A Static Load Balancing Software for Parallel Applications

Transcription

1 SBAC-PAD 99 11th Sympsium n Cmpute Achitectue and High Pemance Cmputing. Natal. Bazil 237 A Static Lad Balancing Stwae Paallel Applicatins Adian Jaquim de Oliveia Cuz 1, Cláudia Rita de Fanc 2, Lenad Silva Vidal 3 1 Depatment Cmpu te Science, Núcle de Cmputaçã Eletônica, Univesidade Fedeal d Ri de Janei Cx Pstal 2324, Ri de Janei, RJ, CEP { adian@nce.uj.b) 2 Depatment Cmpu te Science, Núcle de Cmputaçã Eletônica Cx Pstal2324, Ri de Janei, RJ, CEP Univesidade Fedeal d Ri de Janei { canc@nce.uj.b} 3 Depatment Cmpu te Science, Núcle de Cmputaçã Eletônica Cx Pstal 2324, Ri de Janei, RJ, CEP Univesidade Fedeal d Ri de Janei {lenad@nce.uj.b} Abstact- This wk descibes the implementatin and benchmaks applied t a 1ad balancing stwae designed t impve pemance paallel applicatins unning n netwks hetegeneus and nn-dedicated wkstatins. A use levei mechanism t gathe wklad inmatin abut each nde and the plicy t teat this inmatin in de t geneate a pecise snapsht the wklad each nde the paallel machine ae descibed thughut this wk. An analysis the main issues cncening wklad evaluatin is pvided, alng with a bie explanatin n the suppt eed by cuent peating systems and ways t vecme thei pblems. Finally, esults and intepetatins cmpaative tests made between BECPVM applicatins and PVM applicatins ae pesented. Keywd-Paallel Pcessing, Lad Balancing. l. NTRODUCTON Running paallel applicatins n cmpute netwks is a cheape altenative t expensive paallel cmputes. These netwks ae shaed by a geat numbe uses. But alng the day, nly sme cmputes ae used emain idle lage peids time. -:he ability t identiy idle cmputes and spawn the pcesses a paallel applicatin n them can give a cnsideable pemance impvement t any paallel applicatin. Futheme, mst uses d nt geneate heavy lads t thei cmputes, s thei cmputes can als hst pcesses the paallel applicatins. A mix ast and slw cmputes ms many netwks. By giving peeence t the aste cmputes, the pemance the paallel applicatins is uthe impved. PVM (Paallel Vitual Machine) [GE 94) is a tl that allws a hetegeneus cllectin wkstatins and supecmputes t unctin as a single paallel cmpute, which is called a vitual machine. Each nde the vitual machine is called a hst and the pcesses ming the paallel applicatin ae called tasks. PVM was chsen as u test bed its high levei ptability, simple message passing pgamming and ability t be installed by any use. 11. LOAD BALANCNG The main gal lad balancing a paallel applicatin is t distibute wk amng the ndes a paallel machine in de t btain bette pemance. t is accmplished by giving a geate ptin wk t the aste and less laded ndes. A balanced paallel applicatin may have the beneits. Ali tasks will have cmpaable executin times since each ne will eceive a ptin wk apppiated t its hst capabilities. As a cnsequence, the executin time a paallel applicatin will nt be delayed by the slwe task. We can divide lad balancing in tw categies. The simple ne is called static lad balancing. This type lad balancing takes place bee new tasks ae spawned. t cnsists n chsing the best hsts t spawn the tasks that will wk in the same hst thughut thei executin. Jacksn and Humphes pesent a extensin t PVM that pvides static lad balancing [JAC 97, HUM 95). This extensin equies that uses un thei wn benchmaks in ade t evaluate hadwae pemance and the esults must be given t the system and it is implemented as a mdiicatin t the PVM suce cde.

