Improving Multilevel Approach for Optimizing Collective Communications in Computational Grids

Abstract

Collective operations represent a tool for easy implementation of parallel algorithms in the message-passing parallel programming languages. Efficient implementation of these operations significantly improves the performance of the parallel algorithms, especially in the Grid systems. We introduce an improvement of multilevel algorithm that enables improvement of the performance of collective communication operations. An implementation of the algorithm is used for analyzing its characteristics and for comparing its performance it with the multilevel algorithm.