gmres - restart and Gauss-Seidel

Lukasz Kaczmarczyk lukasz at
Fri Aug 15 04:56:18 CDT 2008

Matthew Knepley wrote:
> That is strange. Your parallel SOR should be a stronger PC than the
> Block-Jacobi SOR that PETSc is using. 
I have such hope and that was a motivation. I like to have algorithm
which is efficient with large number of processors.

However, it is hard to prove
> that restarted GMRES will converge, and thus this behavior is not
> impossible.
It is strange because with Block-Jacobi SOR convergence is ok. May be I
should add some trick, f.e. add shift to diagonal.

> However, if you are using MG and the smoother is doing its job, you
> should not take more than a few iterates. You should never get to
> 30 outer GMRES iterates.
It is true, in practical computations I  never have more than 30
iterations. However for tests, where I set error for small number
(1e-12) number of iterations is more than 30.


More information about the petsc-users mailing list