[mpiwg-tools] Ticket 358

Wed Feb 4 10:11:58 CST 2015

I agree with Bronis that the language is unnecessary and think attempting
to define correct usage may actually prohibit useful libraries from doing
their jobs. As an example from the notes from our discussion on 1/8:

* ULFM proposal: library may do various activities behind the scenes,
e.g., repair communicators, restore lost messages, so that eventually all
processes will be in a recovered state. However, there could be some
amount of time where the original MPI usage semantics were not strictly
followed

I argue for leaving it alone. Attempting to define it could prove to be
complicated enough that we end up accidentally prohibiting functionality.
It has not been a problem to date, and interesting uses of the interface
have come about that don't strictly follow the original MPI semantics.

Kathryn
_________________________________________________________________
Kathryn Mohror, kathryn at llnl.gov, http://scalability.llnl.gov/
Scalability Team @ Lawrence Livermore National Laboratory, Livermore, CA,
USA

On 2/4/15, 6:15 AM, "Bronis R. de Supinski" <bronis at llnl.gov> wrote:

>
>Yes, my answer is no. The only statement that would be
>appropriate for the standard is "Redefining any MPI
>function to provide semantics inconsistent with the
>standard can result in a non-conformant execution."
>However, the statement is a tautology so it is unnecessary.
>
>On Wed, 4 Feb 2015, Marc-Andre Hermanns wrote:
>
>> Bronis,
>>
>> I don't claim that the use of the profiling interface has any influence
>> on the conformance of the MPI implementation, and you are right that it
>> doesn't, as the re-defined code is not part of the implementation but
>> some third-party software.
>>
>> Yet, the standard could very well state what falls under "standard
>> conformant use" (not implementation!) of the interface. Something like
>> "whatever you do, don't make correct MPI applications fail." The
>> question is only: Should it? (And I take your answer to be no.)
>>
>> Cheers,
>> Marc-Andre
>>
>> On 04.02.15 12:48, Bronis R. de Supinski wrote:
>>>
>>> I do not see why your proposal is necessary or apprpriate.
>>> If I want to redefine MPI_Send to be MPI_Recv (or more
>>> likely to send to /dev/null) I can do it. THe MPI standard
>>> cannot prohibit it and it has no impact on the conformance
>>> of the implementation that I am using.
>>>
>>>
>>>
>>> On Tue, 3 Feb 2015, Marc-Andre Hermanns wrote:
>>>
>>>> Hi Martin,
>>>>
>>>> I see Rolf's point, although I am am unsure whether the initial
>>>>wording
>>>> of the change in 453 is correct.
>>>>
>>>> If you see it from the C perspective, declaration and definition are
>>>>two
>>>> separate things. However, the standard only talks about declaration of
>>>> functions in the MPI_ namespace. That, IMHO, should never be allowed.
>>>> What should be allowed is to "define already declared functions in the
>>>> MPI_ namespace."
>>>>
>>>> I am unsure whether Fortran also has this distinction. C/C++ certainly
>>>> has. If we now allow the _declaration_ of calls in the MPI_ namespace,
>>>> then it would allow me to define MPI_Whatever(), which I think should
>>>> still not be allowed.
>>>>
>>>> Off the cuff, I would try to find a wording that goes along the lines
>>>>of:
>>>>
>>>> "... any redefined function in the MPI_ namespace must guarantee the
>>>> semantics defined in the standard, but is allowed to have additional
>>>> side effects".
>>>>
>>>> This is too rough cut, still. Yet, what I want to express is that any
>>>> re-definition should not change the expected behavior of the call.
>>>>
>>>> Examples:
>>>> - A non-blocking call should not suddenly block.
>>>> - A blocking call must still ensure that the user buffer can safely be
