一种优化MPI程序性能的改进方法  

An Improvement Method of MPI Parallel Program Performance Optimization

作　　者：柯鹏[1] 聂鑫[1] 

机构地区：[1]武汉科技大学计算机科学与技术学院,武汉430072

出　　处：《现代计算机（中旬刊）》2011年第8期3-6,共4页Modern Computer

基　　金：国家自然科学基金(No.61070009)

摘　　要：在分布式存储系统上,MPI已被证实是理想的并行程序设计模型。MPI是基于消息传递的并行编程模型,进程间的通信是通过调用库函数来实现的,因此MPI并行程序中,通信部分代码的效率对该并行程序的性能有直接的影响。通过用集群通信函数替代点对点通信函数以及通过派生数据类型和建立新通信域这两种方式,两次改进DNS的MPI并行程序实现,并通过实验给出一个优化MPI并行程序的一般思路与方法。MPI is an ideal parallel programming model which has been confirmed on distributed storage system. Because MPI is based on message passing and adopts the method of message passing to realize the communication between every node, the performance of MPI parallel program is deeply depended on the efficiency of communication. By using the cluster communication to replace the peer to peer communication function and using the derived datatype to create new communication domain, puts forward a common method of optimizing MPI parallel program via the experiment on improving on the MPI parallel program of DNS twice and advancing the performauce of it.

关 键 词：消息传递接口 通信函数 性能优化 动态负载平衡 集群通信 

分 类 号：TP311.1[自动化与计算机技术—计算机软件与理论]