Performance optimization and application of the MUMPS for solving 3-dimensional stationary strength problems on cluster systems

Abstract

We present an approach to performance optimiza-tion of MUMPS, which is one of the most widely used fa-mous open source sparse SLAE solvers. The main idea is to take into account special block structure of sparse matrix-es arising from the finite element analysis of 3-dimensional stationary strength problems. Implemented in the reordering phase, this approach yields about 20 % reduction of total computational time while using fewer cores. Next, we consider application and adjustment of the solver for a specific machine. We demonstrate that the threads/processes configuration significantly influ-ences performance. The results are obtained on the FSUE RFNC – VNIIEF cluster for symmetric positive defined matrixes.