[mpiwg-p2p] Message matching for tools

Tue Feb 9 20:14:54 CST 2016

Hi George,

That's a good point, that tools may also be impacted through the global
communicators.  With this sentence we had wanted to (tactfully) get at the
fact that tools which analyze the communication pattern of the application
(e.g. using the profiling interface to intercept calls and reconstruct the
matching order) may be confounded if the user sets the allow_overtaking
info key.

 ~Jim.

On Tue, Feb 9, 2016 at 6:38 PM, George Bosilca <bosilca at icl.utk.edu> wrote:

> On Wed, Feb 10, 2016 at 12:44 AM, Daniel Holmes <dholmes at epcc.ed.ac.uk>
> wrote:
>
>> Hi George,
>>
>> However, the effect on tools is not limited to global objects such as the
>> predefined communicators - the semantics are changed for *any* communicator
>> on which these info assertions are set. Your "and tools" bit is therefore
>> necessary but not sufficient.
>>
>
> Assuming that the usual precautions are taken, aka. tools and libraries
> create their own communicators, I fail to see how the user hints on it's
> communicators can have any impact on libraries/tools ?
>
> Moreover, as there is no way in MPI to get access to a communicator you
> don't see a reference (except MPI_COMM_WORLD and SELF), the user has no
> practical solution to set hints on a communicator that do no belong to her.
> The opposite is however always true. Because libraries and the tools have
> access to at least some of the application communicators, by setting hints
> on such a communicator (and in the case of tools we are talking about all
> communicators) they can affect the performance of the application.
>
> I believe Jim has now changed the wording of the sentence you are
>> concerned about back to a previous version, i.e. "The usage of info hints
>> may also impact the effectiveness of tools." Does that alleviate your
>> concern?
>>
>
> This is exactly the sentence that raise my concerns. Tools communicators
> are invisible to all other software layers in the application (obviously
> excluding other tools), and as such the particular sentence you mention has
> no meaning.
>
>   George.
>
>
>
>>
>> Cheers,
>> Dan.
>>
>>
>> On 09/02/2016 22:36, George Bosilca wrote:
>>
>> Sorry to jump in the middle (or the end) of the discussion. From my "MPI
>> user" point of view, the last part of the proposed advice is unsettling.
>>
>> "The usage of some info hints may allow the MPI library to change its
>> behavior and, as a result, impact the effectiveness of tools."
>>
>> seems to have a broader scope, and includes application developers, when
>> the most obvious target should have been tools developers. I think removing
>> the last part and emphasizing tools and libraries, makes a more clear point.
>>
>> "Setting info hints on the predefined communicators \const{MPI\_COMM\_
>> WORLD} and \const{MPI\_COMM\_SELF} may have unintended effects, as
>> changes to these global objects may affect all components of the
>> application, including libraries and tools."
>>
>>   George.
>>
>>
>> On Tue, Feb 9, 2016 at 8:02 PM, Jim Dinan <james.dinan at gmail.com> wrote:
>>
>>> We're calling the proposal "info assertions", but the standard still
>>> refers to them as "hints".
>>>
>>> Reading the proposed change again, I think it might be helpful (and more
>>> fair to tools) to say why info could impact tools:
>>>
>>> Setting info hints on the predefined communicators \const{MPI\_COMM\_
>>> WORLD} and \const{MPI\_COMM\_SELF} may have unintended effects, as
>>> changes to these global objects may affect all components of the
>>> application, including libraries.  The usage of some info hints may
>>> allow the MPI library to change its behavior and, as a result, impact the
>>> effectiveness of tools.
>>>
>>>  ~Jim.
>>>
>>> On Tue, Feb 9, 2016 at 12:38 PM, Daniel Holmes < <dholmes at epcc.ed.ac.uk>
>>> dholmes at epcc.ed.ac.uk> wrote:
>>>
>>>> Hi Jim,
>>>>
>>>> I like that change. On a side note: are we really still calling them
>>>> "info hints" in this context rather than "info assertions"?
>>>>
>>>> Cheers,
>>>> Dan.
>>>>
>>>>
>>>> On 09/02/2016 15:51, Jim Dinan wrote:
>>>>
>>>> Hi All,
>>>>
>>>> I'm preparing the updated draft of info assertions for a reading in
>>>> March.  Where did we land on an advice regarding tools?  Do we want advice
>>>> (1) to users, that they info keys may impact tools and/or (2) to tools that
>>>> they should check info?
>>>>
