|
|
- Alison Washington
- 8 years ago
- Views:
Transcription
1 DynamicLoadBalancing ExploitingProcessLifetimeDistributionsfor MorHarchol-Balter and AllenB.Downey tionwhetherpreemptivemigration(migratingactiveprocesses)isnecessary,orwhetherremote WeconsiderpoliciesforCPUloadbalancinginnetworksofworkstations.Weaddresstheques- UniversityofCalifornia,Berkeley execution(migratingprocessesonlyatthetimeofbirth)issucientforloadbalancing.weshow thatresolvingthisisssueisstronglytiedtounderstandingtheprocesslifetimedistribution.our measurementsindicatethatthedistributionoflifetimesforunixprocessispareto(heavy-tailed), withaconsistentfunctionalformoveravarietyofworkloads.weshowhowtoapplythisdistributiontoderiveapreemptivemigrationpolicythatrequiresnohand-tunedparameters.weusea trace-drivensimulationtoshowthatourpreemptivemigrationstrategyisfarmoreeectivethan remoteexecution,evenwhenthememorytransfercostishigh. CategoriesandSubjectDescriptors:unknown[unknown]:unknown unknown bution workloadmodeling,trace-drivensimulation,networkofworkstations,heavy-tailed,paretodistri- GeneralTerms:unknown AdditionalKeyWordsandPhrases:Loadbalancing,loadsharing,migration,remoteexecution, 1.INTRODUCTION Mostsystemsthatperformloadbalancinguseremoteexecution(i.e.non-preemptive migration)basedonaprioriknowledgeofprocessbehavior,oftenintheformofa listofprocessnameseligibleformigration.althoughsomesystemsarecapableof NSFgrantnumberCCR AllenDowneypartiallysupportedbyNSF(DARA)grant MorHarchol-BaltersupportedbyNationalPhysicalScienceConsortium(NPSC)Fellowshipand suchaspreservingautonomy.apreviousanalyticstudybyeageretal.discourages migratingactiveprocesses,mostdosoonlyforreasonsotherthanloadbalancing, DMW AnearlierversionofthispaperappearedintheProceedingsoftheACMSigmetricsConference Permissiontomakedigitalorhardcopiesofpartorallofthisworkforpersonalorclassroomuseis onmeasurementandmodelingofcomputersystems(may23-26,1996)pp grantedwithoutfeeprovidedthatcopiesarenotmadeordistributedforprotordirectcommercial advantageandthatcopiesshowthisnoticeontherstpageorinitialscreenofadisplayalong withthefullcitation.copyrightsforcomponentsofthisworkownedbyothersthanacmmust servers,toredistributetolists,ortouseanycomponentofthisworkinotherworks,requiresprior behonored.abstractingwithcreditispermitted.tocopyotherwise,torepublish,toposton specicpermissionand/orafee.permissionsmayberequestedfrompublicationsdept,acm
2 2implementingpreemptivemigrationforloadbalancing,showingthattheadditional performancebenetofpreemptivemigrationissmallcomparedwiththebenetof M.Harchol-BalterandA.B.Downey simplenon-preemptivemigrationschemes[eageretal.1988].butsimulationstudies,whichcanusemorerealisticworkloaddescriptions,andimplementedsystems haveshowngreaterbenetsforpreemptivemigration[kruegerandlivny1988] 1.1Loadbalancingtaxonomy andatrace-drivensimulationtoinvestigatetheseconictingresults. Onanetworkofsharedprocessors,CPUloadbalancingistheideaofmigratingprocessesacrossthenetworkfromhostswithhighloadstohostswithlowerloads.The andimprovetheutilizationoftheprocessors.analyticmodelsandsimulationstud- motivationforloadbalancingistoreducetheaveragecompletiontimeofprocesses ieshavedemonstratedtheperformancebenetsofloadbalancing,andtheseresults havebeenconrmedinexistingdistributedsystems(seesection1.4). determineswhenmigrationsoccurandwhichprocessesaremigrated.thisisthe questionweaddressinthispaper.theotherhalfofaloadbalancingstrategyis thelocationpolicy theselectionanewhostforthemigratedprocess.previous workhassuggestedthatsimplychoosingthetargethostwiththeshortestcpurun Animportantpartoftheloadbalancingstrategyisthemigrationpolicy,which [Baraketal.1993].Thispaperusesameasureddistributionofprocesslifetimes execution,alsocallednon-preemptivemigration,inwhichsomenewprocessesare relativeunimportanceoflocationpolicy. queueisbothsimpleandeective[zhou1987][kunz1991].ourworkconrmsthe (possiblyautomatically)executedonremotehosts,andpreemptivemigration,in whichrunningprocessesmaybesuspended,movedtoaremotehost,andrestarted. Innon-preemptivemigrationonlynewbornprocessesaremigrated. Processmigrationforpurposesofloadbalancingcomesintwoforms:remote Implicitmigrationpoliciesmayormaynotuseaprioriinformationaboutthe functionofprocesses,howlongtheywillrun,etc. Loadbalancingmaybedoneexplicitly(bytheuser)orimplicitly(bythesystem). mationaboutjoblifetimes.thisinformationisoftenimplementedasaneligibility lifetimeofprocesses,implicitnon-preemptivepoliciesrequiresomeaprioriinfor- listthatspeciesbyprocessnamewhichprocessesareworthmigrating[svensson 1990][Zhouetal.1993]. Sincethecostofremoteexecutionisusuallysignicantrelativetotheaverage sincethisisoftendiculttomaintainandpreemptivestrategiescanperformwell withoutit.thesesystemsuseonlysystem-visibledatalikethecurrentageofeach processoritsmemorysize. Incontrast,mostpreemptivemigrationpoliciesdonotuseaprioriinformation, balancingstrategiesthatassumenoaprioriinformationaboutprocesses. (1)Ispreemptivemigrationworthwhile,giventheadditionalcost(CPUandlatency)associatedwithmigratinganactiveprocess? Thispaperexaminestheperformancebenetsofpreemptive,implicitload Weanswerthefollowingtwoquestions: (2)Whichactiveprocesses,ifany,areworthmigrating?
3 1.2Processmodel Inourmodel,processesusetworesources:CPUandmemory(wedonotconsider ExploitingProcessLifetimeDistributionsforDynamicLoadBalancing 3 scheduling;insection7wediscusstheeectofotherlocalschedulingpolicies.since completion).weassumethatprocessorsimplementtime-sharingwithround-robin I/O).Thus,weuse\age"tomeanCPUage(theCPUtimeaprocesshasused imposedonaprocessis processesmaybedelayedwhileontherunqueueorwhilemigrating,theslowdown thusfar)and\lifetime"tomeancpulifetime(thetotalcputimefromstartto wherewalltimeisthetotaltimeaprocessspendsrunning,waitinginqueue,or migrating. Slowdownofprocessp=wall-time(p) 1.3Outline CPU-time(p) Theeectivenessofloadbalancing eitherbyremoteexecutionorpreemptive balancingstrategies. trace-drivensimulationtoevaluatetheimpactofthisworkloadonproposedload observationsabouttheworkloadonanetworkofunixworkstations,andusesa migration dependsstronglyonthenatureoftheworkload,includingthedistributionofprocesslifetimesandthearrivalprocess.thispaperpresentsempiricabutionispredictablewithgoodnessoftgreaterthan99%andconsistentacross avarietyofmachinesandworkloads.asaruleofthumb,theprobabilitythata workloadsinanacademicenvironment,includinginstructionalmachines,research machines,andmachinesusedforsystemadministration.wendthatthedistri- Section2presentsastudyofthedistributionofprocesslifetimesforavarietyof processwithcpuageofonesecondusesmorethantsecondsoftotalcputime is1=t(seefigure1). 1986],butthispriorworkhasbeenincorporatedinfewsubsequentanalyticand simulatorstudiesofloadbalancing.thisomissionisunfortunate,sincetheresults ofthesearesensitivetothelifetimemodel(seesection2.2). OurmeasurementsareconsistentwiththoseofLelandandOtt[LelandandOtt loadbalancing: Theysuggestthatitispreferabletomigrateolderprocessesbecausetheseprocesseshaveahigherprobabilityoflivinglongenough(eventuallyusingenough CPU)toamortizetheirmigrationcost. eligibilityofaprocessformigrationasafunctionofitscurrentage,migration Ourobservationsoflifetimedistributionshavethefollowingconsequencesfor Afunctionalmodelofthedistributionprovidesananalytictoolforderivingthe withoutmigration.accordingtothiscriterion,aprocessiseligibleformigration slowdownimposedonamigrantprocessislowerinexpectationthanitwouldbe cost,andtheloadsatitssourceandtargethost. onlyifits InSection3wederiveamigrationeligibilitycriterionthatguaranteesthatthe CPUage>1 n?mmigrationcost
