Researchbackgroundandsignificance
Sincethefirstelectroniccomputercameoutin1946,computerscienceandtechnologyhaveundergonetremendouschanges.Thedevelopmentofcomponentsisundoubtedlyanimportantdrivingforceforthedevelopmentofcomputertechnologyinthepastfewdecades,anditisalsothemostactiveandinfluentialfactor.Althoughcomponentmanufacturingtechnologyisconstantlyimprovingandintegratedcircuitdesignmethodsareconstantlybeingupdated,thespeedofcomponentswillnotcontinuetoincreaseatthepreviousrateinthenexttenyears.AccordingtoSIA'sforecast,CMOStechnology,thecurrentleadingtechnologyofintegratedcircuitmanufacturing,mayreachitslimitaround2010.Ofcourse,theuseofnon-CMOSmanufacturingtechnology,suchasquantumtechnology,neuralnetworktechnology,etc.,canincreasethespeedofcomponentstoanewlevel.However,ifthereisnomajorbreakthroughintheoryandtechnology,thespeedofhigh-levelcomponentswillultimatelybelimitedbythespeedoflightandthediameterofhydrogenatoms.Ontheotherhand,humanbeingshavenever-endingdemandsontheperformanceofcomputersystems,andtheproblemsthatneedtobesolvedbycomputersintheapplicationfieldarebecomingmoreandmorecomplexandlargerinscale.Parallelprocessingtechnologyisproposedbasedontheaboveneeds,anditisakeytechnologytoimprovetheperformanceofcomputersystems.Inthepast40years,scientistshaveproposedmanyparallelprocessingtechnologiesanddesignedvariouscomputersystems,suchasSMPandMPP,whichhavegreatlyimprovedtheoverallspeedofthecomputersystem.ButbecausetheprocessorsofSMPandMPPareoftenisomorphic,andapplicationtasksmayincludecomplexinherentparallelism(suchasSIMD,MIMD,andvectors,etc.),differentparallelismmaybesuitablefordifferentprocessorstorun.InSMPandMPPsystems,mostapplicationscanonlyobtainasmallpartoftheirpeakperformance.Traditionalhomogeneouscomputingsystemsusuallyuseaparallelprogrammingmodelonagivenmachine,whichcannotmeettheapplicationrequirementsofmorethanonetypeofparallelism.Anysingletypeofmachineoftenspendsmostofitstimeoncodethatisnotsuitableforitsexecution.Researchanddevelopmentofcomputingsystemsthatsupportavarietyofinherentlyparallelapplicationsisamajorchallengefacingus.Itspurposeistoimprovecomputingefficiencyandobtainmostofthepeakperformanceofapplications.Heterogeneouscomputingsystemssupportapplicationswithmultipleinherentparallelism,whichcanimprovetheratiooftheactualexecutionperformanceoftheapplicationtoitspeakperformance.Aheterogeneoussystemusuallyincludesseveralheterogeneouscomputingnodes,interconnectedhigh-speednetworks,communicationinterfaces,andprogrammingenvironments.Heterogeneoustreatmentwcanbedividedintocoarse-grainedheterogeneoustreatmentandfine-grainedheterogeneoustreatment.Fine-grainedheterogeneityisallocatedandscheduledaccordingtoinstructions,andthecommunicationoverheadistoolarge;whilecoarse-grainedheterogeneityisscheduledaccordingtoinstructionblocks,andthecommunicationoverheadcanbetolerated,whichisakindofheterogeneousprocessingwithdevelopmentprospects.Coarse-grainedheterogeneousprocessingisalsocalledmeta-computing,andthesystemthatsupportsmeta-computingiscalledmeta-computingsystem.Referredtoasmeta-system.
AsapopularInternetnetwork,theInternethasawidedistributionrangeandrelativelyhightransmissionspeed.Fastandotherfeatures.Theresearchofmeta-computingsystemundertheInternetenvironmenthasimportantpracticalsignificance.Ontheonehand,themeta-computingenvironmentenablespeopletogatherscatteredcomputingpowertoformsupercomputingcapabilitiestosolvemajorscientificresearchandtechnicalapplicationproblemssuchasvirtualnuclearexplosions,newdrugdevelopment,weatherforecasting,andenvironmentalissues;ontheotherhand,itenablespeopletosharewideareanetworks.Theheterogeneousresourcesinthenetworkmakefulluseofvariousresources.mycountryhasastrongdemandforahigh-performancecomputingenvironment,whichisnotonlyreflectedinthehigh-endscientificcomputing,butalsointhedemandforhigh-endinformationservicesinvariousindustries.
