individualubuntu12.04

AGPUHeterogeneousClusterSchedulingModelforPreventingTemperatureHeatIslandYun-PengCAO1,2,aandHai-FengWANG1,21SchoolofInformationScienceandEngineering,LinyiUniversity,LinyiShandong,China2760052InstituteofLinyiUniversityofShandongProvincialKeyLaboratoryofNetworkbasedIntelligentComputing,LinyiShandong,China276005Abstract.
WiththedevelopmentofGPUgeneral-purposecomputing,GPUheterogeneousclusterhasbecomeawidelyusedparalleldataprocessingsolutioninmoderndatacenter.
Temperaturemanagementandcontrollinghasbecomeanewresearchhotspotinbigdatacontinuouscomputing.
Temperatureheatislandinclusterhasimportantinfluenceoncomputingreliabilityandenergyefficiency.
InordertopreventtheoccurrenceofGPUclustertemperatureheatisland,abigdatataskschedulingmodelforpreventingtemperatureheatislandwasproposed.
Inthismodel,temperature,reliabilityandcomputingperformancearetakenintoaccounttoreducenodeperformancedifferenceandimprovethroughputperunittimeincluster.
Temperatureheatislandscausedbyslownodesarepreventedbyoptimizingscheduling.
Theexperimentalresultsshowthattheproposedschemecancontrolnodetemperatureandpreventtheoccurrenceoftemperatureheatislandunderthepremiseofguaranteeingcomputingperformanceandreliability.
1IntroductionAfterGPU(GraphicProcessingUnit)wasproposedbyNVIDIAcompanyanditsbirth,ithasbeendevelopingrapidlybeyondthespeedofMoore'sLaw,itscomputingcapabilityhasbeenrisingcontinuously.
AtSIGGRAPHconferencein2003,GPGPU(General-purposecomputingongraphicsprocessingunits)wasintroduced.
GPUsgraduallyshiftedfromdedicatedparallelprocessorsconsistingoffixedfunctionalunitstoarchitectureswithprimarygeneral-purposecomputingresourcesandsecondaryfixedfunctionalunits.
GPUiscomposedofalargenumberofparallelprocessingunitsandmemorycontrolunits,itsprocessingpowerandmemorybandwidthhasobviousadvantagescomparedwithCPU.
However,GPUcannotcompletelyreplaceCPU,alotofoperatingsystems,softwaresandcodescannotrunonGPU.
GPUgeneral-purposecomputingusuallyusesCPU/GPUheterogeneousmode,CPUexecutescomplexlogicandtransactionsandothertasksunsuitableforparallelprocessing,GPUimplementscompute-intensivelarge-scaledataparallelcomputingtasks.
Withitshighperformance,lowenergyconsumptionandotheradvantages,CPU/GPUhybridarchitecturehasbeenwidelyusedingraphicsandimageprocessing,videoencodinganddecoding,matrixcomputingandsimulation,medicalindustryapplication,lifescienceresearch,high-performancecomputing,signalprocessing,databaseanddataminingandmanyotherfields.
Withtechnologyadvancesandbreakthroughs,GPUisplayinganimportantrolecurrentlyinaCorrespondingauthor:lyucyp@163.
comDOI:10.
1051/,711070011ITMWebofConferencesitmconf/201IST201707003(2017)3TheAuthors,publishedbyEDPSciences.
ThisisanopenaccessarticledistributedunderthetermsoftheCreativeCommonsAttributionLicense4.
0(http://creativecommons.
org/licenses/by/4.
0/).
large-scaleparallelcomputing.
Withtherapidincreaseofproblemscalesofvariousapplicationfields,singleGPU'scomputingcapabilityhasbecomeinsufficient,somulti-GPUandGPUclustergeneral-purposecomputinghasbecomeanewresearchhotspot.
Asanimportantapproachofhigh-performancecomputing,GPUclustershavesuchadvantagesaslowcost,highperformanceandlowenergyconsumptionforcompute-intensiveapplications.
InconstructingGPUclusters,CPUandGPUcooperatewitheachother,participateindataprocessing,andformGPUheterogeneouscluster.
GPUheterogeneousclustercanmakefulluseofhardwareresources,improveprocessingspeedandthroughput.
Ithasbecomeanimportantmeansofbigdataprocessing.
Processingbigdata,especiallyreal-timebigdatastreamneedscluster'scontinuouscomputingandprocessing,anditwillinevitablyrequirecomputer'shigh-loadandcontinuouswork,sothetemperatureofCPU,GPUandothercomponentswillcontinuetorise.