4 4wheren(respectivelym)isthenumberofprocessesatthesource(target)host. InSection5weuseatrace-drivensimulationtocompareourpreemptivemigrationpolicywithanon-preemptivepolicybasedonname-lists.Thesimulatoruses starttimesanddurationsfromtracesofarealsystem,andmigrationcostschosen fromameasureddistribution. M.Harchol-BalterandA.B.Downey evenwithsurprisinglylargemigrationcosts,despiteseveralconservativeassumptionsthatgivenon-preemptivemigrationanunfairadvantage. migrationstrategiesinmoredetail.wendthatpreemptivemigrationreduces migration.wealsoproposeseveralalternativemetricsintendedtomeasureusers' themeandelay(queueingandmigration)by35{50%,comparedtonon-preemptive Nextwechooseaspecicmodelofpreemptiveandnon-preemptivemigration migrationcostontherelativeperformanceofthetwostrategies.notsurprisingly, Nevertheless,preemptivemigrationperformsbetterthannon-preemptivemigration wendthatasthecostofpreemptivemigrationincreases,itbecomeslesseective. Weusethesimulatortorunthreeexperiments:rstweevaluatetheeectof costsbasedonrealsystems(seesection4),andusethismodeltocomparethetwo perceptionofsystemperformance.bythesemetrics,theadditionalbenetsof preemptivemigrationcomparedtonon-preemptivemigrationappearevenmore moreeectivethanevenawell-tunednon-preemptivemigrationpolicy.insection5.5weusethesimulatortocompareourpreemptivemigrationstrategywith previouslyproposedpreemptivestrategies. InSection5.4wediscussindetailwhyasimplepreemptivemigrationpolicyis signicant. 1.4Relatedwork insection7andconclusionsinsection8. WenishwithacriticismofourmodelinSection6,adiscussionoffuturework jobs,fewhaveimplementedimplicitloadbalancingpolicies.mostsystemsonly allowforexplicitloadbalancing.thatis,thereisnoloadbalancingpolicy;theuser decideswhichprocessestomigrate,andwhen.examplesincludeaccent[zayas 1987],Locus[Thiel1991],Utopia[Zhouetal.1993],DEMOS/MP[Powelland 1.4.1Systems.Althoughseveralsystemshavethemechanismtomigrateactive preemptivepolicies(activeprocessesareonlymigratedforpurposesotherthanload [DePaoliandGoscinski1995],andMIST[Casasetal.1995]. Miller1983],V[Theimeretal.1985],NEST[AgrawalandEzzet1987],RHODOS [Tanenbaumetal.1990],Charlotte[ArtsyandFinkel1989],Sprite[Douglisand balancing,suchaspreservingworkstationautonomy).examplesincludeamoeba Ousterhout1991],Condor[Litzkowetal.1988],andMach[Milojicic1993].In Afewsystemshaveimplicitloadbalancingpolicies,howevertheyarestrictlynon- aboutprocesses;e.g.,explicitknowledgeabouttheruntimesofprocessesoruserprovidedlistsofmigratableprocesses[agrawalandezzet1987][litzkowandlivny 1990][DouglisandOusterhout1991][Zhouetal.1993]. general,non-preemptiveloadbalancingpoliciesdependonaprioriinformation thattheirschemeiseectiveandrobust. loadbalancingismosix[baraketal.1993].ourresultssupportthemosixclaim Oneexistingsystemthathasimplementedimplementedautomatedpreemptive
5 ExploitingProcessLifetimeDistributionsforDynamicLoadBalancing Studies.Althoughfewsystemsincorporatemigrationpolicies,therehave beenmanysimulationandanalyticalstudiesofvariousmigrationpolicies.most ofthesestudieshavefocusedonloadbalancingbyremoteexecution[livnyand Melman1982][WangandMorris1985][CasavantandKuhl1987][Zhou1987][Pulidasetal.1988][Kunz1991][BonomiandKumar1990][EvansandButt1993][Lin andraghavendra1993][mirchandaneyetal.1990][zhangetal.1995][zhouand Ferrari1987][HacandJin1990][Eageretal.1986]. Onlyafewstudiesaddresspreemptivemigrationpolicies[LelandandOtt1986] [KruegerandLivny1988].TheLelandandOttmigrationpolicyisalsoagebased, butdoesn'ttakemigrationcostintoaccount. Eageret.al.,[Eageretal.1988],concludethattheadditionalperformancebenet ofpreemptivemigrationistoosmallcomparedwiththebenetofnon-preemptive migrationtomakepreemptivemigrationworthwhile.thisresulthasbeenwidely cited,andinseveralcasesusedtojustifythedecisionnottoimplementpreemptive migration,asintheutopiasystem,[zhouetal.1993].ourworkdiersfrom [Eageretal.1988]inbothsystemmodelandworkloaddescription.[Eageretal. 1988]modelaserverfarminwhichincomingjobshavenoanityforaparticular processor,andthusthecostofinitialplacement(remoteexecution)isfree.thisis dierentfromourmodel,anetworkofworkstations,inwhichincomingjobsarrive ataparticularhostandthecostofmovingthemaway,evenbyremoteexecution, issignicantcomparedtomostprocesslifetimes.also,[eageretal.1988]usea degeneratehyperexponentialdistributionoflifetimesthatincludesfewjobswith non-zerolifetimes.whenthecoecientofvariationofthisdistributionmatches thedistributionsweobserved,fewerthan4%ofthesimulatedprocesseshavenonzerolifetimes.withsofewjobs(andbalancedinitialplacement)thereisseldom anyloadimbalanceinthesystem,andthuslittlebenetforpreemptivemigration. Furthermore,the[Eageretal.1988]processlifetimedistributionisexponentialfor jobswithnon-zerolifetimes,theconsequencesofwhichwediscussinsection2.2. Foramoredetailedexplanationofthisdistributionanditseectonthestudy,see [DowneyandHarchol-Balter1995]. KruegerandLivnyinvestigatethebenetsofsupplementingnon-preemptivemigrationwithpreemptivemigrationandndthatpreemptivemigrationisworthwhile.Theyuseahyperexponentiallifetimedistributionthatapproximatesclosely thedistributionweobserved;asaresult,theirndingsarelargelyinaccordwith ours.onedierencebetweentheirworkandoursisthattheyusedasynthetic workloadwithpoissonarrivals.theworkloadweobserved,andusedinourtracedrivensimulations,exhibitsserialcorrelation;i.e.itismoreburstythanapoisson process.anotherdierenceisthattheirmigrationpolicyrequiresseveralhandtunedparameters.insection3.1weshowhowtousethedistributionoflifetimes toeliminatetheseparameters. Likeus,BryantandFinkeldiscussthedistributionofprocesslifetimesandits eectonpreemptivemigrationpolicy,buttheirhypotheticaldistributionsarenot basedonsystemmeasurements[bryantandfinkel1981].alsolikeus,theychoose migrantprocessesonthebasisofexpectedslowdownonthesourceandtargethosts, buttheirestimationofthoseslowdownsisverydierentfromours.inparticular, theyusethedistributionofprocesslifetimestopredictahost'sfutureloadasa functionofitscurrentloadandtheagesoftheprocessesrunningthere.wehave
6 6examinedthisissueandfound(1)thatthismodelfailstopredictfutureloads becauseitignoresfuturearrivals,and(2)thatcurrentloadisthebestpredictorof M.Harchol-BalterandA.B.Downey futureload(seesection3.1).thus,inourestimatesofslowdown,weassumethat thefutureloadonahostisequaltothecurrentload. 2.DISTRIBUTIONOFLIFETIMES Thegeneralshapeofthedistributionofprocesslifetimesinanacademicenvironmenthasbeenknownforalongtime[Rosin1965]:therearemanyshortjobsand afewlongjobs,andthevarianceofthedistributionisgreaterthanthatofan theircurrentage[cabrera1986].thatsameyear,lelandandottproposeda functionalformfortheprocesslifetimedistribution,basedonmeasurementsofthe lifetimesof9.5millionunixprocessesbetween1984and1985[lelandandott 1986].TheyconcludethatprocesslifetimeshaveaUBNE(used-better-than-new- In1986,CabrerameasuredUNIXprocessesandfoundthatover40%doubled exponentialdistribution. aprocess,thegreateritsexpectedremainingcpulifetime.specically,theynd in-expectation)typeofdistribution.thatis,thegreaterthecurrentcpuageof \longprocesseshaveexponentialservicetimes."manysubsequentstudiesassume anexponentiallifetimedistribution. thatfort>3seconds,theprobabilityofaprocess'slifetimeexceedingtseconds isrtk,where?1:25<k<?1:05andrnormalizesthedistribution. Incontrast,Rommel[Rommel1991]claimsthathismeasurementsshowthat functionalformproposedbylelandandotttstheobserveddistributionswell,for processeswithlifetimesgreaterthan1second.thisfunctionalformisconsistent policies,weperformedanindependentstudyofthisdistribution,andfoundthatthe acrossavarietyofmachinesandworkloads,andalthoughtheparameter,k,varies Becauseoftheimportanceoftheprocesslifetimedistributionforloadbalancing from-1.3to-.8,itisgenerallynear-1.thus,asaruleofthumb, Theprobabilitythataprocesswithage1secondusesatleastTsecondsoftotal TheprobabilitythataprocesswithageTsecondsusesatleastanadditional TsecondsofCPUtimeisabout1=2.Thus,themedianremaininglifetimeofa processisequaltoitscurrentage. CPUtimeisabout1=T. tionpolicies. served.section2.2discussesothermodelsforthedistributionoflifetimes,and arguesthattheparticularshapeofthisdistributioniscriticalforevaluatingmigra- Section2.1describesourmeasurementsandthedistributionoflifetimesweob- 2.1Lifetimedistributionwhenlifetime>1s TodeterminethedistributionoflifetimesforUNIXprocesses,wemeasuredthe lifetimesofoveronemillionprocesses,generatedfromavarietyofacademicworkloads,includinginstructionalmachines,researchmachines,andmachinesusedfor systemadministration.weobtainedourdatausingtheunixcommandlastcomm, showsonlyprocesseswhoselifetimesexceedonesecond.thedotted(heavy)line whichoutputsthecputimeusedbyeachcompletedprocess. Figure1showsthedistributionoflifetimesfromoneofthemachines.Theplot
7 ExploitingProcessLifetimeDistributionsforDynamicLoadBalancing Distribution of process lifetimes (fraction of processes with duration > T) Duration (T secs.) 1 Distribution of process lifetimes (log plot) (fraction of processes with duration > T) 1/2 1/4 1/8 1/16 onmachinepomid-semester.thedotted(thicker)lineshowsthemeasureddistribution;thesolid Fig.1.(a)Distributionoflifetimesforprocesseswithlifetimesgreaterthan1second,observed 1/32 (thinner)lineshowstheleastsquarescurvet.(b)thesamedistributiononalog-logscale.the straightlineinlog-logspaceindicatesthatthedistributioncanbemodeledbytk,wherekisthe 1/64 slopeoftheline Duration (T secs.)
