from CS1 to CS2的operator给这三个copyset ④ MDS通过RPC在CS2上创建copyset1,2,3这三个copyset ⑤ 假定三个copyset的leader都是CS3，在CS3的下一次心跳的 response中，下发第三步生成的三个operator ⑥ CS3收到change peer from CS1 to CS2的operator，给CS2同步 raft to CS1的operator给 部分copyset，比如copyset1,2,3 ⑤ MDS通过RPC在CS1上创建copyset1,2,3这些copyset ⑥ 假定三个copyset的leader都是CS3，在CS3的下一次心跳的 response中，下发第四步生成的三个operator ⑦ CS3收到change peer from CS2 to CS1的operator，给CS1同步

function ● Relax type inference/checking for Any at compilation time ● Register a shape function for operator to check the type and compute the output shape© 2019, Amazon Web Services, Inc. or its Affiliates function ● Relax type inference/checking for Any at compilation time ● Register a shape function for operator to check the type and compute the output shape ● Shape function has two modes (op_attrs, input_tensors function ● Relax type inference/checking for Any at compilation time ● Register a shape function for operator to check the type and compute the output shape ● Shape function has two modes (op_attrs, input_tensors

cuDNN Offload to heavily optimized DNN operator library FrameworksLimitations of Existing Approach cuDNN Frameworks New operator introduced by operator fusion optimization potential benefit: System High-level data flow graph and optimizations Directly generate optimized program for new operator workloads and hardware Hardware FrameworksWhy Automation is the Future Clear winner on

service (in Plan) • Config CurveBS by (Cluster and Pool CRDs) in Kubernetes (in Plan) • Support Operator capability level 5 (in Plan) • horizontal / vertical scaling, auto config tuning, abnormal detection Plan) • Config CurveFS by (cluster and storage pools) CRDs in Kubernetes (in Plan) • Support Operator capability level 5 (in Plan) • now support helmCurrent Status • Release 2 major version on CurveBS Cache support on CurveFS • CurveFS cloud native support • csi plugin for CurveFS • support operator capability level 2: automated application provisioning and configuration management and patch and

high performance cloud native file system (Based on CurveBS / S3 compatible Storage)Operator capability Level OPERATOR CAPABILITY LEVEL CURVE COMMENT BASIC INSTALL Y (by Helm) automated application provisioning

3400us (baseline), 40us (target) - 85x speedup - Uh ohEnter, TVM and model co-design - PyTorch operator overhead makes interpreter infeasible - Reduce FLOPs with block-sparsified weight matrices -