dolfin team mailing list archive

["Martin Sandve Alnæs" <martinal@xxxxxxxxx>]

On 12/6/06, Garth N. Wells <g.n.wells@xxxxxxxxxx> wrote:
> >
> >
> > changeset:   2508:edf965fad4ef667f306873fe9194de2ed48bab7f
> > parent:      2504:62956628ca1e84780c27c633a9bae5730a290148
> > user:        "Garth N. Wells <g.n.wells@xxxxxxxxxx>"
> > date:        Wed Dec 06 15:42:42 2006 +0100
> > files:       src/test/passembly/main.cpp
> > description:
> > Apply simple renumbering in parallel assembly test file.
> >
>
> I've implemented some simple renumbering, so it's possible now to
> perform some simple parallel assembly benchmarks. For a touch over 4
> million cells for Poisson with linear triangles on a dual processor
> workstation, the timings (for assembly only) are
>
> One process:                         9.5s
> Two processes (with renumbering):    6.5s
> Two processes (without renumbering): 18s
>
> For the two process case without renumbering, about 10s of the total
> time is spent on communication (calling MatMatAssemblyBegin(),
> MatAssemblyEnd()). For the case with renumbering, this is only about
> 1.5s. The timings are a rough guide only as they're pretty erratic.
>
> During assembly, all cells are looped over but their contribution is
> only computed if they belong to the process. This could obviously be
> done more cleverly which will improve the timings.
>
> Garth

I simply used this Cell constructor:
dolfin::Cell dcell(mesh, cell_indices[iter]);

How did you partition the mesh here?

martin