Parallel block generalized WZ factorization

Abstract

In this paper we first present a block strategy for the generalized WZ factorization, which consists of block factorizing a matrix A in the form A=WZW/sup -1/. This study shows how a block strategy may be used to reduce a large eigenvalue problem into a number of smaller ones. Next, we develop a parallel multi-phase algorithm for this method, which requires processes such as matrices products, Gauss-Jordan elimination, broadcasting, scattering and gathering. To conceive our multi-phase algorithm we have used an informal methodology like the sequential top-down analysis which allows the conception of efficient multiphase parallel algorithms. The experimental tests show a good speed-up and corroborate the theoretical valuations.