网站大量收购闲置独家精品文档,联系QQ:2885784924

WD西部数据Technical Brief Benchmarking GPUDirect and the Western Digital OpenFlex Data24 NVMe-oF Storage Platform (4200 Series)数据中心平台English (US)技术简介.pdf

WD西部数据Technical Brief Benchmarking GPUDirect and the Western Digital OpenFlex Data24 NVMe-oF Storage Platform (4200 Series)数据中心平台English (US)技术简介.pdf

  1. 1、本文档共5页,可阅读全部内容。
  2. 2、原创力文档(book118)网站文档一经付费(服务费),不意味着购买了该文档的版权,仅供个人/单位学习、研究之用,不得用于商业用途,未经授权,严禁复制、发行、汇编、翻译或者网络传播等,侵权必究。
  3. 3、本站所有内容均由合作方或网友上传,本站不对文档的完整性、权威性及其观点立场正确性做任何保证或承诺!文档内容仅供研究参考,付费前请自行鉴别。如您付费,意味着您自己接受本站规则且自行承担风险,本站不退款、不进行额外附加服务;查看《如何避免下载的几个坑》。如果您已付费下载过本站文档,您可以点击 这里二次下载
  4. 4、如文档侵犯商业秘密、侵犯著作权、侵犯人身权等,请点击“版权申诉”(推荐),也可以打举报电话:400-050-0827(电话支持时间:9:00-18:30)。
查看更多

二三|1WesternDigital.

后加三

TechnicalBrief

NVIDIAeGPUDirecteStorage

BenchmarkingGPUDirectandtheWesternDigitalOpenFlexData244200Series

NVMe-oFStoragePlatform

TheFutureofMemoryandStorage(FMS)2024-TornadoDemonstration

Abstract

Thispaperdemonstratesthearchitecturalviabilityofhigh-performancedisaggregatedstorageinfrastructureusingRDMAwithRoCEv2in

acomputeintensiveenvironmentThearchitecturedemonstratessimplicitywhilstallowingaflexibleapproachtothelinearscalingofGPUs,

performance,andstoragerequirements.

Benchmark

Thisbenchmarkwasdesignedtoshowcasethehighsustainablethroughputthatcanbeachievedwhencorrectlyarchitectinga

disaggregatedstoragesolutionforthepurposescfhigh-performancecomputing(suchasMachineLearning)j.Inthisexampletheaimwasto

demonstrateavisualrenderingofa2011OklahomaTornadoinitsformationstage.

Thissimulation(mathematicallyderivedfrominitialconditionsjustbeforethetornadoformed)includes250billiongridpoints,eachwith

overadozenattriputessuchasrain,hailpressure,andwingdspeed.Thisdetailedvisualization,showing6000simulationsteps,provides

unprecedentedinsightintothetornadosdynamicswithastoragerequirementof5.9TBofdata.

TheBenchmarkEnablers

NVIDIAGPUDirectStorage{GDS)

GDSenablesaqirectdatapathfordirectmemoryaccess{(DMA)transfersbetweenGPUmemoryandstorage,whichavoidsabouncebuffer

throughtheCPU.ThisdirectpathincreasessystembandwidthanddecreasesthelatencyandutilizationloadontheCPU.Thisisparticularly

beneficialinhigh-performancecomputingandcomplexdataprocessingtasks.

WithoutGDS,GPUsqirectlyreadtraining/inferencedatafromlocalSSDsvi

文档评论(0)

R的文档库 + 关注
实名认证
内容提供者

相近型号说明书可通用,具体看说明书包括型号

版权声明书
用户编号:7060103150000024

1亿VIP精品文档

相关文档