Two meetings were held, of approximately 2.5 hours each. In addition,
the forum was briefed on our progress.
The name of the chapter has been extended to include "EMBEDDED"
Current strategy, as reported to MPI Forum, in plenary session
* Elect 'working note' status ==> continued aggressive progress needed
for SC'96, but we intend this to be discussed and evolved further beyond
SC'96
* Expect to debut chapter @ SC'96, alongside rest of MPI-2
* chapter is to provide compact, initial realtime functionality and philosophy
* chapter is to provide (almost immediate) basis for implementors to begin
work
* chapter is to provide basis for users to gain experience and provide feedback
* continued evolution of chapter will be an on-going activity (more MPI Forum,
mini MPI/Forum, ...)
The chair of the RT subcommittee, on behalf of the RT subcommittee, requested
plenary time in the July MPI meeting to present the chapter in its state as
of that time.
Progress
--------
Entire chapter was scrutinized by subcommittee over the two meetings.
Realtime requirements are beings spelled out, and debated.
We considered:
(1st mtg) clock sync. issues
(2nd mtg) remainder of chapter
Straw poll 9/0/0:
Require monotonically non-decreasing clocks, forbid "wrapping"
We discussed floating point and fixed-point return values for clock
values. (cf, MPI_Wtime)
Straw poll 7/1/1:
Use Chapter 5 channels + semantic extensions (QOS attributes),
and remove specific RT channels from this chapter.
By concensus:
QOS, including priority, is per communicator
Straw poll 7/0/1:
Defer "more general" buffering schemes and receive window schemes to 2nd
generation RT chapter, where they will be considered.
For the time-driven functionality:
----------------------------------
MPI_START_AT ==> MPI_START_TIME, plus
integrate functionality of MPI_REQUEST_EVENT
Time objects specify interval for execution with notification if timeout.
We discussed degrees of freedom in specifying relative/absolute modes for
ends of interval, and this is to be clarified in next chapter release.
Event-driven:
-------------
RT chapter has not as yet fleshed this out.
We anticipate(d) using Generalized Requests or other mainstream MPI-2
notification strategies. If (when) there are dropped from MPI2, RT chapter
will add analogous functionality for event-driven paradigm.
{It is noted that time-driven is a subset of event-driven (special events).}
Priority-driven:
---------------
QOS is to be per communicator, not per channel or message transmitted.
Commentary on clock synchronization issues and high resolute clock issues (Kanevsky)
------------------------------------------------------------------------------------
We need:
1) fine-grainted and accuration instrumentation
2) RT applications need timing correctness
3) Timing parameters are to be adjusted for portability
4) Time-driven paradigm schedules (shared) resources via time calendars
Parameters: (cached attributes for MPI_COMM_WORLD)
a) Resolution (tick) (cf, MPI_Wtick)
b) Drift - upper bounds on clock rate error
Provides bound on accuracy of measurement for small intervals
c) Skew - upper bound on clock value differences (distributed concept)
d) Accuracy - upper bound on clock value differences w.r.t. an ideal clock
e) Access time - upper bound on MPI_Wtime execution
[Caveats needed here]
Overall, there was a lively discussion about what was possible for an implementation
to specify, and it was agreed that "nominal" or "near quiescent" measures of
MPI_Wtime execution time might be what is provided, so that Access time is actually
in an ideal condition, rather than under a worst case, where it might be arbitrarily
large. An application that works to starve clock access across a network, or
otherwise bogs down a processor might blow away the Access time bound.
First meeting attendees (June 5)
---------------------------------
Robert Babb babb@cs.du.edu Denver U.
Ron Brightwell bright@cs.sandia.gov SNL
Dennis Cottel dennis@nosc.mil NRaD
Arkady Kanevsky arkady@mitre.org MITRE
Lloyd Lewins lewins@msmail4.hac.com Hughes
Brian McCandless bmccand@lsc.nd.edu Notre Dame
Harish Nag harish@ssd.intel.com Intel
Perry Partow partow@nosc.mil NRaD
Tony Skjellum tony@cs.msstate.edu MSU
Jerrell Watts jwatts@scp.caltech.edu Caltech
Second meeting attendees (June 6)
---------------------------------
Robert Babb babb@cs.du.edu Denver U.
Ron Brightwell bright@cs.sandia.gov SNL
Dennis Cottel dennis@nosc.mil NRaD
Arkady Kanevsky arkady@mitre.org MITRE
Lloyd Lewins lewins@msmail4.hac.com Hughes
Harish Nag harish@ssd.intel.com Intel
Perry Partow partow@nosc.mil NRaD
Tony Skjellum tony@cs.msstate.edu MSU
Jerrell Watts jwatts@scp.caltech.edu Caltech