NVIDIA 以太网产品加速数据科学应用.pdf

《NVIDIA 以太网产品加速数据科学应用.pdf》由会员分享，可在线阅读，更多相关《NVIDIA 以太网产品加速数据科学应用.pdf（32页珍藏版）》请在三个皮匠报告上搜索。

1、NVIDIAE2ENVIDIAETHERNETSOLUTION ACCELERATE DATASCIENCEGTC China，Dec 2020#page#WHAT IS DATA SCIENCEData Science is a blend ofvarious tools，algorithms，and machinelearning principleswith the goal todiscover hidden patterns from therawdata，MaintainData Science Life CycleCapturedata acquisition，data entr

2、y，signal reception，datoctionMaintainCaptureProcessdatawarehousing，data cleansing，datastaging，dataProcessdatamining，clustering/classification，datomodeling，datasurAnalyzeexploratory/confirmatory，predictiveonalysisativeanalysCommunicateAndlyzeCommunicatedatareporting.datoyisualization，businessnVID#page

3、#DATA SCIENCE WITH NVIDIA NBUCoptureAnlyzaintainunicateDCI+Remote Data CenteHiohhaandwidth lofrastructureAll flash stoHghdataratesensorscale Dsystem3RnT0hightopologyABod#page#Key TechnologiesSpeed and FeedRDMA and RoCEMonitoring and Management#page#SPEED AND FEED-THE NEED OF BANDWIDTHIntra-layer mod

4、el parallelData parallelIntra-layer model parallel leaves collectives exposedCommunication speedup mustAccelerating math without accelerationmatch math speedup，otherwisecommunication suffers from basic Amadahls lawproblemwe achieve little E2E speedupTypically collectives span NVLink domain onlyAllre

5、duce spans both NVLink and networking domains:bandwidth must be availabble in each#page#NVIDIAS MULTI-GPU，MULTI-NODE NETWORKING AND STORAGE IOOPTIMIZATION STACKBuild larger 8 lower latency resource poolMagnum IOUCXNCCLOpenMPINVLINK FabricGPUDirect P2PGPUDirect RDMAGPUDirect StorageInterconnectTopolo

6、gyStorageTransport日-日INFINIBANDX1同三园多BO出色售团NVLINKRoCEGPU DirectXBARNVLINKNVLINK SwitchOver RoCE orIBOn ChipGPUsGPUSNodes#page#NCCL Ring on multi nodeIB domain（across nodes）NVLINK domain(withinnode）Network domain （acrossnodes）#page#NVLINK CONFIGURATIONS热动区2药2药OBBBBBGPUGPU-6烟66666UMulti-GPUMulti-GPUCP