2 238 SBAC-PAD99 11th Sympsium n Cmpute Achitectue and High Pemance Cmputing- Natal- Bazil Static lad balancing cannt deal with changes in the wklad a hst, since the best way t deal with them is t mve the task t a less laded hst. This mechanism is called task migatin and the lad balancing that uses it is called dynamic Jad balancing. The mechanism task migatin is able t mve a task t a aste hst, i it becmes available, even when the wklad is stable. Task migatin is a cmplex mechanism, which invlves a geat deal peating systems wk. MPVM (CAS 95] and CONDOR [LT 97] ae stwae tls that pvide pcess migatin and dynamic lad balancing.. BEC BEC [FRA 98] is a nmal PVM applicatin that accepts equests m the PVM applicatins in de t spawn thei tasks. BEC neithe changes PVM implementatin, n equies any special pivileges within PVM. t has its wn set unctins that eplaces sme PVM unctins. BEC pvides tanspaent lad balancing t use applicatins. BECs cuent vesin is limited t static lad balancing, but utue vesins may implement this mechanism. The PVMs unctins substituted ae pvm_spawn and pvm_paent. The me is espnsible the spawning new tasks and the late inms a task the identity its paent task. The unctin pvm_spawn is eplaced by bec_spawn and the unctin pvm_paent is eplaced by bec_paent. The unctin bec_spawn is the pint whee PVMs nn-balanced tasks spawning mechanism is substituted by BECs lad balanced task spawning mechanism. BEC is witten in C and has been pted t the llwing peating systems: Linux, SunOS, Slais e A/X. A. Achitectue BEC has thee types daemns: the maste daemn, the pbe daemn and the ceat daemn. Figue shws BEC achitectue with PVM acting as a cmmunicatin laye between BECs daemns. pvmd3 is the pvm daemn, becd is the maste daemn, becpd is the pbe daemn and beccd is the ceat daemn. B. Maste Daemn The maste daemn is the cente BECs achitectue. Thee is nly ne instance it seving equests m tasks ali ve the vitual machine. The maste daemm1 spawns and cntais the.wk dne by the the BEC daemns, eacts t changes in the vitual machine and answes the equests m client applicatins. At the stat up, the maste daemn cnnects t PVM, gets the vitual machine cniguatin, ties t spawn ne pbe daemn and ne ceat daemn in each hst and equests PVM ntiicatin n any change n the vitual machine. Ate stat up, the maste daemn stats listening t equests m client tasks and cmmunicatins m the the BECs daemns. Hst _j, Fig. BECs achitectue C. Ceat Daemns, \ Hst 2 tn The ceat daemns ae espnsible the spawning new tasks in espnse t a call t bec_spawn. The tasks ae spawned by a call t pvm_spawn speciying the lcal hst as the taget hst. The ceat daemns exist t speed up the spawning new tasks n netwks any size. a ceat daemn des nt eceive any message m the maste daemn within a speciied peid time, it sends a special message t the maste daemn in de t discve i BEC is still unning. thee is n answe, it stps. D. Pbe Daemns The pbe daemns gathe inmatin abut hadwae capabilities and wklad inmatin each hst in the vitual machine. Shuld an attempt t send a message cntaining these statistics ails the pbe daemn sel-destucts. V. PERFORMANCE EVALUATON n de t pem lad balancing, BEC must have data abut the pemance ali hsts and must be able t make cmpaisns amng them. The cuent vesin BEC cmbines ali pemance data m a hst int a single numbe called pemance index.

