On Tue, Feb 7, 2012 at 7:34 PM, Roy Stogner <royst...@ices.utexas.edu>wrote:
> Shouldn't the (later) introduction of ghosted PetscVectors have
> ameliorated the problem? They've got all the send_list information
> implicitly built-in to the vector. It seems like it should be
> possible to avoid the excess communication even when users don't
> provide a send_list. Or is it that they have the send_list
> information on sending processors, only on receiving processors, so
> sending processors can't pre-filter what they send?
The problem is that that version of the function explicitly builds full
index sets forcing complete communication. If we would have just said
"current_local_solution = solution" it would actually do the correct
ghosted assignment.
> Some of my huge runs are dying here (trying to do a full localization of
> 400 Million DoFs is not working out!).
>
> We're within 1 OOM of hitting 32-bit int based bugs, huh? Nice.
>
Yep. Dmitry was actually trying to compile libMesh against PETSc built
with 64bit integers just today! Apparently it didn't work out because of
petscint to int conversions in libMesh... oops!
Derek
------------------------------------------------------------------------------
Keep Your Developer Skills Current with LearnDevNow!
The most comprehensive online learning library for Microsoft developers
is just $99.99! Visual Studio, SharePoint, SQL - plus HTML5, CSS3, MVC3,
Metro Style Apps, more. Free future releases when you subscribe now!
http://p.sf.net/sfu/learndevnow-d2d
_______________________________________________
Libmesh-devel mailing list
Libmesh-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/libmesh-devel
