Impact of Blocking Strategies for Sparse Direct Solvers on Top of Generic Runtimes

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

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

Among the preprocessing steps of a sparse direct solver, reordering and block symbolic factorization are two major steps to reach a suitable granularity for BLAS kernels efficiency and runtime management. In this talk, we present a reordering strategy to increase off-diagonal block sizes. It enhances BLAS kernels and allows to handle larger tasks, reducing runtime overhead. Finally, we will comment the resulting gain in the PaStiX solver implemented over StarPU and PaRSEC.

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

Grégoire Pichon, Mathieu Faverge, Pierre Ramet, Jean Roman. Impact of Blocking Strategies for Sparse Direct Solvers on Top of Generic Runtimes. SIAM Conference on Parallel Processing for Scientific Computing (SIAM PP 2016), Apr 2016, Paris, France. ⟨hal-01251808⟩