dolfin team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Fri, Feb 01, 2008 at 07:02:49PM +0100, Gustav Magnus Vikstrøm wrote:
> I've put my paper on parallelization on the wiki. In short - it does
> work. Going from 28 seconds for assembly with one CPU to 4 seconds with
> 16 CPUs. But, from what we learned it seems like we should look into the
> graph/mesh partitioning part a bit more.

I'm very happy with the way it turned out, well done!

> For example, partitioning a UnitCube(100, 100, 100) into 16 partitions
> uses about 8 GB ram and takes 7.5 minutes.

Yes. 100x100x100 (6,000,000 cells) was the biggest mesh we were able
to partition with SCOTCH and it takes a long time. We can ask Francois
Pellegrini about improving these numbers (he has been helpful before).

> In the rush before deadline I forgot the acknowledgements section, but
> Anders really deserves it. Thanks!

Garth has also been helpful and he wrote the original tests.

I'll get back with some more comments on the current status of
parallel assembly and some suggestions in time.

-- 
Anders