PARA'04 State-of-the-Art
in Scientific Computing
June 20-23, 2004 (Home page)
Updated: 19 February 2004

A shared- and distributed-memory parallel sparse direct solver

Anshul Gupta
IBM TJ Watson Research Center
USA

Abstract:
We recently introduced improved symbolic and numerical factorization algorithms for general sparse matrices. The symbolic algorithms are capable of inexpensively computing a minimal task-dependency graph and near-minimal data-dependency graph for factoring a general sparse matrix. These graphs, computed solely from the nonzero pattern of the sparse matrix, are valid for any amount of pivoting induced by the numerical values during numerical factorization.

In this paper, we describe a parallel sparse direct solver based on these algorithms that has recently been included in the Watson Sparse Matrix Package (WSMP). To the best of our knowledge, this is the only sparse solver that utilizes both shared- and distributed- memory parallelism in the same program and is designed for a hierarchical parallel computer with network-interconnected SMP nodes. We compare the WSMP solver with two similar well known solvers: Super_LU and MUMPS. We show that the WSMP solver is more robust and achieves significantly higher performance than other similar solvers based on traditional algorithms.

Home page