[Mpi-forum] MPI user survey

Mon Nov 16 11:34:55 CST 2009

My question does not ask if they want to rewrite their applications.
Since the user base of MPI-2 one-sided is small, it seems more
important to see if there will be new users of MPI-3 one-sided.  One
cannot start using one-sided for the first time without changing some
code.

My belief is that the majority of the people who will benefit from
changes to MPI RMA are those using Send/Recv+threads to fake one-sided
and those using GA/ARMCI.  Those currently using MPI-2 RMA should ask
their vendors to improve the implementation thereof.

Alex - how is your latest suggestion different than what Keith came up with?

Best,

Jeff

On Mon, Nov 16, 2009 at 11:24 AM, Underwood, Keith D
<keith.d.underwood at intel.com> wrote:
> If we ask them if they would like to rewrite their applications, we will get a resounding no.  Again, the goal is to ask what form the new extensions should take.  We should fix RMA, but is that a minimal fix or does it include a very different interface?
>
> Keith
>
>> -----Original Message-----
>> From: mpi-forum-bounces at lists.mpi-forum.org [mailto:mpi-forum-
>> bounces at lists.mpi-forum.org] On Behalf Of Jeff Hammond
>> Sent: Monday, November 16, 2009 8:55 AM
>> To: Main MPI Forum mailing list
>> Subject: Re: [Mpi-forum] MPI user survey
>>
>> Can we completely ignore the performance-richness dichotomy and ask
>> the following?
>>
>> "Would you benefit if the MPI Forum enhances and extends the existing
>> one-sided operations?  That is, would you like to replace MPI
>> two-sided calls in your code with one-sided ones and/or use MPI
>> instead of another one-sided API (e.g. ARMCI)?"
>>
>> As has been pointed out many times, this question is susceptible to
>> the criticism that it is asking "should we make MPI better?" to which
>> almost anyone will answer "yes", which is why it is important to query
>> whether or not such improvements would actually be used in code.
>>
>> Jeff
>>
>> On Mon, Nov 16, 2009 at 10:43 AM, Jeff Squyres (jsquyres)
>> <jsquyres at cisco.com> wrote:
>> > +1
>> >
>> > -jms
>> > Sent from my PDA. No type good.
>> >
>> > ________________________________
>> > From: mpi-forum-bounces at lists.mpi-forum.org
>> > <mpi-forum-bounces at lists.mpi-forum.org>
>> > To: Main MPI Forum mailing list <mpi-forum at lists.mpi-forum.org>
>> > Sent: Mon Nov 16 11:38:51 2009
>> > Subject: Re: [Mpi-forum] MPI user survey
>> >
>> > The important thing to discern is whether we have any user base that
>> wants
>> > that second interface.
>> >
>> >
>> >
>> > Keith
>> >
>> >
>> >
>> > From: mpi-forum-bounces at lists.mpi-forum.org
>> > [mailto:mpi-forum-bounces at lists.mpi-forum.org] On Behalf Of Vinod
>> tipparaju
>> > Sent: Monday, November 16, 2009 8:27 AM
>> > To: mpi-forum at lists.mpi-forum.org
>> > Subject: Re: [Mpi-forum] MPI user survey
>> >
>> >
>> >
>> > I am concerned with the way this question is worded.
>> >
>> >
>> >
>> > I am only concerned for a general user that is doing this survey. I
>> think
>> > this question seems to suggest that MPI-3 RMA will finally be defined
>> in
>> > such a way that either a) rich RMA feature set, or  b) good message
>> rate, is
>> > utimately possible. I disagree with this completely. RMA for MPI-3
>> must
>> > target and deliver both, even if we have to have two interfaces for
>> RMA
>> > (after Keith and Brian do their investigation and conclude that
>> having a
>> > second interface is a must). I think we cannot get away with an
>> interface
>> > with rich feature set that prevents high-message rate or low latency
>> just
>> > because 95% of users who take the survey say they only care about
>> rich
>> > feature set. Alternatively, if we don't care what users say and still
>> are
>> > going to deliver an RMA specification that gives both, why are we
>> asking
>> > this question?
>> >
>> >
>> >
>> > Note that I (as a participant in RMA standardization meetings)
>> understand
>> > that the intention of this question may be to see what people care
>> more
>> > about --latency/message-rate or rich feature set.
>> >
>> >
>> >
>> > I cannot see how this can be worded to address all the concerns with
>> out
>> > including 1/2 a page of explanation. So I think we should hold-off on
>> > including this question in the survey for now.
>> >
>> >
>> >
>> > Thanks,
>> >
>> > Vinod.
>> >
>> >
>> >
>> > --
>> > Vinod Tipparaju ^ http://ft.ornl.gov/~vinod ^ 1-865-241-1802
>> >
>> >
>> >
>> >> From: jsquyres at cisco.com
>> >> To: mpi-forum at lists.mpi-forum.org
>> >> Date: Mon, 16 Nov 2009 07:39:09 -0800
>> >> Subject: Re: [Mpi-forum] MPI user survey
>> >>
>> >> On Nov 16, 2009, at 7:28 AM, Supalov, Alexander wrote:
>> >>
>> >> > Why don't we ask directly: do you want to have subsets in MPI-3 to
>> >> > trade feature richness for performance if you care?
>> >> >
>> >>
>> >> We didn't ask about subsets at all because the WG is "on hold". Do
>> >> you want to introduce a new question? If so, can you suggest the
>> >> specific wording?
>> >>
>> >> Note that Keith suggested an improved wording for the RMA question
>> >> (this was buried in my reply to Dick):
>> >>
>> >> "MPI one-sided communication performance (e.g., message rate and
>> >> latency) is more important to me than supporting a rich remote
>> memory
>> >> access (RMA) feature set (e.g., communicators, datatypes)."
>> >>
>> >> Jeff H. replied to me off-list that he liked this better than his
>> >> suggestion. Does anyone else have any suggestions / comments on this
>> >> one?
>> >>
>> >> --
>> >> Jeff Squyres
>> >> jsquyres at cisco.com
>> >>
>> >> _______________________________________________
>> >> mpi-forum mailing list
>> >> mpi-forum at lists.mpi-forum.org
>> >> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-forum
>> >
>> > _______________________________________________
>> > mpi-forum mailing list
>> > mpi-forum at lists.mpi-forum.org
>> > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-forum
>> >
>> >
>>
>>
>>
>> --
>> Jeff Hammond
>> Argonne Leadership Computing Facility
>> jhammond at mcs.anl.gov / (630) 252-5381
>> http://www.linkedin.com/in/jeffhammond
>> http://home.uchicago.edu/~jhammond/
>>
>> _______________________________________________
>> mpi-forum mailing list
>> mpi-forum at lists.mpi-forum.org
>> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-forum
>
> _______________________________________________
> mpi-forum mailing list
> mpi-forum at lists.mpi-forum.org
> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-forum
>

-- 
Jeff Hammond
Argonne Leadership Computing Facility
jhammond at mcs.anl.gov / (630) 252-5381
http://www.linkedin.com/in/jeffhammond
http://home.uchicago.edu/~jhammond/