Ononehand,computingenergyconsumptionincreases,ontheotherhand,fansandairconditionersareneededforreducingtemperature,therebyincreasingcoolingenergyconsumption.
Whentemperaturerisestoacertainextent,thetemperatureofoneorsomenodeswillbetoohigh.
Thenodewithtoohightemperatureisknownastemperatureheatisland.
Theoccurrenceoftemperatureheatislandwillreducecomputingreliability,rangingfromresulterrortosystem'sparalysisandhalt.
Onceerrorsoccurincomputingresults,recomputingisneeded,resultingintimeandresourcewaste,increasingprocessingcosts.
Inthiscase,wemustreasonablydesignclustertaskschedulingschemetominimizeclusteroverallruntime,controltemperaturetoappropriaterange,preventindividualnodefromrunningsolongthatleadingtooverhightemperatureandformingtemperatureheatisland,toensurereliablecomputingresults,reduceenergyconsumptionasmuchaspossibleandachievegreencomputing.
ThispaperstudiedthetaskschedulingonGPUheterogeneouscluster,andproposedataskschedulingschemeofpreventingtemperatureheatisland.
Thescheme'smainfeaturesandadvantagesare:(1)strongrobustness.
ThestructureofGPUheterogeneousclusteriscomplex,eachnode'sconfigurationisdifferent,andthenodeisoftenchangedandadjusted.
Thistaskschedulingschemecansenseandadapttothiscomplicatedandchangeablesituation.
(2)highprocessingperformance.
Ataskisdividedintosomesub-tasks,andthentheyarescheduledtomultiplenodesforparallelprocessing.
Themainproblemisdeterminingthemodeofdivisionandtreatment.
Theconceptofcomputingscalethresholdandasymmetricpartitioningmethodareproposedinordertoadapttothediversityandheterogeneityofnodeconfiguration,improvetheparallelismandshortenthewholerunningtimeofcluster.
Thisnotonlypreventstemperatureheatislandfromoccurringbecauseofindividualnode'soverlongrunningtime,butalsoimprovesprocessingperformance.
2RelatedresearchesWiththewideapplicationofGPUheterogeneouscluster,itstaskscheduling,temperatureandheatmanagementandenergyconsumptionoptimizationhasbecomearesearchhotspot.
Manyscholarshaveputforwardvariousschedulingschemesandmethodstosolvetheproblemofenergyconsumptionandreliability.
Thishasplayedapositiveroleinreducingclusterenergyconsumptionandensuringthereliabilityofcomputingresults.
In[1]adynamictaskpartitionmethodwasproposed.
Itdividesparallelcomputingtasksaccordingtoexecutionspeedtoachievebestoverallsystemperformance.
In[2]amulti-GPUself-adaptiveloadbalancingmethodwasproposed.
GPUcanself-adaptivelyselecttaskstoexecuteaccordingtolocalfree-busystatebyestablishingtaskqueuemodelbetweenCPUandGPU.
In[3]aloadbalancingstrategythatcombinestaskpartitioningandstealingwasproposed.
IttakesintoaccounttaskaffinityandprocessordiversitytodirecttaskschedulingbetweenCPUandGPU.
In[4]feedbackcontrollingwascombinedwithmixedintegerprogramming,andtheenergyconsumptioncontrollingmodelofWebserverclusterwasconstructed.
In[5]modelpredictivecontrollingstrategywasintroducedfromglobalperspective.
Theenergyconsumptionstateischangedbyadjustingcomputingfrequencyandchangingactivestreammultiprocessor.
ThefeedbackcontrollingandrollingoptimizationmechanismDOI:10.
1051/,711070011ITMWebofConferencesitmconf/201IST201707003(2017)32areusedtopredictfuturecontrollingtoreduceredundantenergyconsumption.
In[6]theenergylossatidlestateisreducedbyaspecificnodeselectionstrategy.
CPUresourceutilizationisimprovedbytasktypedivision,combinationdistributionandDVFS.
Theaboveresearchesmainlyfocusonclustertaskscheduling,changingCPU/GPUcorevoltage,frequency,hardware-basedstatistics,andsoontodesignclusterenergyconsumptionmodel,studytaskschedulingalgorithmandachieveenergy-savingpurpose,butdonotconsidertemperaturemuch.
