I think this is an interesting idea but it will be tricky to get it right.  What must be avoided is any sort of standardization about the details of the internals, for example, a way to find the "eager threshold" assumes that there is such a thing and that it is stable enough to be relevant.  Making this general enough to make sense in the standard while avoiding mandating implementation decisions will be challenging.

Bill

On Apr 14, 2008, at 3:48 PM, Jeff Squyres wrote:
I have a "pre-proposal" for the Forum, meaning that I have an idea but
want to see if others find it valuable before spending the time to
turn it into a formal proposal.

The general idea is to create a flexible mechanism to both pass hints
from the user application to the MPI implementation and to pass
information from the MPI implementation back to the application.  Some
rough ideas:

- Allow hints to be passed from the app to the implementation on a per-
handle basis (e.g., on a specific communicator)
- Allow hints to be passed from the app to the implementation for all
instances of a handle type (e.g., for all communicators)
- Allow hints to be specified before MPI_INIT, or create another
variant of MPI_INIT that allows passing hints
- Create a mechanism for an MPI implementation to pass arbitrary
information back to the user, perhaps in response to an arbitrary
query (maybe something along the lines of an MPI_Info in and and
MPI_Info out)

Comments?

--
Jeff Squyres
Cisco Systems

_______________________________________________
mpi-forum mailing list
mpi-forum@lists.mpi-forum.org
http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-forum

William Gropp
Paul and Cynthia Saylor Professor of Computer Science
University of Illinois Urbana-Champaign