高性能计算通信机制分析与研究  

An Analysis on Communication Mechanisms for HPC

作　　者：岳菲菲[1] 王海军[2] 王新[1] 黄东波[1] 

机构地区：[1]复旦大学计算机科学与技术学院,上海200433 [2]中国科学院计算技术研究所网络重点实验室,北京100190

出　　处：《计算机工程与科学》2009年第A01期27-30,39,共5页Computer Engineering & Science

基　　金：国家自然科学基金资助项目(90812001);国家973计划资助项目(2005CB321807)

摘　　要：高性能计算机在各个领域得到了越来越广泛的应用,而这些系统的体系结构特征却有着巨大的差异。IC工艺的进步使得通信机制逐渐超过计算单元,成为了影响系统性能参数、功耗和系统规模的主要因素,使得高性能计算机的设计从以计算能力为主向以数据传输和通信能力为主,并在系统层次上产生了深远的影响。由于通信机制越来越难设计,所以通信往往成为开发的瓶颈,如何使高性能计算平台运行得更快、更高效一直是研究的热点。本文对高性能计算的通信机制进行了研究和分析,给出了近年来高性能计算系统发展趋势,进一步探讨适应新需求的通信机制。High Performance Computation （HPC） has been shown to be of great use in almost every area of our life As the development of IC technology, the communication mechanisms which have been shown to be superior to regular processor architectures become the most important issue influencing the performance, cost, and scale of HPC. Because of the difficulty in communication design in HPC growing, communication has become the bottleneck of HPC design. Making the communication faster and more efficient has become the most significant part in HPC exploitation. This paper analyzes the performance of several mainstream networks, and predicts the trend in development of communication techniques. This paper also exposes new techniques on communication of HPC.

关 键 词：交换矩阵 MYRINET QsNet INFINIBAND 性能分析 

分 类 号：TP303[自动化与计算机技术—计算机系统结构]