Predicting the Best Mapping for Efficient Exploitation of Task and Data Parallelism

Abstract

The detection and exploitation of different kinds of parallelism, task parallelism and data parallelism often leads to efficient parallel programs. This paper presents a simulation environment to predict the best mapping for the execution of message-passing applications on distributed systems. Using this environment, we evaluate the performance of an image processing application for the different parallelizing alternatives, and we propose the ways to improve its performance.

This work was supported by the MCyT under contract 2001-2592 and partially sponsored by the Generalitat de Catalunya (G. de Rec. Consolidat 2001SGR-00218).