基于FC-TR网络的MPI通信性能优化研究  

Study on MPI communication performance optimization based on FC-based token-routing switch-network

作　　者：曾小荟[1,2,3] 刘清[1] 罗文浪[1] 康继昌[2] 

机构地区：[1]井冈山大学计算机科学与技术系,江西吉安343009 [2]西北工业大学计算机科学与工程系,西安710072 [3]同济大学嵌入式系统与服务计算教育部重点实验室,上海200092

出　　处：《计算机应用研究》2011年第4期1505-1508,共4页Application Research of Computers

基　　金：国家"863"计划资助项目(2003AA001018);江西省教育厅科技计划资助项目(GJJ10538);同济大学嵌入式系统与服务计算教育部重点实验室开放课题(2010-09)

摘　　要：为了给上层的MPI并行应用程序提供更快的基于信令寻径式光纤通道交换网络(FC-TR网络)的通信支持,针对FC-VIA-MPI进行了通信性能优化,对点对点通信的实现提出了新的优化方法,尽量把同一节点内进程之间的网络通信转换为内存操作;同时,根据底层的FC-TR网络软硬件特性,对MPI的部分组通信(如MPI_Bcast、MPI_Allreduce等)的实现也提出了新的优化算法。实验结果表明,优化的FC-VIA-MPI通信性能总体上提高了10%左右。In order to supply the upper layer MPI parallel applications with faster network communications based on FC-based token-routing switch-network（FC-TR network）,this paper optimized the communication performance of FC-VIA-MPI.Applied a new method to point-to-point communication that the network communications between processes in the same computing node were translated into memory operations as much as possible;what＇s more,according to the underlying software and hardware features of FC-TR network,proposed new algorithms to optimize the communication performance of some collective communications（MPI_Bcast,MPI_Allreduce,etc.）.The experimental results show that the communication performance of the optimized FC-VIA-MPI is increased by 10% or so as a whole.

关 键 词：光纤通道通信 消息传递接口 网络通信 通信性能优化 

分 类 号：TP393.07[自动化与计算机技术—计算机应用技术]