简介:Theprocessingspeedofthecommunicationbetweennodesinaparallelprocessorhasbecomethemajorbottleneckoftheprocessor'sperformance.RDMA(RemoteDirectMemoryAccess)technologyhasdrawnmoreattentionrecentlyduetoitscapabilityoftransferringalargeramountofdata,higherspeedandreliability.4DSP(4DigitalSignalProcessing)modulecomprisedofTiger-SHARC201chipisconnectedbyLVDS(LowVoltageDifferentialSignal)circuits.ThispaperproposesageneralandreconfigurableRDMAplatformanditscorrespondingcommunicationprotocolwithalltherouteslinkedbasedonthezerocopy.TheprotocoltransfersmessageofDSPbyinterruptingofDMAandisappliedonmassiveremoteimageimpression,whichreducesmemoryneedsandworkingburdenofCPU.Theexperimentresultsshowthisplatformisefficient,flexible,andexpandableofbeingintegratedtoalargerscaleinthenextdevelopmentstages.
简介:Themulticoreevolutionhasstimulatedrenewedinterestsinscalingupapplicationsonshared-memorymultiprocessors,significantlyimprovingthescalabilityofmanyapplications.Butthescalabilityislimitedwithinasinglenode;thereforeprogrammersstillhavetoredesignapplicationstoscaleoutovermultiplenodes.Thispaperrevisitsthedesignandimplementationofdistributedsharedmemory(DSM)asawaytoscaleoutapplicationsoptimizedfornon-uniformmemoryaccess(NUMA)architectureoverawell-connectedcluster.ThispaperpresentsMAGI,anefficientDSMsystemthatprovidesatransparentsharedaddressspacewithscalableperformanceonaclusterwithfastnetworkinterfaces.MAGIisuniqueinthatitpresentsaNUMAabstractiontofullyharnessthemulticoreresourcesineachnodethroughhierarchicalsynchronizationandmemorymanagement.MAGIalsoexploitsthememoryaccesspatternsofbig-dataapplicationsandleveragesasetofoptimizationsforremotedirectmemoryaccess(RDMA)toreducethenumberofpagefaultsandthecostofthecoherenceprotocol.MAGIhasbeenimplementedasauser-spacelibrarywithpthread-compatibleinterfacesandcanrunexistingmultithreadedapplicationswithminimizedmodifications.WedeployedMAGIoveran8-nodeRDMA-enabledcluster.ExperimentalevaluationshowsthatMAGIachievesupto9.25:4speedupcomparedwithanunoptimizedimplementation,leadingtoasealableperformanceforlarge-scaledata-intensiveapplications.
简介:摘要:在网络的普及下社会发展和人们的生活都得到了更为便捷化的服务,对数据中心RDMA网络的应用需求也大幅增加,如何有效的满足在远程直接数据存取方面的需求,解决网络数据传输延迟问题成为了亟待完成的任务,对此拥塞控制算法和拥塞控制器设计有着针对性的作用。基于此,本文将对面向数据中心RDMA网络的拥塞控制算法研究及控制器设计展开研究。