论文部分内容阅读
本文分析了面向分布存储ＳＩＭＤ／ＭＩＭＤ并行机的并行程序的优化数据安放问题，在ＦＯＲＡＬＬ程序模型和ＭＥＳＨ通信模型之上，研究了数据分解过程中减少通信代价的优化要求．我们使用维偏好图描述并行数组之间的对准需求，通过消除维偏好图中的冲突，可得到维对准图．一个维对准图就对应一个数据安放方案．维对准图的总代价越大，对应的通信代价就越小．文中给出了求最大代价维对准目的一个近似算法．
This paper analyzes the optimal data placement problem of parallel programs for distributed memory SIMD / MIMD parallel machines. Based on the FORALL program model and the MESH communication model, the optimization requirements of reducing the communication costs during data decomposition are studied. We use the dimension preference graph to describe the alignment requirements between parallel arrays, and obtain the dimension alignment map by eliminating the conflicts in the dimension preference graph. A dimensional alignment map corresponds to a data placement program. The larger the total cost of the alignment map, the lower the corresponding communication cost. In this paper, an approximate algorithm for finding the maximum cost dimension is given.