FBOSS 演进以支持生成式 AI 网络工作负载.pdf

1、Evolving FBOSS to support Generative AI Network WorkloadsMeta Platforms Inc.Broadcom Inc.Jasmeet BaggaSoftware Engineer/Meta Platforms Inc.Shrikrishna KhareSoftware Engineer/Meta Platforms Inc.Evolving FBOSS to support Generative AI Network WorkloadsNETWORKINGMehak MahajanSenior Director of Engineer

2、ing/Broadcom Inc.Outline5 Call to Action4 DSF Performance for Gen AI3 Key FBOSS Enhancements for Gen AI2 Key SDK Enhancements for Gen AI1 Disaggregated Scheduled FabricFacebook Open Switching System:FBOSSMetas software stack for managing Network Switches in Metas DCs One of Metas largest network ser

3、vices by deploymentFBOSS uses SAISwitch Abstraction Interface:SAIAn OCP projectOpen source API to control switching elementsVendor independentFBOSSChallengesElephant flows:few extremely large continuous flowsLow entropy:less variation,more likely to cause hash collisionsOscillatory behavior during c

4、ongestionSolution:DSFDisaggregated Scheduled FabricNear-optimal load balancingSmoother bandwidth delivery:credit allocationFlexibility/optionality for endpoints:fabric performs spray/reassemblyNetwork Traffic for AI TrainingNon-blocking Network for 4K GPUsCredit based congestion control,Break packet

5、 into cells and sprayReassembly in hardwareRDSW=Disaggregated Line CardFDSW=Disaggregated Fabric CardChallenges:Generative AI requires much larger number of GPUsJobs spanning multi-K GPUsSolution:Interconnect 4K GPU clusters using Routing and ECMPBut,intra-cluster:non-blocking,inter-cluster:elephant

6、 flows,low entropy,oversubscribed domain*DSF Dual Stage Topology*Hierarchy of Fabric devices,18K GPU non-blocking clusterDefer ECMP decision as late as possibleNetwork Traffic for Generative AI DSF Dual Stage TopologyRDSW=Disaggregated Line CardFDSW=Disaggregated Fabric CardSDSW=2nd stage Fabric Car