通信对机群并行计算性能的影响  

How Does Communication Influence Cluster's Parallel Computing Performance

作　　者：胡明昌[1] 史岗[1] 胡伟武[1] 唐志敏[1] 

机构地区：[1]中国科学院计算技术研究所,北京100080

出　　处：《小型微型计算机系统》2003年第9期1569-1573,共5页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目 (60 0 73 0 18)资助;中国科学院盈科优秀青年学者奖课题 (2 0 0 0 40 2 1)资助中科院;计算所领域前沿青年基金课题 (60 0 0 6B0 ;2 0 0 162 80 -0 1)资助

摘　　要：分析了通信和计算重叠模型及 L og GP模型 ,指出各通信参数对并行计算性能的影响 ,结合并行程序的特征介绍了在机群环境下改善并行计算性能经常采用的五种通信方式 :采用高速网络、采用用户级通信协议、利用 SMP通信、动态预取或迁移数据、消息合并发送 ,详尽测试了各种方式影响性能情况并分析了其特点和适用范围 .采用高速网络是最常用的方法 ,性能提高明显 ,适用于各类应用程序 .需要传送大量小消息的并行程序还应采用用户级通信协议 .对于特定的一类应用程序 ,采用消息合并发送方式提高性能最多 .采用动态预取或迁移数据和利用 SMP通信时要慎重 。This paper introduces the algorithmic communication model and LogGP model and points out how communication parameters affect parallel performance. Considering parallel program communication characteristic, we present five frequently used methods to improve clusters parallel computing performance, including using faster network, adopting user-level network interface protocols, using SMPs as computing nodes, dynamic prefetching data or migrating data home, and merging small messages. The performance of each method is strictly tested and the characteristic of each one is anatomized. Using faster network to reduce network latency is common and effective to all kinds of parallel applications. Those that send a great number of small messages should adopt user-level network interface protocols additionally to reduce overhead. Merging small messages is most efficient for particular applications. Be careful to prefetch data dynamically or migrate datas home, for the optimization is suitable only for some applications. Do not start too many processes in one node when using SMPs to run the program.

关 键 词：LogGP模型 用户级通信 机群 曙光3000 MPI JIAJIA 预取 home迁移 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]