[petsc-dev] making DA more light weight

[Barry Smith]

On May 14, 2014, at 11:26 PM, Jed Brown <jed at jedbrown.org> wrote:

> Barry Smith <bsmith at mcs.anl.gov> writes:
> 
>> http://www.davidketcheson.info/2014/03/11/DMDA_memory.html
> 
> We discussed this here at the time.
> 
> https://plus.google.com/+DavidKetcheson/posts/f5ACSnrXznw
> 
> Matt has refused to implement DMLocalToGlobalBegin_Plex with consistent
> semantics for INSERT_VALUES until I add support for a local
> PetscSFReduce (skipping communication).  With that, we could replace the
> VecScatters with PetscSF, which would amount to 2 or 3 integers worth of
> storage for each leaf, independent of the number of dofs.

   Hmm, why does VecScatter size depend on dof? Since it handles bs > 1 it should not. Perhaps it is things like localtoglobal (not block version) that is taking the memory. Issues of collective construction come up if we make those produced on demand.


>  That could be
> compressed in some cases, though I think we'll quickly run into loss of
> generality if we push that.