On Wed, 21 Apr 2010, Vijay S. Mahadevan wrote:

>> The DofObject encapsulation is pretty good; that should be the only
>> change you require.  If you'd like to make that change a compile-time
>> option (defaulting to unsigned char - the memory usage isn't trivial),
>> we'd appreciate a patch.
>
> When you mean it isn't trivial, are you talking about the case when
> you want to use say unsigned int ? Or is it non-trivial already ?

Depends on your usage case.  The trouble is that, although our typical
big memory sink is the sparse matrix (bandwidth * 8 bytes per dof *
N_dofs_per_proc on each node), and this *should* be much more
expensive than even a 4 byte int * N_dofs_per_proc, on a SerialMesh
your cost becomes a 4 byte int * N_dofs.  If you've got a billion-DoF
problem that you're making tractable via distributed memory, that
might be significant.

> I'll look in to making this a compile time option (and will send you
> the patch)

Thanks!

> but if the memory cost becomes overwhelming, I will need to think of
> a better way to handle this.

It probably won't be overwhelming; if not then the quickest fix would
be moving to ParallelMesh.  Sorry about the adaptive coarsening bug,
and let me know if you run into any others.  ;-)
---
Roy

------------------------------------------------------------------------------
_______________________________________________
Libmesh-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/libmesh-devel

Reply via email to