[Mpi3-hybridpm] MPI-3 Hybrid Telecon Minutes
balaji at mcs.anl.gov
Tue May 5 16:31:12 CDT 2009
Here are the meeting minutes for today's call. I'll be putting this up
on the wiki and sending out the slides and ticket pointers as promised soon.
Attendees: Pavan Balaji, Weikuan Yu, Edgar Gabriel, Brian Smith, Jeff
Squyres, Rajeev Thakur, Steve Oyanagi, Jim Dinan, Marc-Andre Hermanns,
Torsten Hoefler, Sameer Kumar, Jeremiah.
* Discussed some of the proposals that have been submitted for the
Hybrid working group. Currently we have four (treating threads as
processes, dynamic thread levels, thread init/finalize routines for
thread-core mapping, MPI progress query semantics).
* The first proposal was not discussed; we will wait for Alexander to
join to describe it.
* We discussed the dynamic thread levels proposal where an application
can give a hint to the MPI library about the thread-level it needs. For
this, we discussed some of the problems implementations might have
including (i) collective operations that behave in one way when in
THREAD_MULTIPLE mode and in another way in other modes and (ii) dealing
with asynchronous progress threads.
* We discussed the thread init/finalize proposal, where a thread can
optionally explicitly tell the MPI implementation that it exists, so the
MPI library can, for example, bind it to appropriate resources or do
some more intelligent mapping. There was a suggestion that we allow the
application to pass an info argument to the init call, allowing it to
specify what resources it would prefer to be bound to.
* We finally talked about the MPI progress level query function that
allows the application to query the MPI implementation about what level
of progress the MPI implementation provides. The general consensus was
that the ticket was quite vague at this point and needed more
description; I'll take care of fixing this, and will be in touch with
Torsten for feedback.
More information about the mpiwg-hybridpm