Self-tuned parallel processing system for heterogeneous clusters

Abstract

Automatic program execution control is a demanding tool for heterogeneous environments, due to the variable availability of processors and network bandwidth. Issues such as load distribution have to be re-analyzed since strategies developed for homogeneous machines fail to optimize the computation time in heterogeneous networks. The final user should not be involved in deciding which machines should be selected, the data distribution policies that best fit each algorithm, etc. The present work presents such a tool, designed for interactive processing in a distributed computing environment, whose main objective is to minimize the applications computation time. Results are presented for an object tracking algorithm, either in homogeneous and heterogeneous networks showing the advantage of automatic control.