基于HPM模型的CoSMPs系统的算法并行与优化  

Algorithm Parallelization and Optimization of CoSMPs System Based on the HPM Model

作　　者：李玉岗[1] 乔香珍[1] 刘志勇[2] 

机构地区：[1]中国科学院计算技术研究所国家智能计算机研究中心,北京100080 [2]国家自然科学基金委员会信息科学部,北京100085

出　　处：《计算机研究与发展》2004年第4期621-629,共9页Journal of Computer Research and Development

基　　金：国家自然科学基金项目 (6993 3 0 2 0 )

摘　　要：首先 ,基于并行计算模型HPM ,分析了多机机群系统的体系结构特点 ,从并行性和局部性 (存储与通信特性 )两方面分析影响并行应用软件性能的主要因素 ,讨论应用软件并行与优化的相关问题 ;分析了纯MPI和MPI +SMP(或OMP)制导两种编程模式在性能上的优点与不足 然后 ,讨论了在CoSMPs系统上对并行应用软件进行优化的方法 最后 ,对两种不同的通信模式 (循环交换、边界交换 )在CoSMPs系统上的性能进行讨论 ,并在多机机群系统的实例———DW30 0 0超级服务器上进行优化 ;通过计算实例———矩阵乘法和解偏微分方程的五点格式算法加以验证 。First, based on the parallel computing model HPM, the characteristics of the architecture of the SMPS cluster are investigated Focusing on the two aspects: the parallelism and the locality (storage and communication) of the architecture of CoSMPs, the main factors that influence the performance of the parallel applications are analyzed, and the problems of how to parallelize and optimize applications are investigated Both the merits and the demerits of the two programming modes: the MPI mode and the MPI+SMP (OMP) directive mode are investigated Then, some techniques about how to parallelize and optimize applications on CoSMPs are investigated in detail Finally, the performance of the two communicational modes(the loop exchange mode and the border exchange mode) on CoSMPs is investigated on an instance of the cluster of SMPs the Dawning 3000 super computer Via two examples-the matrix multiply algorithm for the loop exchange and the five point algorithm for the border exchange mode the methods are tested and the results are consistent with the theoretical conclusion

关 键 词：CoSMPs HPM SMP 存储层次结构 

分 类 号：TP311.52[自动化与计算机技术—计算机软件与理论]