On Tue, 26 Aug 2014, Miguel Angel Salazar de Troya wrote:

> I'll take a look at the mesh smoothers to see if I can apply something 
> similar.

Whoa - mesh *refinement* smoothers.  I haven't done much with the mesh
smoother codes but I don't think they've got the same sort of
algorithm in them.

> One question related to parallelism. I see that most of the assembly
> routines are implemented using TBB. I configured my libmesh with
> this library. Now, does this mean I'm using it when I run my
> program? Is there a special way to compile/run the program in order
> to take advantage of the TBB perks?

Run your program with "--n_threads=4" if you want 4 threads per MPI
rank.  You'll need to tell your cluster to only give you one MPI rank
per shared-memory system rather than one rank per processor, and there
are sometimes tricks you can play with "processor affinity" settings
that make a huge difference in threading performance, but that's all
processor-dependent.

If all your costly loops are in library code then that's it, you ought
to be threaded.  If you're not using FEMSystem then you'll need to
write a little bit of TBB code yourself to get the system assembly
threaded.

Don't futz with it too much if you don't have to; in my experience
threading has only been extremely helpful when I've got an app code
that's not ParallelMesh compatible but I need to cut down on its
memory usage.
---
Roy

------------------------------------------------------------------------------
Slashdot TV.  
Video for Nerds.  Stuff that matters.
http://tv.slashdot.org/
_______________________________________________
Libmesh-users mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/libmesh-users

Reply via email to