fenics team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Wed, Sep 28, 2005 at 11:24:46AM -0500, Robert C. Kirby wrote:
> >That would be great. If would think DOLFIN/FFC does the form
> >evaluation faster (computing the element stiffness matrix) but
> >
> We don't know that it does.  It has more symbolics, but it uses level  
> 3 BLAS also.  Probably depends heavily on the operator.

Yes, we don't know. That's why we should try.

> >Sundance might very well beat us when you count in insterting into the
> >global matrix (don't know how PETSc compares to Trilinos here) and
> >interacting with the mesh (our mesh is probably a little slow).
> >
> I don't know how efficient Kevin's mesh communication is.  I know he  
> does "SumIntoRows" which is just Trilinos' Matt, Set Values!

I like that. Maybe you could update PETSc to support things like

    Matt, SetValues!
    Matt, Assemble!

You can overload ',' in C++ so it should be possible, I don't know
about '!' though...

> >I think a benchmark would be constructive since it would help both
> >codes find any bottlenecks we're not aware of.
>
> If we could really straighten out who's spending time where and get  
> an accurate comparison of time to build the matrix, it might also  
> tell us how much is lost (if anything) by run-time stuff.  Or, it  
> could very well tell us that even though DOLFIN is compiled, it's  
> embarassingly slow  :)

Yes, that's the point.

/Anders