dolfin team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Mon, Feb 18, 2008 at 08:06:27PM +0100, Niclas Jansson wrote:
> On Mon, 18 Feb 2008 18:31:54 +0100
> Anders Logg <logg@xxxxxxxxx> wrote:
> 
> 
> > > We (Nicklas) have now a preliminary distributed parallel
> > > implementation running for Poisson and a Unicorn flow solver. The
> > > structure is still quite messy and the partitioning is based on
> > > Metis, but we expect it to be integrated with DOLFIN in the coming
> > > months, hopefully before the summer.
> > 
> > Could this be done in small pieces? I'm a little worried about seeing
> > big changesets that add things to the mesh classes.
> > 
> > > When the structure is more clear we can discuss on the mailing list
> > > how to best incorporate this into DOLFIN. Apart from the structure
> > > of the implementation we are also focusing on distributed local mesh
> > > refinement.
> > 
> > Great.
> > 
> > What is the difference from what we have now? Is it the same but with
> > a distributed mesh (so each processor just sees the local part, not
> > the whole mesh)? Or is it something different?
> > 
> 
> Yes, each processor only sees the local part. The entire problem is
> never represented on one processor. Preprocessing, assembly, solution
> and postprocessing are all done without gathering/scattering data
> to/from one processor.
> 
> The only modification to the mesh classes is a new class storing
> information about ghost points and vertex local to global mapping.

What about local-to-global mapping of edges, faces etc?

> Most of the messy parts are inside the function classes.

Sounds scary. It would be good to avoid adding more complexity to the
function classes.

-- 
Anders