Re: [mpi-21] Proposal: Topological Collectives for MPI-3

[Torsten Hoefler <htor@xxxxxxxxxxxxxx>]

Dera all,
> There are surely applications that can benefit from topological collectives
> (one could also imagine use for reduction functions for determining
> min/max over all neighbors etc.), and I agree fully that there are
> several benefits (efficiency, correctness, deadlock-freedom, ...) of 
> leaving the implementation to a library.
Yes, the proposal is open for additions like this. I would like to talk
about this more. 

> Nevertheless, I would probably be against adding this to "core" MPI:
> 
> * the proposal is too limited. There are surely applications which need
>   a much richer set of neighbors (eg. all 8 in a 2-dim Cartesian grid, or
>   also neighbors two "hops" away, etc...). More functionality would be
>   needed for specifying neighborhoods, and this would quickly run out
>   of control and bloat the standard with things that have only very special
>   applicability (yes, some of this can be handled by graphs, but I think 
>   the argument stands)
Yes, see my last email about the modularized approach. At least I would
like to have it in the working group to get an idea about what will be
needed. If we include it in the core-standard or come up with a proposal
that is just supported by the forum can be decided later (with
institutional votes for example). But I think it's worth to think about
it.

> * "collective" is also a restriction, ie. there are applications where only
>   some of the processes are interested in doing the neighbor-exchange, and
>   it could easily become cumbersome to require all to participate. Again:
>   more flexibility probably needed, with high implied costs for the standard
Isn't this what communicators are for? If not all processes want to
participate, shouldn't the communicating processes be in a separate
communicator?

> Instead, I think MPI-3 should concentrate on analysing obstacles to efficiently
> realizing such libraries, and making sure that the real, basic functionality
> is added.
Yes, I agree!

> It would be nevertheless be tempting to standardize some of this as part
> of "auxiliary MPI-3"; conditions for having a chapter in the auxiliary
> standard would be
> - a full proposal in the style of the MPI standard
> - a complete, OPEN implementation in terms of "core" MPI
> - some commitment to maintaining the implementations
> - acceptance from the Forum
Yes, we will try to establish such a proposal in the working group.

> I would suggest setting up subcommittee(s) for dealing with this, eg. 
> proposing basic functionality needed for efficient topology libraries, and
> would be interested in participating
We have the collective working group. We can think about splitting this
group further.

Best,
  Torsten

-- 
 bash$ :(){ :|:&};: --------------------- http://www.unixer.de/ -----
Indiana University    | http://www.indiana.edu
Open Systems Lab      | http://osl.iu.edu/
150 S. Woodlawn Ave.  | Bloomington, IN, 474045-7104 | USA
Lindley Hall Room 135 | +01 (812) 855-3608