Optimisation du produit matrice-vecteur creux sur architecture GPU pour un simulateur de réservoir

Corentin Rossignon 1, 2
1 RUNTIME - Efficient runtime systems for parallel architectures
Inria Bordeaux - Sud-Ouest, UB - Université de Bordeaux, CNRS - Centre National de la Recherche Scientifique : UMR5800
Abstract : For the Total Company, simulating reservoirs is an important step in the process of optimizing production. Nowadays, these simulations run entirely on CPUs. Thus, we have attempted to accelerate the sparse matrix-vector product operators of the simulation by using GPUs. Common GPU libraries for sparse linear algebra use generic formats for sparse matrix storage, that are more or less performant on GPU but that do not allow to fully exploit the specific structure of the matrices used in the reservoir simulator. In order to exploit this structure, we have adapted for our matrices a storage format that enables us to accelerate with a 20x factor the sparse matrix-vector product on 3 GPUs in comparison with a 8-core CPU, and with a 1.5x factor on GPU in comparison with the generic matrix format used by NVIDIA in cuSPARSE.
Mots-clés : solveur creux GPU SpMV CSR
Complete list of metadatas

Cited literature [2 references]  Display  Hide  Download

https://hal.inria.fr/hal-00773571
Contributor : Olivier Aumage <>
Submitted on : Thursday, July 3, 2014 - 3:37:38 PM
Last modification on : Thursday, February 7, 2019 - 5:48:22 PM
Long-term archiving on : Friday, October 3, 2014 - 11:36:12 AM

File

solver_gpu.pdf
Files produced by the author(s)

Identifiers

  • HAL Id : hal-00773571, version 2

Collections

Citation

Corentin Rossignon. Optimisation du produit matrice-vecteur creux sur architecture GPU pour un simulateur de réservoir. ComPAS'13 / RenPar'21 - 21es Rencontres francophones du Parallélisme, Inria Grenoble, Jan 2013, Grenoble, France. ⟨hal-00773571v2⟩

Share

Metrics

Record views

558

Files downloads

1150