[petsc-users] KSP step not accelerating despite good preconditioning in SNES

[Timothée Nicolas]

Thanks a lot, I expected something like that. I will look in this direction.

Best

Timothée

2015-09-28 21:17 GMT+09:00 Matthew Knepley <knepley at gmail.com>:

> On Mon, Sep 28, 2015 at 2:53 AM, Timothée Nicolas <
> timothee.nicolas at gmail.com> wrote:
>
>> Hi all,
>>
>> I have something strange in my application and I don't know what could
>> cause this. I am trying to do an implicit MHD problem and I thought I
>> finally figured out the preconditioner step, but somehow I don't get the
>> expected result, not even close.
>>
>> For the preconditioning, I am using an approximate Schur complement,
>> which requires two relatively easy KSP inversions at each preconditioner
>> application. I apply this algorithm directly to the result function at the
>> end of the routine FormFunction. I have checked that the approximation to
>> the inversion of the Jacobian is good, in the sense that when I multiply
>> the preconditioned vector by the *total* Jacobian matrix, I indeed
>> recover almost the initial unpreconditioned vector. Also, I know that my
>> Jacobian matrix is correct, because (i) I have checked manually that F(X +
>> dX) ~ F(X) + J * dX and (ii) when I don't use -snes_mf and use the provided
>> Jacobian matrix the result is pretty much equivalent to using -snes_mf.
>>
>> In my understanding, this means that what I effectively feed to SNES at
>> the end of my FormFunction routine is a good approximation to J^(-1) F. As
>> a result, I naturally expect that the number of KSP iterations necessary to
>> achieve one SNES iteration be drastically reduced. However, I observe
>> virtually no change whatsoever in the number of iterations.
>>
>> Any thoughts about what I could be missing ? Maybe I forgot to set a SNES
>> or KSP option somewhere ? I can send pieces of code if needs be.
>>
>
> It sounds like you are putting this in the wrong place. If you have the
> action of a good preconditioner for the Jacobian, then you
> should use a PCSHELL and pass it to the KSP. It does not belong in the
> FormFunction.
>
>   Thanks,
>
>      Matt
>
>
>> Best
>>
>> Timothee
>>
>>
>>
>
>
> --
> 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
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20150928/f0f857de/attachment.html>