Systemcomposition
Agoodmeta-computingsystemismainlycomposedofthreeparts:oneisassimpleandreliableaspossibletousetheinterface;thesecondisaresourcemanagementsystem;thethirdisgoodProgrammingmodelandefficientandreliableruntimeenvironment.Amongthem,theresourcemanagementsystemisatthecore.Itisresponsibleforthedescription,change,allocationandprotectionofvariousresourcesintheentiresystem,andworkscloselywiththeupper-levelruntimesystemtoensurethenormaloperationoftheentiresystem.Theresourcemanagementsystemneedstoprovidepowerfultoolsorinterfacesforsystemadministratorsatdifferentlevelstocontrolandprotecttheseresources.
Researchstatus
Inrecentyears,foreigncountrieshavecarriedoutsubstantiveresearchworkinheterogeneouscomputingenvironments,andtheresearchworkintheUnitedStatesisfaraheadofothercountries.Thepre-researchbegantenyearsago,andtheconceptualdesignwasformallyproposedin1992.Nowheterogeneouscomputinghasgrownintoanewresearchhotspotinhigh-performancecomputing.VariousgovernmentdepartmentsintheUnitedStateshaveinitiatedseveralmajorresearchanddevelopmentprojectsrelatedtometa-computing.Theseinclude:(1)ThePAC1project(Gridproject)oftheUSNSFissupportedbytheNaturalScienceFoundationoftheUnitedStates.ItisacontinuationoftheSupercomputingCenterprojectinthe1980s.ItsgoalistotesttheinformationinfrastructureoftheAmericansocietyinthenext10-15years.ThePACIprojectsupportstwocomputingalliances,namelyNCSAledbytheUniversityofIllmoisSupercomputerApplicationCenterandNPACIledbytheUniversityofSanDiegoSupercomputerCenter(UCSD),eachwithmorethan100nodes.(2)TheAS-CIprogramoftheUSDepartmentofEnergyisaprogramdevelopedfornuclearweaponsstorageandvirtualnuclearexplosionresearch.ThethreenationallaboratoriesofLosAlamos,SandiaandLivermoreserveasthesupernodesoftheproject.Thegoalistoreach100trillionfloatingpointcalculationspersecondby2004tomeettheneedsoffullsystem,fullphysics,andfullthree-dimensionalsimulation..(3)Thehigh-performancecomputingmodernHPCMPnetworkoftheUSDepartmentofDefenseisahigh-performancecomputinginfrastructureestablishedbytheUSDepartmentofDefense,whichisusedbymorethan4,000militaryscientificandtechnologicalpersonnelacrossthecountryfornationaldefenseresearchanddevelopment.Thegridincludes4maincomputingcenters,13sub-centers,anddozensofremotecenters.Mostofthecomputingpowerisintegratedinthemaincenter.Thefourmaincenterswereequippedwith23setsofcomputersystemsin1998,withspeedsrangingfrom2billiontimesto500billiontimes.ThisinfrastructurehasgreatlychangedthefaceofdefensescientificresearchintheUnitedStates.