>>>> re-used after the call returns.
>>>> - A broadcast should still be a call where the source of the data is
>>>>the
>>>> root and after completion all other processes have a copy of the
>>>>root's
>>>> buffer. (and not suddenly implement a reduce).
>>>>
>>>> This would still cover MPIEcho, because the additional MPI_Sends on
>>>>the
>>>> slave processes can be regarded as side effects, right?. The
>>>>application
>>>> still just sees the expected behavior of one message being sent from
>>>> process to process.
>>>>
>>>> Cheers,
>>>> Marc-Andre
>>>>
>>>> On 04.02.15 06:17, Schulz Martin wrote:
>>>>> I am still not convinced we need this, but if we find the right
>>>>> language I
>>>>> am not opposed to it. Not sure, how we can write this the right way.
>>>>> This
>>>>> gets fuzzy very easily. Do you have a suggestion?
>>>>>
>>>>> Martin
>>>>>
>>>>> 
>>>>>______________________________________________________________________
>>>>>__
>>>>> Martin Schulz, schulzm at llnl.gov, http://scalability.llnl.gov/
>>>>> CASC @ Lawrence Livermore National Laboratory, Livermore, USA
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> On 2/3/15, 8:37 AM, "Marc-Andre Hermanns" <m.a.hermanns at grs-sim.de>
>>>>> wrote:
>>>>>
>>>>>> Dear all,
>>>>>>
>>>>>> I just got a chance to talk to Rolf on this ticket:
>>>>>> https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/358
>>>>>>
>>>>>> The reason for this ticket was that in ticket 453
>>>>>> (https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/453), which
>>>>>>passed
>>>>>> as 3.0 errata, it is now explicitly allowed to declare functions in
>>>>>>the
>>>>>> MPI_ namespace, as long as the user wants to make use of the
>>>>>>profiling
>>>>>> interface.
>>>>>>
>>>>>> The question now is, whether we should have a clarifying sentence in
>>>>>> the
>>>>>> profiling chapter to tell users of the interface not to tamper with
>>>>>>the
>>>>>> existing semantics of specific calls.
>>>>>>
>>>>>> Rolf understood that the current wording in the ticket may be too
>>>>>> ristrictive, as it remains unclear what 'fully conforms' still
>>>>>>allows.
>>>>>> His suggestion was maybe a less restrictive wording, where tools
>>>>>>like
>>>>>> MPIEcho are still allowed.
>>>>>>
>>>>>> Cheers,
>>>>>> Marc-Andre
>>>>>> --
>>>>>> Marc-Andre Hermanns
>>>>>> Jülich Aachen Research Alliance,
>>>>>> High Performance Computing (JARA-HPC)
>>>>>> German Research School for Simulation Sciences GmbH
>>>>>>
>>>>>> Schinkelstrasse 2
>>>>>> 52062 Aachen
>>>>>> Germany
>>>>>>
>>>>>> Phone: +49 241 80 99753
>>>>>> Fax: +49 241 80 6 99753
>>>>>> www.grs-sim.de/parallel
>>>>>> email: m.a.hermanns at grs-sim.de
>>>>>>
>>>>>
>>>>>
>>>>> _______________________________________________
>>>>> mpiwg-tools mailing list
>>>>> mpiwg-tools at lists.mpi-forum.org
>>>>> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-tools
>>>>>
>>>>
>>>> --
>>>> Marc-Andre Hermanns
>>>> Jülich Aachen Research Alliance,
>>>> High Performance Computing (JARA-HPC)
>>>> German Research School for Simulation Sciences GmbH
>>>>
>>>> Schinkelstrasse 2
>>>> 52062 Aachen
>>>> Germany
>>>>
>>>> Phone: +49 241 80 99753
>>>> Fax: +49 241 80 6 99753
>>>> www.grs-sim.de/parallel
>>>> email: m.a.hermanns at grs-sim.de
>>>>
>>>>
>>>
>>>
>>> _______________________________________________
>>> mpiwg-tools mailing list
>>> mpiwg-tools at lists.mpi-forum.org
>>> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-tools
>>>
>>
>> -- 
>> Marc-Andre Hermanns
>> Jülich Aachen Research Alliance,
>> High Performance Computing (JARA-HPC)
>> German Research School for Simulation Sciences GmbH
>>
>> Schinkelstrasse 2
>> 52062 Aachen
>> Germany
>>
>> Phone: +49 241 80 99753
>> Fax: +49 241 80 6 99753
>> www.grs-sim.de/parallel
>> email: m.a.hermanns at grs-sim.de
>>