8 8showsthemeasureddistribution;thesolid(thinner)lineshowstheleast-squarest tothedatausingtheproposedfunctionalformprflifetime>tg=tk. M.Harchol-BalterandA.B.Downey well.incontrast,figure2showsthatitisimpossibletondanexponentialcurve thattsthedistributionoflifetimesweobserved. Byvisualinspection,itisclearthattheproposedmodeltstheobserveddata oftheformtk,withkvaryingfrom?1:3to?0:8fordierentmachines.table1 showsthevalueoftheparameterforeachmachinewestudied,estimatedbyan iterativelyweightedleast-squarest(withnointercept,inaccordancewiththe functionalmodel).wecalculatedtheseestimateswiththeblsscommandrobust Forallthemachineswestudied,thedistributionofprocesslifetimestsacurve shownhereindicatethatthettedcurveaccountsforgreaterthan99%ofthe ofcertainty).ther2valueindicatesthegoodnessoftofthemodel thevalues thatparameter(alloftheseparametersarestatisticallysignicantatahighdegree [AbrahamsandRizzardi1988]. variationoftheobservedvalues.thus,thegoodnessoftofthesemodelsishigh Thestandarderrorassociatedwitheachestimategivesacondenceintervalfor andconditionaldistributionfunctionforprocesslifetimes.thesecondcolumn (foranexplanationofr2values,see[larsenandmarx1986]). showsthesefunctionswhenk=?1,whichwewillassumeforouranalysisin Section3. Table2showsthecumulativedistributionfunction,probabilitydensityfunction, thatitsmoments(mean,variance,etc.)areinnite.ofcourse,sincetheobserved distributions,though,becausecalculatedmomentstendtobedominatedbyafew distributionshavenitesamplesize,theyhavenitemean(0.4seconds)andcoecientofvariation(5{7).onemustbecautiouswhensummarizinglong-tailed Thefunctionalformweareproposing(thetteddistribution)hastheproperty likethemedian,ortheestimatedparameterk)tosummarizedistributions,rather outliers.inouranalysesweusemorerobustsummarystatistics(orderstatistics thanmoments. second,wedidnotndaconsistentfunctionalform;however,forthemachines westudiedtheseprocesseshadanevenlowerhazardratethanthoseofage>1 second.thatis,theprobabilitythataprocessofaget<1secondslivesanothert secondsisalwaysgreaterthan1=2.thusforjobswithlifetimeslessthan1second, 2.1.1Processwithlifetime<1second.Forprocesseswithlifetimeslessthan1 Manypriorstudiesofprocessmigrationassumeanexponentialdistributionofprocesslifetimes,bothinanalyticalpapers[LinandRaghavendra1993][Mirchandaney andmelman1982][zhangetal.1995][chowdhury1990].thereasonsforthisas- 1991][Pulidasetal.1988][WangandMorris1985][EvansandButt1993][Livny willnotaecttheresultsofloadbalancingstudies. lifetimedistributionisinfactnotexponential,assuminganexponentialdistribution sumptioninclude:(1)analytictractability,and(2)thebeliefthateveniftheactual 2.2Whythedistributioniscritical themedianremaininglifetimeisgreaterthanthecurrentage. etal.1990][eageretal.1986][ahmadetal.1991]andinsimulationstudies[kunz Regardingtherstpoint,althoughthefunctionalformthatweandLelandand
9 ExploitingProcessLifetimeDistributionsforDynamicLoadBalancing 9 NameNumberNumberEstimated of Host po Procs>1sec of 4107 Distrib. T?0:97 Error Std R2 po2 po3 cory pors T?1:22 T?1:27 T?0: bugs T?0:94 T?0: Table1.Theestimatedlifetimedistributionforeachmachinemeasured,andtheassociated faith T?0: goodnessoftstatistics.descriptionofmachines:poisaheavily-useddecserver5000/240, usedprimarilyforundergraduatecoursework.po1,po2,andpo3refertomeasurementsmade onpomid-semester,late-semester,andend-semester.coryisaheavily-usedmachine,usedfor courseworkandresearch.porscheisalessfrequently-usedmachine,usedprimarilyforresearch onscienticcomputing.bugsisaheavily-usedmachine,usedprimarilyformultimediaresearch. Faithisaninfrequently-usedmachine,usedbothforvideoapplicationsandsystemadministration. 1 Observed distribution and two curve fits (fraction of processes with duration > T) 1/2 1/4 k T fit (one parameter) 1/8 exponential fit (two parameters) 1/16 Fig.2.Inlog-logspace,thisplotshowsthedistributionoflifetimesforthe13000processesfrom ourtraceswithlifetimesgreaterthansecond,andtwoattemptstotacurvetothisdata.one 1/32 ofthetsisbasedonthemodelproposedinthispaper,tk.theothertisanexponential curve,ce?t.althoughtheexponentialcurveisgiventhebenetofanextrafreeparameter, 1/64 iteratively-weightedleastsquares. itfailstomodeltheobserveddata.theproposedmodeltswell.bothtswereperformedby Duration (T secs.)
10 10 M.Harchol-BalterandA.B.Downey Distributionoflifetimesforprocesses>1secWhenk=?1 PrfT<L<T+dTsecg=?kTk?1dT PrfL>bsecjage=ag=?bak PrfL>Tsecg=Tk 1=T2dT distributionfunctionfortheprocesslifetimel.thesecondcolumnshowsthefunctionalformof Table2.Thecumulativedistributionfunction,probabilitydensityfunction,andconditional eachforthetypicalvaluek=1. a=b tion,itneverthelesslendsitselftosomeformsofanalysis,asweshowinsection3.1. Ottproposecannotbeusedinqueueingmodelsaseasilyasanexponentialdistribu- migrationpolicydependsonhowtheexpectedremaininglifetimeofajobvaries isimportanttomodelthisdistributionaccurately.specically,thechoiceofa distributionaectstheperformanceofmigrationpolicies,andthereforethatit withage.inourobservationswefoundadistributionwiththeubneproperty Regardingthesecondpoint,wearguethattheparticularshapeofthelifetime theexpectedremaininglifetimeofajobincreaseslinearlywithage.asaresult,we choseamigrationpolicythatmigratesonlyoldjobs. processanditsremaininglifetime.forexample,auniformdistributionhasthe NBUEproperty theexpectedremaininglifetimedecreaseslinearlywithage. choosetomigrateonlyyoungprocesses.inthiscase,weexpectnon-preemptive Thusifthedistributionoflifetimeswereuniform,themigrationpolicyshould Butdierentdistributionsyieldindierentrelationshipsbetweentheageofa migrationtoperformbetterthanpreemptivemigration. withthelowestmigrationcost,regardlessofage. lifetimeofajobisindependentofitsage.inthiscase,sinceallprocesseshavethe sameexpectedlifetimes,themigrationpolicymightchoosetomigratetheprocess Asanotherexample,theexponentialdistributionismemoryless theremaining distribution(auniformdistributioninlog-space)havearemaininglifetimethat increasesuptoapointandthenbeginstodecrease.inthiscase,thebestmigration policymightbetomigratejobsthatareoldenough,butnottooold. Asanalexample,processeswhoselifetimesarechosenfromauniformlog dierentmigrationpolicies.inordertoevaluateaproposedpolicy,itiscritical tochooseadistributionmodelwiththeappropriaterelationshipbetweenexpected remaininglifetimeandage. Thusdierentdistributions,evenwiththesamemeanandvariance,canleadto oflifetimes.thesedistributionsmayormaynothavetherightbehavior,depending onhowaccuratelytheytobserveddistributions.[kruegerandlivny1988]use athree-stagehyperexponentialwithparametersestimatedtotobservedvalues. ThisdistributionhastheappropriateUBNEproperty.Butthetwo-stagehyperexponentialdistribution[Eageretal.1988]useismemoryless;theremaininglifetime ofajobisindependentofitsage(forjobswithnonzerolifetimes).accordingtothis distribution,migrationpolicyisirrelevant;allprocessesareequallygoodcandidates Somestudieshaveusedhyperexponentialdistributionstomodelthedistribution formigration.thisresultisclearlyinconictwithourobservations.
11 preemptivemigration.theheavytailofourmeasuredlifetimedistributionimplies Assumingthewronglifetimedistributionmayalsounderestimatethebenetsof ExploitingProcessLifetimeDistributionsforDynamicLoadBalancing 11 abilitytoidentifythosefewhogs.inalifetimedistributionwithoutsuchaheavy tail,preemptivemigrationmightnotbeaseective. Aswe'lldiscussinSection5.4,partofthepowerofpreemptivemigrationisits thatatinyfractionofthejobsrequiremorecputhanalltheotherjobscombined. 3.MIGRATIONPOLICY Amigrationpolicyisbasedontwodecisions:whentomigrateprocessesandwhich processestomigrate.therstquestionconcernshowoftenoratwhattimesthe Thefocusofthispaperisthesecondquestion,alsoknownastheselectionpolicy: systemchecksforeligiblemigrants.weaddressthisissuebrieyinsection Giventhattheloadatahostistoohigh,howdowechoosewhichprocess tive,migrationtimehasalargeimpactonresponsetime.aprocesswouldchoose lifetimes.themotivationforthisheuristicistwofold.fromtheprocess'sperspec- Ourheuristicistomigrateprocessesthatareexpectedtohavelongremaining tomigrate? tomigrateonlyifthemigrationoverheadcouldbeamortizedoveralongerlifetime. tion),becausetheseprocesseshavenoallocatedmemoryandthustheirmigration Fromtheperspectiveofthesourcehost,ittakesasignicantamountofworkto costislow.theideaofmigratingnewbornprocessesmightalsostemfromthe thatarelikelytobemoreexpensivetorunthantomigrate. packageaprocessformigration.thehostwouldonlychoosetomigrateprocesses processeshaveequalexpectedremaininglifetimesregardlessoftheirage,soone fallacythatprocesslifetimeshaveanexponentialdistribution,implyingthatall Manyexistingmigrationpoliciesonlymigratenewbornprocesses(nopreemp- shouldmigratethecheapestprocesses.theproblemwithonlymigratingnewborn processesisthat,accordingtotheprocesslifetimedistribution,newbornprocesses ourmeasurementsshowthatover70%ofprocesseshavelifetimessmallerthanthe areunlikelytolivelongenoughtojustifythecostofremoteexecution.infact, smallestnon-preemptivemigrationcost(seetable3). Wehavefound,though,thattheabilityofthesystemtopredictprocesslifetimes edgeaboutprocessesandcanselectivelymigrateprocesseslikelytobecpuhogs. bynameislimited(section5.4). Thusanewbornmigrationpolicyisonlyjustiedifthesystemhaspriorknowl- processes. processtorunlongerthanayoungprocess;thus,itispreferabletomigrateold jorityofprocessesareshort,theremightnotbeenougholdprocessestohavea Therearetwopotentialproblemswiththisapproach.First,sincethevastma- Canwedobetter?Thedistributionoflifetimesimpliesthatweexpectandold signicantloadbalancingeect.infact,althoughtherearefewlong-livedprocesses,theyaccountforalargepartofthetotalcpuload.accordingtoour measurements,typicallyfewerthan4%ofprocesseslivelongerthan2seconds,yet thelongtailoftheprocesslifetimedistribution.furthermore,wewillseethatthe abilitytomigrateevenafewlargejobscanhavealargeeectonsystemperfor- theseprocessesmakeupmorethan60%ofthetotalcpuload.thisisdueto
