Iterative algorithms for the solution of nonsymmetric systems in the modelling of weak plasma turbulence (Q1112700)

We compare the efficiency of four conjugate-gradient-like algorithms to solve the nonsymmetric systems generated by this problem in a production- type code: the normal equation algorithm (CGN), the bi-conjugate gradient method (BCG), the conjugate gradient squared method (CGS), and the generalized minimum residual method (GMRES(k)). For all these methods, we precondition the system using an incomplete factorization or a modified incomplete factorization of the coefficient matrix. We show that the iterative solvers we test provide a viable alternative to band Gauss methods, particularly for very large problems. We have tested the solvers with non-symmetric matrices arising from different physical parameters (the electric field) and we show that the iterative solvers perform well also in the presence of a large skew symmetric part if the symmetric part is positive. Our performance data suggest that CGS and GMRES(k) yield the best performances. The matrices we examine are typical of bilinear finite-element discretization of convection-diffusion equations, so that we expect the conclusions drawn in this note to apply to a wider class of problems. All the numerical experiments were performed on the IBM 3090 with vector facility.