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

Wed Jul 14 14:45:26 CDT 2021

On Jul 14, 2021, at 10:01 AM, Tang, Qi <tangqi at msu.edu<mailto:tangqi at msu.edu>> wrote:

Thanks a lot for the explanation, Matt and Stefano. That helps a lot.

Just to confirm, the comment in src/ts/impls/implicit/theta/theta.c seems to indicates TS solves U_{n+1}  in its SNES/KSP solve, but it actually solves the update dU_n in U_{n+1} = U_n - lambda*dU_n in the solve. Right?

SNESSolve yields U_{n+1}. But KSPSolve yields the Newton direction dU_n at each SNES iteration.

It actually makes a lot sense, because KSPSolve in TSSolve reports it uses zero initial guess. So if what I said is true, that effectively means it uses U0 as the initial guess.

Correct. TSSolve uses a warm start SNES so the previous solution is used as the initial guess for the next SNESSolve. Note that TSSolve calls SNESSolve instead of calling KSPSolve directly even when you are solving a linear problem.

Hong (Mr.)

Qi

On Jul 14, 2021, at 2:56 AM, Matthew Knepley <knepley at gmail.com<mailto:knepley at gmail.com>> wrote:

On Wed, Jul 14, 2021 at 4:43 AM Stefano Zampini <stefano.zampini at gmail.com<mailto:stefano.zampini at gmail.com>> wrote:
Qi

Backward Euler is a special case of Theta methods in PETSc (Theta=1). In src/ts/impls/implicit/theta/theta.c on top of SNESTSFormFunction_Theta you have some explanation of what is solved for at each time step (see below). SNES then solves for the Newton update dy_n  and the next Newton iterate is computed as x_{n+1} = x_{n} - lambda * dy_n. Hope this helps.

In other words, you should be able to match the initial residual to

  F(t + dt, 0, -Un / dt)

for your IFunction. However, it is really not normal to use U = 0. The default is to use U = U0
as the initial guess I think.

  Thanks,

     Matt

/*
  This defines the nonlinear equation that is to be solved with SNES
  G(U) = F[t0+Theta*dt, U, (U-U0)*shift] = 0

  Note that U here is the stage argument. This means that U = U_{n+1} only if endpoint = true,
  otherwise U = theta U_{n+1} + (1 - theta) U0, which for the case of implicit midpoint is
  U = (U_{n+1} + U0)/2
*/
static PetscErrorCode SNESTSFormFunction_Theta(SNES snes,Vec x,Vec y,TS ts)

On Jul 14, 2021, at 6:12 AM, Tang, Qi <tangqi at msu.edu<mailto:tangqi at msu.edu>> wrote:

Hi,

During the process to experiment the suggestion Matt made, we ran into some questions regarding to TSSolve vs KSPSolve. We got different initial unpreconditioned residual using two solvers. Let’s say we solve the problem with backward Euler and there is no rhs. We guess TSSolve solves
(U^{n+1}-U^n)/dt = A U^{n+1}.
(We only provides IJacobian in this case and turn on TS_LINEAR.)
So we guess the initial unpreconditioned residual would be ||U^n/dt||_2, which seems different from the residual we got from a backward Euler stepping we implemented by ourself through KSPSolve.

Do we have some misunderstanding on TSSolve?

Thanks,
Qi
T5 at LANL

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$>

--
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!msQrz7__TrpOmaTVhvY1yLAlDQXNJ5jcYVAxF4lcpyLrZqt2lFe22bkbuJMizA$>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20210714/f43d29b4/attachment-0001.html>