[petsc-users] Command line option to set number iterations for Chebyshev eigenvalue estimation?
Garth N. Wells
gnw20 at cam.ac.uk
Thu Jan 23 09:18:30 CST 2014
On 2014-01-23 14:54, Dave May wrote:
> If you don't have any prefix set on the outer KSP, you should be able
> to simply do
>
> -mg_levels_ksp_type chebychev
> -mg_levels_ksp_chebychev_estimate_eigenvalues 0,0.2,0,1.1
> -mg_levels_est_ksp_type cg
> -mg_levels_est_ksp_max_it X
>
Thanks. Last two are what I was looking for and I can confirm that they
work as expected.
It would be helpful is last two options could be (easily) found in the
docs - I couldn't find them.
Garth
> Cheers,
> Dave
>
> On 23 January 2014 15:49, Garth N. Wells <gnw20 at cam.ac.uk> wrote:
>
>> On 2014-01-23 14:32, Matthew Knepley wrote:
>> On Thu, Jan 23, 2014 at 5:35 AM, Garth N. Wells <gnw20 at cam.ac.uk>
>> wrote:
>>
>> Is there are command line option to set the number of iterations
>> used for the eigenvalue estimation inside the Chebyshev
>> preconditioner? "-gamg_est_ksp_max_it" does the trick with GAMG,
>> but I'd also like to change the number of iterations when using
>> Chebyshev smoothing with ML.
>>
>> In GAMG, we use PETSc itself to do the estimate, and give it the
>> prefix gamg_est_. In ML, as far as I can see from the code, we use
>> the
>> internal
>> estimator, and I don't think they give us this knob.
>
> I'm not so interested in the what happens inside ML, but the
> Chebyshev smoother that PETSc applies. With the default 10 GMRES
> iterations, the eigenvalue estimate is not good enough for my problem,
> which leads to the outer CG breaking down with an indefinite operator
> error.
>
> I can get it working by changing
>
> -mg_levels_ksp_chebyshev_estimate_eigenvalues
>
> but I'd rather do a few more iterations to get a better eigenvalue
> estimate. It would also be nice to be able to switch from GMRES to CG.
> I can do this with GAMG via "-gamg_est_ksp_type"
>
> Garth
>
>> Matt
>>
>>
>>> Garth
>>
>> --
>> What most experimenters take for granted before they begin their
>> experiments is infinitely more interesting than any results to
>> which
>> their experiments lead.
>> -- Norbert Wiener
More information about the petsc-users
mailing list