Re: [Libmesh-users] [deal.ii] memory requirements

[Wolfgang B. <ban...@ma...>]

> I'm actually quite interested in the answer to this question.  Has *anyone*
> gotten PETSc working well inside a multithreaded program?  It should be
> possible to call PETSc only from a single thread, but that kinda kills the
> point.

You can call it from several threads, but you have to synchronise with a 
mutex. If you assemble a linear system, copying local to global doesn't 
take that much time (this is the part that needs to be blocked) whereas 
computing local contributions runs nicely in parallel.


> Along those lines, Derek, do you know about threading in Trilinos?

I would be interested in hearing about that too.


> Our new parallel mesh and parallel algorithm implementations are clearly the
> answer for large-scale distributed memory machines, but when we are talking
> about 16+ cores per node threading has its place even inside a distributed
> memory mesh implementation.

And that point isn't going to go away anytime soon :-)

W.

-------------------------------------------------------------------------
Wolfgang Bangerth                email:            ban...@ma...
                                 www: http://www.math.tamu.edu/~bangerth/