A taxonomy and survey of grid resource management

1ATaxonomyandSurveyofGridResourceManagementSystemsKlausKrauter1,RajkumarBuyya2,andMuthucumaruMaheswaran1AdvancedNetworkingResearchLaboratory1DepartmentofComputerScienceUniversityofManitoba,[e-mail],[e-mail]eandSoftwareEngineering2MonashUniversityMelbourne,[e-mail]:UniversityofManitoba(TR-2000/18)andMonashUniversity(TR-2000/80)Abstract–Theresourcemanagementsystemisthecentralcomponentofnetworkcomputingsystems.Therehavebeenmanyprojectsfocusedonnetworkcomputingthathavedesignedandimplementedresourcemanagementsystemswithavarietyofarchitecturesandservices.Inthispaper,wedevelopacomprehensivetaxonomyfordescribingresourcemanagementarchitectures.Weusethistaxonomyinidentifyingapproachesfollowedintheimplementationofrealresourcemanagementsystemsforlarge-scalenetworkcomputingsystemsknownasGrids.Weusethetaxonomyandthesurveyresultstoidentifyarchitecturalapproachesthathavenotbeenfullyexploredintheresearch.Keywords:Metacomputing,Grids,ResourceManagement,Scheduling,andInternetComputing.1.IntroductionAnetworkcomputingsystem(NCS)isavirtualcomputerformedbyanetworkedsetofheterogeneousmachinesthatagreetosharetheirlocalresourceswitheachother.AGridisaverylarge-scalenetworkcomputingsystemthatscalestoInternetsizeenvironmentswithmachinesdistributedacrossmultipleorganizationsandadministrativedomains.MachinesinaGridaretypicallygroupedintoautonomousadministrativedomainsthatcommunicateviahigh-speedcommunicationlinks.AgoodoverviewofGridtechnologyispresentedin[21].Theresourcemanagementsystem(RMS)iscentraltotheoperationofaGrid.ResourcesaretheentitiessuchasprocessorsandstoragethataremanagedbytheRMS.ThesetofservicesprovidedbyaRMSvariesdependingontheintendedpurposeoftheGrid.ThebasicfunctionofaRMSistoacceptrequestsforresourcesfrommachineswithintheGridandassignspecificmachineresourcestoarequestfromtheoverallpoolofGridresourcesforwhichtheuserhasaccesspermission.ARMSmatchesrequeststoresources,schedulesthematchedresources,andexecutestherequestsusingthescheduledresources.JobsaretheentitiesthatutilizeresourcestoexecutenetworkapplicationsintheGridenvironment.Inthispaperwewillusethetermsjobandresourcerequestinterchangeably.Theprocessofmatching,scheduling,andexecutionisperformedsothatsomemetricofaggregatequalityofservice(QoS)deliveredtotherequestorismaximized[23].ThispaperdevelopsataxonomythatdescribestheRMSarchitecturesforaGrid.Inthenextsectionthemotivationforourtaxonomyisdeveloped.WefollowthisbydevelopingaclassificationofGridsystemsbasedonthedesignobjectivesandtargetapplicationssincethesegreatlyinfluencethearchitectureoftheGridRMS.Requirementsandanabstractmodelforresourcemanagementsystemsarepresentedinthenextsection.Therequirementsandtheabstractmodelareusedasabasistodevelopourtaxonomy.Wethendescribeourtaxonomyforheterogeneousnetworkcomputingsystems.A2surveyofrepresentativenetworkcomputingandGridsystemsisprovidedfollowedbysomeconclusionsandsuggestionsforfurtherresearch.2.MotivationfortheTaxonomyAdistributedcomputingschedulingtaxonomyispresentedin[26].Thistaxonomyincludesstaticschedulingtechniquesthatwedonotaddressanddoesnotconsiderscheduling,stateestimation,andresourcemodelsseparatelywhenclassifyingdynamicscheduling.Thetaxonomyfordynamicschedulingpresentedin[27]onlyconsiderstwoaspectsofresourcemanagement,schedulingandstateestimation.Inourtaxonomy,weprovideaclassificationofresourcemodelsandexamineschedulingandstateestimationatafinerlevel.Severaladvancesindistributedresourcemanagementsincethepublicationof[27]havealsobeenincorporatedintoourtaxonomy.Ataxonomyforheterogeneouscomputingenvironmentsispresentedin[29].Thetaxonomycoversapplicationmodel,targetplatformmodel,andmappingheuristicmodel.WedonotcoverapplicationmodelsordifferentiateontargetplatformmodelsinceourfocusisonissuesrelevanttothedesignersofRMSratherthanapplicationandtoolkitdesigners.Severaltaxonomiesforcharacterizingadistributedsystemarepresentedin[25]and[28].TheEM3taxonomyin[25]classifiesaheterogeneouscomputingsystembasedonthenumberofexecutionmodesandmachinemodels.Anextendedversionofthetaxonomydevelopedin[26]isalsopresentedin[25]tocharacterizetheschedulingalgorithmsinheterogeneouscomputingsystems.OurtaxonomyfocusesonRMSdesignissuesandthusdiffersfromthetaxonomiespresentedin[25].Thetaxonomypresentedin[28]providesabroadcharacterizationbasedontheexternalinterfaces,internalsystemdesign,classofhardwareandsoftwareresourcesupport,andresourcemanagementissues.OurtaxonomyofRMSismoredetailedthantheonepresentedin[28].3.GridSystems3.1.SystemElementsAresourcemanagementsystemmanagestheelementsthatcomprisethecomputingenvironment.ThusitisimportanttodescribetheelementsthatcompriseaGrid.Computingsystemsareevolvingfromrelativelyuniformmonolithicelementsuniformlyconnectedbylowspeedcommunicationlinkstohighlydiversespecializedelementsconnectedbydiversehigh-speedcommunicationlinks.Forexample,today’snetworkedsystemscontainpersonalcomputers,personaldigitalassistants,parallelcomputers,networkrouters,networkswitches,clustereds