> From arkady@linus.mitre.org Wed Jan 31 20:03 MET 1996
> To: mpi-realtime@mcs.anl.gov, Bjarne.Herland@ii.uib.no
> Subject: Re: Priorities
>
> Bjarne,
>
> I think that before we consider your model (which is sound)
> we should answer the most critical question which you pose
> in the write-up:
>
> HOW USEFUL IS THIS TO PROGRAMMERS?
> AND HOW DO THEY PROPOSE TO USE PRIORITY MECHANISMS IN MPI?
I completely agree : these questions should be answered before anyone
start doing serious work with this. I will talk to a couple of people
I'm in contact with who are working with real-time problems and post
a summary here when I have something. The actual problem in real-time
computing (as I see it) is to guarantee response-time and/or bandwidth,
and prioritizing messages *might* be the way to deal with this. If we
could build in mechanisms in MPI to *directly* support guaranteed response-
time and/or bandwidth, we probably wouldn't need priorities. Any thoughts
on this?
BTW : some interesting benchmarks can be found on
http://www.cern.ch/pub/combench/
> Another comment is that it is actually very hard to implement
> correctly. Most of the network protocols ignore priorities so
> implementors will have hard time to provide complete/correct
> semantics to programmers.
I have small hopes for efficient implementations of this on standard
network-protocols. However, an MPI-standard describing *desired* behaviour
for real-time systems might inspire someone to do some work on this. And
then we might get something useful. It might be that this is beyond the
scope of MPI, and if so, I assume that the right person will let me know.
Of course - before we know what the desired behaviour is, we need to
talk to some programmers that might use RT-MPI, as you stated above... :-)
>
> Arkady
>
Regards,
- Bjarne
[ rest of message deleted ]