InGPUclustercomputing,especiallycontinuouscomputing,temperaturehasobviousrelationshipwithenergyconsumptionandreliability.
Whentemperatureistoohigh,energyconsumptionincreases,reliabilitydeclines,andtheprobabilityofresulterrorincreases.
Therefore,temperatureshouldbecontrolledinareasonablerangetominimizeenergyconsumptionunderthepremiseofensuringreliability.
Thetaskschedulingschemeproposedinthispaperdistributestasksreasonablyamongcomputingnodestopreventtheoccurrenceoftemperatureheatislandandensurethecorrectnessofcomputingresults.
3TaskschedulingmodelInGPUheterogeneouscluster,CPUandGPUallparticipateindataprocessing.
Theyareregardedascomputingunitsuniformlywhendistributingtasks.
Thecomputersinclusterarecontrollingnodesandcomputingnodes.
Thecontrollingnodecanbesimultaneouslyusedasacomputingnode.
Alltasksformaqueue.
Eachtaskisdecomposedintoseveralsub-taskstoformsub-taskqueue.
Thecontrollingnoderunsthemainschedulingprocess,Scheduler.
Eachcomputingnodehasaschedulingagentprocess,Agent.
SchedulerandAgentcooperatetofinishtaskscheduling.
ThearchitectureisshowninFigure1.
Figure1.
Taskschedulingarchitecture4TaskschedulingalgorithmandstrategyScheduleralgorithmisasfollows:Algorithm1ControllingnodeSchedulerschedulingalgorithm1.
Obtainataskfromtaskqueue2.
Obtainthehardwareconfigurationandrunningstatusinformationofeachcomputingnode3.
Determinethenumberofcomputingnodesparticipatinginparallelprocessing4.
Dividetaskintosub-taskqueueandassignsub-taskstocorrespondingcomputingnode5.
Waitfortheresultsofeachsub-task6.
Modifythestatusofcorrespondingsub-tasksandtheassociatedtasksinqueue.
7.
Reschedulesub-tasksthattimedoutorrequestedtotransfer,modifycorrespondingstatus8.
Goto1Foreachcomputingnode,thesub-tasksthatcontrollingnodedispatchestoitformaqueue.
TheschedulingalgorithmofAgentoncomputingnodeisasfollows:ControllingnodeSchedulerComputingnodeAgent…Sub-taskqueueTaskqueueSub-taskqueueComputingnodeAgentSub-taskqueueComputingnodeAgentSub-taskqueueDOI:10.
1051/,711070011ITMWebofConferencesitmconf/201IST201707003(2017)33Algorithm2ComputingnodeAgentschedulingalgorithm1.
Obtainasub-taskfromthesub-taskqueueoflocalnode2.
Assignthesub-tasktolocalnodeforprocessing3.
Waitfortheresulttobereturnedfromlocalnode4.
Reportresultstocontrollingnode(completion,timeout,orrequestingtransfer)5.
Goto14.
1AcquiringhardwareconfigurationinformationSchedulerfirstobtainsthehardwareconfigurationinformationofeachnodeincluster.
Theinformationcanbemanuallycreatedinadvanceandsavedinfile.
Whenclusterisstarted,Schedulerloadsclusterhardwareconfigurationinformationfile.
Itpollseachcomputingnode,AgentrespondstothepollandreportshardwarechangeinformationtoScheduler.
Or,AgentreportshardwarechangeinformationtoScheduleractively.
ThenSchedulermodifiescluster'shardwareconfigurationinformation.
Inthisway,controllingnodecangraspthelatestchangesinclusterhardwareconfiguration,avoidingunnecessaryacquisitionandreportingofhardwareconfigurationinformation,thusadaptingtoactualhardwareconfigurationchangesandreducingnetworkcommunicationoverhead.
4.
2SchedulingstrategyComputingscaleisusedtomeasuretasksize.
Computingscaleisthenumberofinstructionstobeexecutedortheamountofdatatobeprocessedtocompletethetask.
Ataskcontainsparallelizableandnon-parallelizablepart.
SupposethecomputingscaleofataskisT,TTs+Tp,Tsisthecomputingscaleofnon-parallelizablepart,andTpisthecomputingscaleofparallelizablepart.
LetTtbethecriticalvalueofthecomputingscaleofparallelizablepart,thentaskschedulingstrategyisasfollows:(1)0≤Tp(2)Tp≥Tt,thetaskhasparallelizablepartanditreachesacertainscale.
