[mpiwg-rma] schedule telecon

[Pavan Balaji]

Sorry to keep adding more items, but one more item that needs discussion is the statement in the MPI-3 standard that could potentially disallow loads and PUTs to nonoverlapping locations in SEPARATE windows.  Specifically, see #3 on page 455.

“A put or accumulate must not access a target window once a load/store update or a put or accumulate update to another (overlapping) target window has started on a location in the target window, until the update becomes visible in the public copy of the window."

The first line should either say “load/store access” or “store updates”.  “load/store updates” doesn’t make sense.

Now which of the two it is, is unclear.  Is load + PUTs valid?  This will be fine for hardware managed caches (even noncoherent ones), but not for software managed caches.

One way to decouple this issue is to add more memory models — one that allows this and another that disallows this.  IMO, there should be 4 different memory models that cover all cases:

1. MPI_WIN_SUCKY_MEMORY — for software managed cache type of systems.  Even load + PUTs to nonoverlapping memory regions is not valid.

2. MPI_WIN_SEPARATE — current SEPARATE memory model, but we allow load + PUTs.  This will work as long as the cache is hardware managed.

3. MPI_WIN_UNIFIED — this is what we decided on for our UNIFIED model.  WIN_SYNC is required for overlapping memory regions.  Windows are not always and immediately unified.  They’ll eventually become unified without MPI calls.

4. MPI_WIN_SUPERUNIFIED — no WIN_SYNC required.  Application can assume that the windows are truly unified.  This works on x86 type of architectures.

Of course, all this will be heavily debated.  I’m just queuing up the discussion item.  :-)

  — Pavan

On Nov 4, 2013, at 1:55 PM, Jeff Hammond <jeff.science at gmail.com> wrote:

> 369 seems trivial.  Can we just enumerate the changes on the ticket
> and vote yes/no on moving forward with a reading in December on the
> telecon?
> 
> Jeff
> 
> On Mon, Nov 4, 2013 at 12:51 PM, Jim Dinan <james.dinan at gmail.com> wrote:
>> We also need to make sure that #349 is ready for a reading in December:
>> 
>> MPI_Aint addressing arithmetic:
>> https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/349
>> 
>> I think the main todo here is to update the dynamic window examples.  I also
>> have another open ticket that we should move forward:
>> 
>> RMA needs same_disp info key:
>> https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/369
>> 
>> ~Jim.
>> 
>> 
>> On Sat, Nov 2, 2013 at 8:40 PM, Jeff Hammond <jeff.science at gmail.com> wrote:
>>> 
>>> Yeah, I'll setup a doodle but I wanted to confirm the scope.  Since
>>> Pavan greatly expanded the possible scope, I will propose a doodle
>>> that is oriented at discovering more than one time-slot since I
>>> suspect we will need that.
>>> 
>>> Jeff
>>> 
>>> On Sat, Nov 2, 2013 at 5:04 PM, Rajeev Thakur <thakur at mcs.anl.gov> wrote:
>>>> OK, when?. Do you want to set up a Doodle poll or something?
>>>> 
>>>> Rajeev
>>>> 
>>>> On Nov 2, 2013, at 4:12 PM, Jeff Hammond wrote:
>>>> 
>>>>> I would like to schedule a telecon for the RMA WG to discuss the
>>>>> following three tickets.  We've discussed them all over email but I
>>>>> want to get more conclusive feedback before the Chicago meeting.
>>>>> 
>>>>> - extend the use of MPI_WIN_SHARED_QUERY to all windows
>>>>> (https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/397)
>>>>> - request-based remote completion for RMA
>>>>> (https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/398)
>>>>> - generalize same_op_no_op and allow user to specify all ops to be
>>>>> used (https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/399)
>>>>> 
>>>>> Thanks,
>>>>> 
>>>>> Jeff
>>>>> 
>>>>> --
>>>>> Jeff Hammond
>>>>> jeff.science at gmail.com
>>>>> _______________________________________________
>>>>> mpiwg-rma mailing list
>>>>> mpiwg-rma at lists.mpi-forum.org
>>>>> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-rma
>>>> 
>>>> _______________________________________________
>>>> mpiwg-rma mailing list
>>>> mpiwg-rma at lists.mpi-forum.org
>>>> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-rma
>>> 
>>> 
>>> 
>>> --
>>> Jeff Hammond
>>> jeff.science at gmail.com
>>> _______________________________________________
>>> mpiwg-rma mailing list
>>> mpiwg-rma at lists.mpi-forum.org
>>> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-rma
>> 
>> 
>> 
>> _______________________________________________
>> mpiwg-rma mailing list
>> mpiwg-rma at lists.mpi-forum.org
>> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-rma
> 
> 
> 
> -- 
> Jeff Hammond
> jeff.science at gmail.com
> _______________________________________________
> mpiwg-rma mailing list
> mpiwg-rma at lists.mpi-forum.org
> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-rma

--
Pavan Balaji
http://www.mcs.anl.gov/~balaji