The main reason to have this natural ordering is to read and write data, 
possibly with different numbers of procs.

ALB

On Feb 17, 2014, at 6:21 PM, Jed Brown <[email protected]> wrote:

> Matthew Knepley <[email protected]> writes:
> 
>> On Mon, Feb 17, 2014 at 5:14 PM, Jed Brown <[email protected]> wrote:
>> 
>>> Andrés Alessandro León Baldelli <[email protected]> writes:
>>> 
>>>> Hi all,
>>>> I am working with Blaise Bourdin on implementing a concept of "natural
>>> numbering" for unstructured meshes using DMplex.
>>>> Such natural numbering can be, e.g., that of the mesh prior to parallel
>>> distribution.
>>>> I have a working example in which I compute a field on the distributed
>>> DM and I scatter it back to the serial DM on rank0 but I would like to
>>> create a more general interface. To this regard I ask your advice:
>>>> 1) Is it reasonable to add an SF within the DM typedef for this "natural
>>> to global" scatter?
>>> 
>>> Surely you don't want the (necessarily non-scalable) serial DM to be a
>>> necessary part of this interface?  What do you want to do with this
>>> natural numbering?  Just write output files?
>>> 
>> 
>> Jed, No, no, no. This is not an SF for a serial DM.
> 
> I was responding to the question above.  Yes, we need a genuine parallel
> reader (naive partition) and a way to write back in that ordering.  But
> my question remains: do you want to do something other than IO with that
> "natural" ordering?
> 
> Note that in the natural ordering, you likely have elements for which
> you own zero of the vertices, and vice-versa, so it would be _really_
> bad to compute on it.

Reply via email to