The realtime discussion was centered around the need to do something
specific, with different people re-arguing what was argued last time,
but we essentially agreed to:
* develop priority message extensions for MPI
* develop scheduling hooks for MPI
* look at developing channel abstraction for MPI, reviving
MPI-1 ideas, and looking at virtual channel hardware
* continue to discuss other issues in RT systems
[eg, relating rate monotonic scheduling to MPI]
It was pointed out that we lapsed in our original goal to have a white
paper by the July meeting, and Skjellum indicated that we should try
again to have some specific things written down from about goal set for
September 6.
Lloyd Llewins [Hughes], Arkady Kanevsky [Mitre] and Robert George [NPS] had
lots to say, and their access to real hardware and lots of experience
was duly noted by all. They represent the gamut of viewpoints:
* practical, ad hoc
* carefully scheduled, predictable, deterministic, provable
Furthermore, Arkady argued that MPI/RT should not try to do everything,
but leave that for application-specific meta software [compiler/makefile,
etc] so that MPI/RT should concentrate on providing hooks, whereas
applications will still have to work to be performance portable.
It was generally agreed that providing MPI hooks should enable such
meta software.
* The issue of providing MPI/RT as multiple profiles, or one profile
to MPI-2 continued to come up, and we need to see what happens.
* Skjellum argued that certain parts of MPI/RT committee should impact
MPI-2 itself, and we should, where possible, mainstream proposals
into other parts of MPI-2 forum activities.
* Saha pointed out that subtracting from MPI-1 is also a valid route
to getting to a realtime system.
* Shuler pointed out that PUMA/SUNMOS would address realtime issues in
future [PUMA/SUNMOS is a Sandia operating system for high speed Paragon
programming].
* Robert George agreed to lead the September meeting, as Skjellum will be
absent.
Note:
Please let us know if advance if you will be at September meeting.
We will try to meet on September 7, 9:30-12noon, as during the July
meeting, and otherwise as needed.