12 12 mance,sinceasinglelongprocessonabusyhostimposesslowdownsonmanyshort processes. M.Harchol-BalterandA.B.Downey migratingprocesseswithlongerexpectedlives. activeprocessismuchgreaterthanthecostofremoteexecution.ifpreemptive migrationisdonecarelessly,thisadditionalcostmightoverwhelmthebenetof Asecondproblemwithmigratingoldprocessesisthatthemigrationcostforan 3.1OurMigrationPolicy Forthisreason,weproposeastrategythatguaranteesthateverymigrationimprovestheexpectedperformanceofthemigrantprocessandtheotherprocesses atthesourcehost.thisstrategymigratesaprocessonlyifitimprovestheexpectedslowdownoftheprocess,whereslowdownisdenedasinsection1.2.of course,processesonthetargethostareslowedbyanarrivingmigrant,butona moderately-loadedsystemtherearealmostalwaysidlehosts;thusthenumberof processesatthetargethostisusuallyzero.inanycase,thenumberofprocesses atthetargetisalwayslessthanthenumberatthesource. migrationisdone.ifmigrationcostsarehigh,fewprocesseswillbeeligiblefor migration;intheextremetherewillbenomigrationatall.butinnocaseisthe performanceofthesystemworse(inexpectation)thantheperformancewithout migration. Ifthereisnoprocessonthehostthatsatisestheabovemigrationcriterion,no imposedonamigrantprocess,andusethisresulttoderiveaminimumagefor migrationbasedonthecostofmigration.denotingtheageofthemigrantprocess thenumberofprocessesatthesourcehostbyn;andthenumberofprocessesat bya;thecostofmigrationbyc;the(eventualtotal)lifetimeofthemigrantbyl, Usingthedistributionofprocesslifetimes,wecalculatetheexpectedslowdown thetargethost(includingthemigrant)bym,wehave: Efslowdownofmigrantg t=aprlifetimeof migrantistslowdowngiven =Z1 t=aprftl<t+dtjlagna+c+m(t?a) lifetimeistdt =12ca+m+n t=aat2na+c+m(t?a) t t formigrationonlyifitsexpectedslowdownaftermigrationislessthann(whichis theslowdownitexpectsintheabsenceofmigration). Iftherearenprocessesataheavilyloadedhost,thenaprocessshouldbeeligible Thus,werequire12(ca+m+n)<n,whichimplies Wecanextendthisanalysistothecaseofheterogeneousprocessorspeedsby Minimummigrationage=Migrationcost n?m
13 applyingascalefactortonorm. Thisanalysisassumesthatcurrentloadpredictsfutureload;thatis,thattheload ExploitingProcessLifetimeDistributionsforDynamicLoadBalancing 13 atthesourceandtargethostswillbeconstantduringthemigration.inanattempt calculatingapredictionofsurvivorsandfuturearrivalsbasedonthedistribution toevaluatethisassumption,andpossiblyimproveit,weconsideredanumberof oftime),(2)summingtheagesoftheprocessesrunningonthehost,anda(3) predictor,andthatusingseveralpredictivevariablesincombinationdidnotgreatly alternativeloadpredictors,including(1)takingaloadaverage(overaninterval modelproposedhere.wefoundthatcurrent(instantaneous)loadisthebestsingle improvetheaccuracyofprediction.theseresultsareinaccordwithzhou[zhou 3.2PriorPreemptivePolicies Onlyafewpreemptivestrategieshavebeenimplementedinrealsystemsorproposed 1987]andKunz[Kunz1991]. inpriorstudies.thethreethatwehavefoundare,likeours,basedontheprinciple thataprocessshouldbemigratedifitisoldenough. freeparameterwhosevalueischosenwithoutexplanation,andthatwouldneedto bere-tunedforadierentsystemoranotherworkload. UnderLelandandOtt'spolicy,aprocesspiseligibleformigrationif Inmanycases,thedenitionofoldenoughdependsona\voodoo"constant1:a andlivny'spolicy,likeours,takesthejob'smigrationcostintoaccount.aprocess piseligibleformigrationif wherekisafreeparametercalledmincrit[lelandandott1986].krueger age(p)>agesofkyoungerjobsathost buttheydonotexplainhowtheychosethevalue0:1[kruegerandlivny1988]. TheMOSIXpolicyissimilar[Baraketal.1993];aprocessiseligibleformigration if age(p)>0:1migrationcost(p) ofamigrantprocessisnevermorethan2,sinceintheworstcasethemigrant completesimmediatelyuponarrivalatthetarget. Thechoiceoftheconstant(1:0)intheMOSIXpolicyensuresthattheslowdown age(p)>1:0migrationcost(p): theslowdownthatwouldbeimposedatthesourcehostintheabsenceofmigration(presumablythereismorethanoneprocessthere,orthesystemwouldnot beattemptingtomigrateprocessesaway).second,itisbasedontheworst-case Despitethisjustication,thechoiceofthemaximumslowdown(2)isarbitrary. WeexpecttheMOSIXpolicytobetoorestrictive,fortworeasons.First,itignores 1ThistermwascoinedbyProfessorJohnOusterhoutatU.C.Berkeley. onload. bestchoiceforthisparameter,forourworkload,isusuallynear0.4,butitdepends slowdownratherthantheexpectedslowdown.insection5.5,weshowthatthe
14 14 Migrationcosthassuchalargeeectontheperformanceofpreemptiveloadbalancing;thissectionpresentsthemodelofmigrationcostsweuseinoursimulation Wemodelthecostofmigratinganactiveprocessasthesumofaxedmigration 4.MODELOFMIGRATIONCOSTS M.Harchol-BalterandA.B.Downey proportionaltotheamountoftheprocess'smemorythatmustbetransferred. studies. costformigratingtheprocess'ssystemstateandamemorytransfercostthatis remotehost,logginginorotherwiseauthenticatingtheprocess,andcreatinganew Thecostofremoteexecutionincludessendingthecommandandargumentstothe shellandenvironmentontheremotehost. Wemodelremoteexecutioncostasaxedcost;itisthesameforallprocesses. r:thecostofremoteexecution,inseconds f:thexedcostofpreemptivemigration,inseconds b:thememorytransferbandwidth,inmbpersecond Throughoutthispaper,weusethefollowingnotation: andthus: m:thememorysizeofmigrantprocesses,inmb wherethequotientm=bisthememorytransfercost. costofpreemptivemigration=f+m=b costofremoteexecution=r [DouglisandOusterhout1991]haveanexcellentdiscussionofthisissue,andwe migrationdependsonpropertiesofthedistributedsystem.douglisandousterhout 4.1Memorytransfercosts borrowfromthemhere. Theamountofaprocess'smemorythatmustbetransferredduringpreemptive systemlikesprite,whichintegratesvirtualmemorywithadistributedlesystem, itisonlynecessarytowritedirtypagestothelesystembeforemigration.when thecostofmigrationisproportionaltothesizeoftheresidentsetratherthanthe theprocessisrestartedatthetargethost,itwillretrievethesepages.inthiscase Atthemost,itmightbenecessarytotransferaprocess'sentirememory.Ona sizeofmemory. ferredwhiletheprogramcontinuestorunatthesourcehost.whenthejobstops creased,thedelayimposedonthemigrantprocessisgreatlydecreased.additional dirtyduringtheprecopy.althoughthenumberofpagestransferredmightbein- executionatthesource,itwillhavetotransferagainanypagesthathavebecome Insystemsthatuseprecopying,suchasV[Theimeretal.1985],pagesaretrans- techniquescanreducethecostoftransferringmemoryevenmore[zayas1987]. 4.2Migrationcostsinrealsystems Thespecicparametersofmigrationcostdependnotonlyonthenatureofthe system(asdiscussedabove)butalsoonthespeedofthenetwork.tables3and4 showreportedcostsfromavarietyofrealsystems.laterwewilluseatrace-driven simulatortoevaluatetheeectoftheseparametersonsystemperformance.we
15 System Sprite ExploitingProcessLifetimeDistributionsforDynamicLoadBalancing Hardware SPARCstation1 0.33sec Costofrexec,r 15 [DouglisandOusterhout1991] GLUNIX [Vahdatetal.1994][Vahdat1995]ATMnetwork MIST [Prouty1996] 10Mb/secEthernet Utopia HPworkstations HP9000/720 10Mb/secEthernet 0.25to0.5sec [Zhouetal.1993] DEC3100andSPARCIPC0.1sec 0.33sec Table3.Costofnon-preemptivemigrationinvarioussystems.Someofthesenumberswere obtainedfrompersonalcommunicationwiththeauthors. System Sprite [DouglisandOusterhout1991] MOSIX 10Mb/secEthernet Hardware SPARCstation1 Fixed 0.33sec Cost,f 2.00sec/MB Inverse [Baraketal.1993][Braverman1995]and48666MHz IntelPentium90MHz0.006sec0.44sec/MB Bandwidth,1=b Table4.Valuesforpreemptivemigrationcostsfromvarioussystems.Manyofthesenumbers wereobtainedfrompersonalcommunicationwiththeauthors.thememorytransfercostisthe MIST [Prouty1996] HP9000/720s 10Mb/secEthernet 0.24sec 0.99sec/MB willmakethepessimisticsimplicationthatamigrant'sentirememorymustbe transferred,although,aspointedoutabove,thisisnotnecessarilythecase. productoftheinversebandwidth,1=b,andtheamountofmemorythatmustbetransferred,m. gration.wecomparetwomigrationstrategies:ourproposedage-basedpreemptive Inthissectionwepresenttheresultsofatrace-drivensimulationofprocessmi- 5.TRACE-DRIVENSIMULATION migrationstrategy(section3.1)andanon-preemptivestrategythatmigratesnewbornprocessesaccordingtotheprocessname(similartostrategiesproposedby [Wangetal.1993]and[Svensson1990]).Withtheintentionofndingaconservativeestimateofthebenetofpreemptivemigration,wegivethename-based usethesimulatortorunthreeexperiments.first,insection5.2,weevaluate thesensitivityofeachstrategytothemigrationcostsr,f,b,andmdiscussed strategythebenetofseveralunrealisticadvantages;forexample,thename-lists arederivedfromthesametracedatausedbythesimulator. insection4.next,insection5.3,wechoosevaluesfortheseparametersthatare representativeofcurrentsystemsandcomparetheperformanceofthetwostrategies Section5.1describesthesimulatorandthetwostrategiesinmoredetail.We indetail.insection5.4wediscusswhythepreemptivepolicyoutperformsthe non-preemptivepolicy.lastly,insection5.5,weevaluatetheanalyticcriterion formigrationageproposedinsection3.1,comparedtocriteriausedinprevious studies.
