----------
X-Sun-Data-Type: default
X-Sun-Data-Description: default
X-Sun-Data-Name: minutes_apr_97
X-Sun-Content-Lines: 131
Minutes of the MPI/RT working group on March 5 in Chicago
Attendees:
Dennis Cottel NRaD
Thomas Robey Khoral Research
Anna Rounbehler SKY Computers
Arkady Kanevsky MITRE
Ron Brightwell Sandia National Labs
Robert Babb U. of Denver
Anthony Skjellum MSU
Harish Nag Intel,SSPD
Jerrell Watts Syracuse U.
Zhenqian Cui MSU
Peter Brennan MPI Software Technology Inc.
Tarek Abdelzaher U. of Michigan
The following topics were discussed:
1. Gordon High Performance Conference was announced.
See http://www.erc.msstate.edu/conferences/gordon97.
2. Tony will explore copyright transfer from U. of Tennessee
to MSU for future MPI/RT drafts and MPI documents.
3. A long discussion on the simulators of QoS for various platforms
including:
how to achieve it, how to use it, how to optimize it,
was conducted. This has a direct affect on implementation strategies.
4. Multiple levels of implementation activities were proposed:
API prototype with empty QoS guarantees, a generic QoS guarantee
for a neutral platform, and optimized QoS guarantee for real-time
platform.
5. The channel section was discussed.
A proposal was presented and discussed on creating objects to
reduce the number of parameters for channel management operations
and simplify their manipulations. The six page proposal was
written and distributed at the meeting and will be discussed at
the next meeting.
A split of collective channel operations into two stage process:
reserve and commit was proposed. A joint operation which
allows all channels (collective and point-to-point)
to get a complete picture on all needed resources collectively
over a communicator was also discussed.
The above mentioned object proposal supports this concept.
6. Several topics on Buffers were discussed. Buffers addresses as
IN/OUT parameters proposal was reintroduced. The vote was passed to pursue
the following model:
Nondestructive buffer send, with explicit wait and freeing the buffer.
All buffer operations are explicit (including all real-time data
transfer operations. Buffer iterators specified via strategy.
The other issues discussed were: who owns the buffers
application/implementation, buffers replenishment for buffer pools -
buffer passing between application and implementation, sharing buffer
pools between channels, buffer pool strategies and their specifications.
7. A discussion on mismatches of QoS for two endpoints of the
point-to-point channels, and multiple endpoints of collective
channels was conducted.
For the client-server paradigm only one end specifies QoS and the
other end accepts/rejects it. An analogous situation exists
for the root of a collective operation.
8. The current MPI/RT chapter will become a stand-alone document.
There was a discussion on the structure of the new document.
The new chapters will follow the logical breakdown of the
sections in the current document. The exact mapping of sections
to chapters will be discussed at the next meeting.
Additional chapters could include: introduction, dependencies on
MPI-1,2, collective operations, objects/handlers/request and other
underlying principles, threads.
Fault-tolerance will be renamed to fault handling.
9. Future of the MPI/RT was discussed.
We agreed to continue meetings on the 6 week basis.
Here is the tentative schedule of the future meetings:
Dates Place Local Contact
June 5-6 MITRE, Boston Arkady Kanevsky,
August 7-8 MITRE, Boston Arkady Kanevsky,
September 25-26 MSU Anthony Skjellum,
November 6-7 MITRE, Boston Arkady Kanevsky,
December 11-12 NRaD, San Diego Dennis Cottel,
January 15-16 Sandia, Albuquerque Ron Brightwell,
February 26-27 NRaD, San Diego Dennis Cottel,
April 9-10 U. of Denver, Robert Babb,
May 21-22 INTEL, Portland Harish Nag,
June 29-30 MITRE, Boston Arkady Kanevsky.
All meeting are 2 days: Full day first day and half day the second.
The BOF meetings or tutorials will probably be held at
SC'97 (High Performance Networking and Computing Conference) and
RTSS'97 (Real-Time Systems Symposium) conferences.
We agreed to change our meeting name to MPI/RT Forum.
A more rigorous voting and appeal rules will be introduced next time.
The JOD will be published at the same time as the MPI-2 standard document
as a stand-alone document that contains a chapter on MPI/RT.
The current version will be presented there with the following changes:
A. The first figure in the JOD MPI/RT chapter will be replaced by the
current figure in the stand-alone chapter.
B. Delete C++ binding (The group voted on it unanimously).
C. Add disclaimer and the pointer to the ftp/html site of MPI/RT.
D. Revert Buffer Management section to the previous version.
Add the state diagram to it.
E. Fix Appendices. Only the second appendix should be present that
will be numbered A.
F. Change Fault-Tolerance to Fault-Handling (voted in unanimously for
JOD and all future versions).
G. Put the two new figures from the buffer section that describe the
sequence of application commands to the profile section.
F. Add a figure to the profiles sections that illustrates relationships
between, MPI-1, MPI-2, and MPI/RT (Venn Diagram).