dolfin team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Sun, Mar 04, 2007 at 08:40:28AM -0600, Matthew Knepley wrote:
> On 3/4/07, Garth N. Wells <g.n.wells@xxxxxxxxxx> wrote:
> >
> >Quoting Anders Logg <logg@xxxxxxxxx>:
> >
> >> On Fri, Mar 02, 2007 at 05:19:43PM +0100, Garth N. Wells wrote:
> >> >
> >> >
> >> > Anders Logg wrote:
> >> > > A first version of the new UFC-based assembly is now implemented.
> >> > > A bunch of features are still missing, but we can now assemble the
> >> > > bilinear form for Poisson's equation and get correct results... :-)
> >> > >
> >> > > Here are some notes if anyone wants to dig in. For example, it might
> >> > > be a good idea to start thinking of merging in the parallel assembler
> >> > > at this point.
> >> > >
> >> >
> >> > Are there any short-term plans to work on partitioning of the mesh and
> >> > distributing the mesh data across processors? This is where the real
> >> > work is for parallel assembly and solve.
> >> >
> >> > Garth
> >>
> >> Yes, I have a student working on mesh partitioning in DOLFIN and we
> >> will start adding parts of it in a while, but I can't promise when it
> >> will be ready. The parallel assembly/distribution will probably be
> >> implemented before the partitioning has been implemented (and until
> >> then we can rely on METIS).
> >>
> >
> >Might be worth considering SCOTCH
> >
> >  http://www.labri.fr/perso/pelegrin/scotch/
> >
> >rather than implementing a partitioning algorithm. Unlike METIS, SCOTCH is
> >released under GPL.
> 
> The thing that METIS has over Scotch, Jostle, and Chaco is better handling
> of the objective function. It is MUCH better at load balancing due to the
> complicated multiobjective you want to have. Zoltan is a nice hypergraph
> partitioner, which seems to work even better than graph partitioning for
> load balancing. I have recently incorporated this into PETSc.
> 
>  Matt

Nice. Looks very promising.

/Anders