文档介绍：High-PerformanceScienti?cDataManagementSystem JaechunNo RajeevThakur SejongUniversity ArgonneNationalLaboratory Seoul,RepublicofKorea Argonne,IL60439,USA AlokChoudhary NorthwesternUniversity Evanston,IL60208,USA Abstract Manyscienti?capplicationshavelargeI/Orequirements,intermsofboththesizeofdataandthe numberof?,storage,ess,andanalysisofthisdatapresentan ,twodifferentsolutionshavebeenusedforthistask:?leI/O ?les ,?exible,andpowerfulbutdonotperform ,called Scienti?cDataManager(SDM),binesthegoodfeaturesofboth?leI/ providesahigh-levelAPItotheuserand,internally,usesaparallel?lesystemtostorerealdata(using variousI/OoptimizationsavailableinMPI-IO)andadatabasetostoreapplication- ordertosupportI/Oinirregularapplications,SDMmakesextensiveuseofMPI-IO’snoncontiguous collectiveI/,SDMusestheconceptofahistory?letooptimizethecostofthe SDMandpresentperformanceresultswithtworegularapplications,ASTRO3DandanEulersolver,and withtwoirregularapplications,aCFDcodecalledFUN3DandaRayleigh-Taylorinstabilitycode. Keywords:scienti?cdatamanagement,parallelI/O,MPI-IO,database,metadata Proposedrunninghead:High-PerformanceScienti?cDataManagementSystem ThisworkwassupportedinpartbytheMathematical,Information,putationalSciencesDivisionsubprogramofthe putingResearch,,underContractW-31-109-Eng-38,andinpartby aWork-for-,underNSFCooperativeAgreement#ACI-9619019. 1 Introduction Manylarge-scalescienti?cexperimentsandsimulationsgenerateverylargeamountsofdata[2,9](onthe orderofseveralhundredgigabytestoter