[Mpi3-hybridpm] Minutes of 10/27 concall

Darius Buntinas buntinas at mcs.anl.gov
Tue Oct 27 11:33:43 CDT 2009


I apologize for not getting a list of all attendees.

A few questions were raised about Marc's proposal:

Can the max number of threads be different on different nodes?  We
figured that they can be different, but this should be mentioned in the
proposal.

How would an MPI process start UPC threads?  UPC has a separate
launching mechanism, so how would that interact with MPI?  One proposal
was that UPC would start the threads outside of MPI, then be reparented
to the MPI process.  So UPC would give the EWORLD communicator to MPI.
Another proposal is that MPI would create the EWORLD comunicator then
pass that to the UPC starting mechanism.  In either case some interface
glue would be needed between the UPC and MPI implementations.

A request was made that we should go through each existing ticket and
indicate how Marc's proposal addresses the issues in the ticket.
Similarly, there should be a ticket opened up for Marc's proposal.

If I missed anything, or misrepresented something please let me know.

-d



More information about the mpiwg-hybridpm mailing list