[Mpi-forum] Datatypes for all-to-all collectives on intercomms

[HOLMES Daniel]

Hi Michael,

Personally, I would read and summarise the current MPI text about this as these two restrictions:

The type map described by {sendcount, sendtype} in A must be consistent with the type map described by {recvcount, recvtype} in B, e.g. {4, MPI_INTEGER} in A and {1, strided_4_integers}.

and, at the same time,

The type map described by {sendcount, sendtype} in B must be consistent with the type map described by {recvcount, recvtype} in A, e.g. {4, MPI_DOUBLE} in B and {1, strided_4_doubles}.

For an intra-communicator, or for processes in an inter-communicator with dual-membership (erroneous but somehow supported according to section 6.6.2), one of the send/recv communications for each process is back to that same process. Thus, these two statements constrain the operation to have only one type map, which is then applicable to both directions at all processes. That type map could be specified using different but consistent {count, type} parameters at each process.

So, in short, yes.

Cheers,
Dan.
—
Dr Daniel Holmes PhD
Applications Consultant in HPC Research
d.holmes at epcc.ed.ac.uk<mailto:d.holmes at epcc.ed.ac.uk>
Phone: +44 (0) 131 651 3465
Mobile: +44 (0) 7940 524 088
Address: Room 2.09, Bayes Centre, 47 Potterrow, Central Area, Edinburgh, EH8 9BT
—
The University of Edinburgh is a charitable body, registered in Scotland, with registration number SC005336.
—

On 30 Apr 2019, at 10:07, Michael Knobloch via mpi-forum <mpi-forum at lists.mpi-forum.org<mailto:mpi-forum at lists.mpi-forum.org>> wrote:

Hi all,

when looking at the semantics of all-to-all collectives on intercomms I
came to a point that isn't entirely clear from the standard: Am I'm
allowed to use different datatypes for each communication direction,
i.e. send MPI_INTEGERS from group A to group B and MPI_DOUBLES from B to A?

The standard says (page 168, line 38 for MPI_Alltoall, similar
statements for the other all-to-all collectives): The type signature
associated with sendcount, sendtype, at a process must be equal to
the type signature associated with recvcount, recvtype at any other
process. This implies that the amount of data sent must be equal to the
amount of data received, pairwise between every pair of processes. As
usual, however, the type maps may be different.

This however is only true for an intracommunicator as the statement only
holds for every pair of *communicating* processes, which in the case of
an intercommunicator is not every pair of processes (though the name
implies a different semantics). And the advice to users (page 169, line
23) contradicts that by saying that the counts may be different, but
says nothing about the datatypes.

-Michael
--
Michael Knobloch
Juelich Supercomputing Centre
Institute for Advanced Simulation
Forschungszentrum Juelich GmbH
52425 Juelich, Germany

Phone:  +49-2461-61-3546
Fax:    +49-2461-61-6656
E-mail: m.knobloch at fz-juelich.de<mailto:m.knobloch at fz-juelich.de>
WWW:    http://www.fz-juelich.de/jsc/

_______________________________________________
mpi-forum mailing list
mpi-forum at lists.mpi-forum.org
https://lists.mpi-forum.org/mailman/listinfo/mpi-forum

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mpi-forum.org/pipermail/mpi-forum/attachments/20190430/58f9b7be/attachment-0001.html>