An Improvement to Affine Decomposition on Distributed Memory Architecture

Abstract

Automatic decomposition is an optimization technique that distributes computation and data onto different processors. The consequence of decomposition directly affects the performance of parallel program. Since every computing node has its own memory in distributed memory parallel computers (DMPCs), false dependence does not hinder the parallelism. Affine decomposition is an effective method to represent and derive computation partition and data distribution, and its principle of adding dependence constraint is too strict to gain more parallelism. Some loop nests do not satisfy the affine condition, and are prohibited from parallelism by affine decomposition. However, if only the irregular access is caused by indirect array, loop and array reference can be partitioned at compile time. To tackle above problems of affine decomposition, an improved static decomposition algorithm of DMPCs proposed in this paper. The experimental results show that this algorithm can improve the performance of parallel programs.