(j3.2006) (SC22WG5.5283) [ukfortran] Alessandro's slides

[N.M. Maclaren]

On Jun 20 2014, Dan Nagle wrote:
>
>The idea is to have libcaf work with any mpi,
>and it's been tested with the mpis on several systems.
>
>I don't know exactly which mpi was used on which system.

Yes.  I wrote a draft of a proposed design, which was used as a basis,
though I have not been involved recently (or with the actual development).
As people will know, one of my main hobby-horses is to ensure that the
standard does not specify anything that is not reasonably implementable
on top of any plausible MPI.

Bill, Reinhold and I had an, er, robust debate by Email over 14-158,
and I drafted a paper during the flight on the progress issue.  My point
is that this is a problem as deceptively simple and actually fiendishly
tricky as the data consistency one, and we need a proper debate (and
probably decision) at the WG5 level over what Fortran can reasonably
assume of an implementation.

The problems are, of course, mainly the atomics and multi-count events.
I am 99% sure that the rest of the design is reasonably implementable,
though some codes may be efficient on a specialist coarray implementation
and direly inefficient on the general ones.

Regards,
Nick.