[petsc-users] [EXTERNAL] Re: Problem with PCFIELDSPLIT

[Tang, Qi]

Matt,

We are confident that the modified Schur complement works well based on our subksp and ksp tests. Let me summarize what you suggest us to do regarding to our original question of TSSolve. Instead of calling subksp, you suggest we should provide two matrices J and Jpre through TSSetIJacobian, where J is the original Jacobian and Jpre is the one we modified and both are 2x2 blocks. Then if we call fieldsplit, petsc will automatically use Jpre to construct its Schur complement in the preconditioner stage. Is that what you suggested?

Thanks a lot!

Qi





On Jul 7, 2021, at 3:54 PM, Matthew Knepley <knepley at gmail.com<mailto:knepley at gmail.com>> wrote:

On Wed, Jul 7, 2021 at 2:33 PM Jorti, Zakariae <zjorti at lanl.gov<mailto:zjorti at lanl.gov>> wrote:

Hi Matt,


Thanks for your quick reply.

I have not completely understood your suggestion, could you please elaborate a bit more?

For your convenience, here is how I am proceeding for the moment in my code:


TSGetKSP(ts,&ksp);

KSPGetPC(ksp,&pc);

PCSetType(pc,PCFIELDSPLIT);

PCFieldSplitSetDetectSaddlePoint(pc,PETSC_TRUE);

PCSetUp(pc);

PCFieldSplitGetSubKSP(pc, &n, &subksp);

KSPGetPC(subksp[1], &(subpc[1]));

I do not like the two lines above. We should not have to do this.

KSPSetOperators(subksp[1],T,T);

 In the above line, I want you to use a separate preconditioning matrix M, instead of T. That way, it will provide
the preconditioning matrix for your Schur complement problem.

  Thanks,

      Matt

KSPSetUp(subksp[1]);

PetscFree(subksp);

TSSolve(ts,X);


Thank you.

Best,


Zakariae

________________________________
From: Matthew Knepley <knepley at gmail.com<mailto:knepley at gmail.com>>
Sent: Wednesday, July 7, 2021 12:11:10 PM
To: Jorti, Zakariae
Cc: petsc-users at mcs.anl.gov<mailto:petsc-users at mcs.anl.gov>; Tang, Qi; Tang, Xianzhu
Subject: [EXTERNAL] Re: [petsc-users] Problem with PCFIELDSPLIT

On Wed, Jul 7, 2021 at 1:51 PM Jorti, Zakariae via petsc-users <petsc-users at mcs.anl.gov<mailto:petsc-users at mcs.anl.gov>> wrote:

Hi,


I am trying to build a PCFIELDSPLIT preconditioner for a matrix

J =  [A00  A01]

       [A10  A11]

that has the following shape:


M_{user}^{-1} = [I   -ksp(A00) A01] [ksp(A00)           0] [I                        0]

                          [0                        I]  [0               ksp(T)] [-A10 ksp(A00)  I ]


where T is a user-defined Schur complement approximation that replaces the true Schur complement S:= A11 - A10 ksp(A00) A01.


I am trying to do something similar to this example (lines 41--45 and 116--121): https://www.mcs.anl.gov/petsc/petsc-current/src/snes/tutorials/ex70.c.html<https://urldefense.com/v3/__https://www.mcs.anl.gov/petsc/petsc-current/src/snes/tutorials/ex70.c.html__;!!HXCxUKc!hoEfgnaraTfQoSgAiplsc6GJ_HuPXN88m5AJVy1gb7WVMNkGENDnJ3zToOGlhw$>


The problem I have is that I manage to replace S with T on a separate single linear system but not for the linear systems generated by my time-dependent PDE. Even if I set the preconditioner M_{user}^{-1} correctly, the T matrix gets replaced by S in the preconditioner once I call TSSolve.

Do you have any suggestions how to fix this knowing that the matrix J does not change over time?

I don't like how it is done in that example for this very reason.

When I want to use a custom preconditioning matrix for the Schur complement, I always give a preconditioning matrix M to the outer solve.
Then PCFIELDSPLIT automatically pulls the correct block from M, (1,1) for the Schur complement, for that preconditioning matrix without
extra code. Can you do this?

  Thanks,

    Matt

Many thanks.


Best regards,


Zakariae



--
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/<https://urldefense.com/v3/__http://www.cse.buffalo.edu/*knepley/__;fg!!HXCxUKc!hoEfgnaraTfQoSgAiplsc6GJ_HuPXN88m5AJVy1gb7WVMNkGENDnJ3wB3dcMFw$>


--
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/<https://urldefense.com/v3/__http://www.cse.buffalo.edu/*knepley/__;fg!!HXCxUKc!hoEfgnaraTfQoSgAiplsc6GJ_HuPXN88m5AJVy1gb7WVMNkGENDnJ3wB3dcMFw$>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20210716/3d1bb7ae/attachment.html>