《鲁蔚征-面向AI应用的文件系统与评测基准.pdf》由会员分享,可在线阅读,更多相关《鲁蔚征-面向AI应用的文件系统与评测基准.pdf(16页珍藏版)》请在三个皮匠报告上搜索。
1、鲁蔚征 中国人民大学Weizheng LuRenmin University of ChinaFile Systems and Benchmark Tools for AI StorageOutlinesML/AL WorkloadsDistributed File Systems for AIBenchmark Tools&ResultsML/AI WorkflowsTrainingPreprocessingData Loadersmall or big filesTFRecord or raw fileCheckpointInferenceTypical AI DatasetsImages
2、 and VideosImageNet:14M small filesyoutube-8M:1.53TBTextC4The PileFalcon-RefinedWebRecommendation SystemsML/AI Training CharacterizationsBig DataSame Data Multiple Training Jobshyper-parameter tuningdifferent model archietcure,different parameters(learning rate,loss function)Fluid,Microsoft QuiverCo
3、mpute Nodes SSD or RAMglobal storage pute nodes local storageLustre PCC,Alluxio,JuiceFS,Preprocessing:Falcon-RefinedWeb2.8TB extractedFalcon 180B LLMPipelinesPreparationFilteringDeduplicationhttps:/arxiv.org/abs/2306.01116Timeline of Distributed File Systemsl20032003Lustre Initial Releasel20062006Ha
4、doopHDFSl20062006AmazonIntroduced S3ll20CephShown on OSDI06 and SC06l20072007BeeGFSbeta on ISC07Timeline of Distributed File Systems(cont.)l20122012AlexNetImageNetl20142014Alluxiol20162016YRCloudFilel20172017JuiceFSl20202020GPT3l20222022ChatGPTl20142014Wekal20152015Kubernetes1.0 Relea
5、sedPOSIX or notProsConsFile SystemsPOSIXdevs,ops,and software rely on POSIXPortableoverheadLustre,JuiceFSnon-POSIXlow costlimited abilitiesadditional codeHDFS,S3Case Study:AlluxioGoal:Data OrchestrationUnder StoreS3,HDFS,POSIX FSWorkersCache on RAM or SSDClientfuseCase Study:JuiceFSGoal:high-perform
6、ance,cloud nativedata is chunked on S3,HDFSmetadata is in redis,MySQL,PostgreSQLclient mount fuseCommon Benchmark Toolstraditional tools IOPS&BandWidth(BW)fiomdtestiozonereal-world workloadsML benchmarkMLPerfMLPerfa suite contains mainstream AI workloadsMLPerf TrainingMLPerf StorageMLPerf Storagesyn
7、thetic random datasimulate AI acceleratorsMLPerf Training WorkloadsBenchmark ResultsLustre+all flashLustre+HDDJuiceFS+S3xfs+local SSDfio IOPS READ2700k20k14k40kfio BW READ30GB/s12GB/s2.6GB/s0.9GB/sImageNet PyTorch1600s1640s1570s1570sLLM checkpoint(LLaMA 70B)1 min10 minMLPerf Storage UNet3Dfio results are based on a script file from DDNall flush:24*NVMe(DDN AI400)+IBHDD:Metadata-7*SSD,Object-50*HDD(DDN 7990)+IBJuiceFS:Metadata redis,Object-S3+10Gb EthDiscussionWorkload FilesystemBenchmark Result Real PerformanceCost PerformanceThanks