[petsc-users] Example of SNES using matrix free jacobian

[Barry Smith]

  This
  [0]PETSC ERROR: #1 SNES callback function
   indicates the crash is in your computeResidual function and thus you need to debug your function

   Barry


> On Oct 29, 2024, at 4:28 PM, Daniel Pino Munoz <daniel.pino_munoz at mines-paristech.fr> wrote:
> 
> I ran it with -malloc_debug and it does not change anything.
> 
> The output is the following:
> 
> he absolute tolerance is 0.001
> The relative tolerance is 0.001
> The divergence tolerance is 10000
> The maximum iterations is 10000
> Initial load !
> [0]PETSC ERROR: ------------------------------------------------------------------------
> [0]PETSC ERROR: Caught signal number 11 SEGV: Segmentation Violation, probably memory access out of range
> [0]PETSC ERROR: Try option -start_in_debugger or -on_error_attach_debugger
> [0]PETSC ERROR: or see https://urldefense.us/v3/__https://petsc.org/release/faq/*valgrind__;Iw!!G_uCfscf7eWS!eQxllmSvw4WuVrXfPz5uK3h9cSjVYg1fCeWM8bhtm4NwaGhghQVhAMg18ppPqOQxe2lkX-UQyOXgJx91MV6-KqE$  and https://urldefense.us/v3/__https://petsc.org/release/faq/__;!!G_uCfscf7eWS!eQxllmSvw4WuVrXfPz5uK3h9cSjVYg1fCeWM8bhtm4NwaGhghQVhAMg18ppPqOQxe2lkX-UQyOXgJx910VEZbZQ$ 
> [0]PETSC ERROR: ---------------------  Stack Frames ------------------------------------
> [0]PETSC ERROR: The line numbers in the error traceback are not always exact.
> [0]PETSC ERROR: #1 SNES callback function
> [0]PETSC ERROR: #2 SNESComputeFunction() at /home/daniel-pino/Software/Dependencies/petsc/src/snes/interface/snes.c:2489
> [0]PETSC ERROR: #3 SNESSolve_KSPONLY() at /home/daniel-pino/Software/Dependencies/petsc/src/snes/impls/ksponly/ksponly.c:27
> [0]PETSC ERROR: #4 SNESSolve() at /home/daniel-pino/Software/Dependencies/petsc/src/snes/interface/snes.c:4841
> --------------------------------------------------------------------------
> MPI_ABORT was invoked on rank 0 in communicator MPI_COMM_WORLD
>   Proc: [[27669,1],0]
>   Errorcode: 59
> 
> NOTE: invoking MPI_ABORT causes Open MPI to kill all MPI processes.
> You may or may not see output from other processes, depending on
> exactly when Open MPI kills them.
> --------------------------------------------------------------------------
> 
> 
> 
> On 29/10/2024 20:17, Barry Smith wrote:
>>    Hmm, cut and paste the output when it crashes.
>> 
>>     Also run with -malloc_debug and see what happens
>> 
>> 
>>     Barry
>> 
>> 
>>> On Oct 29, 2024, at 3:13 PM, Daniel Pino Munoz <daniel.pino_munoz at mines-paristech.fr> wrote:
>>> 
>>> Hi Barry,
>>> 
>>> Thanks for getting back to me!
>>> 
>>> I tried replacing KSPSetOperators(ksp, J, J); by SNESSetJacobian(snes,J,J, MatMFFDComputeJacobian)
>>> 
>>> and I get the same result = It works in Debug mode but not in Release. I also ran valgrind and it did not catch any memory problem.
>>> 
>>> Any ideas?
>>> 
>>> PS : You are right regarding the number of iterations of the non preconditioned problem. In the previous version of the code that only used a KSP, I already had to set -ksp_gmres_restart 100. But thanks for the heads up.
>>> 
>>> Best,
>>> 
>>>   Daniel
>>> 
>>> On 29/10/2024 20:01, Barry Smith wrote:
>>>>    Don't call
>>>> 
>>>>    KSPSetOperators(ksp, J, J);
>>>> 
>>>> 
>>>>     instead call
>>>> 
>>>>     SNESSetJacobian(snes,J,J, MatMFFDComputeJacobian)
>>>> 
>>>>     but I am not sure that would explain the crash.
>>>> 
>>>>     BTW: since you are applying no preconditioner if the matrix is ill-conditioned it may take many iterations or not converge. You can try something like -ksp_gmres_restart 100 or similar value to try to improve convergence (default is 30).
>>>> 
>>>>     Barry
>>>> 
>>>> 
>>>> 
>>>> 
>>>>> On Oct 29, 2024, at 12:37 PM, Daniel Pino Munoz <daniel.pino_munoz at mines-paristech.fr> wrote:
>>>>> 
>>>>> Dear all,
>>>>> 
>>>>> I have a linear problem that I am currently solving with a KSP matrix-free.
>>>>> 
>>>>> I would like to move on to a non linear problem, so figure I could start by solving the same linear problem using SNES. So I am setting the problem as follows:
>>>>> 
>>>>> SNESCreate(PETSC_COMM_WORLD, &snes);
>>>>> MatCreateShell(PETSC_COMM_WORLD, n_dofs, n_dofs, PETSC_DETERMINE, PETSC_DETERMINE, &ctx, &J);
>>>>> MatCreateShell(PETSC_COMM_WORLD, n_dofs, n_dofs, PETSC_DETERMINE, PETSC_DETERMINE, &ctx, &B);
>>>>> MatShellSetOperation(J, MATOP_MULT, (void (*)(void))(Multiplication));
>>>>> MatCreateVecs(J, &x_sol, &b);
>>>>> VecDuplicate(x_sol, &r);
>>>>> SNESSetFromOptions(snes);
>>>>> SNESSetFunction(snes, r, &(computeResidual), &ctx);
>>>>> SNESSetUseMatrixFree(snes, PETSC_FALSE, PETSC_TRUE);
>>>>> SNESGetLineSearch(snes, &linesearch);
>>>>> SNESGetKSP(snes, &ksp);
>>>>> KSPSetOperators(ksp, J, J);
>>>>> KSPSetInitialGuessNonzero(ksp, PETSC_TRUE);
>>>>> 
>>>>> I tested it with a small problem (compiled in debug) and it works.
>>>>> 
>>>>> When I compiled it in Release, it crashes with a segfault. I tried running the Debug version through valgrind, but even for a small problem, it is too slow. So I was wondering if you guys could see any rocky mistake on the lines I used above?
>>>>> 
>>>>> Otherwise, is there any example that uses a SNES combined with a matrix free KSP operator?
>>>>> 
>>>>> Thank you,
>>>>> 
>>>>>   Daniel
>>>>>