[petsc-users] Performance of mumps vs. Intel Pardiso

[Hong]

Faraz:

> >Mumps uses parmetis or scotch for parallel symbolic factorization. For
> sequential symbolic >factorization, it has several matrix orderings, which
> you can experiment with the option >'-mat_mumps_icntl_7 <>'.
> > I doubt any of these ordering would match the performance of Pardiso.
>
> Thanks, I have been experimenting with the different ordering options in
> Mumps. So far I have not seen any speed difference among them. Do you feel
> Pardiso ordering is so superior to result in 2X increase in speed? I find
> this hard to believe since my problem is fairly classic and can't imagine
> why a 10+ year old algorithm would be so much better than what's available
> today.
>
> I do not think Pardiso has much algorithmic superior over mumps.

>
> > Again, how large is your matrix? How do you run Pardiso in parallel? Can
> you use Pardiso on large matrices as mumps?
>
> My matrix is 3million^2 with max 1000 non-zeroes per line. Pardiso
> supports multi-thread, so I just do export OMP_NUM_THREADS=24 to use all
> available cpus . When running Petsc/Mumps I have to do export
> OMP_NUM_THREADS=1, otherwise I get very weird cpu utilization. Perhaps that
> has something to do with the speed difference?
>

Taking advantage of computer architecture likely makes Pardiso superior.

Hong
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20160629/419164f3/attachment-0001.html>