Reducing Communication Costs in Collective I/O in Multi-core Cluster Systems with Non-exclusive Scheduling

Abstract

Asthenumberofnodesinhighperformancecomputing(HPC)systemsincreases,collectiveI/ObecomesanimportantissueandI/OaggregatorsarethekeyfactorsinimprovingtheperformanceofcollectiveI/O.WhenanHPCsystemusesnon-exclusivescheduling,adifferentnumberofCPUcorespernodecanbeassignedforMPIjobsthus,I/Oaggregatorsexperienceadisparityintheirworkloadsandcommunicationcosts.BecausethecommunicationbehaviorsareinfluencedbythesequenceoftheI/OaggregatorsandbythenumberofCPUcoresinneighbornodes,changingtheorderofthenodesaffectsthecommunicationcostsincollectiveI/O.Therearefewstudies,however,thatseektoincorporatestepstoadequatelydeterminethenodesequence.Inthisstudy,itwasfoundthataninappropriateorderofnodesresultsinanincreaseinthecollectiveI/Ocommunicationcosts.Inordertoaddressthisproblem,weproposetheuseofspecificheuristicmethodstoregulatethenodesequence.WealsodevelopapredictionfunctioninordertoestimatetheMPI-IOperformancewhenusingtheproposedheuristicfunctions.TheperformancemeasurementsindicatedthattheproposedschemeachieveditsgoalofpreventingtheperformancedegradationofthecollectiveI/Oprocess.Forinstance,inamulti-coreclustersystemwiththeLustrefilesystem,thereadbandwidthofMPI-Tile-IOwasimprovedby7.61%to17.21%andthewritebandwidthofthebenchmarkwasalsoincreasedby17.05%to26.49%.