itaps-parallel To-do list for next phone conference
Onkar Sahni
osahni at scorec.rpi.edu
Thu Jun 5 09:54:30 CDT 2008
Karen,
See my comments on part-handles below (inline with your points):
>
> Part handles ...
> a) are globally unique (so that remote part handles make sense and Zoltan
> doesn't go insane);
On any process, only local part-handles and neighboring (remote-)
part-handles are accessible through iMeshP interface. And it seems
reasonable that any iMeshP implementation will provide unique handles
for local and neighboring parts on any single process.
> b) allow identification (without communication or O(#parts) storage) of
> what
> process the part lives on [note: communication is allowed in
> iMeshP_syncPartitionAll to updated info about parts that could be used
> here];
I would say allow identification, comparison etc., whatever is needed by
applications (like mesh-modification or flow-solver) and services (like
Zoltan). I do not know what is the set of necessary operators required
in this category and this has been my concern, especially for services
like Zoltan (as they might need more than local and neighboring
part-handles).
> c) can be computed knowing only process rank and local part number (that
> is,
> iMeshP_createPart can return a part handle without communication);
I am not sure about part-handle being computed knowing only process rank
and local part number. I agree iMeshP_createPart must return a part
handle without communication but it doesn't need to know process rank
and local part number as input of partition_handle will provide this
information. In other words we do not have iMeshP_computePartHandle(int
rank, int localID, iMeshP_PartHandle *part_handle) and I think it is
difficult to support this unless we work with iMeshP_PartID (not
handle).
> d) can be distinguished from pointers and entity set handles (so that our
> overloading works).
I agree part-handle should be overloaded in necessary iMesh functions
(as I understand all handles are typedefed to void* so I do not know
what it means by can be distinguished from pointers).
- Onkar
More information about the itaps-parallel
mailing list