On Sun, Jul 28, 2019 at 2:54 AM Pierre Jolivet via petsc-dev <
[email protected]> wrote:

> Hello,
> I’m facing multiple issues with PCREDUNDANT and MATMPISBAIJ:
> 1)
> https://www.mcs.anl.gov/petsc/petsc-current/src/mat/impls/sbaij/mpi/mpisbaij.c.html#line3354
>  shouldn’t
> this be sum != N? I’m running an example where it says that sum (4) !=
> Nbs (60), with a bs=15.
>

Clearly a bug.


> 2) when I’m using MATMPIBAIJ, I can do stuff like: -prefix_mat_type baij
> -prefix_pc_type redundant -prefix_redundant_pc_type ilu, and in the
> KSPView, I have "package used to perform factorization: petsc”, so the
> underlying MatType is indeed MATSEQBAIJ.
>
However, with MATMPISBAIJ, if I do: -prefix_mat_type sbaij -prefix_pc_type
> redundant, first, it looks like you are hardwiring a PCLU (MatGetFactor()
> line 4440 in src/mat/interface/matrix.c
>

Using LU as a default for symmetric matrices does seem wrong.


> Could not locate a solver package.), then, if I
> append -prefix_redundant_pc_type cholesky, I end up with an error related
> to MUMPS: MatGetFactor_sbaij_mumps() line 2625 in
> src/mat/impls/aij/mpi/mumps/mumps.c Cannot use PETSc SBAIJ matrices with
> block size > 1 with MUMPS Cholesky, use AIJ matrix instead. Why isn’t this
> call dispatched to PETSc Cholesky for SeqSBAIJ matrices?
>
>
Generally, we don't like to switch parameters under the covers like this.
We would rather you get your inputs right so you know what is going on.


> Thanks,
> Pierre
>
> 1) I don’t think this is tested right now, at least not in
> src/ksp/ksp/examples/tutorials
> 2) reproducer: src/ksp/ksp/examples/tutorials/ex2.c
> $ mpirun -np 2 ./ex2 -pc_type redundant -mat_type sbaij
> // error because trying to do LU with a symmetric matrix
> $ mpirun -np 2 ./ex2 -pc_type redundant -mat_type sbaij -redundant_pc_type
> cholesky -ksp_view
> // you’ll see: that MUMPS is being used, but since bs=1, it’s working, but
> it won’t for the general case
> //                  the MatType is mpisbaij with "1 MPI processes" whereas
> with baij, it’s seqbaij
>
>

Reply via email to