[Mpi3-rma] MPI-3 UNIFIED model clarification

Pavan Balaji balaji at mcs.anl.gov
Sat Aug 3 20:25:12 CDT 2013


We say that the public and private windows are the same.  We also say 
that after a flush the data is available in the target public window.

I guess the question here is -- when can the target process read the 
data written by the origin process?  In the separate model, the answer 
is simple -- we do the necessary synchronization and then do a WIN_SYNC 
to synchronize the two windows.  In the unified model, it's unclear 
since the user does not need to a WIN_SYNC.

  -- Pavan

On 08/03/2013 07:02 PM, Underwood, Keith D wrote:
> Do we make a statement about the state of the memory? Or, do we make a statement about the ordering of loads at the target?  It seems that we make the former statement and not the latter.  We similarly cannot guarantee that the compiler won't optimize away a loop spinning on an RMA variable if the user doesn't do something.
>
> Keith
>
>> -----Original Message-----
>> From: Pavan Balaji [mailto:balaji at mcs.anl.gov]
>>
>> While that statement is true in general, some of us thought that those "some
>> places" included all cache-coherent architectures, and not just x86-like
>> architectures.  Clearly that was not what *all* of the WG thought.
>>
>> Basically, the situation right now is that other cache coherent architectures
>> that require a memory barrier on the target side cannot support UNIFIED at
>> all (at least, not without considerable overhead).
>> This means that other use cases of UNIFIED, such as concurrently accessing
>> nonoverlapping portions of memory through load/store and PUT/GET, are
>> not available to users of such architectures either (even though the
>> hardware is perfectly capable of providing this ability).
>>
>> In short --
>>
>> SEPARATE is too general.  It doesn't take advantage of some hardware
>> features provided by cache-coherent hardware.
>>
>> UNIFIED is too specific.  It is too hard for non-x86 architectures to provide.
>>
>>    -- Pavan
>>
>> On 08/02/2013 09:56 PM, Underwood, Keith D wrote:
>>> The point of UNIFIED was that we add a more useful model that worked in
>> *some* places.
>>>
>>>> -----Original Message-----
>>>> From: Pavan Balaji [mailto:balaji at mcs.anl.gov]
>>>> Sent: Friday, August 02, 2013 7:30 PM
>>>> To: Underwood, Keith D
>>>> Cc: Jed Brown; mpi3-rma at lists.mpi-forum.org
>>>> Subject: Re: [Mpi3-rma] MPI-3 UNIFIED model clarification
>>>>
>>>>
>>>> On 08/02/2013 08:27 PM, Underwood, Keith D wrote:
>>>>> That was specifically *not* the point of unified.  The point of
>>>>> unified was that "if the platform can support it, the user can get
>>>>> something better".  Remember all those conversations about
>>>>> synchronization using hands waving across the machine room?  Unless
>>>>> you guys want to go get rid of UNIFIED?  There really isn't a point
>>>>> to unified if you have to call MPI_Win_sync, is there?
>>>>
>>>> Yes, I remember that discussion.  However, what I took away from that
>>>> discussion was that the user *could* do something outside of MPI, not
>>>> *must*, to ensure correct behavior.
>>>>
>>>> If that was not the intention, then there needs to be a memory model
>>>> where that's the case.
>>>>
>>>>     -- Pavan
>>>>
>>>> --
>>>> Pavan Balaji
>>>> http://www.mcs.anl.gov/~balaji
>>
>> --
>> Pavan Balaji
>> http://www.mcs.anl.gov/~balaji

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



More information about the mpiwg-rma mailing list