A direct parallel sparse matrix solver Original Research Article

The direct sparse matrix solver is based on a domain decomposition technique to achieve data and work parallelization. Geometries that have long and thin structures are specially efficiently tractable with this solver, provided that they can be decomposed mainly in one direction. Due to the separation of the algorithm into a factorization stage and a solution stage, time-dependent problems with a constant coefficient matrix are particularly well suited for this solver. The parallelization performances obtained on a Cray T3D show that the method scales up to at least 256 processors.