[petsc-dev] Petsc Support for Bandwidth Minimization

[Jed Brown]

On Mon, Jan 2, 2012 at 23:10, Dave Nystrom <dnystrom1 at comcast.net> wrote:

> Thanks.  I just tried a run with -ksp_type cg -pc_type cholesky
> -pc_factor_mat_ordering_type rcm -log_summary and did not see a performance
> difference.
>

Maybe that factorization was done under a different prefix? You should set
a different prefix for each solver so that you can address them
independently.

Also note that any reasonable ordering should produce essentially no fill
for your reduced matrix. The reason to use RCM is that it is specifically
billed as a bandwidth-reducing ordering and it is very inexpensive.


>  I'm trying to reproduce the results of our home grown cg with
> cholesky preconditioner but for some reason, I am getting about 3x the
> iteration count as I get with our home grown cg with cholesky
> preconditioner
> using the z-coupling.  Not sure what is wrong.  Maybe I have a bug.
>

This is likely a semantic bug. The ordering above should not affect the
iteration count when using "Cholesky".
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20120103/2e2cc3e7/attachment.html>