dolfin team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Wed, Jan 17, 2007 at 11:11:18AM +0100, Martin Sandve Alnæs wrote:
> On 1/17/07, Anders Logg <logg@xxxxxxxxx> wrote:
> >On Wed, Jan 17, 2007 at 08:43:04AM +0100, Martin Sandve Alnæs wrote:
> >> On 1/16/07, Garth N. Wells <g.n.wells@xxxxxxxxxx> wrote:
> >> > Martin Sandve Alnæs wrote:
> >> > > But if we are to merge the projects it would be nice to know as soon 
> >as
> >> > > possible, to avoid wasting time on duplicated code.
> >> > >
> >> >
> >> > I don't see a problem with two codes as it gives each project the
> >> > freedom to pursue different interests/goals/visions. It gets 
> >complicated
> >> > when a lot of developers get involved. It would/will be nice if/when
> >> > PyCC is opened up so we can look at each others work, exchange ideas 
> >and
> >> > code, and collaborate more.
> >>
> >> True. I think keeping two separate Assembler projects while
> >> collaborating is the best plan, to avoid too strong dependencies.
> >>
> >> martin
> >
> >I agree, let's have two separate assemblers for a while. But let's set
> >up a set of benchmarks that we can run with both codes to learn how to
> >do things best.
> >
> >We need to choose a set, say 3-4 different combinations, of
> >
> >     (form, mesh, repetitions)
> >
> >and then measure the CPU time (wall-clock time) for assembling the
> >sparse tensor (matrix).
> >
> >Suggestions?
> 
> The forms should at least include a simple mass matrix, such that the
> element tensor computation is not dominating, and something with many
> coefficients, so efficient coefficient handling is tested.

>From my experience, you need to try hard to have something where the
element tensor computation is dominating. Even with high-degree
polynomials and say a convective term like v*dot(b, grad(u)), the
insertion dominates (but hopefully this can change).

> The target matrix format (petsc, epetra, pycc.crs, etc) should also be
> a parameter. And perhaps dofmap properties, like numbering scheme?

Sounds good, but maybe the target matrix format and the sorting of the
dofmap are not attributes of a benchmark, but rather parameters you
optimize over to minimize assembly time (so they can be chosen freely
to get the best result).

/Anders