Advanced Charm++ Tutorial课件.ppt

Short Message Optimizations Direct all to all communication is α dominated Message combining for small messages Reduce the total number of messages Multistage algorithm to send messages along a virtual topology Group of messages combined and sent to an intermediate processor which then forwards them to their final destinations AAPC strategy may send same message multiple times 精品文档 Virtual Topology: Mesh Organize processors in a 2D (virtual) Mesh Phase 1: Processors send messages to row neighbors Message from (x1,y1) to (x2,y2) goes via (x1,y2) Phase 2: Processors send messages to column neighbors 2* messages instead of P-1 精品文档 AAPC Performance 精品文档 Large Message Issues Network contention Contention free schedules Topology specific optimizations 精品文档 Ring Strategy for Collective Multicast Performs all to all multicast by sending messages along a ring formed by the processors Congestion free on most topologies 0 1 2 i i+1 P-1 …… …….. 精品文档 Streaming Messages Programs often have streams of short messages Streaming library combines a bunch of messages and sends them off Stripping large charm++ header Short array message packing Effective message performance of about 3us 精品文档 Using communication library Communication optimizations embodied as strategies EachToManyMulticastStrategy RingMulticast PipeBroadcast Streaming MeshStreaming 精品文档 Bracketed vs. Non-bracketed Bracketed Strategies Require user to give specific end points for each iteration of message sends Endpoints declared by calling ComlibBegin() and ComlibEnd() Examples: EachToManyMulticast Non bracketed strategies No such end points necessary Examples: Streaming, PipeBroadcast 精品文档 Accessing the Communication Library From mainchare::main Creating a strategy Strategy *strat = new EachToManyMulticastStrategy(USE_MESH) Strat = new StreamingStrategy(); Strat-enableShortMessagePacking(); Associating a proxy with a Strategy ComlibAssociateProxy(strat, myproxy); myproxy should be passed to all array