[petsc-dev] testing code which use partitioners

[Fande Kong]

It is definitely valuable to have partitioners produce machine independent
results.  We right now have to generate different outputs on different
machines in the testing system.

So far we found that just a few partitioners (literally Metis, ParMETIS and
PTScotch ) are interesting. Is it possible to have these packages to take a
user-provided random generator ?

BTW, a while ago, Jed recommended to take a look at KaHIP. I actually
invested a little bit on that a few months ago, and I found that KaHIP does
not even provide a parallel partitioning interface and then I gave up to
spend more time on it. We are definitely interested in the parallel
partitioning since it can potentially partition a large mesh.


Thanks,

Fande Kong



On Sun, Jan 6, 2019 at 9:49 PM Jed Brown via petsc-dev <
petsc-dev at mcs.anl.gov> wrote:

> METIS will use its internal implementation if USE_GKRAND is defined.
> Maybe other partitioners have similar functionality.
>
> I'm not wild about the complexity of providing our own definitions.
> Note that libmetis.so does not link to libpetsc.so so it would be
> nontrivial to make that work with both shared and static libraries.
>
> "Smith, Barry F. via petsc-dev" <petsc-dev at mcs.anl.gov> writes:
>
> >    It really irks me that we can't use the various partitioners in our
> nightly tests because they use rand() and produce different results on
> > different machines.
> >
> >    Here is what I propose
> >
> > 1) add ./configure flag --with-petsc-rand
> >
> > 2) if the flag is set then provide our own rand() and srand() (by
> crudely wrapping PetscRandom()) in pinit.c which is compiled if the flag is
> set
> >
> > 3) add the --with-petsc-rand to all the nightlybuilds
> >
> >   Problems with this plan?
> >
> >     Barry
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20190106/ae676c4f/attachment.html>