16 16 5.1Thesimulator Wehaveimplementedatrace-drivensimulationofanetworkofsixidenticalworkstations.2 M.Harchol-BalterandA.B.Downey eachfrom9:00a.m.to5:00p.m.fromthesixtracesweextractedthestarttimes Weselectedsixdaytimeintervalsfromthetracesonmachinepo(seeSection2.1), andcpudurationsoftheprocesses.wethensimulatedanetworkwhereeachof thedaytimetraces. sixhostsexecutes(concurrentlywiththeothers)theprocessarrivalsfromoneof fractionofprocesses(0:1%)arrivetondallhostsbusy.inordertoevaluatethe activityduringtheeight-hourtrace.formostofthetraces,everyarrivingprocess mixanddistributionoflifetimes,thereisconsiderablevariationinthelevelof ndsatleastoneidlehostinthesystem,butinthetwobusiesttraces,asmall Althoughtheworkloadsonthesixhostsarehomogeneousintermsofthejob lowesttohighestload.run0hasatotalof15000processessubmittedtothesix intervals.werefertotheseasruns0through7,wheretherunsaresortedfrom simulatedhosts;run7has30000processes.theaveragedurationofprocesses(for eectofchangesinsystemload,wedividedtheeight-hourtraceintoeightone-hour allruns)is0.4seconds.thusthetotalutilizationofthesystem,,isbetween0.27 and0.54. givenrunandagivenhost,theserialcorrelationininterarrivaltimesistypically between.08and.24,whichissignicantlyhigherthanonewouldexpectfroma Poissonprocess(uncorrelatedinterarrivaltimesyieldaserialcorrelationof0.0; perfectcorrelationis1.0). ThebirthprocessofjobsatourhostsisburstierthanaPoissonprocess.Fora randomlyfromameasureddistribution(seesection5.2).thissimplicationobliteratesanycorrelationsbetweenmemorysizeandotherprocesscharacteristics,but thememorysizeofeachprocess,whichdeterminesitsmigrationcost,ischosen itallowsustocontrolthemeanmemorysizeasaparameterandexamineitseect Althoughthestarttimesanddurationsoftheprocessescomefromtracedata, onsystemperformance. areneverblockedoni/o.duringagiventimeinterval,wedividecputimeequally amongtheprocessesonthehost(processorsharing). Inoursystemmodel,weassumethatprocessesarealwaysreadytorun;i.e.they migrationtothesourcehost.thissimplicationisconservativeinthesensethat thetransferredpages,partintransitinthenetwork,andpartonthetargethost unpackingthedata.thesizeofthesepartsandwhethertheycanbeoverlapped dependondetailsofthesystem.inoursimulationwechargetheentirecostof Inrealsystems,partofthemigrationtimeisspentonthesourcehostpackaging itmakespreemptivemigrationlesseective. thepoliciesassimpleandassimilaraspossible.forbothtypesofmigration,we considerperformingamigrationonlywhenanewprocessisborn,eventhougha tion3.1withanon-preemptivemigrationstrategy,wherethenon-preemptivestrat- egyisgivenunfairadvantages.forpurposesofcomparison,wehavetriedtomake 5.1.1Strategies.WecomparethepreemptivemigrationstrategyproposedinSec- 2The trace-driven simulator and the trace data are available at
17 preemptivestrategymightbenetbyinitiatingmigrationsatothertimes.also, forbothstrategies,ahostisconsideredheavily-loadedanytimeitcontainsmore ExploitingProcessLifetimeDistributionsforDynamicLoadBalancing 17 gration.finally,weusethesamelocationpolicyinbothcases:thehostwiththe lowestinstantaneousloadischosenasthetargethost(tiesarebrokenbyrandom thanoneprocess;inotherwords,anytimeitwouldbesensibletoconsidermi- selection). areconsideredeligibleformigration: Thustheonlydierencebetweenthetwomigrationpoliciesiswhichprocesses isbornataheavily-loadedhost,theprocessisexecutedremotelyontheselected itsnameisonalistofprocessesthattendtobelong-lived.ifaneligibleprocess host.processescannotbemigratedoncetheyhavebegunexecution. Name-basednon-preemptivemigration.Aprocessiseligibleformigrationonlyif choseathresholdonmeandurationthatisempiricallyoptimal(forthissetofruns). Addingmorenamestothelistdetractsfromtheperformanceofthesystem,asit durationandselectingthe15commonnameswiththelongestmeandurations.we Wederivedthislistbysortingtheprocessesfromthetracesaccordingtonameand Theperformanceofthisstrategydependsonthelistofeligibleprocessnames. allowsmoreshort-livedprocessestobemigrated.removingnamesfromthelist detractsfromperformanceasitbecomesimpossibletomigrateenoughprocesses agedforsomefractionofitsmigrationcost.basedonthederivationinsection3.1, list,ourresultsmayoverestimatetheperformancebenetsofthisstrategy. thisfractionis1 tobalancetheloadeectively.sinceweusedthetracedataitselftoconstructthe Age-basedpreemptivemigration.Aprocessiseligibleformigrationonlyifithas causeitdoesnotallowthesystemtoinitiatemigrationsexceptwhenanewprocess processesthatsatisfythemigrationcriterionaremigratedaway. source(target)host.whenanewprocessisbornataheavily-loadedhost,all Thisstrategyunderstatestheperformancebenetsofpreemptivemigration,be- n?m,wheren(respectivelym)isthenumberofprocessesatthe arrives. morecomplicatedpredictorsoffutureloads,butnoneofthesepredictorsyielded signicantlybetterperformancethantheinstantaneousloadweusehere. AsdescribedinSection3.1,wealsomodeledotherlocationpoliciesbasedon theperformanceofthemigrantwillimproveinexpectation).oneofthestrategies wheneveraprocessbecomeseligible(sincetheeligibilitycriterionguaranteesthat thanwhenanewprocessarrives.ideally,onewouldliketoinitiateamigration weconsideredperformsperiodicchecksofeachprocessonaheavily-loadedhost Wealsoconsideredtheeectofallowingpreemptivemigrationattimesother followingperformancemetrics: toseeifanysatisfythecriterion.theperformanceofthisstrategyissignicantly betterthanthatofthesimplerpolicy(migratingonlyatprocessarrivaltimes) Metrics.Weevaluatetheeectivenessofeachstrategyaccordingtothe metricofsystemperformance.whenwecomputetheratioofmeanslowdowns(as fromdierentstrategies)wewillusenormalizedslowdown,whichistheratioof (thus,itisalwaysgreaterthanone).theaverageslowdownofalljobsisacommon Meanslowdown.Slowdownistheratioofwall-clockexecutiontimetoCPUtime
18 18 M.Harchol-BalterandA.B.Downey Fraction of procs 0.5 Distribution of slowdowns smallslowdowns,buttheprocessesinthetailofthedistributionaremorenoticeableandannoying Fig.3.Distributionofprocessslowdownsforrun0(withnomigration).Mostprocessessuer CPUtime.Forexample,ifthe(unnormalized)meanslowdowndropsfrom2:0to tousers. inactivetime(theexcessslowdowncausedbyqueueingandmigrationdelays)to :5,theratioofnormalizedmeanslowdownsis0:5=1:0=0:5:a50%reductionin Slowdown ofthetwostrategies;itunderstatestheadvantagesofthepreemptivestrategyfor thesetworeasons: delay. Skeweddistributionofslowdowns:Evenintheabsenceofmigration,themajority Meanslowdownaloneisnotasucientmeasureofthedierenceinperformance Userperception:Fromtheuser'spointofview,theimportantprocessesarethe Thevalueofthemeanslowdownisdominatedbythismajority. onesinthetailofthedistribution,becausealthoughtheyaretheminority,they ofprocessessuersmallslowdowns(typically80%arelessthan3.0.seefigure3). haveasmalleectonthemeanslowdown,butalargeeectonauser'sperception causethemostnoticeableandannoyingdelays.eliminatingthesedelaysmight ofperformance. Therefore,wewillalsoconsiderthefollowingthreemetrics: tailofthedistribution;i.e.thenumberofjobsthatexperiencelongdelays.(see maybemoremeaningfultointerpretthismetricasameasureofthelengthofthe tryingtoscheduletasks.inlightofthedistributionofslowdowns,however,it dictabilityofresponsetime[silberschatzetal.1994],whichisanuisanceforusers Varianceofslowdown.Thismetricisoftencitedasameasureoftheunpre- severelyimpactedbyqueueingandmigrationpenalties.(seefigures5cand5d). abledelaysexplicitly,weconsiderthenumber(orpercentage)ofprocessesthatare Figure5b). Numberofseverelyslowedprocesses.Inordertoquantifythenumberofnotice-
19 than:5seconds)aremoreperceivabletousersthandelaysinshortjobs.(see Meanslowdownoflongjobs.Delaysinlongerjobs(thosewithlifetimesgreater ExploitingProcessLifetimeDistributionsforDynamicLoadBalancing 19 Figure6). 5.2Sensitivitytomigrationcosts Inthissectionwecomparetheperformanceofthenon-preemptiveandpreemptive strategiesoverarangeofvaluesofr,f,bandm(themigrationcostparameters process)andb(thebandwidthofthenetwork).wechosethememorytransfercost Weconsideredarangeforthexedmigrationcostof:1<f<10seconds. denedinsection4). fromadistributionwiththesameshapeasthedistributionofprocesslifetimes, Thememorytransfercostisthequotientofm(thememorysizeofthemigrant Forthefollowingexperiments,wechosetheremoteexecutioncostr=:3seconds. featureofthisdistributionisthattherearemanyjobswithsmallmemorydemands seconds.theshapeofthisdistributionisbasedonaninformalstudyofmemory-use patternsonthesamemachinesfromwhichwecollectedtracedata.theimportant settingthemeanmemorytransfercost(mmtc)toarangeofvaluesfrom1to64 distributiondoesnotaecttheperformanceofeithermigrationstrategystrongly, butofcoursethemean(mmtc)doeshaveastrongeect. andafewjobswithverylargememorydemands.empirically,theexactformofthis migrationstrategiesusingnormalizedslowdown.specically,foreachoftheeight one-hourrunswecalculatethemean(respectivelystandarddeviation)oftheslowdownimposedonallprocessesthatcompleteduringthehour.foreachrun,we Figures4aand4barecontourplotsoftheratiooftheperformanceofthetwo thentaketheratioofthemeans(standarddeviations)ofthetwostrategies.lastly wetakethegeometricmeanoftheeightratios(fordiscussionofthegeometric mean,see[hennessyandpatterson1990]). tivemigration,namelythexedcost,f,andthemmtc,m=b.thecostofnon- preemptivemigration,r,isxedat0:3seconds.asexpected,increasingeitherthe xedcostofmigrationorthemmtchurtstheperformanceofpreemptivemigration.thecontourlinemarked1:0indicatesthecrossoverwheretheperformance ThetwoaxesinFigure4representthetwocomponentsofthecostofpreemp- non-preemptivemigration.whenthexedcostofmigrationorthemmtcare ofpreemptiveandnon-preemptivemigrationisequal(theratiois1:0).forsmaller valuesofthecostparameters,preemptivemigrationperformsbetter;forexample, ifthexedmigrationcostis0.3secondsandthemmtcis2seconds,thenormalizedmeanslowdownwithpreemptivemigrationisalmost40%lowerthanwith unaectedbythesecostssothenon-preemptivestrategycanbemoreeective. veryhigh,almostallprocessesareineligibleforpreemptivemigration;thus,the preemptivestrategydoesalmostnomigrations.thenon-preemptivestrategyis aregionwherepreemptivemigrationyieldsahighermeanslowdownthannonpreemptivemigration,butalowerstandarddeviation.thereasonforthisisthadowns.thecrossoverpoint wherenon-preemptivemigrationsurpassespreemp- Figure4bshowstheeectofmigrationcostsonthestandarddeviationofslowtivemigration isconsiderablyhigherherethaninfigure4a.thusthereis turnsouttobeshort-lived.theseprocessessuerlargedelays(relativetotheir non-preemptivemigrationoccasionallychoosesaprocessforremoteexecutionthat
20 20 M.Harchol-BalterandA.B.Downey (a) 10 Ratio of mean slowdowns Fixed migration cost (sec.) X NON PREEMPTIVE BETTER HERE 1.0 PREEMPTIVE BETTER HERE (b) Mean memory transfer cost (sec.) 10 Ratio of std of slowdowns 1.0 Fixed migration cost (sec.) Fig.4.(a)Theperformanceofpreemptivemigrationrelativetonon-preemptivemigrationdeterioratesasthecostofpreemptivemigrationincreases.Thetwoaxesarethetwocomponentsof PREEMPTIVE BETTER EXCEPT UPPER RIGHT X ofslowdownmaygiveabetterindicationofauser'sperceptionofsystemperformancethanmean theparticularsetofparameterswewillconsiderinthenextsection.(b)thestandarddeviation thepreemptivemigrationcost.thecostofnon-preemptivemigrationisheldxed.thexmarks slowdown.bythismetric,thebenetofpreemptivemigrationisevenmoresignicant Mean memory transfer (sec.)
