MPI_Alltoallw is inconsistent in its definition because it has byte displacements.

[tonyskjellum]

Problem

MPI_Alltoallw is inconsistent in its definition because it has byte displacements.

Proposal

MPI_Alltoallw is inconsistent in its definition because it has byte displacements, yet they are defined as int, not MPI_Aint.

So, we need a ticket as an "Advice to Users" in the Collective Chapter. Notice that a solution from Big MPI Ticket #80 will address this in an _x API. But, we also want to document the limitation of the existing API, which remains in use.

Changes to the Text

An advice to users will be added beneath the MPI_Alltoallw definition.

Impact on Implementations

None.

Impact on Users

It is just a warning. It may induce users to use MPI_Neighbor_alltoallw or to use MPI_Alltoallw_x (if adopted).

References

See Ticket #80 and https://github.com/mpi-forum/mpi-standard/pull/34 .

[jeffhammond]

I don’t see why this is necessary. The standard is not a user guard and this advice seems pretty obvious to the type of people who read the standard. Hopefully the advice will be present in tutorials and man pages. -- Jeff Hammond [email protected] http://jeffhammond.github.io/

[tonyskjellum]

Reopen?

[dholmes-epcc-ed-ac-uk]

@tonyskjellum re-visit for consideration when MPI-4k (pronounced MPI-fork) is in full swing.

[jsquyres]

Please don't use the term "fork" -- per #80 (comment).