3 SBAC-PAD 99 1 lth Sympsium n Cmpue Achitectue and High Pemance Cmputing -Natal- Bazil 239 The ange pssible paametes used t evaluate the pemance is vey bad. Hweve, the tw mst imptant BEC ae the wklad, which shws hw much a hst is being used, and the hadwae speed, which pevents BEC m chsing empty but slw hsts. Ali UNX vesins have sme way t pbe the wklad, but the plicy t access this inmatin usually invlves sme special pivileges. The hadwae speed is tickie t btain m the peating systems. Theee, BEC uses time benchmaks t btain bth values. BEC als cnsides sme the paametes: the numbe uses, the numbe teminais and the numbe active tasks spawned by BEC n a hst. These ae cmplementay paametes used t mdiy the pemance index ali the hsts. The set paametes used might be changed in utue vesins BEC. The benchmak, the meaning, and lhe methds t btain each paamete ae explained belw. E. Benchmak This benchmak measues the executin time a set tests. The tests ae a summatin an intege seies, a lating pint summatin and the cpy data blcks in memy. The summatin invlves ali basic aithmetic instuctins. These tests epesent the mst cmmn pcessing tasks caied ut by paallel applicatins. The span the tests is limited by the hsts esuces cnsumed duing the tests and the need t exceed the usual Unix timeslice. Lng duatin tests wuld imply geate esuce cnsumptin, aecting the tasks and degading the hsts pemance. Fast tests wuld nt exceed the pcess timeslice and the pcess wuld nt be peempted, pducing a alse wklad measue. The benchmak is caied ut by the pbe daemns n thei espective hsts and sent t the maste daemn. The benchmak is epeated peidically t give an updated snapsht the hsts pemance. The pbing inteval can be changed by the unctin bec_pbeinteval allwing the use t deine the best pbing inteval his needs. Wklad ndex (W) - it is the measue hw laded with active pcesses a hst is. t is pptinal t the time elapsed duing the executin the benchmak. The time elapsed cntains the time the pcess was eectively unning and the time the pcess was stpped while the the active pcesses executed. t is a dynamic paamete cllected peidically by the pbe daemns. Hadwae lndex (H) - this paamete measues hw ast the cmbinatin hadwae and peating system is, it is pptinal t the pcessa hlding time necessay t accmplish a task. The aste cmbinatins shuld need less time t un, hence they have the Jwe values. t is btained by the sum the use time and the system time the pcess uses t cmplete the benchmak. The hadwae index is a cnstant paamete taken nce at the pbe daemns stat up. Numbe Uses (NUse) - it is the numbe uses Jgged in a hst. Each lgged use cnsumes pat the esuces a hst, s it is a gd plicy t avid hsts being used by many uses. t als pevents the paallel applicatin m distubing the uses. This paamete is cllected peidically by the pbe daemns eading the systems use lgin ecds. Numbe Teminais (NTty) - it is the ttal numbe teminais pened by the uses a hst, this numbe includes teminais used in X sessins. Me pen teminais incease the pssibility wklad peaks. This paamete is cllected by the pbe daemns by peidically eading the systems use lgin ecds. Numbe Tasks Spawned by BEC (NTask) - it is an imptant paamete that pevents BEC m spawning lage numbes tasks thugh successive calls t bec_spawn between pemance data gatheing. t is updated by the maste daemn when task is spawned by a ceat daemn when a spawned task exits. By inceasing this paamete ate a task is spawned, the pemance index gets wse and the hst is less likely t be ehsen again. When the task exits the hst etuns t its nmal pemance index. The updates in the paamete avid the need t cllect pemance data again the hst. F. Pemance Jndex The Pemance ndex a hst is the cmbinatin the esults the benehmak n that hst. The maste daemn eceives the esults m the pbe daemns and uses the llwing equatin t evaluate the pemance index: Pemance lndex = Scale x H x WJ x (1 + wu x NUse + wy x NTty + wt x NTask) Scale is a ange adjustment t the pemance index. The acts wu, wy and wt ae weights t the last thee paametes. BEC pvides the unctin bec_weights t allw the use t ine-tune these weights. The lwe the hadwae index and the wklad index ae, the bette is the pemance the hst, theee the bette hsts ae thse with lwe pemance indexes. V. T ASK SPAWNJNG The pcess task spawning begins with a call t the unctin bec_spawn. This unctin wads the equest t the maste daemn in a message cntaining ali paametes passed by the use, sme PVM envinment vaiables and shell envinment vaiables expted with PVM_EXPORT.

