[Mpi3-tools] Meeting minutes

[Pavan Balaji]

Attendees: Chris Gottbrath (Totalview), Jeff Squyres (Cisco), John 
DelSignore (Totalview), Marty Itzkowitz (Sun)

Minutes:

We discussed the PMI initial draft to see if the concept itself	was of
interest to the	working	group. The main	concern	was that it was	quite
restrictive for	different MPI implementations to follow	a similar
API. For example, Open MPI internally uses a binary representation
which means that they have to convert to strings on each end to	work
with the PMI API, which	might not be optimal.

On the positive	side, we discussed how PMI can be used as a tool to
provide information about dynamic processes to other tools (e.g., the
debugger) or as	a mechanism to collect other system information	(such
as the topology, and other node/system information such	as
uptime/load).

The general consensus was to have an API that is higher	level than PMI
that can be used by different tools to query the MPI implementation
for information. This is slightly different from the PMI scope in that
each process cannot add	any information key for	that process and any
other process within MPI can read it. So, it is not clear if/how this
can be done in a common way, but Pavan will talk to John @ Totalview
to see if this is possible.

  -- Pavan

-- 
Pavan Balaji
http://www.mcs.anl.gov/~balaji