曙光1000上矩阵乘积算法的性能分析  

PERFORMANCE ANALYSIS OF MATRIX MULTIPLICATION ALGORITHMS ON DAWNING 1000

作　　者：谢幸[1] 顾乃杰[1] 陈国良[1] 

机构地区：[1]中国科学技术大学计算机科学技术系

出　　处：《计算机研究与发展》1999年第7期848-852,共5页Journal of Computer Research and Development

摘　　要：矩阵乘积算法在科学计算中应用十分广泛．文中给出了典型矩阵乘积算法在曙光１０００上的性能比较和分析，并针对ＳＵＭＭＡ算法研究了分块尺寸对其通信性能的影响，指出分块尺寸是影响其通信性能的一个重要因素．原算法并没有给出其分块尺寸的具体选取方法，文中通过理论和实验的分析提出了一个选取最优分块尺寸的标准．实验结果显示ＳＵＭＭＡ算法按文中的标准选取最优分块尺寸后性能得到大幅度提高，可达机器峰值的５０．７％．Matrix multiplication is widely used in scientific computing. The performance of typical matrix multiplication algorithms on Dawning 1000 are compared and analyzed in this paper. The analysis shows that matrix block size is an important factor in the performance of the SUMMA algorithm, but how to choose the best block size is not considered in the original algorithm. A criterion is proposed here for choosing the best block size in the SUMMA algorithm after theoretic and experimental analysis. Higher performance is achieved by improving the SUMMA algorithm using this criterion. Experimental results show that it can reach 50.7% of the peak performance.

关 键 词：矩阵乘积 性能分析 算法 并行计算机 

分 类 号：TP338.602[自动化与计算机技术—计算机系统结构]