Attendees:
Dennis Cottel NRaD (Host)
Thomas Robey Khoral Research
Anna Rounbehler SKY Computers
Arkady Kanevsky MITRE
Ron Brightwell Sandia National Labs
Robert Babb U. of Denver
Harish Nag Intel,SSPD
Jerrell Watts Syracuse U.
Tony Skjellum MSU
Marvin Kuznitz NUWC
Daniel Kim NUWC
Zhenqian Cui MSU
Jin Li MSU
-----------------------------------------------------------------
The meeting began approximately 8am on each day (20/21 Feb).
The following topics were discussed.
1. real-time CORBA RFI response
OMG requests input on what to make real-time CORBA. Arkady will
send current MPI/RT draft, and we will keep loose communication open.
Consensus is to proceed.
2. channel operations:
quiescent state,
atomic semantics, split semantics pro and cons,
structural retrofit.
Lots of discussion on the particulars ensued, including but not limited
to the advantages and disadvantages of the merged interface. It was
pointed out that the channel constructors/destructors (akin to MPI-1
communicator operations) are themselves without QoS, so that MPI/RT
will work in phases of setup-realtime-unsetup/resetup..., but not
be able to transform resource utilization in a predictable manner. This
seems to be an adequate to all.
The various proposals to delete MPIRT_Channel_modify and
enhance MPIRT_Channel_init were discussed. It was agreed that
MPIRT_Channel_init should not give up channels, if the requested
channel configuration cannot be satsified. Buffer pools will
not be damaged or destroyed. An action was taken to redefine
error array for MPIRT_Channel_init.
A vote was taken to delete the MPIRT single channel functions.
The single channel case is a subset of the multiple
channel functions. Vote to delete:
12 yes
2 abstain
Tony and Arkady will rewrite appropriately.
3. QoS constructors and destructors,
Objects, operation formats.
A vote was taken to have the following QoS portable arguments.
QoS_time
QoS_event
QoS_priority
and various implementation-specific tuning paramters.
12 yes
2 abstain.
Tony will write section.
4. Voted to delete QoS advice to users in the Appendix and
replace it with a reference to TENET papers.
5. It was agreed that a mixed paradigm section would be added,
with minimal text.
6. A web site for MPI/RT is planned at MSU. A link to
MPIRT examples will be available.
likely URL http://www.cs.msstate.edu/mpirt (Ed: REVISED!)
Items on here to include
Revised IEEE lecture (3-4 hour tutorial)
Examples of MPI/RT programs
Links to other resources and other MPI/RT pages
Eventually an FAQ
Copies of up-to-date drafts
7. Prioritization of the activities.
The following list represents the minimal set needed to
start prototype implementation:
1. Buffer Management
2. Channel Management
3. Clocks
4. Specification of QoS
5. Lower Level Event Driven
6. Underlying objects and structures
7. Examples
8. Subset of Collective operations and their QoS and
buffers specifications.
After that either Time Driven or High Level Event Driven
will be implemented. Items 1-8 are to be refined by the
April MPI Forum Meeting.
8. Buffer management was discussed in detail.
The new Transition State diagram with API and system
strategies were created and discussed. The requirements
for data fusion and flexibility were addressed.
Arkady, Tony and Jerrell will rewrite the section and changes
to all other sections that are affected by it.
A state diagram was developed for the buffer pools.
9. A brief discussion on MPIRT nomenclature was presented.
10. A short discussion on MPI/RT objects and their relationship
to the MPI (1.0 and 2.0) objects.
Concerns about useful things to be dropped from MPI-2 (eg, into JOD),
and need to preserve them for MPI/RT were mentioned.
11. A schedule was proposed
A. March MPI forum meeting
B. No MPI/RT technical meeting between MPI forum meetings scheduled.
People individually and in small groups will make technical
progress and exchange email with the real time group..
[Ed: This is revised because of March MPI meeting to be a small meeting
at MSU on 3/24/97]
C. The April MPI forum meeting will be augmented by adding
half a day on Tuesday and Wednesday and half a day on Friday
to work on MPi/RT technical issues.
[Ed: This is revised to be half-day Friday and part/whole of Saturday.]
The group thanked the organizer, Dennis Cottel and his organization, NRaD,
for hosting this meeting, and the meeting adjourned approximately 11:30am
on Friday, 21st Feb.