21 runtimes)andaddtothetailofthedistributionofslowdowns.inthenextsection, weshowcasesinwhichthestandarddeviationofslowdownsisactuallyworsewith ExploitingProcessLifetimeDistributionsforDynamicLoadBalancing 21 non-preemptivemigrationthanwithnomigrationatall(threeoftheeightruns). tems(seesection4.2)andusethemtoexaminemorecloselytheperformanceof thetwomigrationstrategies.thevalueswechoseare: Inthissectionwechoosemigrationcostparametersrepresentativeofcurrentsys- 5.3Comparisonofpreemptiveandnon-preemptivestrategies r:thecostofremoteexecution,0.3seconds f:thexedcostofpreemptivemigration,0.3seconds m:themeanmemorysizeofmigrantprocesses,1mb b:thememorytransferbandwidth,0.5mbpersecond point(comparedtothecaseofnomigration). withanx.figure5showstheperformanceofthetwomigrationstrategiesatthis InFigures4aand4b,thepointcorrespondingtotheseparametervaluesismarked bylessthan20%formostruns(and40%forthetworunswiththehighestloads). Preemptivemigrationreducesthenormalizedmeanslowdownby50%formost runs(andmorethan60%fortwooftheruns).theperformanceimprovementof preemptivemigrationovernon-preemptivemigrationistypicallybetween35%and Non-preemptivemigrationreducesthenormalizedmeanslowdown(Figure5a) 50%. metricstotrytoquantifythesebenets.figure5bshowsthestandarddeviation and5dexplicitlymeasurethenumberofseverelyimpactedprocesses,accordingto statestheperformancebenetsofpreemptivemigration.wehaveproposedother ofslowdowns,whichreectsthenumberofseverelyimpactedprocesses.figures5c Asdiscussedabove,wefeelthatthemeanslowdown(normalizedornot)under- twodierentthresholdsofacceptableslowdown.bythesemetrics,thebenetsof non-preemptivemigrationappearsmuchgreater.forexampleinfigure5d,inthe migrationingeneralappeargreater,andthediscrepancybetweenpreemptiveand absenceofmigration,7{18%ofprocessesareslowedbyafactorof5ormore. Non-preemptivemigrationisabletoeliminate42{62%ofthese,whichisasignicantbenet,butpreemptivemigrationconsistentlyeliminatesnearlyall(86{97%) severedelays. nomigrationatall.forthepreemptivemigrationstrategy,thisoutcomeisnearly allprocessesinvolved(inexpectation).intheworstcase,then,thepreemptive migrationactuallymakestheperformanceofthesystemworsethaniftherewere impossible,sincemigrationsareonlyperformediftheyimprovetheslowdownsof AnimportantobservationfromFigure5bisthatforseveralruns,non-preemptive formanceasloadincreases(asshowninfigure5).inthepresenceofpreemptive strategywilldonoworsethanthecaseofnomigration(inexpectation). regardlessoftheoverallloadonthesystem. migration,boththemeanandstandarddeviationofslowdownarenearlyconstant, Anotherbenetofpreemptivemigrationisgracefuldegradationofsystemper-
22 22 M.Harchol-BalterandA.B.Downey 4.0 Mean slowdown no migration non-preemptive, name-based migration preemptive, age-based migration run number Standard deviation of slowdown no migration non-preemptive, name-based migration preemptive, age-based migration % run number Processes slowed by a factor of 3 or more no migration non-preemptive, name-based migration preemptive, age-based migration 20% 10% 0% 20% run number Processes slowed by a factor of 5 or more no migration non-preemptive, name-based migration preemptive, age-based migration 15% Fig.5.(a)Meanslowdown.(b)Standarddeviationofslowdown.(c)Percentageofprocesses slowedbyafactorof3ormore.(d)percentageofprocessesslowedbyafactorof5ormore. 10% 5% 0% run number
Exploiting Process Lifetime Distributions for Dynamic Load Balancing
Exploiting Process Lifetime Distributions for Dynamic Load Balancing MOR HARCHOL-BALTER and ALLEN B. DOWNEY University of California, Berkeley We consider policies for CPU load balancing in networks of
More informationIn Proceedings of Performance Tools 98 Lecture Notes in Computer Science, Vol. 1468, pp. 231-242, September 1998
In Proceedings of Performance Tools 98 OnChoosingaTaskAssignmentPolicyfora Lecture Notes in Computer Science, Vol. 1468, pp. 231-242, September 1998 MorHarchol-Balter?;1,MarkE.Crovella??;2,andCristinaD.Murta???;2
More informationWeiping Zhu C.F. Steketee. processes into account. to the potential performance gain from this service.
An Experimental Study of Load Balancing on Amoeba Weiping Zhu C.F. Steketee School of Computer and Information Science University of South Australia Adelaide, Australia SA5095 Abstract This paper presents
More informationLecture Outline Overview of real-time scheduling algorithms Outline relative strengths, weaknesses
Overview of Real-Time Scheduling Embedded Real-Time Software Lecture 3 Lecture Outline Overview of real-time scheduling algorithms Clock-driven Weighted round-robin Priority-driven Dynamic vs. static Deadline
More informationImproving Distributed Workload Performance by Sharing Both CPU and Memory Resources
Improving Distributed Workload Performance by Sharing Both CPU and Memory Resources Xiaodong Zhang Yanxia Qu Li Xiao Department of Computer Science College of William and Mary Williamsburg, VA 387-8795
More informationChapter 19: Real-Time Systems. Overview of Real-Time Systems. Objectives. System Characteristics. Features of Real-Time Systems
Chapter 19: Real-Time Systems System Characteristics Features of Real-Time Systems Chapter 19: Real-Time Systems Implementing Real-Time Operating Systems Real-Time CPU Scheduling VxWorks 5.x 19.2 Silberschatz,
More informationOperating Systems. III. Scheduling. http://soc.eurecom.fr/os/
Operating Systems Institut Mines-Telecom III. Scheduling Ludovic Apvrille ludovic.apvrille@telecom-paristech.fr Eurecom, office 470 http://soc.eurecom.fr/os/ Outline Basics of Scheduling Definitions Switching
More informationReal-Time Scheduling (Part 1) (Working Draft) Real-Time System Example
Real-Time Scheduling (Part 1) (Working Draft) Insup Lee Department of Computer and Information Science School of Engineering and Applied Science University of Pennsylvania www.cis.upenn.edu/~lee/ CIS 41,
More informationDeciding which process to run. (Deciding which thread to run) Deciding how long the chosen process can run
SFWR ENG 3BB4 Software Design 3 Concurrent System Design 2 SFWR ENG 3BB4 Software Design 3 Concurrent System Design 11.8 10 CPU Scheduling Chapter 11 CPU Scheduling Policies Deciding which process to run
More informationSimplest Scalable Architecture
Simplest Scalable Architecture NOW Network Of Workstations Many types of Clusters (form HP s Dr. Bruce J. Walker) High Performance Clusters Beowulf; 1000 nodes; parallel programs; MPI Load-leveling Clusters
More informationRoad Map. Scheduling. Types of Scheduling. Scheduling. CPU Scheduling. Job Scheduling. Dickinson College Computer Science 354 Spring 2010.
Road Map Scheduling Dickinson College Computer Science 354 Spring 2010 Past: What an OS is, why we have them, what they do. Base hardware and support for operating systems Process Management Threads Present:
More informationTrust Oriented Cooperative Resource Scheduling in Grid Computing
CHAPTER 4 Trust Oriented Cooperative Resource Scheduling in Grid Computing This chapter presents a Trust Oriented Cooperative Resource Scheduling (TOCRS) strategy in grid computing. This strategy integrates
More informationAn Economic Perspective of! Disk vs. Flash Media! in Archival Storage Preeti Gupta Avani Wildani Ethan L. Miller! Daniel Rosenthal Ian F. Adams,!
An Economic Perspective of Disk vs. Flash Media in Archival Storage Preeti Gupta Avani Wildani Ethan L. Miller Daniel Rosenthal Ian F. Adams, Christina Strong Andy Hospodor What s the trouble with media
More informationAachen Summer Simulation Seminar 2014
Aachen Summer Simulation Seminar 2014 Lecture 07 Input Modelling + Experimentation + Output Analysis Peer-Olaf Siebers pos@cs.nott.ac.uk Motivation 1. Input modelling Improve the understanding about how
More information3.2 LOGARITHMIC FUNCTIONS AND THEIR GRAPHS. Copyright Cengage Learning. All rights reserved.
3.2 LOGARITHMIC FUNCTIONS AND THEIR GRAPHS Copyright Cengage Learning. All rights reserved. What You Should Learn Recognize and evaluate logarithmic functions with base a. Graph logarithmic functions.
More informationScheduling. Scheduling. Scheduling levels. Decision to switch the running process can take place under the following circumstances:
Scheduling Scheduling Scheduling levels Long-term scheduling. Selects which jobs shall be allowed to enter the system. Only used in batch systems. Medium-term scheduling. Performs swapin-swapout operations
More informationBetting with the Kelly Criterion
Betting with the Kelly Criterion Jane June 2, 2010 Contents 1 Introduction 2 2 Kelly Criterion 2 3 The Stock Market 3 4 Simulations 5 5 Conclusion 8 1 Page 2 of 9 1 Introduction Gambling in all forms,
More informationReordering of IP Packets in Internet
Reordering of IP Packets in Internet Xiaoming Zhou and Piet Van Mieghem Network Architectures and Services TU Delft January 29, 2004 Xiaoming Zhou 1 Introduction to Reordering Problem description and definitions
More informationSIMULATION OF LOAD BALANCING ALGORITHMS: A Comparative Study
SIMULATION OF LOAD BALANCING ALGORITHMS: A Comparative Study Milan E. Soklic Abstract This article introduces a new load balancing algorithm, called diffusive load balancing, and compares its performance
More informationScheduling Allowance Adaptability in Load Balancing technique for Distributed Systems
Scheduling Allowance Adaptability in Load Balancing technique for Distributed Systems G.Rajina #1, P.Nagaraju #2 #1 M.Tech, Computer Science Engineering, TallaPadmavathi Engineering College, Warangal,
More informationLogs Transformation in a Regression Equation
Fall, 2001 1 Logs as the Predictor Logs Transformation in a Regression Equation The interpretation of the slope and intercept in a regression change when the predictor (X) is put on a log scale. In this
More informationInternational Journal of Innovative Research in Science, Engineering and Technology Vol. 2, Issue 5, May 2013
ISSN: 2319-8753 International Journal of Innovative Research in Science, Engineering and Technology Vol. 2, Issue 5, May 2013 of vibration are 0.14 rad/s and 0.42 rad/s respectively. The dynamic response
More informationSUMAN DUVVURU STAT 567 PROJECT REPORT
SUMAN DUVVURU STAT 567 PROJECT REPORT SURVIVAL ANALYSIS OF HEROIN ADDICTS Background and introduction: Current illicit drug use among teens is continuing to increase in many countries around the world.
