《20230614_B-102_Davis.PDF》由會員分享,可在線閱讀,更多相關《20230614_B-102_Davis.PDF(28頁珍藏版)》請在三個皮匠報告上搜索。
1、SmartNIC(DPU)Storage Solutions and Use CasesRob DavisVP Storage Technology,NVIDIA Networking PlatformsWhat Are DPUs?Cloud Native HPC Accelerated ComputingCloud ComputingBare Metal as a Service5GIndustrial/Core/Edge Cloud GamingCybersecurityAI-PoweredEnterprise AIFlash Based StorageNVIDIA DPU BlueFie
2、ld RoadmapBlueField-29.8 SPECint*200 Gbps BlueField-342 SPECint*400 Gbps2020 2022 2024 DOCA ONE DEVELOPMENT ARCHITECTURE*SPECint2017-rate2021 2023 BlueField-4136 SPECint*800 GbpsPERFORMANCE,SCALE,IO2025 BlueField-1100 Gbps Storage Companies Implementing DPU Storage SolutionsDPU Storage Use Cases5All
3、 Flash Array/JBOF for Storage Area NetworksCompute and Storage DisaggregationServer Based Storage/SDSTarget StorageNetworkDPUDPUDPUComputational StorageDPUComputational StorageBackend NVMe-oF ClusterDPUDPU-ACCELERATED SERVERTRADITIONAL SERVERDPU with Arm Cores&AcceleratorsCPUSoftware-defined Storage
4、Software-defined NetworkingNICCPUSoftware-defined StorageSoftware-defined NetworkingInfrastructure ManagementSoftware-defined SecurityVMsContainersVMsContainersInfrastructure ManagementSoftware-defined SecurityAcceleration EnginesAcceleration EnginesHow:DPUs Offload,Accelerate and Secure Storage NVI
5、DIA DPU ArchitectureCPU(s)ARM/RISC-VHW OffloadsCPU cycle intensive or latency sensitive functionsIOPCIe,Ethernet,IB SDKDOCA(OPI)DPU Storage AcceleratorNVMe-oF Target OffloadDDR4DDR4PCIe SwitchNVME SSDNVME SSDInitiator x86ConnectX-5Initiator x86ConnectX-5SOCSOC8M IOPs,512B block size5M IOPs,4K block
6、side5 usec latency(not including SSD)OffloadTarget2 100Gb InitiatorsDDR4DDR4PCIe SwitchNVME SSDNVME SSDInitiator x86ConnectX-5Initiator x86ConnectX-5no OffloadTarget2 100Gb Initiators6M IOPs,512B block size2M IOPs,4K block side15 usec latency(not including SSD)Server Side Offload SNAP ExamplePhysica
7、l Local NVMe StoragePhysical Local StorageOS/HypervisorNVMe Standard DriverPCIe BUSNVMeHost ServerLocal Physical Storage to Hardware Emulated StorageNVMe SSD EmulationHost ServerNVMe SNAPOS/HypervisorNVMe Standard DriverNVMeEmulatedStoragePCIe BUSRemote StoragevirtIO-blkvirtIO-fileEncryption Offload
8、At the Target JBOFSecure data at rest on drives with AESAt the Front-End ControllerSecure data on drives and in-flight to JBOFs and AFAs with AES,TLS,MACsec,and IPsecAt the Initiator(Server)Secure data across entire storage data pathOwner of the data controls the keysStorage HeadCPUInitiatorsStorage
9、 HeadCPUInitiatorsJBOFControllerStorage HeadCPUInitiatorsJBOFControllerJBOFControllerPrograming DPUs13|2022 Flash Memory Summit.All Rights Reserved.Open Programmable Infrastructure ProjectCommunity Driven Open Ecosystem for frameworks based on DPU enabled systems Originated as Diamond Bluff in late
10、2021 rapid growth in 2022,20+companies,100+individuals OPI Project is now a Linux Foundation project https:/opiproject.org/https:/ Announcement 21-JUN-2022Current Working AreasOrganization&Administration+Legal/GovernanceVision Statement/Goals+requirementsEvents,Outreach,OrientationProvisioning and P
11、latform ManagementOpen Programmable Infrastructure API and Behavioral ModelUse CaseDeveloper Platform/POC/Reference PlatformOPI Project GoalsCreate community-driven standards-based open ecosystem for DPU/IPU-like technologiesCreate vendor agnostic framework and architecture for DPU/IPU-based softwar
12、e stacksReuse existing or define a set of new common APIs for DPU/IPU-like technologies when requiredProvide implementation examples to validate the architectures/APIs“Founding members of OPI include Dell Technologies,F5,Intel,Keysight Technologies,Marvell,NVIDIA and Red Hat with a growing number of
13、 contributors representing a broad range of leading companies in their fields ranging from silicon and device manufactures,ISVs,test and measurement partners,OEMs to end users.”OPI Project DeliverablesOpen Source ProjectsSpecifications/StandardsReference PlatformsTest Suites&CasesPOC/PrototypesDPUEx
14、ample of Embedded Scale Out Back EndDPU17Autonomous Vehicle High Speed StorageAutonomous Vehicle Training Data Collection An autonomous vehicle can generate as much as 64 TB/daySSDs are removed by hand from vehicles and sent to the data centerHere they are installed into servers and uploaded for tra
15、ining and analysis Sometimes days later1920DPU NVMe Storage2x200 GbEDPU Autonomous Vehicle Training Data Collection SolutionUp to 64 TB/dayData is downloaded directly to the data center in minutesDPUExample of a SDS DPU Solution Attack vectors are significantly reducedIntroduce strict security barri
16、er that separates applications from the storage environmentReduce the API between app and storage to that of a Lustre aware Router(LNET)Only the DPU can communicate with Backend StorageMuch stronger security than any storage and network in isolation has todayClassic DesignBlueField Based EXAScalerEx
17、ample of a DPU Computational Storage PlatformGPU DPU Example of a DPU Computational Storage SolutionsStorage GatewayDPUPCIe switchDPUPCIe switchSDS on CNAStorage Class SSDHigh-Capacity SSDsDPUPCIe switchSDS on CNAStorage Class SSDHigh-Capacity SSDsDPUPCIe switchJBOF becomes Smart Storage BladeDPUs in Use Cases Across Storage Today!All Flash Array/JBOF for Storage Area NetworksCompute and Storage DisaggregationServer Based Storage/SDSTarget StorageNetworkDPUDPUDPUComputational StorageDPUComputational StorageBackend NVMe-oF ClusterDPUThank You!