Thanks Mikhail,

You have a good point.


With the current semantic used in the IMB benchmark, this cannot be equivalent to

MPI_Reduce() of N bytes followed by MPI_Scatterv() of N bytes.


So this is indeed a semantical question :

what should be a MPI_Reduce_scatter() of N bytes equivalent to ?

1) MPI_Reduce() of N bytes followed by MPI_Scatterv() in which each task receives N/commsize bytes

2) MPI_Reduce() of N*commsize bytes followed by MPI_Scatterv() in which each task receives N bytes.


I honestly have no opinion on that, and as long as there is no memory corruption, I am happy with both options.



Cheers,


Gilles

On 12/5/2018 12:25 PM, Mikhail Kurnosov wrote:
Hi,

The memory manager of IMB (IMB_mem_manager.c) do not support the MPI_Reduce_scatter operation. It allocates too small send buffer: sizeof(msg), but the operation requires commsize * sizeof(msg).
There are two possible solutions:

1) Fix computations of recvcounts (as proposed by Gilles)
2) Change memory allocation for send buffer in the memory manager of IMB. That approach was consistent with IMB style (for example, buffer allocation for MPI_Scatter operation)

WBR,
Mikhail Kurnosov
On 04.12.2018 17:06, Peter Kjellström wrote:
On Mon, 3 Dec 2018 19:41:25 +0000
"Hammond, Simon David via users" <users@lists.open-mpi.org> wrote:

 > Hi Open MPI Users,
 >
 > Just wanted to report a bug we have seen with OpenMPI 3.1.3 and 4.0.0
 > when using the Intel 2019 Update 1 compilers on our
 > Skylake/OmniPath-1 cluster. The bug occurs when running the Github
 > master src_c variant of the Intel MPI Benchmarks.

I've noticed this also when using intel mpi (2018 and 2019u1). I
classified it as a bug in imb but didn't look too deep (new
reduce_scatter code).

/Peter K

--
Sent from my Android device with K-9 Mail. Please excuse my brevity.

_______________________________________________
users mailing list
users@lists.open-mpi.org
https://lists.open-mpi.org/mailman/listinfo/users

_______________________________________________
users mailing list
users@lists.open-mpi.org
https://lists.open-mpi.org/mailman/listinfo/users
_______________________________________________
users mailing list
users@lists.open-mpi.org
https://lists.open-mpi.org/mailman/listinfo/users

Reply via email to