[Mpi3-rma] MPI-3.1 consideration slides

[Jeff Hammond]

Yes, and it is unfortunate that there is a dearth of representative
platforms.  However, MPI-2 RMA worked to get this right and it's
generally accepted that those semantics support non-coherent
platforms, so in the absence of an argument that MPI-2 RMA semantics
don't work for non-coherent platforms, I would argue that we can just
use those.

Jeff

On Mon, Dec 3, 2012 at 7:55 AM, Jim Dinan <dinan at mcs.anl.gov> wrote:
> On 12/3/12 9:44 AM, Jeff Hammond wrote:
>>
>> On Mon, Dec 3, 2012 at 7:39 AM, Jim Dinan <dinan at mcs.anl.gov> wrote:
>>>
>>> On 12/2/12 7:44 AM, Jeff Hammond wrote:
>>>>>>
>>>>>>
>>>>>> (1) Define semantics of shared memory windows in separate model.
>>>>>>
>>>>>> In Section 11.2.3 it says "MPI does not define semantics for accessing
>>>>>> shared memory windows in the separate memory model."
>>>>>>
>>>>>> I would like to try to define these semantics in MPI 3.1.  Shouldn't
>>>>>> it at least be possible to use MPI_WIN_(UN)LOCK with
>>>>>> MPI_LOCK_EXCLUSIVE here?
>>>>>>
>>>>>> I believe that supporting RMA for weakly- or non- coherent memory
>>>>>> architectures is vital for future systems.
>>>>>
>>>>>
>>>>> I agree. We had this discussion and decided to postpone it because we
>>>>> had other more pressing issues. Now may be the time. A discussion would
>>>>> be good.
>>>>
>>>>
>>>>
>>>> Yes, there was no urgency.  I'm glad the standard left this issue wide
>>>> open instead of defining something restrictive.
>>>
>>>
>>>
>>> IMHO, defining shared memory windows in the separate model is not a
>>> worthy
>>> use of the Forum's time.  This will be hard to get right, and AFAIK no
>>> (reasonable) system requires it.  As we've currently left it, if this
>>> functionality is needed, one can define the semantics as an MPI
>>> extension.
>>> I think this is a better option that defining a new model in the absence
>>> of
>>> a platform that uses it.
>>
>>
>> Intel SCC is non-coherent but supports POSIX shared memory.  It's
>> really quite simple to use there, but one has to have synchronization
>> primitives.  I just want the Forum to define those in the same manner
>> as for MPI-2 RMA, which supports non-coherent platforms.
>>
>> What's the harm in saying that MPI_WIN_UNLOCK approach causes global
>> visibility of the window to all processes?
>
>
> I think that what you suggested would be a useful and reasonable extension
> to an MPI implementation.  However, the SCC is still a research prototype; I
> don't think it has reached the level of maturity needed to define semantics
> in a standards document.
>
>
>  ~Jim.
> _______________________________________________
> mpi3-rma mailing list
> mpi3-rma at lists.mpi-forum.org
> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma



-- 
Jeff Hammond
Argonne Leadership Computing Facility
University of Chicago Computation Institute
jhammond at alcf.anl.gov / (630) 252-5381
http://www.linkedin.com/in/jeffhammond
https://wiki.alcf.anl.gov/parts/index.php/User:Jhammond