on the proceor n a certan group. Group and 2 deal wth more applcaton whle they hold le proceor than group 3, o the average load of proceor are heaver that thoe of group 3, a demontrated n (a), (b) and (c) repectvely. Becaue ome applcaton may not be able to be executed on the proceor n group 3, they cannot be tranferred to the proceor n group and 2 to make a load balance. Occuped torage (MB) Occuped torage (MB) 80 60 40 20 00 80 60 40 20 0 2.5 2.5 0.5 3 x 04 (a) Storage-aware data upply 0 (b) Non-torage-aware data upply Fgure 7. Storage uage 5. Concluon Data treamng applcaton are of the novel type of grd cenaro for own ther charactertc, uch a requrng of real-tme data upply and ntegrated reource allocaton cheme. Dfferent from extng reource management and chedulng cheme that jut focu on computng reource, the ytem propoed n th paper take computatonal reource, bandwdth and torage nto account multaneouly and make ntegrated management and chedulng cheme, whch are proved to be feable wth excellent performance. Up to now, requrement of qualty of ervce (QoS) for applcaton have not been pad enough attenton, and th derable character wll be the empha of further reearch. Schedulng for ppelned applcaton wll be tuded whch more complex wth requrement of balance among tage and approprate data upply. Ongong work nclude the conderaton of data harng cenaro among multple data
proceng applcaton. Alo ome heurtc chedulng algorthm under development for refned performance optmzaton. proceor 5 proceor 4 proceor 3 proceor 2 proceor proceor 4 proceor 3 proceor 2 proceor proceor 7 proceor 6 proceor 5 proceor 4 proceor 3 proceor 2 proceor 3 5 7 4 9 20 tme 8 2 5 6 4 2 23 27 (a) Group 9 0 000 2000 3000 4000 5000 6000 7000 8000 tme 0 3 7 2 (b) Group 2 8 22 0 000 2000 3000 4000 5000 6000 tme (c) Group 3 Fgure 8. roceor agnment for job Acknowledgement Th work upported by Mntry of Educaton of Chna under the hgher educaton qualty engneerng project Natonal Open Coure Integrated Sytem, and Mntry of Scence and Technology of Chna under the natonal 863 hgh-tech &D program (grant 24 30 26 6 28 25 29 No. 2006AA0Z237, No. 2007AA0Z79 and No. 2008AA0Z8). eference []. E. Deelman, C. Keelman, G. Mehta, L. Mehkat, L. earlman, K. Blackburn,. Ehren, A. Lazzarn,. Wllam, and S. Koranda, GrhyN and LIGO, Buldng a Vrtual Data Grd for Gravtatonal Wave Scentt, roc. th IEEE Int. Symp. on Hgh erformance Dtrbuted Computng, pp. 225-234, 2002. [2].. orde for the Open Scence Grd Conortum, The Open Scence Grd, roc. Computng n Hgh Energy and Nuclear hyc Conf., Interlaken, Swtzerland, 2004. [3]. I. Foter and C. Keelman, The Grd: Blueprnt for a New Computng Infratructure, Morgan Kaufmann, San Francco, 998. [4]. S. J. Chapn, D. Katramato, J. Karpovch and A. S. Grmhaw, The Legon eource Management Sytem, Job Schedulng Stratege for arallel roceng, Sprnger Verlag, pp.62-78, 999. [5].. Buyya, D. Abramon, and J. Gddy, Nmrod/G: An Archtecture for a eource Management and Schedulng Sytem n a Global Computatonal Grd, roc. Hgh erformance Computng ASIA, 2000. [6]. M. Ltzkow, M. Lvny, and M. Mutka, Condor A Hunter of Idle Worktaton, roc. 8 th Int. Conf. on Dtrbuted Computng Sytem, pp. 04-, 988. [7]. L. Chen and G. Agrawal, A Statc eource Allocaton Framework for Grd-baed Streamng Applcaton, Concurrency and Computaton: ractce and Experence, 8:653 666, 2006. [8]. B. Agarwalla, N. Ahmed, D. Hlley, and U. amachandran, Streamlne: a Schedulng Heurtc for Streamng Applcaton on the Grd, roc.3 th Annual Multmeda Computng and Networkng Conf., 2006. [9]. A. amakrhnan, G. Sngh, H. Zhao, E. Deelman,. Sakellarou, K. Vah, K. Blackburn, D. Meyer, and M. Samd, Schedulng Data-ntenve Workflow onto Storage-Contraned Dtrbuted eource, roc. 7 th IEEE Int. Symp. on Cluter Computng and the Grd, o de Janero, Brazl, pp. 40-409, 2007. [0]. L. Battetll, et al., EnLIGHTened Computng: An Archtecture for Co-allocatng Network, Compute, and other Grd eource for Hgh-End Applcaton, roc. HONET2007. []. A. Takefua, et al., G-lambda: coordnaton of a grd cheduler and lambda path ervce over GMLS, roc. Grd2005. [2]. I. Foter and C. Keelman, Globu: A Metacomputng Infratructure Toolkt, Int. J. Supercomputer Applcaton, vol., No. 2, pp.5-28, 997. [3]. J. H. Holland, Adaptaton n Natural and Artfcal Sytem, Unverty of Mchgan re, 975. [4]. K. Kar, S. Sarkar, L. Taula, A Smple ate Control Algorthm for Maxmzng Total Uer Utlty, roc. Infocom 200.