A Comparison of Preconditioned Krylov Subspace Methods for Large-Scale Nonsymmetric Linear Systems

TL;DR

This paper compares various preconditioned Krylov subspace methods for solving large, sparse, nonsymmetric linear systems from PDEs, providing practical guidelines for method selection based on empirical and theoretical analysis.

Contribution

It offers a comprehensive comparison of KSP methods and preconditioners, including new insights into their performance and convergence for large-scale PDE problems.

Findings

GMRES performs best with effective multigrid preconditioners.

BoomeAMG converges faster than ML but may fail on ill-conditioned problems.

Right preconditioning is generally more effective.

Abstract

Preconditioned Krylov subspace (KSP) methods are widely used for solving large-scale sparse linear systems arising from numerical solutions of partial differential equations (PDEs). These linear systems are often nonsymmetric due to the nature of the PDEs, boundary or jump conditions, or discretization methods. While implementations of preconditioned KSP methods are usually readily available, it is unclear to users which methods are the best for different classes of problems. In this work, we present a comparison of some KSP methods, including GMRES, TFQMR, BiCGSTAB, and QMRCGSTAB, coupled with three classes of preconditioners, namely Gauss-Seidel, incomplete LU factorization (including ILUT, ILUTP, and multilevel ILU), and algebraic multigrid (including BoomerAMG and ML). Theoretically, we compare the mathematical formulations and operation counts of these methods. Empirically, we compare the convergence and serial performance for a range of benchmark problems from numerical PDEs in 2D and 3D with up to millions of unknowns and also assess the asymptotic complexity of the methods as the number of unknowns increases. Our results show that GMRES tends to deliver better performance when coupled with an effective multigrid preconditioner, but it is less competitive with an ineffective preconditioner due to restarts. BoomerAMG with proper choice of coarsening and interpolation techniques typically converges faster than ML, but both may fail for ill-conditioned or saddle-point problems while multilevel ILU tends to succeed. We also show that right preconditioning is more desirable. This study helps establish some practical guidelines for choosing preconditioned KSP methods and motivates the development of more effective preconditioners.