[petsc-users] Fwd: SNES with Fieldsplit + MatNest in parallel
Matthew Knepley
knepley at gmail.com
Tue Nov 9 05:25:53 CST 2021
On Tue, Nov 9, 2021 at 6:18 AM Nicolas Barnafi <nabw91 at gmail.com> wrote:
> Whoops, I think the notation wasn't very good, sorry. This should have
> been my previous message:
>
> Rank-0, is_0: [0 1 2] ... [633 634 635]
> Rank-0, is_1 [636 637 638] ... [729 730 731]
> Rank-1, is_0: [732 733 734] ... [1368 1369 1370]
> Rank-1, is_1 [1371 1372 1373] ... [1464 1465 1466]
>
> where I am printing the three first and last elements of each IS object,
> so in fact the first processor has the range [0,635] (is0) + [636,731]
> (is1) and the second one has [732, 1370] (is0) + [1371, 1466] (is1). This
> is actually the smallest (3D) problem I have access to without changing
> anything, as I have hard-coded some elements (i.e. boundary conditions). It
> still runs super quickly, so it is not much of a problem to debug it as it
> is.
>
The next step would be to verify that the operators you get in those slots
are correct in parallel. I normally just print out the matrix
for a small problem. This is the only way I know to debug assembly.
To check the solver, you set it to an exact solver and see that it takes 1
iterate. That means using LU as the subsolver
for the first block, using the full Schur complement, and solving the Schur
system with a residual tolerance of something like 1e-10.
Thanks,
Matt
> Thanks you,
> NB
>
>
> I am now completely confused. The way PCFIELDSPLIT works, when doing Schur
> complements, is to take
> in two ISes, which can be thought of as lists of vector rows. These two
> ISes must partition [0, N) if N is the
> global size of the problem. So
>
> 1) The numbers you have above do not appear to be a partition
>
> 2) The problem appears to be huge. Use a problem with 2-4 to debug.
>
> Thanks,
>
> Matt
>
>
>> Thanks!
>> NB
>>
>>
>>
>
> --
> 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
>
> https://www.cse.buffalo.edu/~knepley/
> <http://www.cse.buffalo.edu/~knepley/>
>
>
>
--
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
https://www.cse.buffalo.edu/~knepley/ <http://www.cse.buffalo.edu/~knepley/>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20211109/b7782626/attachment-0001.html>
More information about the petsc-users
mailing list