OpenXLA：面向异构硬件时代的便携式性能解决方案.pdf

《OpenXLA：面向异构硬件时代的便携式性能解决方案.pdf》由会员分享，可在线阅读，更多相关《OpenXLA：面向异构硬件时代的便携式性能解决方案.pdf（11页珍藏版）》请在三个皮匠报告上搜索。

1、Gaurav Agarwal,Distinguished Engineer,MarvellSrihari Karnam,Sr.Principal Engineer,MarvellOpenXLA:Portable Performance for the Heterogeneous Hardware EraAgenda1The Vision:Composable Data CentersThe software driven flexible,scalable,cost and power efficient2The Software ChallengeCore requirements for

2、programming Composable systems3The Foundation:OpenXLAThe right platform to build upon4The Path ForwardExtending OpenXLA for Composability5A Call to ActionHow we can unlock this potential togetherCo-Optimized Solutions With Composable SystemsComposable Rack Scale SystemSoftware DrivenWorkload Optimiz

3、edDynamically ScalableImproved Resource UtilizationPower OptimizedEnables Dynamic CompositionEvolutionLegacy Architecture Pools of general-purpose serversRigid,Siloed,Inefficient for modern diverse workloadsServer 1CPUDRAMGPUSSD Server 2CPUDRAMGPUSSD Server nCPUDRAMGPUSSD Software ChallengesCPUxPUPC

4、Ie/CXLxPUxPUAcceleratorSCMStorage w/ComputeMemory ApplianceFabric-attached Comp MemoryHybrid SwitchingUALinkHeterogenous SystemsHeterogenous Programming ModelDisaggregated ComputingHeterogeneity-aware compute partitioningDeep-memory hierarchiesTopology variabilityControl Plane(Operational Complexity

5、)DiscoveryConfigurationResource provisioningComposable Computing demands a Software Evolution for Portable Performance in Heterogenous EraComposable Computing:Key RequirementsA Unified Programming ModelPortable,high-level abstractions for heterogeneous hardwareWhats NeededDynamic Resource Orchestrat

6、ionRuntime discovery,Scheduling,and provisioning of pooled resourcesTopology&Performance AwarenessOptimize placement for latency,bandwidth,and costAdaptive Compilation&ExecutionJIT optimization for specific hardware compositionsHow To Achieve It Prevent a new era of vendor lock in Standardized IRs E