分布存储并行计算机优化数据安放的一种方法

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．

DATA OPTIMIZATION OF DISTRIBUTED MEMORY PARALLEL COMPUTERS

This paper addresses the problem of optimizing data placement of data parallel programs targeted at distributed memory SIMD/MIMD computers.A dimensional preference graph(DPG) is invented to describe the requirement of alignment between parallel arrays.By eliminating conflicts in DPG,dimensional alignment graph(DAG) can be obtained,which indicates the resulting data placement scheme.The larger the weight of DAG,the smaller the corresponding communication cost.A heuristic algorithm is devised to compute the DAG.