面向MPP Fortran的自动数据分布

Automatic Data Distribution for MPP Fortran

自动数据分布是面向大规模并行处理ＭＰＰ（ｍａｓｓｉｖｅｌｙｐａｒａｌｅｌｐｒｏｃｅｓｓｉｎｇ）系统程序自动并行化的一项关键技术．数据分布方式直接影响着应用程序在ＭＰＰ系统上的并行执行性能．本文以ＭＰＰＦｏｒｔｒａｎ为例，详细探讨了自动数据分布的有关技术，如对准分析、分布方式的产生、静态性能评估和数据重新分布等，并提出了相应的算法．这些算法将在作者研制的面向ＭＰＰＦｏｒｔｒａｎ的程序自动并行化工具中实现．

Automatic data distribution is a key technique in the area of automatic program parallelization for MPP(massively parallel processing) systems. Data distribution schemes directly influence the parallel execution performance of application programs in MPP systems. Taking MPP Fortran as an example, this paper discusses techniques of automatic data distribution in detail such as alignment analysis, shared array distribution, performance estimation for shared array distribution schemes, and shared array redistribution. The corresponding algorithms are presented and they will be implemented in these automatic program parallelization tools for MPP Fortran.