摘要
计算与数据划分问题是影响并行程序在分布主存多处理机中执行性能的重要因素,也是并行编译优化的重点.针对该问题,提出了一套关于数据空间融合的理论框架,并基于该框架给出了一种有效的全局计算与数据划分方法,用于分布主存计算环境中的计算与数据划分问题的求解.该方法能够尽量开发计算空间的并行度,利用数据融合技术优化数据分布,并能搜寻优化的全局计算与数据划分.该方法还能很自然地与数据复制以及偏移常量的对准结合在一起,从而使得数据通信量尽可能地小.实验结果表明了所提出方法的有效性.
Computation and data decompositions are key factors of affecting the performance of parallel programs running on distributed memory multicomputers. This paper presents a theoretical framework of data space fusion and an effective global computation and data decomposition approach based on it, which can be used to solve computation and data decomposition problems on distributed memory multicomputers. The approach can exploit the parallelism of computation space as high as possible, use the technique of data space fusion to optimize data distribution, and search the optimizing global computation and data decompositions. The approach can also be integrated with data replication and offset alignment naturally, and therefore can make the communication overhead as low as possible. Experimental results show that the approach presented in the paper is effective.
出处
《软件学报》
EI
CSCD
北大核心
2004年第9期1311-1327,共17页
Journal of Software
基金
国家自然科学基金~~
关键词
分布主存多处理机
并行编译器
计算划分
数据划分
数据融合
distributed memory multicomputer
parallel compiler
computation decomposition
data decomposition
data fusion