[petsc-dev] PaStiX solution

[Alexander Grayver]

On 28.12.2011 16:27, Jed Brown wrote:
> On Wed, Dec 28, 2011 at 09:10, Alexander Grayver 
> <agrayver at gfz-potsdam.de <mailto:agrayver at gfz-potsdam.de>> wrote:
>
>     So if I understood it right, I put diff you gave in file
>     pastix_patch in petsc-dev dir and called:
>
>     lib/petsc-dev> patch -Np1 < pastix_patch
>
>     And built petsc again. Now
>
>     lib/petsc-dev> hg diff src/mat/impls/aij/mpi/pastix/pastix.c
>
>
> Yes, the patch applied correctly.
>
>
>     So I guess it's patched correctly?
>     Because I don't see much difference in output. Please find it
>     attached.
>     Did I make something wrong?
>
>
> Hmm, it is hard-coded to say that the matrix is symmetric. We are 
> setting API_FACT_LU (you could check that it has not been changed by 
> attaching a debugger in MatFactorNumeric_PaStiX and printing 
> lu->iparm[IPARM_FACTORIZATION]). Maybe PaStiX is ignoring this and 
> always doing LLt instead.

No. I checked, the API_FACT_LU is set correctly.

> We probably need to set API_PARM_LDLT (and possibly API_PARM_LU, but 
> they claim that only does static pivoting, which might still be a 
> problem). You could try adding this to 
> src/mat/impls/aij/mpi/pastix/pastix.c or ask the PaStiX developers 
> what they suggest for symmetric indefinite systems.

I tried API_PARM_LDLT, but it has no effect.

Thanks for help, I will consider writing PaStiX team.

Regards,
Alexander
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20111228/564fee4e/attachment.html>