Reordering Strategy for Blocking Optimization in Sparse Linear Solvers

Grégoire Pichon 1 Mathieu Faverge 1 Pierre Ramet 1 Jean Roman 1
1 HiePACS - High-End Parallel Algorithms for Challenging Numerical Simulations
LaBRI - Laboratoire Bordelais de Recherche en Informatique, Inria Bordeaux - Sud-Ouest
Abstract : Solving sparse linear systems is a problem that arises in many scientific applications, and sparse direct solvers are a time-consuming and key kernel for those applications and for more advanced solvers such as hybrid direct-iterative solvers. For this reason, optimizing their performance on modern architectures is critical. The preprocessing steps of sparse direct solvers—ordering and block-symbolic factorization—are two major steps that lead to a reduced amount of computation and memory and to a better task granularity to reach a good level of performance when using BLAS kernels. With the advent of GPUs, the granularity of the block computation has become more important than ever. In this paper, we present a reordering strategy that increases this block granularity. This strategy relies on block-symbolic factorization to refine the ordering produced by tools such as Metis or Scotch, but it does not impact the number of operations required to solve the problem. We integrate this algorithm in the PaStiX solver and show an important reduction of the number of off-diagonal blocks on a large spectrum of matrices. This improvement leads to an increase in efficiency of up to 20% on GPUs. 1. Introduction. Many scientific applications, such as electromagnetism, astrophysics , and computational fluid dynamics, use numerical models that require solving linear systems of the form Ax = b. In those problems, the matrix A can be considered as either dense (almost no zero entries) or sparse (mostly zero entries). Due to multiple structural and numerical differences that appear in those problems, many different solutions exist to solve them. In this paper, we focus on problems leading to sparse systems with a symmetric pattern and, more specifically, on direct methods which factorize the matrix A in LL t , LDL t , or LU , with L, D, and U, respectively, unit lower triangular, diagonal, and upper triangular according to the problem numerical properties. Those sparse matrices appear mostly when discretizing partial differential equations (PDEs) on two-(2D) and three-(3D) dimensional finite element or finite volume meshes. The main issue with such factorizations is the fill-in—zero entries becoming nonzero—that appears in the factorized form of A during the execution of the algorithm. If not correctly considered, the fill-in can transform the sparse matrix into a dense one which might not fit in memory. In this context, sparse direct solvers rely on two important preprocessing steps to reduce this fill-in and control where it appears. The first one finds a suitable ordering of the unknowns that aims at minimizing the fill-in to limit the memory overhead and floating point operations (Flops) required to complete the factorization. The problem is then transformed into (P AP t)(P x) = P b,
Complete list of metadatas

Cited literature [29 references]  Display  Hide  Download

https://hal.inria.fr/hal-01485507
Contributor : Mathieu Faverge <>
Submitted on : Wednesday, April 5, 2017 - 11:03:19 AM
Last modification on : Thursday, December 13, 2018 - 6:48:02 PM
Long-term archiving on : Thursday, July 6, 2017 - 1:15:10 PM

File

M106245.pdf
Publisher files allowed on an open archive

Identifiers

Citation

Grégoire Pichon, Mathieu Faverge, Pierre Ramet, Jean Roman. Reordering Strategy for Blocking Optimization in Sparse Linear Solvers. SIAM Journal on Matrix Analysis and Applications, Society for Industrial and Applied Mathematics, 2017, SIAM Journal on Matrix Analysis and Applications, 38 (1), pp.226 - 248. ⟨10.1137/16M1062454⟩. ⟨hal-01485507v2⟩

Share

Metrics

Record views

554

Files downloads

390