NVMe™ PCIe™虚拟SSD的互操作迁移.pdf

1、Lee Prewitt,MicrosoftMike Allison,SamsungNVMe Interoperable Migration of PCIe virtual SSDsHyperscaler Use CaseLee PrewittDirector Cloud Hardware StorageMicrosoftSTORAGEHyperscaler PerspectiveChallenge:Consistent,Predictable,&Performant VM Guest ExperienceLive migration should not be observable by th

2、e guest VMLive migration should be efficient and performance isolatedHardware iteration should not be observable by the guest VMIncludes SSD upgrade,platform change,etc.Historical Approach to Achieve these GoalsHyperscaler-specific logic to make changes invisible to the guest VMFor lower latency,ven

3、dor specific changes to SSDs were introducedStandards Solve this Challenge in a Durable Manner Samsung,Google,Microsoft,and others have led standards activities in NVMe and OCP since 2023 to solve this challengeFeatures are delivered piece by piece to advance the ecosystem Hardware&software are test

4、ed at scale in a robust mannerPCIeHypervisorVMGuest KernelHost Kernel NVMe DriverPCI NVMe DriverNVMe SSDRobust standards needed to broadly deliver the consistent,predictable,performant VM storage experience Why is Live Migration Needed?Customers hate to be interrupted when they are working Need very

5、 low interruption rates on imperfect hardware Measured as the number of customer VM events per yearAnnual Interruption Rate(AIR)Live Migration is required to meet AIR goalsUse cases for Live Migration Actual hardware failure on the node(may not affect all VMs)Predicted hardware failure(allows timely

6、 migration before issues occur)Scheduled node maintenance(Hardware or firmware upgrade)Load balancing(allows a mixture of large and small VMs to coexist)Why is Para-virtualization Not Enough?SSDs are fast Para-virtualization is slow Leaves IOPs on the table Para-virtualization is expensive Reduces t