[MOAB-dev] MOAB handles and processor rank

[Jed Brown]

On Thu 2008-09-25 11:25, Tim Tautges wrote:
>
>
> Jed Brown wrote:
>> On Thu 2008-09-25 10:43, Jason Kraftcheck wrote:
>>> MOAB currently embeds the rank of the owning processor in each entity
>>> handle.  Tim and I were just discussing removing this 'feature', as we had
>>> decided that it wasn't the best solution for MOAB's parallel data management
>>> and no longer use it.  If this will be a problem for anyone else (i.e. it is
>>> used in code I'm not familiar with in MOAB or is essential to some external
>>> code), please let me know soon.
>>
>> I was curious why those existed.  I have no use for tags on individual
>> entities.  My only requirement is that the order of entities in each
>> shared set be consistent between processors.
>>
>> Jed
>
> I'm not sure that's a safe assumption.  The order depends on the local  
> handles, which can be different.  You can get the handle of a given  
> entity on the other processor(s) sharing that entity.

Oh, hmm.  I implemented the reduction using MPI_Type_indexed rather than
packing a buffer.  The access pattern is much less regular if they are
not in the same order.  I can sort the entities on all non-owning procs
to match the owner order.

In early testing, it looked like the order was consistent, but I didn't
check carefully.  Of course it might be hard to preserve order in the
presence of mesh modification.  I can get remote handles using
iMeshP_getOwnerCopy.  Is it safe to just sort these locally?  Certainly
I shouldn't have to send the array of handles to the owner just to get
the permutation.

Jed
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 197 bytes
Desc: not available
URL: <https://lists.mcs.anl.gov/mailman/private/moab-dev/attachments/20080925/208c60a6/attachment.pgp>