[petsc-users] PC introduce errors at processor limits?

[Matthew Knepley]

On Wed, May 8, 2013 at 3:33 AM, Frederik Treue <frtr at fysik.dtu.dk> wrote:

> On Tue, 2013-05-07 at 09:23 -0500, Matthew Knepley wrote:
> > On Tue, May 7, 2013 at 9:20 AM, Frederik Treue <frtr at fysik.dtu.dk>
> > wrote:
> >         Hi,
> >
> >         I may be overlooking something very obvious here, but:
> >
> >         I'm trying to solve a convection-diffusion problem. The method
> >         I use
> >         requires me to solve some Helmholtz and Poisson equations,
> >         which is the
> >         time consuming part. In order to reduce this time, I try to
> >         use
> >         preconditioners. However, no matter which preconditioner I use
> >         (I've
> >         tried PCJACOBI,PCBJACOBI,PCPBJACOBI,PCMG with 2 levels,
> >         galerkin) it
> >         introduces errors along the edges of the local domains when
> >         using
> >         multiple processors. These errors are small, but they don't
> >         converge to
> >         0 as a function of resolution. I have checked with PCNONE,
> >         which
> >         eliminates the problem, but becomes unbearably slow.
> >
> >         Is this somehow unavoidable? Or am I making some silly
> >         mistake? The code
> >         is somewhat complicated, but if desired, I can try to cook up
> >         a
> >         proof-of-(non)concept.
> >
> >
> > You have made a mistake somewhere, probably in your parallel function
> > evaluation.
>
> But how can this be? If it works without a preconditioner (which it
> does), doesn't that prove that the operators are correctly implemented?
> Or am I missing something something here?


Here is my reasoning:

1) Examples work for you

2) Jacobi is identical in serial and parallel

3) This part of PETSc is tested by thousands of people every day, not to
mention all
    regression tests, and has been stable for a decade at least.

Thus my conclusion is that you have an error in the parallel residual
evaluation.

My advice is to take a working example and slowly change it to get to your
equation.

   Matt


>
> /Frederik Treue
>
>
>
>


-- 
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/20130508/eee360b5/attachment-0001.html>