分布式系统中数据分解代码的自动产生

Automatic Code Generation of Data Decomposition in Distributed Shared Memory System

为分布内存系统开发的并行编译器碰到的第一个问题就是如何分解一个应用程序中的数据。由于访问非本地节点上数据的代价是昂贵的,所以数据分解必须仔细考虑。尽管数据分解的定义已被提出,但是文献并没有给出相应的算法.本文介绍了在一个已被证明且功能强大的数学模型下如何产生数据分解代码的算法,并在SUIF(Stanforduniversityintermediateformat)系统中的Paraguin编译器上得到实现。

The fist problem of developing parallelizing compiler in DSM（Distributed Shared Memory） system is how to decompose the data of a program.. The non-local data accesses result in communication that can greatly impact performance, thus the mapping or decomposition of data must be carefully considered. Although a concept of data decomposition has been introduced in previous work,the algorithm was not shown. In this paper, an algorithm to decompose data based on a proven and powerful mathematical model is introduced and is implemented in Paraguin compiler of SUIF.