[mpiwg-rma] Fence + threads

William Gropp wgropp at illinois.edu
Sat Jun 21 10:24:02 CDT 2014


I agree with Rajeev, and I suspect the text that only mentions communicators was an oversight - it probably came from the language in MPI-1.1 where only communicators existed.  I do think that this needs to be clarified.  After all, it is (as it should be) incorrect for two threads in the same process to call MPI_Barrier on the same communicator, even though we could define the semantics in this specific situation.

Bill

William Gropp
Director, Parallel Computing Institute
Thomas M. Siebel Chair in Computer Science
University of Illinois Urbana-Champaign





On Jun 20, 2014, at 1:36 PM, Rajeev Thakur wrote:

> I always assumed that concurrent calls from multiple threads to collective functions on the same comm, win, or fh are not allowed. For example, there is also a collective MPI_File_sync(fh).
> 
> Rajeev
> 
> 
> On Jun 20, 2014, at 1:32 PM, "Balaji, Pavan" <balaji at anl.gov> wrote:
> 
>> 
>> Right.  We don’t say something similar for windows.  Was it intentional or just an accidental ommission?
>> 
>> — Pavan
>> 
>> On Jun 20, 2014, at 1:29 PM, Rajeev Thakur <thakur at mcs.anl.gov> wrote:
>> 
>>> FWIW, the collective communication chapter says this on 217:24-27.
>>> 
>>> "Finally, in multithreaded implementations, one can have more than one, concurrently
>>> executing, collective communication call at a process. In these situations, it is the user's responsibility
>>> to ensure that the same communicator is not used concurrently by two dierent
>>> collective communication calls at the same process."
>>> 
>>> 
>>> 
>>> On Jun 20, 2014, at 1:18 PM, "Balaji, Pavan" <balaji at anl.gov> wrote:
>>> 
>>>> 
>>>> That doesn’t disallow it.  It just says the ordering is arbitrary, which is OK in my example.
>>>> 
>>>> — Pavan
>>>> 
>>>> On Jun 20, 2014, at 1:16 PM, Rajeev Thakur <thakur at mcs.anl.gov> wrote:
>>>> 
>>>>> There is this text on pg 484, ln 18-22
>>>>> 
>>>>> "Collective calls: Matching of collective calls on a communicator, window, or file handle is done according to the order in which the calls are issued at each process. If concurrent threads issue such calls on the same communicator, window or file handle, it is up to the user to make sure the calls are correctly ordered, using interthread synchronization."
>>>>> 
>>>>> Rajeev
>>>>> 
>>>>> On Jun 20, 2014, at 1:11 PM, "Balaji, Pavan" <balaji at anl.gov>
>>>>> wrote:
>>>>> 
>>>>>> 
>>>>>> The standard itself doesn’t seem to disallow it.  This means that, at least, MPICH is incorrect here.  I’d expect most MPICH derivatives to be incorrect for this as well.
>>>>>> 
>>>>>> However, I’m trying to understand if that was the intention (and just forgotten to describe in the standard).  We never discussed this in the MPI-3 working group.
>>>>>> 
>>>>>> — Pavan
>>>>>> 
>>>>>> On Jun 20, 2014, at 12:46 PM, Dave Goodell (dgoodell) <dgoodell at cisco.com> wrote:
>>>>>> 
>>>>>>> On Jun 20, 2014, at 11:16 AM, "Balaji, Pavan" <balaji at anl.gov> wrote:
>>>>>>> 
>>>>>>>> 
>>>>>>>> Is the following code correct?  Two threads of a process both do the following:
>>>>>>>> 
>>>>>>>> MPI_PUT
>>>>>>>> MPI_PUT
>>>>>>>> MPI_WIN_FENCE
>>>>>>>> 
>>>>>>>> Collectives are not allowed on the same communicator simultaneously.  But we don’t say anything similar for windows.
>>>>>>> 
>>>>>>> As I read it, it should be fine according to the standard as long as you haven't specified any assertions that would cause you to be lying to the implementation.
>>>>>>> 
>>>>>>> When we don't specifically disallow concurrent multithreaded operations, then the general threading text seems to allow this.  See MPI-3, p.483, l.1-11:
>>>>>>> 
>>>>>>> ----8<----
>>>>>>> The two main requirements for a thread-compliant implementation are listed below.
>>>>>>> 
>>>>>>> 1. All MPI calls are thread-safe, i.e., two concurrently running threads may make MPI calls and the outcome will be as if the calls executed in some order, even if their execution is interleaved.
>>>>>>> 
>>>>>>> 2. Blocking MPI calls will block the calling thread only, allowing another thread to execute, if available. The calling thread will be blocked until the event on which it is waiting occurs. Once the blocked communication is enabled and can proceed, then the call will complete and the thread will be marked runnable, within a finite time. A blocked thread will not prevent progress of other runnable threads on the same process, and will not prevent them from executing MPI calls.
>>>>>>> ----8<----
>>>>>>> 
>>>>>>> So, absent some external inter-thread synchronization, it will be undefined exactly which epoch the MPI_PUTs end up in.
>>>>>>> 
>>>>>>> -Dave
>>>>>>> 
>>>>>>> _______________________________________________
>>>>>>> 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
>>>>> 
>>>>> _______________________________________________
>>>>> 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
>>> 
>>> _______________________________________________
>>> 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
> 
> _______________________________________________
> mpiwg-rma mailing list
> mpiwg-rma at lists.mpi-forum.org
> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-rma

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mpi-forum.org/pipermail/mpiwg-rma/attachments/20140621/2cd2167c/attachment-0001.html>


More information about the mpiwg-rma mailing list