Performance of Three Krylov Subspace Methods on a PARAGON System.

Abstract

For the solution of a system of linear equations with general non-Hermitian nonsingular coefficient matrix, three Krylov subspace methods, CGS, TFQMR and QMR, are applicable offering the advantage that the coefficient matrix is solely involved in the form of matrix-vector products. If the coefficient matrix is sparse these iterative methods are attractive in contrast to direct methods, provided that they converge sufficiently fast. In this note, the performance of the three methods is investigated on an Intel PARAGON XPS 10, a parallel machine with distributed memory.