Application of AVX (Advanced Vector Extensions) for improved performance of the PARFES - finite element Parallel Direct Solver

The paper considers application of the AVX (Advanced Vector Extensions) technique to improve the performance of the PARFES parallel finite element solver, intended for finite element analysis of large-scale problems of structural and solid mechanics using multi-core computers. The basis for this paper was the fact that the dgemm matrix multiplication procedure implemented in the Intel MKL (Math Kernel Library) and ACML (AMD Core Math Library) libraries, which lays down the foundations for achieving high performance of direct methods for sparse matrices, does not provide for satisfactory performance with the AMD Opteron 6276 processor, Bulldozer architecture, when used with the algorithm required for PARFES. The procedure presented herein significantly improves the performance of PARFES on computers with processors of the above architecture, while maintaining the competitiveness of PARFES with the Intel MKL dgemm procedure on computers with Intel processors.