Afterseveralyearsofresearch,meta-computingalsohasacertainfoundationabroad,andprototypesorexperimentalsystemswithdifferentdesigngoalsareproduced.I-WayandGlobusaremoretypicalmeta-computingexperimentalsystems.Theirgoalistoprovideremoteuserswithavisualinterfacetothehigh-performancecomputingplatform.Itshigh-performancecomputingplatformisconnectedbytheexistingparallelcomputingsystemsofseveralnationallaboratoriesanduniversitiesintheUnitedStatesthroughahigh-speednetwork.Globusisthebasiccomponentofthemeta-computingsystem.ItprovidesC++libraryexpansionfunctionsforusers,andsupportsservicessuchasremotefiles,remoteI/O,resourcemanagement,andNexuscommunicationmanagement.Globusmakeseffectiveuseofexistinghigh-performancecomputingsystemsthroughcollaborativeoperationsandextendedexecutionprogramsinawide-areadistributedenvironment.However,asthedemandforservicesincreases,thelackofaunifiedprogramminginterfaceandmodelmakesitrelativelydifficultforuserstouseGlobus.TheLeglonsystememphasizestheefficientoperationofparalleltasksonworkstationclusters(orclusters).LegionisimplementedusingC++'sextensiblelanguageMentat,andsupportsargument-basedusersecurityservices,object-orientedvirtualfileservicesandMMPScommunicationservices.Globeisameta-computingsystemdesignprojectdevelopedatVrijeUniversityintheNetherlands.ItsharesmanyofthesamedesigngoalsasGlobusandLegion.Globeisamiddlewaresystemrunningontheexistingmainoperatingsystemandcommonnetworkprotocols.Globedoesnotprovidepowerfulinformationsecurityservicesandusersecuritydemonstrationservices.InChina,researchonhigh-performancecomputingalsohasaconsiderablefoundation,andtheNationalHigh-performanceComputingCenterhasbeenestablished.However,theresearchonmeta-computingisstillinthepreliminaryresearchstage,andthereisstillabiggapbetweenitandforeigncountries.
Keytechnology
1.Networktechnology
Researchtoimprovenetworkperformance,reliabilityandfaulttolerance,andprovideahigh-bandwidthcommunicationsystemfordesktopcomputers.Inaddition,high-speedcommunicationprotocolsandcommunicationservicequalityassurancetechnologiesareveryimportanttotheresearchofmeta-computingsystems.
2.Efficientandscalabledistributedparallelcomputingsystem
Inordertoenableuserstotransparentlyusenetworkresources,itisnecessarytoestablishanefficientandscalabledistributedparallelcomputingenvironment,andprovidevariousadvancedService.Theunderlyinginfrastructureshouldinclude:resourcemanagementandallocation,communicationsecuritymechanism,fault-tolerantmechanism,unifiedresourceinformationservice3.User-orientedprogrammingmodel
Inordertomakefulluseofthecharacteristicsofthemeta-computingsystem,itisnecessarytodesignaneffectiveprogrammingmodelforthemeta-systemtomakefulluseofresources.Object-orientedtechnologywillbeoneofthemainimplementationtechnologies.DistributedobjecttechnologybasedonCORBA,JavaorWebtechnologiesissuitableforbuildingcomplexcomputingobjectsandmobilecomputingobjects.
4.Visualizationtechnology
Visualizationtechnologyinmeta-computingusesgraphics,imagesandothermeanstodescribethebehavior,statusandstructureofresourcesandtasksinthemeta-computingsystem,whichistheresourcemanagement,jobschedulinganddatatransmissionofthemeta-computingsystemprovidehelp.
5.Large-scalescientificapplicationcomputingresearch
Thepurposeofthemeta-computingsystemistomakefulluseofvariousresourcesonthenetworktosupportlarge-scaleparalleldistributedcomputing.Therefore,theresearchonapplicationshouldalsobecarriedoutatthesametime,andthefunctionandtechnicalrequirementsofthemeta-computingsystemshouldbeputforwardbyapplication-driven,andtheeffectivenessofitstechnicalapproachandtechnicalrealizationshouldbeverified.
6.EfficientprogramcompilationmodelandexecutionengineInthetraditionalcase,theapplicationprogramneedstobecompiledintoexecutablecodeforacertainplatformbeforeexecution.Theplatformrelevanceofthesenativecodesdeterminesthattheycannotbeexecutedacrossplatforms.Thisrestrictstheexecutionofprogramsandtheuseofresourcesinaheterogeneousenvironment.Thesolutiontothisproblemistouseplatform-independentintermediatecodes,suchasbytecodesformedaftercompilationofJavaprograms.Theseintermediatecodescanbeeasilystoredandspreadonthenetwork.Sincetheintermediatecodeisusuallyinterpretedandexecuted,theexecutionefficiencyisrelativelylow.Inordertosolvetheseproblems,itisnecessarytostudythecompilationmodelandexecutionengineoftheprogramtosolvethecontradictionbetweenplatformindependenceandexecutionefficiency.