>>>> For instance, we could extend the current advice with the following
>>>> sentence:
>>>>
>>>> Setting info hints on the predefined communicators \const{MPI\_COMM\_
>>>> WORLD} and \const{MPI\_COMM\_SELF} may have unintended effects, as
>>>> changes to these global objects may affect all components of the
>>>> application, including libraries.  The usage of info hints may also
>>>> impact the effectiveness of tools.
>>>>
>>>>  ~Jim.
>>>>
>>>> On Fri, Dec 18, 2015 at 5:24 AM, Marc-Andre Hermanns <
>>>> <hermanns at jara.rwth-aachen.de>hermanns at jara.rwth-aachen.de> wrote:
>>>>
>>>>> Hi Jeff,
>>>>>
>>>>> >     at the moment we don't handle MPI_THREAD_MULTIPLE at all. But we
>>>>> want
>>>>> >     to get there ;-)
>>>>> >
>>>>> >
>>>>> > You should vote for endpoints, as this may help you out here,
>>>>> > particularly if users start mapping endpoints 1:1 w/ threads.
>>>>>
>>>>> That would certainly ease things for us in these situations.
>>>>> Unfortunately endpoints force use to adapt other infrastructure in our
>>>>> measurement system.
>>>>>
>>>>> >     b) Creating a derived datatype on the fly to add tool-level data
>>>>> to
>>>>> >     the original payload may induce a large overhead in practically
>>>>> >     _every_ send & receive operation and perturb the measurement.
>>>>> >
>>>>> >
>>>>> > You should evaluate this experimentally.  I wrote a simple test
>>>>> > (
>>>>> https://github.com/jeffhammond/BigMPI/blob/master/test/perf/typepiggy.c
>>>>> )
>>>>> > and measured 1.5 us per call of overhead to create a datatype.  That
>>>>> > is not significant except for very small messages.
>>>>>
>>>>> Thanks for the pointer. You are right. I should evaluate this further.
>>>>> 1.5us does indeed seem tolerable. I wonder how the influence of the
>>>>> derived datatype is on overall messaging performance, though.
>>>>>
>>>>> This is also something I should evaluate in the process.
>>>>>
>>>>> Cheers,
>>>>> Marc-Andre
>>>>>
>>>>> --
>>>>> Marc-Andre Hermanns
>>>>> Jülich Aachen Research Alliance,
>>>>> High Performance Computing (JARA-HPC)
>>>>> Jülich Supercomputing Centre (JSC)
>>>>>
>>>>> Schinkelstrasse 2
>>>>> 52062 Aachen
>>>>> Germany
>>>>>
>>>>> Phone: +49 2461 61 2509 | +49 241 80 24381
>>>>> Fax: +49 2461 80 6 99753 <%2B49%202461%2080%206%2099753>
>>>>> www.jara.org/jara-hpc
>>>>> email: <hermanns at jara.rwth-aachen.de>hermanns at jara.rwth-aachen.de
>>>>>
>>>>>
>>>>> _______________________________________________
>>>>> mpiwg-p2p mailing list
>>>>> mpiwg-p2p at lists.mpi-forum.org
>>>>> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-p2p
>>>>>
>>>>
>>>>
>>>>
>>>> _______________________________________________
>>>> mpiwg-p2p mailing listmpiwg-p2p at lists.mpi-forum.orghttp://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-p2p
>>>>
>>>>
>>>> --
>>>> Dan Holmes
>>>> Applications Consultant in HPC Research
>>>> EPCC, The University of Edinburgh
>>>> James Clerk Maxwell Building
>>>> The Kings Buildings
>>>> Peter Guthrie Tait Road
>>>> Edinburgh
>>>> EH9 3FD
>>>> T: +44(0)131 651 3465
>>>> E: dholmes at epcc.ed.ac.uk
>>>>
>>>> *Please consider the environment before printing this email.*
>>>>
>>>>
>>>> The University of Edinburgh is a charitable body, registered in
>>>> Scotland, with registration number SC005336.
>>>>
>>>> _______________________________________________
>>>> mpiwg-p2p mailing list
>>>> mpiwg-p2p at lists.mpi-forum.org
>>>> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-p2p
>>>>
>>>
>>>
>>> _______________________________________________
>>> mpiwg-p2p mailing list
>>> mpiwg-p2p at lists.mpi-forum.org
>>> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-p2p
>>>
>>
>>
>> --
>> Dan Holmes
>> Applications Consultant in HPC Research
>> EPCC, The University of Edinburgh
>> James Clerk Maxwell Building
>> The Kings Buildings
>> Peter Guthrie Tait Road
>> Edinburgh
>> EH9 3FD
>> T: +44(0)131 651 3465
>> E: dholmes at epcc.ed.ac.uk
>>
>> *Please consider the environment before printing this email.*
>>
>>
>> The University of Edinburgh is a charitable body, registered in
>> Scotland, with registration number SC005336.
>>
>>
>
> _______________________________________________
> mpiwg-p2p mailing list
> mpiwg-p2p at lists.mpi-forum.org
> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-p2p
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mpi-forum.org/pipermail/mpiwg-p2p/attachments/20160209/a383b15e/attachment-0001.html>