[petsc-dev] PCFIELDSPLIT with inner -ksp_reuse_preconditioner

[Matthew Knepley]

On Wed, Jul 13, 2022 at 11:02 AM Barry Smith <bsmith at petsc.dev> wrote:

>
>   Some of the ugliness of SetFromOptions in some of the nested solvers :-(
>
> * thread #1, queue = 'com.apple.main-thread', stop reason = breakpoint 1.1
>   * frame #0: 0x00000001132795f0
> libpetsc.3.017.3.dylib`KSPSetFromOptions(ksp=0x00007f899f149a70) at itcl.c
> :324:19
>     frame #1: 0x0000000112eddda4
> libpetsc.3.017.3.dylib`PCSetUp_FieldSplit(pc=0x00007f899f08d870) at
> fieldsplit.c:1087:32
>     frame #2: 0x0000000112dfb874
> libpetsc.3.017.3.dylib`PCSetUp(pc=0x00007f899f08d870) at precon.c:991:5
>     frame #3: 0x0000000113285834
> libpetsc.3.017.3.dylib`KSPSetUp(ksp=0x00007f899e95f670) at itfunc.c:401:3
>     frame #4: 0x000000011328b5e9
> libpetsc.3.017.3.dylib`KSPSolve_Private(ksp=0x00007f899e95f670,
> b=0x00007f899f0dec70, x=0x00007f899f0ce870) at itfunc.c:835:3
>     frame #5: 0x000000011328aa66
> libpetsc.3.017.3.dylib`KSPSolve(ksp=0x00007f899e95f670,
> b=0x00007f899f0dec70, x=0x00007f899f0ce870) at itfunc.c:1066:3
>     frame #6: 0x000000010e7c94e8 ex1111`main(argc=12,
> args=0x00007ff7b1739328) at ex1111.c:33:3
>     frame #7: 0x000000011703751e dyld`start + 462
>
> The SetFromOptions doesn't always traverse all the sub-methods suitably
> calling SetFromOptions on them because they may not have been created by
> the time that
> the SetFromOptions is first called. In this case, the PCSetUp_FieldSplit()
> creates the sub solvers and calls the options database on them after the
> initial call to PCSetFromOptions_FieldSplit()
>
> I've attached a patch that should fix this particular issue.
> I ran it with your test code, and it seemed to resolve the issue and ran
> the test suite, and it did not break anything.
>
> A redesigned/implemented PETSc would need some thought on how to improve
> the SetFromOptions model to prevent this kind of difficulty.
>

My solution, which is used elsewhere in PETSc, would be for all
PetscObjects to have a setfromoptions flag. If it creates a subobject and
its own
flag is set, it calls SetFromOptions on it.

  Thanks,

     Matt


> Barry
>
> On Jul 13, 2022, at 8:40 AM, Pierre Jolivet <pierre at joliv.et> wrote:
>
> Hello,
> A user(*) wants to switch back and forth with the option
> -fieldsplit_1_ksp_reuse_preconditioner, but it’s generating results that I
> don’t understand.
> In the following example I solve three successive linear systems.
> If I run
> $ ./ex1111 -pc_type fieldsplit -fieldsplit_pc_type lu -log_view -info
> -options_view -switch false
> -fieldsplit_1_ksp_reuse_preconditioner is always false, and I properly get:
> 1) convergence in a single iteration every time
> 2) [0] PCSetUp(): Setting up PC with same nonzero pattern for the last
> solve
> 3) MatLUFactorNum         3 in -log_view
>
> Now, if I run
> $ ./ex1111 -pc_type fieldsplit -fieldsplit_pc_type lu -log_view -info
> -options_view -switch true
> -fieldsplit_1_ksp_reuse_preconditioner is set to true for the second solve
> and switched back to false for the last solve.
> I (wrongfully, IMHO?) get:
> 1) convergence in 2 iterations for the last solve (instead of 1?)
> 2) [0] PCSetUp(): Leaving PC with identical preconditioner since reuse
> preconditioner is set (which should not be set for the last solve?)
> 3) MatLUFactorNum         1 (instead of 2?)
> 4) #PETSc Option Table entries:
> -fieldsplit_1_ksp_reuse_preconditioner false (though from -info, cf. point
> #2 just above, this option value is true?)
>
> Does it make sense or do you agree that something is off?
>
> Thanks,
> Pierre
>
> <ex1111.c>
>
> (*)
> https://community.freefem.org/t/supply-petsc-numbering-for-shared-memory-block-matrix/486/18
>
>
>

-- 
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-dev/attachments/20220713/bb954876/attachment-0001.html>