[petsc-users] reuse of KSP is sometimes harmful

[Matthew Knepley]

2011/1/28 Cristian Tibirna <ctibirna at giref.ulaval.ca>

> Hello
>
> We use PETSc-2.3.3-p15 (long story...) embedded inside a large FE code from
> which it would be difficult for me to extract exact PETSc code. Sorry for
> this. But we have a rather ugly problem and I would like to know just if it
> is
> our use or there might be some intrinsic problem.
>
> What we do, in principle, is
> 1) create a KSP, a Mat and some Vec
> 2) start a loop (e.g. a home-made nonlinear solver)
>        inside the loop:
>        2.1)  reassemble Mat either through one of two variants:
>                a) sometimes the nonzero structure is constant
>                        only the assembled values are changing
>                b) sometimes the nonzero structure changes too,
>                        which brings us to create and use a new Mat
>        2.2) reassemble the rhs Vec
>        2.3) on fait un KSPSetOperators avec
>        2.4) solve the system and get the solution (or correction)
>
> NOTE: As you see, the KSP is created only once then reused.
> NOTE: the point b) -- the main issue of this email -- consists most usually
> of
> a MatPtAP operation, which naturally gives us a new Mat with new non-zero
> structure.
>
> If we do the above with the variant a), all works perfectly.
> if we do the variant b), the solution is completely aberant.
>

That should work. This is not really enough of a description for us to debug
the problem.
I suggest using -pc_type lu -ksp_type preonly for an exact solution, and
then check the
solution for doing a MatMult(). Usually, this problem occurs because you
forget to call
KSPSetOperators().

   Matt


> The only workaround we found was to recreate the KSP at each iteration in
> the
> loop. Then the solution vector is correctly computed.
>
> What gives?
>
> --
> Cristian Tibirna                                (1-418-) 656-2131 / 4340
>  Laval University - Quebec, CAN ... http://www.giref.ulaval.ca/~ctibirna
>  Research professional at GIREF ... ctibirna at giref.ulaval.ca
>



-- 
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/20110128/36b10b25/attachment.htm>