Automatic partitioning of data and computations on scalable shared memory multiprocessors

Abstract

This paper describes an algorithm for deriving data and computation partitions on scalable shared memory multiprocessors. The algorithm establishes affinity relationships between where computations are performed and where data is located based on array accesses in the program. The algorithm then uses these affinity relationships to determine both static and dynamic partitions for arrays and parallel loops. Experimental results from a prototype implementation of the algorithm demonstrate that it is computationally efficient and that it improves the parallel performance of standard benchmarks. The results also show the necessity of taking shared memory effects (memory contention, cache locality, false-sharing and synchronization) into account-partitions derived to minimize only interprocessor communications do not necessarily result in the best performance.