Theparallelizablepartoftaskisdividedintosmallersub-tasks,manycomputingunitswithstrongestcomputingcapabilityareselectedfromidleprocessingunitstoprocessthem.
4.
2.
1DeterminingTtandthenumberofcomputingunitsTheprocessingcapabilityisassumedtobeCswhentaskisprocessedseparatelybyasinglecomputingunit.
Withoutlossofgenerality,assumingthatwhenparallelprocessing,thenumberofcomputingunitsparticipatinginprocessingisn,theirprocessingcapabilityisallCp.
Inordertoobtainbetterperformance,then:ttspTTQCnC≥+(1)WhereQistheadditionaltimeoverheadrequiredforparallelprocessing,includingparallelcomputingpreparation,resultmerging,synchronization,networktransmission,andsoon.
Atthesametimeinordertoensurehighprocessingefficiency,then:tpTQnC≥(2)Solvingtheinequalitygroupconsistingofabovetwoinequalitieswillget:DOI:10.
1051/,711070011ITMWebofConferencesitmconf/201IST201707003(2017)34max{,}pstppsnCCQTnCQnCC≥(3)ThevalueofQcanbedeterminedexperimentallyorbyaccumulatinghistoricalempiricaldata.
Cpcanbetakenastheaverageofthecurrentcomputingcapabilityofallcomputingunits,andCsistheaverageofthecurrentcomputingcapabilityofallCPUsincluster.
Letmax{,}psmppsmCCQTmCQmCC=,m=1,2,…,Nidle,Nidleisthenumberofallidlecomputingunitsincurrentcluster.
Inordertoincreasetheparallelizationdegreeoftaskprocessing,changefromNidleindescendingmanneruntilthefirstnumberkwhichletsTp≥Tkisfound,thenkisthenumberofunitsinvolvedinparallelprocessing,thealgorithmtodetermineitisasfollows:Algorithm3Determiningthenumberofparallelprocessingunits1.
getNidle2.
i←Nidlek←13.
ifi≤1goto74.
Ti←max{,}psppsiCCQiCQiCC5.
ifTp≥Tithenk←igoto76.
i←i-1goto37.
endIfk<2orqualifiedkvaluecannotbefound,thetaskishandledbyoneCPUandnotscheduledinparallelmanner.
4.
2.
2PartitioningparallelpartAssumingthatthecurrentcomputingcapabilityofkcomputingunitsinvolvedinparallelcomputingisC1,C2,….
,Ck,thescaleofsub-tasksassignedtoeachprocessingunitisT1,T2,…,Tk,thenthetimetocompletethetaskis:1212max{kkTTTCCCt=(4)WhereT1+T2+…+Tk=Tp.
ItcanbeproventhatwhenipCTiCT=(i=1,2,…,k),tisminimumandpTCt=,whereC=C1+C2+…+Ck.
Therefore,theproportionofallocatedtasktototaltaskscalebeingequaltotheratioofthecurrentcomputingcapabilityofthecomputingunittothesumofthecurrentcomputingcapabilitiesofallcomputingunitsparticipatinginparallelprocessing,caneffectivelyreduceoverallprocessingtime,balanceload,andavoidthecasethatsomeunitsareidleandsomeunitsrunforlongtimeandcausetemperatureheatislandstooccur.
4.
3EstimatingcurrentcomputingcapabilityCurrentcomputingcapabilityisrelatedtoitsownhardwareconfigurationandhardware'scurrentstateofutilization.
Forcomputingunitswithsameconfiguration,thebusieroneshavestrongercurrentcomputingcapabilitythantheidleones.
Byreferencing[7]andimproving,thecurrentcomputingcapabilityisestimated.
ForanycomputingnodeNi,consideritsfivehardwareconfigurationparameters:CPUfrequencyrate_cpui,memorysizememi,cachesizecachei,GPUfrequencyrate_gpui,GPUmemorysizemem_gpuiandfivestateparameters:CPUutilizationutlz_cpui,memoryDOI:10.
1051/,711070011ITMWebofConferencesitmconf/201IST201707003(2017)35utilizationutlz_memi,cacheutilizationutlz_cachei,GPUutilizationutlz_gpui,GPUmemoryutilizationutlz_gpumemi.
ThecurrentcomputingcapabilityofnodeNiis:1122334455iCkQkQkQkQkQ5)k1,k2,k3,k4andk5representsthelevelproportionweightofinfluenceonnodecurrentcomputingcapabilityofCPU,memory,Cache,GPUandGPUmemoryrespectively.
