[mpich-discuss] determinism in collectives

[Jeff Hammond]

Yes, what I was saying before is that we know that MPI on Blue Gene/Q
drops into the MPICH2 algorithm for MPI_Reduce on for
MPI_DOUBLE_COMPLEX with MPI_SUM on irregular communicators and that
MPIC_Send and MPIC_Recv do not change their behavior on this machine,
so whatever is true of stock MPICH2 1.4 is true for MPI on Blue
Gene/Q.

This confirms what we suspected, which is that BGQ-MPI has
deterministic reductions in this case because MPICH2 does.

Jeff

On Mon, Nov 12, 2012 at 9:43 PM, Pavan Balaji <balaji at mcs.anl.gov> wrote:
>
> The default machine-independent implementation of MPICH is
> deterministic.  However, each device can (and most times does) implement
> its own version of MPI_Reduce.  We have no control on them.
>
>  -- Pavan
>
> On 11/12/2012 09:40 PM US Central Time, Jeff Hammond wrote:
>> We are primarily interested in MPICH2, as the motivation for this
>> question relates to a particular supercomputer that uses MPICH2 and,
>> in particular, the MPICH2 implementation of MPI_Reduce for
>> (comm,type,op) of interest to us.
>>
>> Jeff
>>
>> On Mon, Nov 12, 2012 at 9:35 PM, Pavan Balaji <balaji at mcs.anl.gov> wrote:
>>>
>>> The MPI standard does not guarantee this.  For example, an
>>> implementation could use topology-aware reductions or an MPI_ANY_SOURCE
>>> while waiting for messages from peers making the order in which the
>>> reduction operations are applied different for each run.
>>>
>>>  -- Pavan
>>>
>>> On 11/12/2012 09:00 PM US Central Time, Nichols A. Romero wrote:
>>>> Hi,
>>>>
>>>> I apologize if this has been asked a trillion times, if its documented somewhere just point it out to me online.
>>>>
>>>> Are collective operations, such as sums, using MPI_Reduce deterministic? In other words, if I run MPI_Reduce N times on P tasks, will the sums always occur in the same order. Thus it would be reproducible to the expected precision.
>>>>
>>>
>>> --
>>> Pavan Balaji
>>> http://www.mcs.anl.gov/~balaji
>>> _______________________________________________
>>> discuss mailing list     discuss at mpich.org
>>> To manage subscription options or unsubscribe:
>>> https://lists.mpich.org/mailman/listinfo/discuss
>>
>>
>>
>
> --
> Pavan Balaji
> http://www.mcs.anl.gov/~balaji



-- 
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