dear all, isn't the a large part of all these problems the self-imposed deadlines and self-imposed yearly releases? Is there really consensus that this is the way to proceed? As some of you know, I'm not so happy with the state of 2.2 and think we would have done better by not being so strict about the Helsinki/SC deadline, and I do not think that we do any good (to the user community, to ourselves) by continuing this way. I also agree that waiting for 3 years (until MPI 3.0) for something like NBC is not at all a good idea. Best regards, Jesper On Wed, Sep 02, 2009 at 04:35:03PM -0500, Martin Schulz wrote:

On Sep 2, 2009, at 2:53 PM, Marc-Andre Hermanns wrote:

...

Hi all,

maybe just to add to Torsten, who wrapped up the discussion quite well.

- we (here in Helsinki) agreed that it is vital to get some things out to the user soon (like the NBCs) and not have them wait all the way, till MPI 3.0 is settled

While I see the reasoning for getting something out early, I do agree with Bronis: this schedule seems very aggressive and rushed and we might not get out of it what we want.

...

- we did not agree, yet, what name such a release would have: 2.3, 2.9, 3.0, ... (that is to be discussed)

Personally, I think calling anything with 3.* is a mistake: introducing just a few new things in 3.0 and then the big additions in a 3.1, potentially even things that break applications, will be confusing and not consistent to previous MPI versions.

...

I think the main message of Jeff's earlier mail was: If you want to get your stuff out soon, you will have to get it to the first reading by march, or you will have to wait till the "big package" is passed in 3

This seems to mean that proposals will have to get feedback from the forum in January or, if you want to allow for at least one round of feedback, even in November. This would, in my opinion, rush the development of proposals as well as their discussion.

Also, by cramping this so much together and assuming that most proposals that could make this deadline and should be included sooner than later make it, the forum will have a large number of big proposals to go through at a single meeting.

...

years, as I took it from the discussions that everybody present was reluctant to have another intermediate release.

What was the reasoning behind that? Why wouldn't the same reason, to release important additions early, apply in 2011?

Martin

...

Best regards, Marc-Andre

Torsten Hoefler wrote:

...

Bronis, all,

...

Re:

...

Note to everyone who is not in the Helsinki meeting: there was consensus that we want some kind of release in November 2010. The exact name of that release is still TBD (likely "3.0", but there's still some discussion about this), but the point is that WG's need to have text ready for 1st readings by the ***MARCH 2010*** Forum meeting in order to be included in the Nov 2010 release.

This is not too far away. No, it is not. It does not seem realistic to me. We would clearly have the nonblocking collectives ready but I don't think much else would be. I think we could have a LITTLE of the tools stuff ready. From what I have seen, I doubt other WGs will get beyond that... we talked about a staged release, possibly also MPI 2.3. I don't think that we rush anything. I'm positive that the FT group could have a proposal for consistent error codes and smaller things that are very important for FT by this time. Other groups like the Hybrid WG could also have something.

And even if we don't have anything, what can you say against releasing a smaller incremental version with just NBCs and minor things.

There was an objection that doing another MPI 2.2-like effort would cost a lot of the MPI 3 energy. While I fully agree to this, I'd say that this seems to be a question of policy. We decide what we discuss (and we should have a good clean base now).

I would also like to have NBCs in an MPI release that does *not* break API compatibility. The burden on the user who wants to use NBC would be too big (imagine that you had to rewrite your 10^6 SLOC MPI program to run with MPI 3.0 just because you want to use NBCs). A good example for such an API breaker is the necessary MPI_Count discussion.

A staged release would also ensure easy implementation and the timely availability of the compliant MPI implementations (I suspect that we will see MPI 2.2 adoption rather soon).

I am really against rush here, we have to evaluate every new proposal as rigorously as we evaluated NBCs. If a proposal fails the March "deadline", then it will simply float into the next release.

All the Best, Torsten

-- Marc-Andre Hermanns Juelich Supercomputing Centre Institute for Advanced Simulation Forschungszentrum Juelich GmbH D-52425 Juelich Germany

Phone : +49-2461-61-2054 Fax : +49-2461-61-6656 eMail : [email protected] WWW : http://www.fz-juelich.de/jsc/

JSC is the coordinator of the John von Neumann Institute for Computing and member of the Gauss Centre for Supercomputing

Sitz der Gesellschaft: Juelich Eingetragen im Handelsregister des Amtsgerichts Dueren Nr. HR B 3498 Vorsitzende des Aufsichtsrats: MinDir'in Baerbel Brumme-Bothe Geschaeftsfuehrung: Prof. Dr. Achim Bachem (Vorsitzender), Dr. Ulrich Krafft (stellv. Vorsitzender), Prof. Dr. Harald Bolt, Prof. Dr. Sebastian M. Schmidt _______________________________________________ mpi-forum mailing list [email protected] http://*lists.mpi-forum.org/mailman/listinfo.cgi/mpi-forum

_______________________________________________________________________ Martin Schulz, [email protected], http://people.llnl.gov/schulzm CASC @ Lawrence Livermore National Laboratory, Livermore, USA

_______________________________________________ mpi-forum mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-forum