The October 9-10 meetings were attended by
Tony Skjellum, MSU
Nathan Doss, Sanders
Lloyd Lewins, Hughes
Steve Kubica, KRI
Ron Brightwell, Sandia
Dennis Cottel, NRaD
Harish Nag, Intel
Andy Sherman, SCA
Arkady Kanevsky, MITRE
Jerrell Watts, Caltech
Robert Babb, U. of Denver
Anna Roenbehler, SKY
The following topics where discussed:
1. Modify a discussion of QoS for portability in the introduction.
2. We corrected the text (based on previous meeting's agreement) to
delete wildcard receives from exception for embedded MPI 1.2 (aka
resource constrained). That is, wildcard receives are to be allowed in
the "resource constrained" profiles of MPI/RT.
3. We deleted profiles for initialization. The relationship between
MPI_INIT and MPIRT_INIT was discussed. We agreed to remove current
MPIRT functionality from RT Initialization and Termination section. We
also agreed to have more discussion on applications that use MPIRT
functionality on a subset of nodes while others use standard MPI 1.2.
4. We had an extensive discussion on accuracy and access time for
synchronized clocks. The implementors may have a hard time coming up
with useful bounds, while users can not derive them by themselves. We
agreed to look into this further.
5. We decided to have only one form of creating buffer pools
(MPIRT_BUFFER_POOL_CREATE). The address of all buffers should be passed
to the MPIRT_BUFFER_POOL_CREATE. This will replace base argument.
Bufpool now becomes only OUT parameter.
Tags were replaced by Indices that provide one-to-one correspondence
to buffers.
We discussed the addition of out channel parameter in MPIRT_BUFFER_GET
to find out on which channel a message arrived. The extensions of the user
defined buffer strategy were discussed.
We did not decide if it would be allowed to create channels only
using buffer pools, or with the non-buffer pool approach also.
For now, we retain both interfaces. (See item 7 below also.)
6. It was decided that we need more examples demonstrating the use
of various MPIRT functionality and paradigms.
7. For collective RT operations we decided to add a non buffer pool
version analogous to the existing MPI functionality but with QoS.
It was also decided and voted on that there are no new collective
operations specific QoS. The existing QoS defined for 3 paradigms
are sufficient.
There was also a discussion about combining under one function
collective channel deletion with existing point-to-point channel
deletion.
8. We discussed QoS one more time. We agreed to add Venn Diagram for
different paradigms QoS. We also discussed mixed paradigms QoSs.
Most of the discussion will appear as advice to implementors.
There was a prolonged discussion about separating lower level event-driven
QoS from the others, since it requests platform QoS rather than MPIRT QoS
guarantee. Hence, this measure is applicable for all paradigms. No decision
was made.
We agreed that most of the current advice to the implementors will
be removed. We will add an appendix that will include reference to several
existing and research solutions to QoS guaranteed implementations.
9. We need to add specific formats for QoSs. The time-driven format
appears in the example at the end of the sections but does not appear
in the QoS section. This needs to be remedied at once.
10. For the event-driven paradigm we agreed to add QoS parameter
to each request handle. Also the same lower level QoS request handle
parameter may appear in other places, like timeout handles.
11. We discussed high level event-driven paradigm. There were several
informal different proposals that deal with different ways this
functionality can be implemented. This section will be modified and
most of the discussion will appear as advice to the implementors.
12. Ad hoc section was discussed. The current write-up mixes up Ad Hoc
and soft quality of service models. We agreed to split it into Ad
hoc, soft QoS real-time, and mixed model real-time.
13. All references to "embedded" become "Resource Constrained". The
current write-up will be updated to reflect implications of profiles
from 11.3.
14. The instrumentation proposal was discussed at length. It was
decided to remove old introduction (page 39 line 45 - page 40 line
30).
It was agreed that we need more info about profiling and existing platform
specific instrumentation. Various updates to the existing functionality
will be posed in view of this additional background study.
15. Timing correctness section (11.15) will be merged into the introduction.
16. We decided to move section on deadlock avoidance and recovery (11.16)
into appendix (real-time JOD).
17. We agreed to meet at MITRE for the next MPI/RT meeting, tentatively
December 5/6. Arkady Kanevsky, arkady@linus.mitre.org, will organize and
host.
Submitted, October 10, 1996
Arkady Kanevsky, Anna Roenbehler, Tony Skjellum