Researchtrendsandgoals
Asasupportplatformfordistributedandparallelcomputing,meta-computingsystemshavemanycommoncharacteristicsofdistributedenvironmentsandparallelcomputing.Theexistingdesigntechnologyofdistributedenvironmentandparallelsystemprovidesastrongfoundationfortheresearchofmeta-computingenvironment.Theresearchofparallelcomputingmodel,thematurityofparallelexecutionenvironmentandparallelcompilationtechnology,andtheuseofgeneralparallelprogrammingenvironmentMPIandPVMprovideastrongguaranteeforthedevelopmentofmeta-computing.Theresearchofdistributedfilesystemandnetworkfilesystem,especiallytheresearchofOSF/DCEandCORBA,providespowerfulguidanceandreferencefortheresearchofmeta-computingsystem.Butthedesignprinciplesofmeta-computingaredifferentfromthem.Atpresent,DCEandCORBA,whichhavebecomeindustrystandards,arealsodistributedcomputingtools,buttheyonlysupportdistributedcomputinganddonotsupportparallelcomputing;theyuseaclient/servermodelinsteadofaparallelcomputingmodel;theyaremainlydevelopedforbusiness,Itisnotforhighperformancecalculations.Themeta-computingenvironmentisdifferent.Ithastheadvantagesofdistributedcomputingandparallelcomputing.Itcanbeusedforbusinessdevelopmentaswellashigh-performancecomputing.
Theresearchofmeta-computingsystemswilltendtoawide-areaheterogeneousenvironmentbasedontheInternet,andsupportthousandsofcomputernodesbelongingtodifferentmanagementdomainstoparticipateincomputingandresourcesharing.Inthefuture,thedevelopmentofameta-computingsystemwithexcellentperformancewillfollowthedesigngoalsasfollows:
Provideasingleimagespace:Userscantransparentlyoperateandusethemeta-computingenvironmentthroughasingleimagespace.Thesingleimagespaceshieldstheboundaryofthehardware,realizestheconversionofheterogeneousdataformats,providesvirtualfileservices,andmapsthescatteredhostnodesintoaunifiedentryvirtualmachine.
Ensureusersiteautonomy:MetacomputingnodesareheterogeneousnodesontheInternet,andthesenodesparticipatinginthemetasystemmaybelongtodifferentmanagementdomains.Inordertoensurethelegitimaterightsandinterestsofnodeowners,meta-computingmustensuretheautonomyofeachnode.Whendesigningameta-computingsystem,itisimportanttoconsidertheminimuminstallationofeachnode.Itisveryimportanttorealizethinservices,anditisalsothekeytotheexpansionofthesystem.
Providepowerfulsecurityservices:Becausethenodesofthemetasystemaredecentralizedandliquid,thesecurityfunctionsofeachtypeofnodearedifferent,soitismoredifficulttomanagethenodesofthemetasystem.Inthedesign,thesecurityofeachnodemustbestrengthened,atleastnottodestroyorreducetheoriginalsecurityfeaturesofthenode.
Supporthigh-efficiencyfault-tolerantfunction:Thedesignofmeta-computingsystemshouldconsiderthesystem'sfault-tolerantfunction,systemexceptionhandlingfunction,andsoftwaredebuggingandtroubleshootingfunctions.
Supportforunifiedresourcemanagement:Duetothevarietyofnodesinaheterogeneousenvironment,themetasystemmustsupportmultiplearchitectures.Inthemeta-system,itisbetternottohaveacentralnode,andtheentiresystemisfullydistributed.Throughmiddlewaretoensuretheinteroperabilityofheterogeneoushardwareandsoftware,theunifiedmanagementandschedulingofresources.
Supporttherationalschedulingofapplicationsonthemeta-system:Inthemeta-system,inadditiontoassigningtaskstomachines,theheterogeneouscharacteristicsofthemachinesmustalsobeconsideredtoachieveselectivemappingandschedulingoftasks.Thereareoftendependenciesamongsubtasksdecomposedbyanapplication,andtheschedulingofdependenttasksonheterogeneousmachinesisthefocusofmeta-systemdesign.
Meta-computingisarelativelynewresearchfield,andtherearestillsomedifficultiesindevelopinganefficientandeasy-to-usemeta-computingsystem.Ameta-computingsystemmusthaveafriendlyhuman-computerinteractioninterface,aunifiedprogrammingenvironment,areliablecommunicationprotocolandanefficienttaskschedulingalgorithm,justlikeahomogeneoussystem,andprovideaninterfacewithpopularparallelcodingsoftwaresuchasPVM.