1: \begin{abstract}

2: We numerically analyze the possibility of turning off post-smoothing

3: (relaxation) in geometric multigrid when used as a preconditioner in conjugate

4: gradient linear and eigenvalue solvers for the 3D Laplacian.  The geometric

5: Semicoarsening Multigrid (SMG) method is provided by the \emph{hypre} parallel

6: software package.   We solve linear systems using two variants (standard and

7: flexible) of the preconditioned conjugate gradient (PCG) and preconditioned

8: steepest descent (PSD) methods. The eigenvalue problems are solved using the

9: locally optimal block preconditioned conjugate gradient (LOBPCG) method

10: available in \emph{hypre} through BLOPEX software.  We observe that turning off

11: the post-smoothing in SMG dramatically slows down the standard PCG-SMG. For

12: flexible PCG  and LOBPCG, our numerical results show that post-smoothing can be

13: avoided, resulting in overall acceleration, due to the high costs of smoothing

14: and relatively insignificant decrease in convergence speed.    We numerically

15: demonstrate for linear systems that PSD-SMG and flexible PCG-SMG converge

16: similarly if SMG post-smoothing is off.  We experimentally show that the effect

17: of acceleration is independent of memory interconnection.  A theoretical

18: justification is provided. 

19: \end{abstract}