4 240 SBAC-PAD 99 llth Sympsium n Cmpute Achitectue and High Pemance Cmputing- Natal- Bazil The maste daemn selects the best hsts that match the use cnditins and issue a equest the ceat daemns the selected hsts t spawn the desied numbe tasks. Ali the equest data is sted in a intenai list task spawning equests in executin, s the maste daemn can cntinue t listen t the messages while it waits the espnses m the ceat daemns. The algithm t chse hsts will be explained latte. The ceat daemns egiste the equesting tasks PVM envinment in its wn PVM envinment t guaantee that the spawned tasks will inheit the cect envinment, and then call pvm_spawn t lcally spawn the tasks its hst. The esulting tasks identiicatins ae sent back t the maste daemn. When the maste daemn eceives a espnse m a ceat daemn, it updates the cespnding entance n the list task spawning equests in executin and updates the pemance index the hst. Once ali the espnses pending the equest aive, the maste daemn sends the numbe spawned tasks and thei identiicatins t the equesting task. BECs heuistics t distibute tasks amng the hsts uses the idea that it is bette t ceate me tasks n a aste hst then adding a slwe hst t the set hsts. Theee, i hst A has a pemance index that is hal the pemance index a hst B, then hst B shuld get hal less the numbe tasks assigned t hst A. Task spawning thugh BEC is slwe than task spawning via PVM, but the gains in pemance ully cmpensate this disadvantage. V. COMPARA TVE TESTS This sectin pesents the esults tests cmpaing the pemance PVM paallel applicatins unning with withut BECs suppt. The data the ist and uth tests wee extacted m a pgam that geneates pi. These tests wee designed t evaluate wkstatins glbal pemance aviding inluences netwk lad. The secnd test was a pgam that calculates a scala pduct and the thid a pgam that cmputes the aveage and the standad deviatin a lage set pints. These tests, besides evaluate the wkstatins glbal pemance, measue the delays caused by lage exchange data n cmmn netwks. Each test was un 50 times n tw dieent netwks and pduced the aveage executin times shwn hee. The ist netwk cmpised tive Sun4m wkstatins unning Slais and u Sun4c wkstatins unning SunOS. Nine Pentium PCs equal cniguatin unning Linux cmpsed the secnd ne. Bth netwks emained ully peatinal. Duing the tests, the eseaches and students used an aveage u wkstatins. The wklad pbing was epeated in a 120 secnds inteval. The main gal the tests n the SUN netwk was t pve that BEC is able t chse the aste machines amng a pl hsts dieent pemance. Wheeas the main gal the tests n the Linux netwk was t shw that BEC can chse the least aded hsts t impve pemance even when ali machines have the same hadwae chaacteistics. G. Numeic Cmputatin Pi Vaying the Numbe Pints The pupse this test is t evaluate the behaviu a pcessa intensive paallel applicatin. The netwk lad has min inluence in the executin times since nly the bdes the intevals assigned t the slave tasks and the patia! esults ae exchanged between the tasks. Fig. 2 Numeic Cmputatin Pi n a Hetegeneus Netwk Fig. 3 Numeic Cmputatin Pi n a Hmgeneus Netwk As Figue 2 shws, BECs pemance gains n the hetegeneus netwk inceased m 47% t 70% as the numbe pints inceased. Similaly, Figue 3 shws that BECs pemance gains n the hmgeneus netwk

5 SBAC-PAD99 11th Sympsium n Cmpute Achitectue and High Pemance Cmputing- Natal- Bazil 241 inceased m 15% t 44% as the numbe pints inceased. The pemance gains inceased with the numbe pints, because the time taken with data exchanging between tasks is cnstant, s the ttal executin time becmes me dependent n the pemance the hst as the numbe pints incease. BEC chse the aste hsts, thus giving me pemance gains. H. Scala Pduct Vaying the Size the Vects This is a data intensive test, invlving sending lage vects t the slave tasks and eceiving a lating pint numbe as the patia!. esult. The pcessing dne by the slave tasks cmpised tw pducts and a sum pe pint. S this test measues the behaviu BEC with an applicatin equiing lage data exchange and lw pcessing. Mst the executin time is taken by the time t tanse data, s this applicatin is highly netwk dependent. BECs pemance gains n the hetegeneus netwk anged between 20% t 33%, as shwn in Figue 4. ;:;,; E i=,; ;; t!!:. < lo Fig. 4 Scala Pduct Tw Vects n a Hetegeneus Netwk Figue 5 ees t the same tests un n a hmgeneus netwk. BECs