We had the first reading of our real-time chapter at the plenary MPI.
There were two major questions that MPI at large wants us to answer.
1. What will be the status of our chapter for the Supercomputing symposium
and for the final standard draft of the MPI-II for the spring of 1997?
2. How much of the MPI-I and MPI-II will we include in the MPI/RT,
and what is the relationship of the MPI/RT to the rest of the MPI?
We had two working meetings.
The attendees were:
Arkady Kanevsky - Mitre
Robert Babb - U. of Denver
Anna Roubehler - SKY computers
Ron Brightwell - Sandia
Dennis Cottel - NRaD
Andy Sherman - Scientific Computing Associates
Thom McMahon - MSU
Shane Hebert - MSU
David DiNucci - NASA Ames
The following topics were discussed:
1. The group unanimously voted to move the MPI/RT chapter to
the Journal of Development (JOD) of MPI-II.
The journal is part of MPI-II, but not a required standard.
Given the current status of our chapter, the fact that it will
not be ready as a standard by November, and that we would
like to continue to be a part of MPI-II, JOD is our only option.
2. We agreed that we need one extra meeting before the October MPI meeting.
The current dates for the meeting are Sept. 19-20.
Arkady will take care of the local arrangements and will email
hotel info, directions, and agenda later.
Meanwhile, buy the plane tickets now before it is too late.
3. We had a discussion addressing the second question of the plenary session.
We agreed that MPI/RT will have several ``profiles''.
For example, strict MPI/RT without any other MPI functions,
MPI/RT with other MPI functions, and a ``small'' embedded version of MPI/RT.
What profiles we will have and what they will encompass will be discussed
in the future meeting.
We will have an operation to initialize MPI/RT:
MPIRT_ INITIALIZE(parameter)
The parameter will specify the profile.
Also synchronized clock parameters will become accessible after the
initialization.
Arkady will write this section.
4. We discussed Anna's instrumentation proposal.
The issues discussed included: persistent monitors, monitoring of
the channels, how to fit the existing monitoring abilities of the
platforms, format of gathered monitoring information, and how an
application will get access to it.
Anna will update her proposal and present the new version at the next
meeting.
5. We discussed QOS without having a proposal in place yet.
We separated QOS parameters into two sets: the ones specific to the
paradigms (time, event, priority-driven) like integer priority,
time interval, and the ones that are common
to all paradigms, like bandwidth, latency, and traffic load in general.
The relationship between the two sets and the format of individual
parameters will be discussed at the next meeting.
Arkady will write the outline and everybody should
send their set of QOS parameters for the discussion at the next meeting
via email.
6. We discussed the Buffer management section.
We voted not to include a single operation to create several buffers
since the same functionality can be achieved by adding one buffer
at a time and since this avoids the allocation boundary issue (word, double
word boundary or other boundary for the beginning of each buffer allocation).
There are still unresolved issues with buffer management policies,
channel queue policies and their relationship.
These will be discussed at the next meeting.
Ron will update this section and present it at the next meeting.