Comparação do desempenho spMv entre formatos de armazenamento de matrizes esparsas provenientes do método AIM de simulação de reservatórios

Abstract

In this work we aim to compare two sparse matrix storage formats, ievaluating the performance of isparse matrix dense vector multiplication (spMv). The kind of matrices presented in this paper comes from numerical simulations of oil reservoirs, specifically on AIM (Adaptive Implicit Method) aproach. In one format, as a previous step, the mesh of the problem is reordered, so the matrix is divided in four block submatrices, with all blocks of the same size. The other format uses the natural order for the mesh, which results in a block matrix withs blocks of different sizes. It is shown that the second format has little loss of performance when compared to the first one, but the first exhibit problems to improve the utilization of iterative linear solvers through application of preconditioners. The profiling software VTune had been utilized to identify the behavior and hotspots of each format.