針對 AI 工作負載優化的存儲架構.pdf

《針對 AI 工作負載優化的存儲架構.pdf》由會員分享，可在線閱讀，更多相關《針對 AI 工作負載優化的存儲架構.pdf（28頁珍藏版）》請在三個皮匠報告上搜索。

1、1|2024 SNIA.All Rights Reserved.Storage Architecture Optimized for AI WorkloadsPresented by Paul McLeodProduct Director,StorageSupermicro2|2024 SNIA.All Rights Reserved.Storage Architecture Optimized for AI Workloads About Supermicro Storage challenges with AIOps and MLOps Moving beyond legacy stora

2、ge Solution approach Hardware innovation with EDSFF Summary3|2024 SNIA.All Rights Reserved.ABOUT SUPERMICRORevenueRevenue$14B+(FY2024 guidance)(FY2024 guidance)$7.1B(FY2023)$7.1B(FY2023)$5.2B(FY2022)$5.2B(FY2022)Worldwide Worldwide PresencePresence6M+Sq ft.Facilities Worldwide6M+Sq ft.Facilities Wor

3、ldwide1.Silicon Valley(HQ),2.Taiwan,3.The Netherlands,4.Malaysia and othersProductionProduction$18B/$18B/yryr Production Capacity(CY23)Production Capacity(CY23)Top 5 Largest Server System Provider Worldwide(IDC&Gartner 2022),1.3M units annuallyHuman Human Resource inResource in4 Campuses4 Campuses60

4、00+headcount Worldwide,50%Technical/R&DKey Growth Key Growth MatrixMatrix#1 in Generative AI and LLM Platforms500%+YoY Growth in Accel.Computing4|2024 SNIA.All Rights Reserved.AI/ML Implementation5|2024 SNIA.All Rights Reserved.Challenges for AI/ML Storage projectsLarge scale,rapid growthMixed data

5、sizes High concurrency of I/O PipelinesCentralized management Integration of emerging technologiesSource:WEKA6|2024 SNIA.All Rights Reserved.Source:SNIA7|2024 SNIA.All Rights Reserved.Source:SNIA8|2024 SNIA.All Rights Reserved.AI data pipeline:Multiple pipelines heating storageSource:WEKA9|2024 SNIA

6、.All Rights Reserved.9GPU Direct Storage(GDS)with WEKASupermicro+WEKA GDS provides RDMA with the GPU Memory Lowest latency for the AI Pipeline File-based single namespace for Flash and HDD Transparent file level access to S3 objects Scale-up from 138GB/s with an entry cluster*The performance number

7、is based on six node PCIe4 WEKA storage clusterSupermicro Offers Tiered Storage Building Blocks for WEKAFile80-90%HDD10-20%Flash102023 Supermicro102024 SupermicroSupermicro Powered Application TierNo provider offers more choices for GPU-accelerated computing112023 Supermicro112024 SupermicroHigh-Per

8、formance All-Flash TierWe worked with Weka engineers to optimize for Supermicro storage servers122023 Supermicro122024 SupermicroMulti-Tier Storage Architecture for AI and ML WorkloadsThe key to cost-effectively storing all your data,safely,on premises13|2024 SNIA.All Rights Reserved.AI Storage Refe

9、rence ArchitectureSSESSE-MQM9700MQM9700-NS2FNS2FDeltaDelta-Next GPUNext GPUSYSSYS-821GE821GE-TNHRTNHRCustomer Network400G IB Network-SSE-MQM9700-NS2F25G or 100G Object FrontEnd Ethernet NetworkWeka ClusterWeka ClusterFor reference only,layout varies depending on actual system quantityFor reference o

10、nly,layout varies depending on actual system quantitySSESSE-SN3700SN3700-CS2FCCS2FC25G or 100G Object BackEnd Ethernet NetworkASG ASG-WK1032CWK1032C-GEN5GEN5SSGSSG-640SP640SP-DE1CR90DE1CR9025G or 100GbE SW25G or 100GbE SW25G or 100GbE SW25G or 100GbE SWObject ClusterObject ClusterSSESSE-MQM9700MQM97

11、00-NS2FNS2FGPU ClusterGPU ClusterSSESSE-F3548SF3548SSSESSE-MQM9700MQM9700-NS2FNS2F14|2024 SNIA.All Rights Reserved.AI customer#1 IO Pattern Millions of Tiny IOs Reads/WritesSource:WEKA16|2024 SNIA.All Rights Reserved.Hardware Innovation17|2024 SNIA.All Rights Reserved.Embracing Emerging StandardsEDS

12、FF and CXLEDSFF E1.S,E3.S,and E3.L form factors,as well as AICs,have been integrated into the Compute Express Link(CXL)ecosystem,underscoring their utility in high-performance,high-capacity server environments promoting robust,scalable,and efficient designs.201520162017201820192020202120222023202410

