[petsc-dev] fieldsplit + composite + ksp

[Matthew Knepley]

On Wed, Sep 18, 2019 at 2:49 AM Pierre Jolivet via petsc-dev <
petsc-dev at mcs.anl.gov> wrote:

> Hello,
> I’m solving the following dummy system
> http://jolivet.perso.enseeiht.fr/composite_ksp.tar.gz
> [A, B];[C, D], with a PCFIELDSPLIT. For the PC of D, I’m using a
> PCCOMPOSITE with two sub PCs. One of which is a PCKSP.
> Could you please help me figure out what is wrong in the following piece
> of code, that may be launched with the following arguments:
> $ mpirun -n 1 ./a.out -ksp_type preonly -pc_type fieldsplit
> -fieldsplit_1_pc_type composite -fieldsplit_1_sub_1_pc_type ksp
> -fieldsplit_1_sub_1_ksp_ksp_type gmres -fieldsplit_1_sub_1_ksp_pc_type gamg
> -fieldsplit_1_sub_1_ksp_ksp_converged_reason
> -fieldsplit_1_sub_1_ksp_pc_gamg_sym_graph 1
> -fieldsplit_1_sub_1_ksp_pc_gamg_square_graph 10
> -fieldsplit_1_sub_1_ksp_ksp_rtol 1e-8
>
> It solves the dummy system twice, with a varying block D.
>

Its not the PC, its the matrix. Everything in the PC gets resetup just like
you want.

I did MatEqual(S2_1, S2_001, &equal) and equal was false.

  Thanks,

     Matt


> It should give you:
>       Linear fieldsplit_1_sub_1_ksp_ solve converged due to CONVERGED_RTOL
> iterations 8
> solve #0: 16098.3
>       Linear fieldsplit_1_sub_1_ksp_ solve did not converge due to
> DIVERGED_PC_FAILED iterations 0
>                      PC_FAILED due to SUBPC_ERROR
> solve #1: inf
>
> If I switch line 70 to #if 0, I get the expected output:
>       Linear fieldsplit_1_sub_1_ksp_ solve converged due to CONVERGED_RTOL
> iterations 8
> solve #0: 16098.3
>       Linear fieldsplit_1_sub_1_ksp_ solve converged due to CONVERGED_RTOL
> iterations 8
> solve #1: 325.448
>
> I’m realizing that this has probably nothing to do with the outer
> PCFIELDSPLIT, but this comes from a rather large FSI solver, so reproducing
> this behavior in “only” 97 SLOC is good enough for your I hope.
>
> Thanks in advance,
> Pierre
>


-- 
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/20190918/b296fcdd/attachment.html>