On the peculiarities of solving large systems of linear algebraic equations on high performance computing systems of different architecture

Abstract

The results of testing a set of the Krylov subspace iterative methods (CGS, BiCGStab) with algebraic multigrid preconditioner for solving large sparse systems of linear algebraic equations are discussed. The scalability characteristics for the MPI and hybrid versions of the code on three HPC-systems are given. The peculiarities of using these methods on computer systems of different processor architecture (Intel Harpertown, Intel Nehalem, and AMD Magny-Cours) are analyzed.