[mpiwg-rma] Proposed Info Key
balaji at anl.gov
Tue Oct 14 11:19:41 CDT 2014
Sorry, I was clearly smoking something, so both my previous statements were incorrect:
1. Current mpich/master does check this unconditionally, but we are trying to make it conditional on an info key: alloc_shared. The patch will be in soon.
2. Converting it to an AVL tree might not be as easy as I thought initially. We do a bunch of Group_translate_ranks and Allgathers. I’ll need to think more carefully to see if it’ll be possible to do this in an inexpensive way to avoid the info key altogether.
On Oct 14, 2014, at 11:06 AM, Balaji, Pavan <balaji at anl.gov> wrote:
> If we can optimize the search with an AVL tree, this info might be unnecessary. It’d be cheap enough that we’d do it everytime anyway.
> — Pavan
> On Oct 14, 2014, at 10:58 AM, Jeff Hammond <jeff.science at gmail.com> wrote:
>> Thanks, Pavan, for sharing the details. I forgot about the MPICH info
>> key used here.
>> Now that I understand the issue, I think we should pursue both
>> tickets, especially since my ticket is definitely 4.0 material.
>> Standardizing the info key that makes the MPICH solution portable in
>> the 3.1 timeframe is a good idea.
>> And I should note that this is not only an academic problem - there
>> are a number of projects making use of the (ppn+1)-windows trick
>> already and will benefit from the info key Mikhail proposed.
>> On Tue, Oct 14, 2014 at 8:54 AM, Balaji, Pavan <balaji at anl.gov> wrote:
>>> On Oct 14, 2014, at 10:43 AM, Jim Dinan <james.dinan at gmail.com> wrote:
>>>> IIUC, MPICH has started detecting this automatically without the info key, so this ticket may be moot. Hopefully an MPICH developer familiar with this optimization can comment.
>>> Yes, but we rely on an internal info key as well in order to avoid the search time. Though I think we can optimize this search with an AVL tree or something.
>>> Note that Jeff’s ticket gives more capability that this — it allows the user to do load/store accesses. Your info key only allows the MPI implementation to internally optimize communication.
>>> — Pavan
>>> Pavan Balaji ✉️
>>> mpiwg-rma mailing list
>>> mpiwg-rma at lists.mpi-forum.org
>> Jeff Hammond
>> jeff.science at gmail.com
>> mpiwg-rma mailing list
>> mpiwg-rma at lists.mpi-forum.org
> Pavan Balaji ✉️
> mpiwg-rma mailing list
> mpiwg-rma at lists.mpi-forum.org
Pavan Balaji ✉️
More information about the mpiwg-rma