[mpi3-coll] MPI-3 Collectives

[Jesper Larsson Traeff]

Hi,

I'd like to participate in this collectives group. Depending on time of
day (European), I'd also try to take part in the phone conference next week.

To the proposed list of issues:

1) non-blocking collective operations
2) topological collective operations (associated with topological
   communicators)
3) sparse collective communication (extended definition of parnership
   relations and communicated data)
4) new collective operations (such as MPI_Reduce_scatter_block and
   MPI_Exscan_all proposed by Jesper)

I'd like to add

5) Persistent collectives (as argued by Tony)

I think 4) could possibly be moved to 2.2 and I will consider making a
proposal for a wording.

My general viewpoint is still that instead of adding all sorts of nice 
things to MPI-3 (with the danger that they may neither be implemented 
nor widely used), this working group should analyse what is really fundamental
and needed in order to make these things implementable (with another personal
viewpoint that I don't think it should be necessary to "expose" any of
the inner workings of an MPI implementation - which of course comes with 
some, small(?) price in possible efficiency).

As I proposed earlier, maybe some means of adding "packages" to MPI could
be defined (a "package" being something which has been discussed and
endorsed by the MPI Forum, but does not belong to the standard and is therefore
not required to be implemented natively, but for which an open source
implementation of a certain quality on top of MPI exists, and for which it
is allowed by implementers to do native, more efficient implementations)?

best regards

Jesper