DYNAMIC BLOCK DATA DISTRIBUTION FOR PARALLEL SPARSE GAUSSIAN ELIMINATION
El Mostafa Daoudi, Pierre Manneback, Mostapha Zbakh · 2001
This article is devoted to describe a new dynamic block data distribution algorithm over a grid of processors for sparse Gaussian elimination in order to improve the load balance compared to the classical static block-cyclic distribution. In order to assure a numerical stability and to separate the ordering and the symbolic factorizations, Demmel and al. [2,3] presented a new method for sparse Gaussian Elimination with Static Pivoting called GESP where the data structure and the communication graph are known before the numerical factorization. In this work, we assume that the ordering and the symbolic factorizations are already performed and we are interesting by the numerical factorization of the final structure of the matrix to be computed. The experimental results show the advantages of our new approach.