[petsc-dev] field split/schur complement, incompressible flow

[Matthew Knepley]

On Wed, May 12, 2010 at 3:11 PM, Jed Brown <jed at 59a2.org> wrote:

> On Wed, 12 May 2010 14:13:37 -0400, Matthew Knepley <knepley at gmail.com>
> wrote:
> > Yes, definitely. I have been too lazy to add names, but this is exactly
> > what is needed.
>
> Okay, so what about naming splits on the command line?  We usually try
> to make the command line mimic the API.  Accepting a dict might be neat:
>
>  -pc_fieldsplit_splits "{'velocity': [0,1,2], 'pressure': [3]}"
> -fieldsplit_velocity_pc_type ml
>
> As a possible extension for the options database to more easily support
> very complex simulations with many nested objects, and long redundant
> prefixes, I've been thinking about YAML support.  YAML files can have
> references so you can define a collection of solver profiles, then use
> them in multiple places in the same simulation (for instance if you have
> two similar systems for which you would want to use the same or similar
> solver parameters).  Would anyone else like this feature?  Note that a
> YAML parser would also trivially be able to handle the dict string
> suggested as a command-line option above.
>

I guess I have no problem with this if:

  a) Its completely backwards compatible with the options-style

  b) the YAML part is not going to break the build

This will likely mean revamping the Options*() API.


> Backing up slightly, this thread
>
>  http://lists.mcs.anl.gov/pipermail/petsc-dev/2010-March/002439.html
>
> proposed having the names for field associations on the DM (so that
> visualization and other diagnostics could also use names).  In that
> case, the DM would have (or be able to provide) the index sets, and
> PCFieldSplit would only need to know the names of the groups.  On the
> other hand, there are certainly users that don't use DM at all, but
> still want to use PCFieldSplit.  One approach would be to have a
> super-lightweight DM that just kept a collection of index sets with
> names and groupings (assuming no further structure).  The other is to
> have PCFieldSplit retain a low-level interface to set the splits.


I favor the low-level interface to FieldSplit taking a name and IS.

   Matt


>
> Jed

-- 
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
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20100512/cf5e428e/attachment.html>