[petsc-dev] proposed minor PetscPartitioner changes

[Matthew Knepley]

On Mon, Nov 6, 2017 at 8:34 AM, Smith, Barry F. <bsmith at mcs.anl.gov> wrote:

> > On Nov 6, 2017, at 7:27 AM, Matthew Knepley <knepley at gmail.com> wrote:
> >
> > On Mon, Nov 6, 2017 at 8:24 AM, Smith, Barry F. <bsmith at mcs.anl.gov>
> wrote:
> >
> >    Vaclav,
> >
> >       Actually you should not just do this! PETSc already has a full
> class for managing partitioning (that Matt ignored for no good reason)
> >
> > Obviously, a good reason existed. All sorts of horrible Mat-specific
> crap was in these.
>
>   Matt,
>
>    There is no "Mat-specific crap" it in there. A Mat is used to represent
> the graph (you do exactly the same thing in your horrible code, convert to
> the matrix format that each external package uses, why have all the
> duplicate code for no reason.)
>
> > It was impossible to use for mesh partitioning.
>
>   Bullshit, you didn't even try.


A useful comment if based on fact or experience.


>
> > In addition,
> > when partitioning meshes, you need extra things (like calculation of the
> dual graph), which is not used in the matrix case.
>
>    MatPartitioning is NOT about partitioning MATRICES it is about
> partitioning GRAPHS, so it can support anything related to graphs including
> dual graphs etc. Again, you didn't even try to extend/improve MatPartioning
> you just wrote all your own duplicative code, and not even particularly
> well.
>

By "not particularly well, I assume there is some sort of deficiency there.
However, I do not believe that there is.

   Matt


>    Once Vaclav takes a look at MatPartitioning he'll realize quickly he
> can do a proper refactorization of all of the partitioning code that is
> simpler and better. It is about time this rathole gets cleaned up.
>
>    Barry
>
> >
> >   Matt
> >
> >
> > see MatPartitioningCreate(). Please look at all the functionality before
> doing anything.
> >
> >      Any refactorization you do needs to combine, simplify, and cleanup
> the two interfaces to create one simpler one. And please let us know your
> design ideas (with for example an issue on bitbucket) before you go ahead
> and write a lot of code we may end up not liking.
> >
> >      Thanks for looking at this,
> >
> >    Barry
> >
> >
> > > On Nov 6, 2017, at 7:09 AM, Vaclav Hapla <vaclav.hapla at erdw.ethz.ch>
> wrote:
> > >
> > > Hello
> > >
> > > The whole PetscPartitioner class sources are part of src/dm/impls/plex/plexpartition.c,
> mixed together with some DMPlex* functions.
> > > If you don't mind, I would move the PetscPartitioner* stuff into the
> separate file petscpartitioner.c
> > > (in future, it could be even moved to a separate directory).
> > >
> > > I would also like to
> > > * add PetscPartitioner{Add,Set,Get}OptionsPrefix,
> > > * create the partitioner lazily in DMPlexGetPartitioner, followed by
> PetscObjectIncrementTabLevel, PetscLogObjectParent and
> PetscPartitionerSetOptionsPrefix calls,
> > > * edit src/dm/impls/plex/examples/tutorials/ex5.c so that it calls
> DMSetFromOptions right after DMPlexCreateFromFile, in order to make the
> partitioner changeable from options.
> > >
> > > Do you have anything against?
> > >
> > > Vaclav
> >
> >
> >
> >
> > --
> > What most experimenters take for granted before they begin their
> experiments is infinitely more interesting than any results to which their
> experiments lead.
> > -- Norbert Wiener
> >
> > https://www.cse.buffalo.edu/~knepley/
>
>


-- 
What most experimenters take for granted before they begin their
experiments is infinitely more interesting than any results to which their
experiments lead.
-- Norbert Wiener

https://www.cse.buffalo.edu/~knepley/ <http://www.caam.rice.edu/~mk51/>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20171106/9ce03888/attachment.html>