More informationSource-domain DDoS Prevention
bhattacharjee, LTS S 05 Page: 0 Source-domain DDoS Prevention Bobby Bhattacharjee Christopher Kommareddy Mark Shayman Dave Levin Richard La Vahid Tabatabaee University of Maryland bhattacharjee, LTS S
More informationE- Encryption in Unix
UNIVERSITY of WISCONSIN-MADISON Computer Sciences Department CS 537 A. Arpaci-Dusseau Intro to Operating Systems Spring 2000 Security Solutions and Encryption Questions answered in these notes: How does
More informationCPU Scheduling. CPU Scheduling
CPU Scheduling Electrical and Computer Engineering Stephen Kim (dskim@iupui.edu) ECE/IUPUI RTOS & APPS 1 CPU Scheduling Basic Concepts Scheduling Criteria Scheduling Algorithms Multiple-Processor Scheduling
More information2.500 Threshold. 2.000 1000e - 001. Threshold. Exponential phase. Cycle Number
application note Real-Time PCR: Understanding C T Real-Time PCR: Understanding C T 4.500 3.500 1000e + 001 4.000 3.000 1000e + 000 3.500 2.500 Threshold 3.000 2.000 1000e - 001 Rn 2500 Rn 1500 Rn 2000
More informationOPERATING SYSTEMS SCHEDULING
OPERATING SYSTEMS SCHEDULING Jerry Breecher 5: CPU- 1 CPU What Is In This Chapter? This chapter is about how to get a process attached to a processor. It centers around efficient algorithms that perform
More informationLOAD BALANCING TECHNIQUES
LOAD BALANCING TECHNIQUES Two imporatnt characteristics of distributed systems are resource multiplicity and system transparency. In a distributed system we have a number of resources interconnected by
More informationModule 6. Embedded System Software. Version 2 EE IIT, Kharagpur 1
Module 6 Embedded System Software Version 2 EE IIT, Kharagpur 1 Lesson 30 Real-Time Task Scheduling Part 2 Version 2 EE IIT, Kharagpur 2 Specific Instructional Objectives At the end of this lesson, the
More informationThe Case for Massive Arrays of Idle Disks (MAID)
The Case for Massive Arrays of Idle Disks (MAID) Dennis Colarelli, Dirk Grunwald and Michael Neufeld Dept. of Computer Science Univ. of Colorado, Boulder January 7, 2002 Abstract The declining costs of
More informationCTERA Cloud Care. Support Services. Mar 2015. Version 2.0. 2015, CTERA Networks. All rights reserved.
CTERA Cloud Care Support Services Mar 2015 Version 2.0 2015, CTERA Networks. All rights reserved. Contents CTERA Support Services -------------------------------------------------------------------------------
More informationScheduling. Yücel Saygın. These slides are based on your text book and on the slides prepared by Andrew S. Tanenbaum
Scheduling Yücel Saygın These slides are based on your text book and on the slides prepared by Andrew S. Tanenbaum 1 Scheduling Introduction to Scheduling (1) Bursts of CPU usage alternate with periods
More informationChapter 12: Multiprocessor Architectures. Lesson 01: Performance characteristics of Multiprocessor Architectures and Speedup
Chapter 12: Multiprocessor Architectures Lesson 01: Performance characteristics of Multiprocessor Architectures and Speedup Objective Be familiar with basic multiprocessor architectures and be able to
More informationCorporate Defaults and Large Macroeconomic Shocks
Corporate Defaults and Large Macroeconomic Shocks Mathias Drehmann Bank of England Andrew Patton London School of Economics and Bank of England Steffen Sorensen Bank of England The presentation expresses
More informationEmbedded Systems Lecture 9: Reliability & Fault Tolerance. Björn Franke University of Edinburgh
Embedded Systems Lecture 9: Reliability & Fault Tolerance Björn Franke University of Edinburgh Overview Definitions System Reliability Fault Tolerance Sources and Detection of Errors Stage Error Sources
More informationDynamic Load Balancing of SSH Sessions Using User-Specific Selection Policies
Dynamic Load Balancing of SSH Sessions Using User-Specific Selection Policies Paul Z. Kolano NASA Advanced Supercomputing Division, NASA Ames Research Center M/S 258-6, Moffett Field, CA 9435 U.S.A. E-mail:
More informationObjectives. Chapter 5: Process Scheduling. Chapter 5: Process Scheduling. 5.1 Basic Concepts. To introduce CPU scheduling
Objectives To introduce CPU scheduling To describe various CPU-scheduling algorithms Chapter 5: Process Scheduling To discuss evaluation criteria for selecting the CPUscheduling algorithm for a particular
More informationBus u i s n i e n s e s s s Cas a e s, e, S o S l o u l t u io i n o n & A pp p r p oa o c a h
Work Load Modeling and Work Load Modeler in Performance Testing Business Case, Solution & Approach Case An application is made ready to go-live in the next 2 months, but the application performance behavior
More informationThe Association of System Performance Professionals
The Association of System Performance Professionals The Computer Measurement Group, commonly called CMG, is a not for profit, worldwide organization of data processing professionals committed to the measurement
More information35 th Design Automation Conference Copyright 1998 ACM
Design Reliability - Estimation through Statistical Analysis of Bug Discovery Data Yossi Malka, Avi Ziv IBM Research Lab in Haifa MATAM Haifa 3195, Israel email: fyossi, aziv@vnet.ibm.comg Abstract Statistical
More informationSOFTWARE FOR GENERATION OF SPECTRUM COMPATIBLE TIME HISTORY
3 th World Conference on Earthquake Engineering Vancouver, B.C., Canada August -6, 24 Paper No. 296 SOFTWARE FOR GENERATION OF SPECTRUM COMPATIBLE TIME HISTORY ASHOK KUMAR SUMMARY One of the important
More informationMathematical Harmonies Mark Petersen
1 Mathematical Harmonies Mark Petersen What is music? When you hear a flutist, a signal is sent from her fingers to your ears. As the flute is played, it vibrates. The vibrations travel through the air
More informationThe Effect of Network Cabling on Bit Error Rate Performance. By Paul Kish NORDX/CDT
The Effect of Network Cabling on Bit Error Rate Performance By Paul Kish NORDX/CDT Table of Contents Introduction... 2 Probability of Causing Errors... 3 Noise Sources Contributing to Errors... 4 Bit Error
More informationKappa: A system for Linux P2P Load Balancing and Transparent Process Migration
Kappa: A system for Linux P2P Load Balancing and Transparent Process Migration Gaurav Mogre gaurav.mogre@gmail.com Avinash Hanumanthappa avinash947@gmail.com Alwyn Roshan Pais alwyn@nitk.ac.in Abstract
More informationVirtualization: Concepts, Applications, and Performance Modeling
Virtualization: Concepts, s, and Performance Modeling Daniel A. Menascé, Ph.D. The Volgenau School of Information Technology and Engineering Department of Computer Science George Mason University www.cs.gmu.edu/faculty/menasce.html
More informationSEQUENCES ARITHMETIC SEQUENCES. Examples
SEQUENCES ARITHMETIC SEQUENCES An ordered list of numbers such as: 4, 9, 6, 25, 36 is a sequence. Each number in the sequence is a term. Usually variables with subscripts are used to label terms. For example,
More informationCharacterizing Task Usage Shapes in Google s Compute Clusters
Characterizing Task Usage Shapes in Google s Compute Clusters Qi Zhang 1, Joseph L. Hellerstein 2, Raouf Boutaba 1 1 University of Waterloo, 2 Google Inc. Introduction Cloud computing is becoming a key
More informationPerformance Analysis of Web based Applications on Single and Multi Core Servers
Performance Analysis of Web based Applications on Single and Multi Core Servers Gitika Khare, Diptikant Pathy, Alpana Rajan, Alok Jain, Anil Rawat Raja Ramanna Centre for Advanced Technology Department
More informationCharacterizing Task Usage Shapes in Google s Compute Clusters
Characterizing Task Usage Shapes in Google s Compute Clusters Qi Zhang University of Waterloo qzhang@uwaterloo.ca Joseph L. Hellerstein Google Inc. jlh@google.com Raouf Boutaba University of Waterloo rboutaba@uwaterloo.ca
More informationExercises : Real-time Scheduling analysis
Exercises : Real-time Scheduling analysis Frank Singhoff University of Brest June 2013 Exercise 1 : Fixed priority scheduling and Rate Monotonic priority assignment Given a set of tasks defined by the
More informationFault Characteristics in Electrical Equipment
1. Introduction Proper design and installation of electrical equipment minimizes the chance of electrical faults. Faults occur when the insulation system is compromised and current is allowed to flow through
More informationChapter 5 Process Scheduling
Chapter 5 Process Scheduling CPU Scheduling Objective: Basic Scheduling Concepts CPU Scheduling Algorithms Why Multiprogramming? Maximize CPU/Resources Utilization (Based on Some Criteria) CPU Scheduling
More informationCPU Scheduling. Core Definitions
CPU Scheduling General rule keep the CPU busy; an idle CPU is a wasted CPU Major source of CPU idleness: I/O (or waiting for it) Many programs have a characteristic CPU I/O burst cycle alternating phases
More informationSIDN Server Measurements
SIDN Server Measurements Yuri Schaeffer 1, NLnet Labs NLnet Labs document 2010-003 July 19, 2010 1 Introduction For future capacity planning SIDN would like to have an insight on the required resources
More informationPre-Session Review. Part 2: Mathematics of Finance
Pre-Session Review Part 2: Mathematics of Finance For this section you will need a calculator with logarithmic and exponential function keys (such as log, ln, and x y ) D. Exponential and Logarithmic Functions
More informationCHAPTER 5 WLDMA: A NEW LOAD BALANCING STRATEGY FOR WAN ENVIRONMENT
81 CHAPTER 5 WLDMA: A NEW LOAD BALANCING STRATEGY FOR WAN ENVIRONMENT 5.1 INTRODUCTION Distributed Web servers on the Internet require high scalability and availability to provide efficient services to
More informationDetecting Flooding Attacks Using Power Divergence
Detecting Flooding Attacks Using Power Divergence Jean Tajer IT Security for the Next Generation European Cup, Prague 17-19 February, 2012 PAGE 1 Agenda 1- Introduction 2- K-ary Sktech 3- Detection Threshold
More informationA Policy-Based Admission Control Scheme for Voice over IP Networks
Journal of Computer Science 5 (11): 817-821, 2009 ISSN 1549-3636 2009 Science Publications A Policy-Based Admission Control Scheme for Voice over IP Networks Sami Alwakeel and Agung Prasetijo Department
More informationEnsuring Collective Availability in Volatile Resource Pools via Forecasting
Ensuring Collective Availability in Volatile Resource Pools via Forecasting Artur Andrzejak andrzejak[at]zib.de Derrick Kondo David P. Anderson Zuse Institute Berlin (ZIB) INRIA UC Berkeley Motivation
More informationVertical Scaling of Oracle 10g Performance on Red Hat Enterprise Linux 5 on Intel Xeon Based Servers. Version 1.0
Vertical Scaling of Oracle 10g Performance on Red Hat Enterprise Linux 5 on Intel Xeon Based Servers Version 1.0 March 2009 Vertical Scaling of Oracle 10g Performance on Red Hat Enterprise Linux 5 on Inel
More informationPerformance Test Process
A white Success The performance testing helped the client identify and resolve performance bottlenecks which otherwise crippled the business. The ability to support 500 concurrent users was a performance
More informationIn Proceedings of the First IEEE Workshop on Real-Time Applications, New York, NY, May 1993.
