[petsc-users] PCFIELDSPLIT with MATSBAIJ

Zhang, Hong hzhang at mcs.anl.gov
Mon Aug 28 09:35:31 CDT 2023


Carl-Johan,
________________________________
Thanks for the advice. Our code is well-tested for AIJ. Is SBAIJ inherently significantly slower than AIJ, or is it just that it's not so much used and thus not as developed?
SBAIJ only stores upper half triangular part of matrix. When it needs a lower triangular part of entry, it has to jump around searching for that entry (column search instead of row accessing in AIJ), causing overhead for data-accessing. In parallel computation, it leads to extra inter-processor communication.
Hong
On Sat, Aug 26, 2023 at 6:27 PM Zhang, Hong via petsc-users <petsc-users at mcs.anl.gov<mailto:petsc-users at mcs.anl.gov>> wrote:
I would suggest avoiding using SBAIJ matrices, at least in the phase of application code development. We implemented SBAIJ for saving storage, not computational efficiency. SBAIJ does not have as many support as AIJ. After your code works for AIJ, then you may consider taking advantage of smaller storage of SBAIJ (could at cost of communication overhead).
Hong
________________________________
From: petsc-users <petsc-users-bounces at mcs.anl.gov<mailto:petsc-users-bounces at mcs.anl.gov>> on behalf of Pierre Jolivet via petsc-users <petsc-users at mcs.anl.gov<mailto:petsc-users at mcs.anl.gov>>
Sent: Saturday, August 26, 2023 10:20 AM
To: Carl-Johan Thore <carl-johan.thore at liu.se<mailto:carl-johan.thore at liu.se>>
Cc: petsc-users <petsc-users at mcs.anl.gov<mailto:petsc-users at mcs.anl.gov>>
Subject: Re: [petsc-users] PCFIELDSPLIT with MATSBAIJ



On 27 Aug 2023, at 12:14 AM, Carl-Johan Thore <carl-johan.thore at liu.se<mailto:carl-johan.thore at liu.se>> wrote:

“Well, your A00 and A11 will possibly be SBAIJ also, so you’ll end up with the same issue.”
I’m not sure I follow. Does PCFIELDSPLIT extract further submatrices from these blocks, or is there
somewhere else in the code that things will go wrong?

Ah, no, you are right, in that case it should work.

For the MATNEST I was thinking to get some savings from the block-symmetry at least
even if symmetry in A00 and A11 cannot be exploited; using SBAIJ for them would just be a
(pretty big) bonus.

“I’ll rebase on top of main and try to get it integrated if it could be useful to you (but I’m traveling
right now so everything gets done more slowly, sorry).”
Sound great, Thanks again!

The MR is there https://gitlab.com/petsc/petsc/-/merge_requests/6841.
I need to add a new code path in MatCreateRedundantMatrix() to make sure the resulting Mat is indeed SBAIJ, but that is orthogonal to the PCFIELDSPLIT issue.
The branch should be usable in its current state.

Thanks,
Pierre


From: Pierre Jolivet <pierre.jolivet at lip6.fr<mailto:pierre.jolivet at lip6.fr>>
Sent: Saturday, August 26, 2023 4:36 PM
To: Carl-Johan Thore <carl-johan.thore at liu.se<mailto:carl-johan.thore at liu.se>>
Cc: Carl-Johan Thore <carljohanthore at gmail.com<mailto:carljohanthore at gmail.com>>; petsc-users at mcs.anl.gov<mailto:petsc-users at mcs.anl.gov>
Subject: Re: [petsc-users] PCFIELDSPLIT with MATSBAIJ




On 26 Aug 2023, at 11:16 PM, Carl-Johan Thore <carl-johan.thore at liu.se<mailto:carl-johan.thore at liu.se>> wrote:

"(Sadly) MATSBAIJ is extremely broken, in particular, it cannot be used to retrieve rectangular blocks in MatCreateSubMatrices, thus you cannot get the A01 and A10 blocks in PCFIELDSPLIT.
I have a branch that fixes this, but I haven’t rebased in a while (and I’m AFK right now), would you want me to rebase and give it a go, or must you stick to a release tarball?"

Ok, would be great if you could look at this! I don't need to stick to any particular branch.

Do you think MATNEST could be an alternative here?

Well, your A00 and A11 will possibly be SBAIJ also, so you’ll end up with the same issue.
I’m using both approaches (monolithic SBAIJ or Nest + SBAIJ), it was crashing but I think it was thoroughly fixed in https://gitlab.com/petsc/petsc/-/commits/jolivet/feature-matcreatesubmatrices-rectangular-sbaij/
It is ugly code on top of ugly code, so I didn’t try to get it integrated and just used the branch locally, and then moved to some other stuff.
I’ll rebase on top of main and try to get it integrated if it could be useful to you (but I’m traveling right now so everything gets done more slowly, sorry).

Thanks,
Pierre


My matrix is
[A00 A01;
A01^t A11]
so perhaps with MATNEST I can make use of the block-symmetry at least, and then use MATSBAIJ for
A00 and A11 if it's possible to combine matrix types which the manual seems to imply.

Kind regards
Carl-Johan



On 26 Aug 2023, at 10:09 PM, Carl-Johan Thore <carljohanthore at gmail.com<mailto:carljohanthore at gmail.com>> wrote:

Hi,

I'm trying to use PCFIELDSPLIT with MATSBAIJ in PETSc 3.19.4.
According to the manual "[t]he fieldsplit preconditioner cannot
currently be used with the MATBAIJ or MATSBAIJ data formats if the
blocksize is larger than 1". Since my blocksize is exactly 1 it would seem that I can use PCFIELDSPLIT. But this fails with "PETSC ERROR: For symmetric format, iscol must equal isrow"
from MatCreateSubMatrix_MPISBAIJ. Tracing backwards one ends up in
fieldsplit.c at

/* extract the A01 and A10 matrices */ ilink = jac->head;
PetscCall(ISComplement(ilink->is_col, rstart, rend, &ccis)); if
(jac->offdiag_use_amat) { PetscCall(MatCreateSubMatrix(pc->mat,
ilink->is, ccis, MAT_INITIAL_MATRIX, &jac->B)); } else {
       PetscCall(MatCreateSubMatrix(pc->pmat, ilink->is, ccis,
MAT_INITIAL_MATRIX, &jac->B)); }

This, since my A01 and A10 are not square, seems to explain why iscol is not equal to isrow.
From this I gather that it is in general NOT possible to use
PCFIELDSPLIT with MATSBAIJ even with block size 1?

Kind regards,
Carl-Johan

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20230828/7fd4f7bc/attachment-0001.html>


More information about the petsc-users mailing list