Theirsumis1.
Q1-Q5respectivelydenotesCPUcurrentcapability,memorycurrentcapability,cachecurrentcapability,GPUcurrentcapabilityandGPUmemorycurrentcapabilityafternormalizationofnodeNi.
Q1iscalculatedas:11_(1_)(_(1_))iiNjjjratecpuutlzcpuQratecpuutlzcpu=*=*∑(6)TheformulasforQ2-Q5aresimilar.
Foracertainnode,Ci,Q1,Q2,Q3,Q4andQ5canbedeterminedexperimentally,andthentheapproximatevalueofk1,k2,k3,k4andk5canbedeterminedbyregressionmethod.
5ExperimentandanalysisTheschemeproposedinthispaperwasverifiedexperimentally.
Twoexperimentswereconductedonsamecluster.
Theexperimentprogramis:Somerelativesoftwares(suchasCPU-Z,HWMonitor,CoreTemp,etc.
)wereusedtomeasuretemperaturesofCPUandGPUofeachcomputingunitatdifferenttimeduringcluster'srunning,andthetemperaturecurveofeachcomputingunitwasdrawnaccordingtothem.
SevencomputerswereusedtoconstituteGPUheterogeneouscluster.
Fiveofthemhavetheconfiguration:modelisLenovoErazerX700,memoryis16G,CPUisInteli7-3930k,GPUisNVIDIAGTX660i,operatingsystemisUbuntu12.
04LTS,clusterenvironmentishadoop2.
2.
0,JavaversionisJDK1.
7.
Theothertwohavelowerconfiguration:CPUisIntelPentium(R)Dual-CoreE53002.
60GHz,memoryis4G,GPUisNVIDIAGeForce9400GT,operatingsystemisWindows764-bitUltimate.
TheexperimentaldataistaxiGPSdataanddatageneratedcontinuouslybyloadrunner.
5.
1ConventionalschedulingmethodFirstly,conventionalschedulingmethodwasused.
Onlytaskbalancedschedulingwasconsidered,regardlessoftemperaturechanges.
Every1minutetemperaturewassampledonce.
TheresultisshowninFigure2,whereC1,C2,.
.
.
,C7iseachcomputingnode.
0123456789101135363738394041424344454647484950Temprature(oC)SamplingIntervalC1C2C3C4C5C6C7Figure2.
TemperaturechangeinconventionalschedulingmethodDOI:10.
1051/,711070011ITMWebofConferencesitmconf/201IST201707003(2017)36TheclusterprocessestaxiGPSdatafirstly.
Thedataamountislarger,butbecauseitishistoricaldata,itdoesnottakelongtimetoprocessit.
Temperatureandpoweraremeasuredwithmeasuringinstruments,temperaturesofCPU,GPUandsoonaremonitoredwithsoftwares.
ItisfoundthatthetemperatureandpowerofCPUandGPUareincreasingduringprocessing,butthetaskhasbeenfinishedbeforetemperaturerisestothesetthreshold,andtheproblemoftemperatureheatislandandreliabilitydoesnotoccur.
Thensimulationdatathatloadrunnersoftwarecontinuestogenerateisprocessed.
Atthistime,CPUandGPUtemperaturecontinuestorise,energyconsumptioncontinuestoincrease.
Afteracertaintime,temperatureexceedsthethresholdandtemperatureheatislandisformed,computingresulterroroccurs.
Thedifferencebetweenthelowestandhighesttemperaturesofvariouscomputingnodesisabout12°C.
5.
2SchedulingschemeproposedinthispaperInthesecondexperiment,thesameexperimentalenvironmentanddatawereused,buttheschedulingschemepreventingtemperatureheatislandproposedinthispaperwasused.
Duringprocessingtask,temperatureiscollected.
TheresultisshowninFigure3.
0123456789101135363738394041424344454647484950Temprature(oC)SamplingIntervalC1C2C3C4C5C6C7Figure3.
TemperaturechangeinschedulingmethodpreventingtemperatureheatislandTheresultofprocessingtaxiGPSdataissimilartotheprevious,buttheresultshowsthatthedifferenceoftemperatureandenergyconsumptionofeachnodetendstodecrease.
Thisshowsthatthisschemeismoretime-balancedintaskschedulingtopreventtemperatureheatislandfromoccurringandguaranteeoverallstability.
