[petsc-dev] Problems with MatGetSubMatrices_SeqSBAIJ

[Alejandro Cosimo]

Hi,

After updating petsc from version 3.6.2 to 3.7, I had the following problem:

The result that I get from doing something like
MatMult(schurComplement,v1,v2); differs between the petsc versions that I
mentioned. I traced back the problem to the function
MatGetSubMatrices_SeqSBAIJ(). What I did in order to temporarily solve my
problem was to use in petsc 3.7 the code of MatGetSubMatrices_SeqSBAIJ()
corresponding to petsc 3.6.2.

How I solved the problem is not a real solution, but it allows me to keep
working. My question is: is it a bug or do you think that there could be
something wrong with my code from which I use petsc?

I think it should be a bug. If I run the ksp example 59 which uses FETI-DP
(whose implementation also uses something like
MatMult(schurComplement,v1,v2)) with petsc 3.7, I get the following
unexpected answer

mpiexec -n 4 ./ex59 -npx 2 -npy 2 -nex 60 -ney 60 -ksp_max_it 3
------------------FETI-DP stats-------------------------------
Number of degrees of freedom               :      355
Number of iterations                       :        3
Eigenvalues preconditioned operator        : 1.49e+00 8.86e+00
Error betweeen exact and computed solution : 1.91e+00
--------------------------------------------------------------

If I run the same problem with petsc 3.7 but with
MatGetSubMatrices_SeqSBAIJ() corresponding to petsc 3.6.2, I get the
expected answer

mpiexec -n 4 ./ex59 -npx 2 -npy 2 -nex 60 -ney 60 -ksp_max_it 3
------------------FETI-DP stats-------------------------------
Number of degrees of freedom               :      355
Number of iterations                       :        3
Eigenvalues preconditioned operator        : 1.08e+00 2.08e+00
Error betweeen exact and computed solution : 3.60e-02
--------------------------------------------------------------


Thanks,
Alejandro
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20160502/300b2a16/attachment.html>