dolfin team mailing list archive

["Garth N. Wells" <gnw20@xxxxxxxxx>]

Niclas Jansson wrote:
> Niclas Jansson <njansson@xxxxxxxxxx> writes:
>
> > dolfin@xxxxxxxxxx (DOLFIN) writes:
> >
> > > One or more new changesets pushed to the primary dolfin repository.
> > > A short summary of the last three changesets is included below.
> > >
> > > changeset:   6855:51f268bf79d79b0de5d3cb20a7cacb35357c81ce
> > > tag:         tip
> > > user:        "Garth N. Wells <gnw20@xxxxxxxxx>"
> > > date:        Wed Aug 26 14:30:18 2009 +0100
> > > files:       demo/pde/elasticity/cpp/main.cpp dolfin/fem/Assembler.cpp
> > > description:
> > > Add OpenMP code to assembler (commented out for now).
>
> If insertion is done in a critical section there would virtually be
> no scope for speedup (for larger number of threads)

I have a problem where the element tensors are very complicated and 
costly to compute and the linear solve is cheap - with a small number of 
threads I get a modest speed up. Annoyingly, PETSc dies when calling 
VecGetValues with more than one thread.

> A better idea would probably be to either
>
> 1) Partition the mesh, and only have a critical section for shared
> cells.
>
> 2) Assemble into a private matrix and either add them together A_1 +
> ... + A_n or add them together with some fancy tree reduction algorithm.

The best idea is for us to complete the message passing-based 
parallelistion ;). I have interior facet integrals which are not yet 
supported which is why I've tried OpenMP inside the assembler.

Garth

> Niclas
> _______________________________________________
> DOLFIN-dev mailing list
> DOLFIN-dev@xxxxxxxxxx
> http://www.fenics.org/mailman/listinfo/dolfin-dev