In Proceedings of the First IEEE Workshop on Real-Time Applications, New York, NY, May 1993. Probabilistic Job Scheduling for Distributed Real-time Applications Azer Bestavros Dimitrios Spartiotis Computer
More informationMathematics Review for MS Finance Students
Mathematics Review for MS Finance Students Anthony M. Marino Department of Finance and Business Economics Marshall School of Business Lecture 1: Introductory Material Sets The Real Number System Functions,
More informationMultiple Optimization Using the JMP Statistical Software Kodak Research Conference May 9, 2005
Multiple Optimization Using the JMP Statistical Software Kodak Research Conference May 9, 2005 Philip J. Ramsey, Ph.D., Mia L. Stephens, MS, Marie Gaudard, Ph.D. North Haven Group, http://www.northhavengroup.com/
More informationWeb Application Testing. Web Performance Testing
Web Application Testing Web Performance Testing Objectives of Performance Testing Evaluate runtime compliance to performance requirements Check different properties such as throughput (bits/sec, packets/sec)
More informationLoad Balancing in Distributed Web Server Systems With Partial Document Replication
Load Balancing in Distributed Web Server Systems With Partial Document Replication Ling Zhuo, Cho-Li Wang and Francis C. M. Lau Department of Computer Science and Information Systems The University of
More informationEvaluating Trading Systems By John Ehlers and Ric Way
Evaluating Trading Systems By John Ehlers and Ric Way INTRODUCTION What is the best way to evaluate the performance of a trading system? Conventional wisdom holds that the best way is to examine the system
More informationMOSIX: High performance Linux farm
MOSIX: High performance Linux farm Paolo Mastroserio [mastroserio@na.infn.it] Francesco Maria Taurino [taurino@na.infn.it] Gennaro Tortone [tortone@na.infn.it] Napoli Index overview on Linux farm farm
More informationOperating System: Scheduling
Process Management Operating System: Scheduling OS maintains a data structure for each process called Process Control Block (PCB) Information associated with each PCB: Process state: e.g. ready, or waiting
More informationCHAPTER 3 CALL CENTER QUEUING MODEL WITH LOGNORMAL SERVICE TIME DISTRIBUTION
31 CHAPTER 3 CALL CENTER QUEUING MODEL WITH LOGNORMAL SERVICE TIME DISTRIBUTION 3.1 INTRODUCTION In this chapter, construction of queuing model with non-exponential service time distribution, performance
More informationMarket Technician, TradeStation Labs TSLabs@TradeStation.com
Market Technician, TradeStation Labs TSLabs@TradeStation.com Standard deviation is a common statistical calculation that is often used in the world of finance to measure risk. The higher the standard deviation,
More informationReal-time PCR: Understanding C t
APPLICATION NOTE Real-Time PCR Real-time PCR: Understanding C t Real-time PCR, also called quantitative PCR or qpcr, can provide a simple and elegant method for determining the amount of a target sequence
More informationQoS-Driven Server Migration for Internet Data Centers
QoS-Driven Server Migration for Internet Data Centers S. Ranjan, J. Rolia, H. Fu, and E. Knightly Abstract Many organizations have chosen to host Internet applications at Internet Data Centers (IDCs) located
More informationHow To Test A Web Server
Performance and Load Testing Part 1 Performance & Load Testing Basics Performance & Load Testing Basics Introduction to Performance Testing Difference between Performance, Load and Stress Testing Why Performance
More informationSIMPLIFIED PERFORMANCE MODEL FOR HYBRID WIND DIESEL SYSTEMS. J. F. MANWELL, J. G. McGOWAN and U. ABDULWAHID
SIMPLIFIED PERFORMANCE MODEL FOR HYBRID WIND DIESEL SYSTEMS J. F. MANWELL, J. G. McGOWAN and U. ABDULWAHID Renewable Energy Laboratory Department of Mechanical and Industrial Engineering University of
More informationWeb Load Stress Testing
Web Load Stress Testing Overview A Web load stress test is a diagnostic tool that helps predict how a website will respond to various traffic levels. This test can answer critical questions such as: How
More informationReal-time KVM from the ground up
Real-time KVM from the ground up KVM Forum 2015 Rik van Riel Red Hat Real-time KVM What is real time? Hardware pitfalls Realtime preempt Linux kernel patch set KVM & qemu pitfalls KVM configuration Scheduling
More informationUNIVERSITY of MASSACHUSETTS DARTMOUTH Charlton College of Business Decision and Information Sciences Fall 2010
UNIVERSITY of MASSACHUSETTS DARTMOUTH Charlton College of Business Decision and Information Sciences Fall 2010 COURSE: POM 500 Statistical Analysis, ONLINE EDITION, Fall 2010 Prerequisite: Finite Math
More informationDiscrete-Event Simulation
Discrete-Event Simulation Prateek Sharma Abstract: Simulation can be regarded as the emulation of the behavior of a real-world system over an interval of time. The process of simulation relies upon the
More informationProcess Migration and Load Balancing in Amoeba
Process Migration and Load Balancing in Amoeba Chris Steketee Advanced Computing Research Centre, School of Computer and Information Science, University of South Australia, The Levels SA 5095 Email: Chris.Steketee@cis.unisa.edu.au
More informationAdvanced Statistical Analysis of Mortality. Rhodes, Thomas E. and Freitas, Stephen A. MIB, Inc. 160 University Avenue. Westwood, MA 02090
Advanced Statistical Analysis of Mortality Rhodes, Thomas E. and Freitas, Stephen A. MIB, Inc 160 University Avenue Westwood, MA 02090 001-(781)-751-6356 fax 001-(781)-329-3379 trhodes@mib.com Abstract
More informationQUICKCLOSE HP35S Surveying Programs - Version 1.1
QUICKCLOSE HP35S Surveying Programs - Version 1.1 USER REFERENCE GUIDE - for Evaluation Important Note: Program steps are only included in the version of User Reference Guides available from resellers.
More informationSo in order to grab all the visitors requests we add to our workbench a non-test-element of the proxy type.
First in oder to configure our test case, we need to reproduce our typical browsing path containing all the pages visited by the visitors on our systems. So in order to grab all the visitors requests we
More informationHybrid processing of SCADA and synchronized phasor measurements for tracking network state
IEEE PES General Meeting, Denver, USA, July 2015 1 Hybrid processing of SCADA and synchronized phasor measurements for tracking network state Boris Alcaide-Moreno Claudio Fuerte-Esquivel Universidad Michoacana
More informationFault-Tolerant Framework for Load Balancing System
Fault-Tolerant Framework for Load Balancing System Y. K. LIU, L.M. CHENG, L.L.CHENG Department of Electronic Engineering City University of Hong Kong Tat Chee Avenue, Kowloon, Hong Kong SAR HONG KONG Abstract:
More informationHuawei Enterprise Warranty for Barebone Servers
Huawei Enterprise Warranty for Barebone Servers 2015-07-03 Copyright Huawei 2015. Page 1 of 5 Dear Partner, Thank you for choosing the products and services of Huawei. This document describes the warranty
More informationENSC 427: Communication Networks. Analysis of Voice over IP performance on Wi-Fi networks
ENSC 427: Communication Networks Spring 2010 OPNET Final Project Analysis of Voice over IP performance on Wi-Fi networks Group 14 members: Farzad Abasi (faa6@sfu.ca) Ehsan Arman (eaa14@sfu.ca) http://www.sfu.ca/~faa6
More informationLecture 3 Theoretical Foundations of RTOS
CENG 383 Real-Time Systems Lecture 3 Theoretical Foundations of RTOS Asst. Prof. Tolga Ayav, Ph.D. Department of Computer Engineering Task States Executing Ready Suspended (or blocked) Dormant (or sleeping)
More informationBiDAl: Big Data Analyzer for Cluster Traces
BiDAl: Big Data Analyzer for Cluster Traces Alkida Balliu, Dennis Olivetti, Ozalp Babaoglu, Moreno Marzolla, Alina Sirbu Department of Computer Science and Engineering University of Bologna, Italy BigSys
More informationHow To Improve Performance On A Single Chip Computer
: Redundant Arrays of Inexpensive Disks this discussion is based on the paper:» A Case for Redundant Arrays of Inexpensive Disks (),» David A Patterson, Garth Gibson, and Randy H Katz,» In Proceedings
More informationCAPACITY MANAGEMENT A METHODOLOGY Phase 2: Service Level Agreements
CAPACITY MANAGEMENT A METHODOLOGY Phase 2: Service Level Agreements Adrian Heald Capacity Reporting Services Primitive Capacity Planning measures the "fullness" of each IT processing component and recommends
More informationAdaptive Scheduling for QoS-based Virtual Machine Management in Cloud Computing
Yang Cao, Cheul Woo Ro : Adaptive Scheduling for QoS-based Virtual Machine Management in Cloud Computing 7 http://dx.doi.org/10.5392/ijoc.2012.8.7 Adaptive Scheduling for QoS-based Virtual Machine Management
More informationDistinguishing between FE and DDoS using Randomness Check
Distinguishing between FE and DDoS using Randomness Check Hyundo Park, Peng Li, Debin Gao, Heejo Lee and Robert Deng Presented by Hyundo Park Korea University Singapore Management University Index Introduction
More informationHSR HOCHSCHULE FÜR TECHNIK RA PPERSW I L
1 An Introduction into Modelling and Simulation 4. A Series of Labs to Learn Simio af&e Prof. Dr.-Ing. Andreas Rinkel andreas.rinkel@hsr.ch Tel.: +41 (0) 55 2224928 Mobil: +41 (0) 79 3320562 Lab 1 Lab
More information2. The Vector Network Analyzer
ECE 584 Laboratory Experiments 2. The Vector Network Analyzer Introduction: In this experiment we will learn to use a Vector Network Analyzer to measure the magnitude and phase of reflection and transmission
More informationHP Windows 7 Onsite Upgrade Service
HP Windows 7 Onsite Upgrade Service HP Care Pack Services Technical data The Windows 7 Onsite Upgrade Service provides your company with onsite installation, setup, and training for the Microsoft Windows
More informationChapter 10: Fixed Income Analysis. Joel Barber. Department of Finance. Florida International University. Miami, FL 33199
Chapter 10: Fixed Income Analysis Joel Barber Department of Finance Florida International University Miami, FL 33199 Mortgage Mortgage Backed Security Amortization of Mortgage timet interest: im t 1 timet
More informationWaiting Times Chapter 7
Waiting Times Chapter 7 1 Learning Objectives Interarrival and Service Times and their variability Obtaining the average time spent in the queue Pooling of server capacities Priority rules Where are the
More information