Datastreamsgeneratedcontinuouslybyprogramareprocessedbycluster.
Itwasfoundthat,althoughthetemperatureofCPUandGPUincreased,thetemperatureandpowerdidnotincreasecontinuouslywhentemperaturerisednearlytothresholdvalue,andnotemperatureheatislandandcomputingerroroccurred.
Whendatasupplyamountwasincreased,thephenomenonthattemperatureandpowerincreasedidnotoccur.
Thisshowsthattheschedulingschemetrystobalancerunningtime,inhibittheincreasingoftemperatureandenergyconsumptiontopreventtemperatureheatislandfromoccurring.
Thedifferencebetweenthelowestandhighesttemperaturesamongvariouscomputingnodesisabout9°C.
Analyzingaboveexperimentalresults,itisshownthat,ifconventionalmethodisadopted,thetemperatureofeachcomputingnodeincreasescontinuouslywiththeprocessingoftask,thetemperatureofsomenodesexceedsthreshold,andthetemperaturefluctuatesgreatly.
However,whentheschedulingmethodproposedinthispaperisused,temperatureisalsorising,butbecausetaskdivisionmakesnoderunningtimebeconsistentasfaraspossible,therangeoftemperaturefluctuationissmall,theoveralltemperaturechangeisrelativelycalm,thusitisavoidedthatthetemperatureheatislandoccurs.
DOI:10.
1051/,711070011ITMWebofConferencesitmconf/201IST201707003(2017)37ConclusionTheGPUheterogeneousclustertaskschedulingschemeproposedinthispaperavoidslongrunningtimeofindividualnodesasfaraspossible,preventstemperatureheatislandfromoccurring,guaranteescomputingreliability,controlsenergyconsumptioninacertainrange,andalsoconsiderstheconstraintsamongtemperature,reliability,performanceandenergyconsumption,minimizesenergyconsumptionorimprovesprocessingspeedasfaraspossibleunderthepremiseofensuringreliability.
ThemainworkofnextstepistostudyhowGPUheterogeneousclusterperceivesandpredictsclustertemperatureanditsvariation,andapplyittoclustertaskscheduling.
AcknowledgmentsThisresearchprojectissupportedbythejointspecialprojectofShandongProvincialNaturalScienceFoundation(ProjectNo.
:ZR2015FL014)andthespecialprojectofShandongProvincialIndependentInnovationandAchievementTransformation(ProjectNo.
:2014ZZCX02702).
References1.
C.
Q.
Yang,F.
Wang,Y.
F.
Du,etal.
AdaptiveoptimizationforpetascaleheterogeneousCPU/GPUcomputing.
The2010IEEEInt'lConf.
onClusterComputing.
(2010)2.
L.
Chen,O.
Villa,S.
Krishnamoorthy,G.
R.
Gao.
Dynamicloadbalancingonsingle-andmulti-GPUsystems.
The2010IEEEInt'lSymp.
onParallel&DistributedProcessing(IPDPS).
(2010)3.
E.
Hermann,B.
Raffin,F.
Faure,T.
Gautier,J.
Allard.
Multi-GPUandmulti-CPUparallelizationforinteractivephysicssimulations.
The16thInt'lEuro-ParConf.
onParallelProcessing:PartII(Euro-Par2010).
Berlin,Heidelberg:Springer-Verlag.
(2010)4.
L.
Bertini,C.
B.
Julius,D.
Mosse.
Poweroptimizationfordynamicconfigurationinheterogeneouswebserverclusters.
JournalofSystemsandSoftware,83(4):585-598.
(2010)5.
H.
F.
Wang,Y.
P.
Cao.
GPUPowerConsumptionOptimizationControlModelofGPUClusters.
ActaElectronicaSinica,43(10):1904-1910.
(2015)6.
H.
P.
Huo,X.
M.
Hu,C.
C.
Sheng,B.
F.
Wu.
Anenergyefficienttaskschedulingschemefornode-layerheterogeneousGPUclusters.
ComputerApplicationsandSoftware,30(3):283-286.
(2013)7.
H.
Liu,J.
G.
Wang,Z.
Z.
Ge,etal.
Self-learningLoadBalancingSchedulingAlgorithmforGPUHeterogeneousCluster.
JournalofXi'anShiyouUniversity,30(3):105-111.
(2015)DOI:10.
1051/,711070011ITMWebofConferencesitmconf/201IST201707003(2017)38