On the use of H-Matrix Arithmetic in PaStiX: a Preliminary Study

Mathieu Faverge (1, 2) , Grégoire Pichon (2) , Pierre Ramet (1, 2) , Jean Roman (2)

1 LaBRI - Laboratoire Bordelais de Recherche en Informatique
2 HiePACS - High-End Parallel Algorithms for Challenging Numerical Simulations

When solving large sparse linear systems, both the amount of memory needed and the computational cost represent a burden to efficiency. In order to solve larger systems, low-rank strategies are used to reduce the overall complexity of a solver. In this talk, we present a preliminary study of the use of H-Matrix arithmetic in a supernodal solver. We also present a new feature in PaStiX, a reordering strategy to reduce the number of off-diagonal blocks in the symbolic factorization. It allows BLAS kernels to be more efficient, and those ideas could be explored in the context of a low-rank strategy.

Mots clés: sparse direct solver; h-matrix

Domaines: Calcul parallèle, distribué et partagé [cs.DC]

Mathieu Faverge, Grégoire Pichon, Pierre Ramet, Jean Roman. On the use of H-Matrix Arithmetic in PaStiX: a Preliminary Study. Workshop on Fast Direct Solvers, Jun 2015, Toulouse, France. ⟨hal-01187882⟩