pemance gains anged between 6% and 12%, except 10,000 pints whee BEC inceased executin time by 28%. This is explained by the executin time being t small t vecme the exta time t spawn tasks equied by BEC.. Aveage and Standad Deviatin a Se Vaying the Numbe Pints This is anthe data intensive test. t cnsists n sending pats a set pints t the slave tasks and eceiving back the patia) sum t cmpute the aveage. The aveage is sent t the slave tasks t cmpute the patia! standad deviatin. As Figue 6 shws, the pemance gains attained by BEC, n the hetegeneus netwk, wee between 22% and 32%. Figue 7 shws that 0,000 pints BEC inceased the executin time by 28% n the hmgeneus netwk. F highe numbe pints, pemance gains wee between 5% and 13% lo H -- J " t a t-- t-- -j ["lg lo s SO DPVM.9n b.orj lc:jbec 0.66S S M275 J Numbc u( Pulnts (xiooo) Fig. 6 Aveage and Standad Deviatin a Set Pints n a Hetegenus Nctwk =- lo - - R,-- - < 6 t-- t t-- t t j lo SO icpvm MX 4.}9() S} labec.sm 1 2.7RR s.m um l Numb Plnts (xjooo) Fig. 5 Scala Pduct tw Vects n a Hmgeneus Netwk -1 Numlw u Plnts (dooo) Fig. 7 A veage and Standad Deviatin a Set Pints n a Hmgeneus Netwk

6 242 SBAC-PAD99 11th Sympsium n Cmpute Achitectue and High Pemance Cmputing- Natal- Bazil J. Numeic Cmputatin Pi Vaying the Numbe Tasks This test shws hw BECs pemance gains ae aected by inceasing the numbe tasks, which ces BEC t chse slwe hsts. Figue 8 shws the test made n the hetegeneus netwk. BEC btained pemance gains until the numbe tasks inceased t tive, abve this value BEC was ced t use the hsts with slwe hadwae. The pue PVMs applicatins culd gct pemance gains a little uthe because BECs task spawning time. BEC attained the best esult with ive tasks distibuted ve the tive Slais wkstatins S O.S s 6 7 g T 9 DPVM 2.23R Y 1.2R R079 DBEC O.K5) O.KZ04 O.KJ4Z 1.1) ShY Nulll.bc utasks (JdOOO) Fig. 8 Numeic Cmputatin Pi n a Hetegeneus Netwk i l Figue 9 shws the test made n the hmgeneus netwk. Pue PVM applicatins gained pemance as the numbe tasks inceased because the amunt wk pe task deceased. BEC als gained pemance while thee wee ee hsts. As the laded hsts wee added, the pemance was educed because the task spawning time n these hsts is lnge. 2S l. 2.S ; J 2 ) 4 s 6 7 g T 9 J DPVM :BEC OS J a.nn Y NumbCTuks(xiOOO) V. CONCLUS ONS The esults btained shw the va\ue lad balancing t squeeze even me pemance m cmputatinal esuces. These pemance gains can be used t slve even me cmplex pblems. BEC is cmpletely unctinal and has been successully used in undegaduate classes leaning paallel pgamming and eseach activities. The main advantage BEC is the eductin the executin time paallel applicatins. Anthe esult is the ability BEC t avid busy wkstatins, dcceasing the impact paallel applicatins t the uses. Wk esults als shwed that the standad deviatin the cxecutin times was educed by 80% by cmpaisn t pue PVM applicatins. This allws bette pcdictin the cmpletin times when paallel applicatins ae executed epeatedly. The implementatin BEC allws easy pting PVM applicatins, which can be dne by eplacing a small numbe unctin calls, with n changes n pgamming methds. BEC des nt equie special pivileges t be installed and used, allwing any use access t the gains pemance pvided by lad balancing. REFERENCES [CAS 95] CASAS, Jeemy: CLARK, Dan: KONURU, Ravi; OTO, Steve; PROUTY, Rbet: WALPOLE, Jnathan. MPVM: A Migatin Tanspaent Vesin PVM, Technical Rept CSE , Dept. Cmpute Science and Engineeing, Oegn Gaduate lnstitute Science & Technlgy, Febuay [FRA 98] FRANCO, Cláudia Rita de: YDAL, Lenad Silva. BEC Balancead Estátic de Caga paa PVM, B. Se. Pject, Fedeal Uni vesity Ri de Janei, [GE 94] GEST, A: BEGUELN, Adam: DONGARRA, Jack; JANG, Weicheng: MANCHECK, Rbet: SUNDERAM, Yaidy. PVM: Paa/lei Vitual Machine A Uses Guide and Tutial Netwked Paai/e/ Cmputing, MT Pess. [HUM 95] HUMPHRES. Chis W. A Lad Balancing Extensin the PVM Stwae System, M. Se. Thesis, Univesity Alabama, [JAC 97] JACKSON, David J; HUMPHRES, Chis W. A Simple Yet Eective Lad Balancing Extensin t the PVM Stwae System. Paallel Cmputing, vl. 22, lssue 12, Febuay 1997, pp , Nth Hlland [LT 97] LTZKOW, Michael: TANNENBAUM, Tdd; BASNEY, Jim: LYNY, Min. Checkpint and Migatin UNX Pcesses in the Cnd Distibuted Pcessing System, Technical Rept #1346, Univesity Wiscnsin-Madisn Cmpute Sciences, Apil1997. Fig. 9 Numeic Cmputatin Pi n a Hmgeneus Netwk