###
DOI:
Journal of Software:2000.11(12):1674-1680

负载平衡无关的并行程序最适处理器网格选择
张云泉,施巍松
(中国科学院 软件研究所 并行软件研究开发中心,北京,100080)
Optimal Processor Grid Selection for Parallel Program Independent of Load Balance
ZHANG Yun-quan,SHI Wei-song
()
Abstract
Chart / table
Reference
Similar Articles
Article :Browse 2591   Download 2897
Received:June 18, 1999    Revised:September 28, 1999
> 中文摘要: 用户在编写并行程序时,通常是把物理处理器看成逻辑的处理器(进程)网格,以便于算法的实现.随着用户可用处理器的不断增多,可选择的网格形状也随之增加,如何为基于消息传递的并行程序选择合适的、能发挥出并行机潜在性能的处理器网格形状,是一个迫切需要解决的问题.在提出基于通信点概念的最小度数通信点集合法之后,通过对并行程序通信模式的分析,试图解决与负载平衡无关的并行程序的最适处理器网格选择问题.通过对ScaLAPACK软件包中的一个并行测试程序——并行Cholesky(对称正定矩阵分解)通信点集合度的分析,此方法成功地选择了最适处理器网格形状,并与实验结果相一致.
Abstract:Physical processors are often viewed as a logical processor grid or process grid to ease the parallel algorithm implementation and to provide useful coordination information among parallel processes. However, the shape of processor grid has great impact on the final performance of user's parallel programs. How to select a suitable or even optimal processor grid for an parallel algorithm on certain parallel machines becomes an urgent problem. In this paper, a novel method named MDCPS (minimum degree of communication point set) is proposed, which tries to find out the optimal processor grid for parallel program independent the impaction of load balance through analysis on its communication pattern. The analysis results on ScaLAPACK parallel Cholesky factorization program match the experimental results well and show that the proposed method can select the optimal processor grid for parallel program successfully.
文章编号:     中图分类号:    文献标志码:
基金项目:国家攀登计划B资助项目;国家863高科技项目基金资助项目(863-306-ZT06-02-01;863-306-ZD01-03-2);国家自然科学基金资助项目(NSF69883006) 国家攀登计划B资助项目;国家863高科技项目基金资助项目(863-306-ZT06-02-01;863-306-ZD01-03-2);国家自然科学基金资助项目(NSF69883006)
Foundation items:
Reference text:

张云泉,施巍松.负载平衡无关的并行程序最适处理器网格选择.软件学报,2000,11(12):1674-1680

ZHANG Yun-quan,SHI Wei-song.Optimal Processor Grid Selection for Parallel Program Independent of Load Balance.Journal of Software,2000,11(12):1674-1680