Evaluation of linear system equations solvers on multicore architectures

Abstract eng:
The aim of this paper is the evaluation of an existing parallel, linear equation solver to solve largescale, sparse, non-symmetric systems of linear equations as a part of the finite element software. Traditional finite element solvers run in a serial mode and are limited by available resources. The parallel approach allows to take an advantage of the distributed memory to forming large system matrices and multiple processing units to achieve significant speedups. In this paper, we study differences between the sequential and parallel solutions. Parallel approach contributing shared and distributed memory is represented by SuperLU DIST solver. SuperLU DIST is primary based on distributed memory model with using Message Passing Interface. SuperLU DIST can be also used as the hybrid parallel model combining shared memory model or with graphical computing units. The performance of parallel solvers is tested on problems from solid mechanics. The two type of benchmark problems discuss in this paper and computations of this problems are based on OOFEM which is a free finite element code with object oriented architecture for solving mechanical, transport and fluid mechanics problems that operates on various platform. The efficiency of hybrid SuperLU solver is com pared to the efficiency of pure shared memory SuperLU solver and to distributed memory solver in PETSc.

• Export as: BibTeX | MARC | MARCXML | DC | EndNote | NLM | RefWorks
• Add to your basket