1: \begin{abstract}

2: Preconditioned Krylov subspace (KSP) methods are widely used for solving

3: very large and sparse linear systems arising from PDE discretizations.

4: For modern applications, these linear systems are often nonsymmetric

5: due to the nature of the PDEs, the boundary or jump conditions, or

6: the discretization methods. While a number of preconditioned KSP methods

7: and their implementations are readily available, it is often unclear

8: to users which ones are the best choices for different classes of

9: problems. In this work, we present a systematic comparison of some

10: representative KSP methods and preconditioners. We consider four KSP

11: methods, namely restarted GMRES, TFQMR, BiCGSTAB, and QMRCGSTAB, coupled

12: with three preconditioners, namely Gauss-Seidel, incomplete LU factorization

13: (ILU), and algebraic multigrid (AMG). We assess these preconditioned

14: KSP methods using large, sparse, nonsymmetric linear systems arising

15: from PDE discretizations in 2D and 3D. We compare the convergence

16: and timing results of these different combinations, and assess the

17: scalability of these methods with respect to the number of unknowns.

18: Our results show that GMRES tends to deliver the best performance

19: when coupled with AMG preconditioners, but it is far less competitive

20: than the other three methods due to restarts when using Gauss-Seidel

21: or ILU preconditioners. Our results also show that the smoothed-aggregation

22: AMG delivers better performance and exhibits better scalability than

23: classical AMG, but it is less robust, especially for ill-conditioned

24: systems. The study helps establish some practical guidelines for choosing

25: preconditioned KSP methods. It also motivates the further development

26: of more effective and robust multigrid preconditioners for large,

27: sparse, nonsymmetric, and potentially ill-conditioned linear systems.

28: \end{abstract}

29: