[mvapich-discuss] (no subject)

[Ivan Raikov]

Hello Hari,

  Thanks a lot for your reply. Following are the answers to your questions:
1. Size of job: I have been able to reproduce this issue in jobs
ranging from 2 processes running on the same node, to 256 processes
running across 16 nodes.
2. Message size: reproducible for messages of size 3 bytes or larger
and datatype MPI_BYTE.
3. Processor type: Intel Xeon E5-2680v3 (This is on the SDSC Comet cluster)
4. Interconnect type: Mellanox FDR InfiniBand
5. OFED version: I am attaching the output of ofed_info, does this
answer your question?

Thanks and best regards,
Ivan





On Thu, Jan 28, 2016 at 9:07 PM, Hari Subramoni <subramoni.1 at osu.edu> wrote:
> Hello Ivan,
>
> Your understanding of the semantics is correct. MVAPICH2 also follows these
> semantics. However there could be some interactions happening with the
> underlying algorithm being used for the Scatter operation in MVAPICH2 on
> your system.
>
> To further debug this, could you please provide us with the following
> details
>
> 1. The size of the job you're running
> 2. The message size being used for the Scatter operation
> 3. Processor type
> 4. Interconnect type
> 5. OFED version
>
> Thx,
> Hari.
>
> On Jan 28, 2016 11:40 PM, "Ivan Raikov" <ivan.g.raikov at gmail.com> wrote:
>>
>> Hi all,
>>
>>     I recently tried to use some of my OpenMPI code in an MVAPICH2
>> environment, and I noticed a discrepancy in the semantics of
>> MPI_Scatter and other collective operations, so I thought I would ask
>> about this here.
>>
>> I have been following the pattern below when calling collective
>> operations in OpenMPI:
>>
>> if (myrank == root) {
>>     MPI_Scatter(vector, N, MPI_INTEGER, recvbuf, N, MPI_INTEGER, root,
>> MPI_COMM_WORLD);
>>   }
>>   else {
>>     MPI_Scatter(NULL, 0, MPI_DATATYPE_NULL, recvbuf, N, MPI_INTEGER,
>> root, MPI_COMM_WORLD);
>>   }
>>
>> My understanding of the MPI_Scatter documentation was that the
>> sendbuf, sendcount, and sendtype arguments are "significant only at
>> root" meaning that their values should be ignored by all non-root
>> processes. This certainly works in OpenMPI. However, the same code
>> results in no data being sent to the non-root processes with MVAPICH2
>> 2.1. Have I completely misunderstood the specification, or does
>> MVAPICH2 have different collective operation semantics for a reason?
>> Thanks in advance for any clarification and best regards,
>>
>>
>>   -Ivan Raikov
>> _______________________________________________
>> mvapich-discuss mailing list
>> mvapich-discuss at cse.ohio-state.edu
>> http://mailman.cse.ohio-state.edu/mailman/listinfo/mvapich-discuss
-------------- next part --------------
A non-text attachment was scrubbed...
Name: ofed_info.log
Type: text/x-log
Size: 10742 bytes
Desc: not available
URL: <http://mailman.cse.ohio-state.edu/pipermail/mvapich-discuss/attachments/20160129/146dbd25/attachment-0001.bin>