Performance analysis of the MHM simulator in a petascale machine
Antonio Tadeu Azevedo Gomes1; Diego Paredes2; Weslley da Silva Pereira1; Roberto Pinto Souto1; Frédéric Valentin1
1 LNCC; 2 PUCV
doi:10.20906/CPS/CILAMCE2017-0381
Resumo
We present the performance analysis of a parallel simulator that implements the multiscale hybrid-mixed (MHM) finite element method for the elastostatic equation. The formulation behind MHM comprises a global problem defined over the skeleton of a coarse mesh and a set of independent elasticity problems posed over element-wise submeshes. The performance analysis considers an isotropic elastic 3D model. For the strong scaling, we characterize whether it is more efficient to refine the submeshes and use polynomial degrees of lower order or vice-versa while keeping the same approximation error. We obtained the better efficiency---greater than 70% on 3,072 cores---when using higher-order polynomials. For the weak scaling, we fixed the high-order polynomials and varied the level of the refinement of the submeshes. We then increased the size of the problem by refining the global mesh at the same rate as the increase in the number of cores. The measurements suggest the configurations with refined submeshes offer the best efficiency---greater than 80% on 3,072 cores.
Palavras-chave: Scientific computing; Advanced numerical algorithms; Parallel and distributed computing