An Efficient Data and Computation Decomposition Technique for Nested Loops on NUMA Multiprocessor Systems

Abstract

This paper presents an automatic computation/data decomposition technique for nested loops on NUMA (Non-Uniform Memory Access) systems. In NUMA systems, the remote memory access time is longer than the local one, and computation/data decomposition affects the amount of remote accesses incurred by parallel processing. Therefore, the system performance is dependent on how to decompose computation/data onto parallel processors. Here, we propose a modified locality algorithm to improve the one in [6] for the case when the decomposition is not communication-free. In addition, a new performance estimating method is also presented. The whole method has been implemented on SUIF [8]. Experimental results demonstrate the superiority of our proposed algorithm over that in previous littature.