Computing the inverse of a triangular matrix on heterogeneous clusters

Abstract

The aim of this paper is to study the impact of heterogeneity on the parallelization of an algorithm for computing the inverse of a triangular matrix using a divide and conquer algorithmic paradigm. The target parallel and distributed system is composed of p heterogeneous processors of different speeds linked by a homogeneous interconnection network. The methodology used in this paper has been thought to be extended to other algorithms with similar structures. The theoretical analysis leads to an optimal two-phases parallel algorithm by designing a scheduling procedure where the processors remain always active.