On Wed, 17 Mar 2010, Lorenzo Botti wrote:
Oh yes,you are definitely right.
With r3697 I get
Memory 1 process = 3.6 GB
Memory 4 processes = 1.1 GB * 4 = 4.4 GB
Huge improvement. Thank you.
Whew! Good to hear.
Was there a problem with mesh, dofs, ghosted vectors or... what else?
The DofMap send_list. There was a series of very subtle (e.g. only
shows up after dozens of adapted steps on dozens of processors) bugs
that were revealed (not caused!) by ghosted vectors. One of the fixes
for one of those bugs went too far and ended up filling the send_list
with even non-ghost dofs.
Let us know if something so important happens!
Have I missed something on the mailing list?
My mistake, I suspect; since it wasn't an accuracy bug I probably only
mentioned it on libMesh devel.
A question about parallel mesh.
Is the geometry supposed to be parallel?
I mean all the nodes and elements instantiations.
Or just iterators will be local?
Even with SerialMesh, iterators over local_*elements will be local.
With ParallelMesh, non-ghost elements don't even have to exist on the
local processor. A warning, though: there's a long-standing adaptive
coarsening bug with ParallelMesh that has prevented any of the main
developers from using it for real runs, which means that even if you
avoid adaptive coarsening I wouldn't be confident that it's well
tested enough to trust yourself right now.
---
Roy
------------------------------------------------------------------------------
Download Intel® Parallel Studio Eval
Try the new software tools for yourself. Speed compiling, find bugs
proactively, and fine-tune applications for parallel performance.
See why Intel Parallel Studio got high marks during beta.
http://p.sf.net/sfu/intel-sw-dev
_______________________________________________
Libmesh-users mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/libmesh-users