13、th Generation11th Generation12th Generation13th GenerationSource:SNIA18|2024 SNIA.All Rights Reserved.Gen5 EDSFF Petascale Platform Innovation Superior Signal Integrity Mainboard direct connection to SFF-TA-1002 1C connectors/SSDs and reduce the backplane routing signal loss Reduce 40%of the signal

14、loss Better Air Flow No vertical backplane blocking the air flow 75%increase in front opening 20%improvement system CFMReduce Backplane Trace Layout Signal Loss&Improve Air FlowE3.S SSD+EDSFF BPNU.2 SSD+BPN19|2024 SNIA.All Rights Reserved.Unified Chassis Support Intel DP and AMD UP 1U up to 24 E1 SS

15、D 1U up to 16 E3 SSD and CXL 2U up to 32 E3 SSD Less than 31”chassis depth Balanced Architecture Front storage IO and rear networking Eliminate the processor NUMA complexity Gen5 EDSFF Petascale Platform InnovationGen5 EDSFF Petascale Platform InnovationPurposed Built for New All-NVMe and Software-D

16、efined Data Center x64 x64 AMD Single ProcessorIntel Dual Processorx32 x48 x32 x48 31”chassis depth1U16,1U24,2U32 and CXL20|2024 SNIA.All Rights Reserved.CXL(Compute Express Link)A high-speed interconnect,industry-standard interface for communications between processors,accelerators,memory,storage,a

17、nd other IO devices.CXL Memory Expansion Enabling memory cache coherency between CPU memory and attached memory devices 1st system supports 4x E3 CXL 2T device.Petascale 1U system (both AMD and Intel)1st industry E3 CXL PoC Partner with Micron CXL team and demonstrate great performance improvement w

18、hen use Micron CXL CMM and SMC Petascale system.Petascale+CXL Memory Expansion SolutionPetascale+CXL Memory Expansion SolutionNext Gen Memory Expansion Solution for Next Gen Data Center212023 Supermicro212024 SupermicroH13H13 2U E3.S 2U E3.S Petascale Petascale AllAll-FlashFlashPCIe 5.0 Slots 2 x16

19、slots&2 AIOMsSingle ProcessorUp to 350W TDPE3.S 1T SSDUp to 32 E3 1T slotsDDR5 SlotsUp to 24 DIMMsHot swap U.2PCIe 5.0 x16 AIOMsPCIe 5.0 x8 or x16 FHHL slotsRedundant Power Supply 2000W(Titanium level)222023 Supermicro222024 SupermicroPCIe 5.0 Slots2 x16 slots&2x AIOMsX13X13 1U E1.S 1U E1.S Petascal

20、e Petascale AllAll-FlashFlashDual ProcessorUp to 270W TDPE1.S SSDUp to 24 E1 slotsDDR5 SlotsUp to 32 DIMMsRedundant Power Supply 2000W(Titanium level)PCIe 5.0 x16 AIOMsPCIe 5.0 x16 FHHL slotsSupport 9.5mm or 15mm E1.S 232023 Supermicro232024 SupermicroPCIe 5.0 SlotsUp to 2 x16 slots+2 AIOMsH13 H13 1

21、U 1U Petascale CXL Petascale CXL&E3 SSD&E3 SSD Server Server E3.S 2T(x8)CXL Type 3 ModuleE3.S 1T(x4)SSDRedundant Power Supply 1600W(Titanium level)PCIe 5.0 Slots x16 AIOMsPCIe 5.0 Slots x16 slotsDDR5 DRAMUp to 24 DIMMsSingle AMD Genoa Processor24|2024 SNIA.All Rights Reserved.Summary Conventional st

22、orage approaches arent well suited to AI and ML workloads The“I/O Blender”effect in the data pipeline mixes read/write on small files and multiple simultaneous pipelines produce mix of I/O patterns A two-tier storage architecture with a Parallel File System on Supermicros Petascale All-Flash storage

23、 server enables high performance E3.S flash from multiple partners An object tier using Supermicros high-density disk-based SuperStorage storage server provides 90 drives and over 2PB*raw capacity in 4U This solution has been deployed with a multinational high tech manufacturing customer with 25PBBe

24、tter Faster Greener 2024 Supermicro.*Raw value is based on vendor raw base capacity of 24TB.TB is base-10 decimal.25|2024 SNIA.All Rights Reserved.AI Storage White PaperThis paper is available at SNIA.All Rights Reserved.26For More InformationSupermicro: Info: SNIA.All Rights Reserved.Thank You!28|2024 SNIA.All Rights Reserved.Please take a moment to rate this session.Your feedback is important to us.29|2024 SNIA.All Rights Reserved.29