[Mpi3-hybridpm] Hybrid WG meeting at the Forum

[Pavan Balaji]

We had the plenary session for the Hybrid WG, where I presented the 
attached slides.

For the shared memory proposal, we discussed the shared memory 
communicator creation function. The following comments were raised:

1. The communicator creation function should be broader than shared 
memory alone. It can just take an info key and create any type of 
communicator with information from the MPI library. E.g., processes 
sharing an I/O node, or processes on a socket, etc.

2. The function should return an MPI_Group rather than a communicator, 
in order to avoid using up context IDs.

3. There was some more discussion on whether MPI_Shm_sync is actually 
unusable, which I can summarize during the next telecon.

For the helper threads proposal, the following comments were made:

1. OpenMP tasks vs. threads. What happens when the user's view of OpenMP 
threads is not the same as the number of actual system threads 
available. Some tasks might be waiting in a MPI_Team_leave function, so 
other tasks might not get scheduled at all.

2. Why does this need to be in MPI? Why can't this be a separate library 
that provides thread resources to applications, MPI, math libraries, etc.?

3. The Forum wants to see a better use case for it than just 
micro-benchmarks and demonstrate that it actually helps on multiple 
platforms, not just BG/P.

4. The naming has to be improved.

  -- Pavan

-- 
Pavan Balaji
http://www.mcs.anl.gov/~balaji
-------------- next part --------------
A non-text attachment was scrubbed...
Name: hybrid.pptx
Type: application/vnd.openxmlformats-officedocument.presentationml.presentation
Size: 84381 bytes
Desc: not available
URL: <http://lists.mpi-forum.org/pipermail/mpiwg-hybridpm/attachments/20110510/4ba33465/attachment.pptx>