Re: Fortran 2008 (was Re: New style DO syntax?)

ejko123@xxxxxxxxx wrote:

The problem with co-arrays is that there's only one implementation,
and that is by a vendor whose market share is minuscule. There can't
be more than a couple dozen people who have significant code
development
experience with co-arrays, and the spectrum of applications with which
they
have worked is necessarily limited...

You may be suprised. Co-arrays should be familiar to most folks who have
used the Cray T3D and T3E systems of just a few years ago. CAF is based
on the 1-sided shmem_get/shmem_put model of those machines. The T3D and
T3E were distributed memory MPP machines, but allowed individual processors
to directly load/store data on remore memories. (The T3D was especially
entertaining, when using Bob Numrichs original "F--" library.)

If you ask me, I think it would be much better for purposes of F2008 to
put
some effort into a better set of Fortran bindings to MPI, say as a
technical report.

The F-- model was so effective that it forced the MPI guys to include
the 1-sided "RMA" ops in MPI-2. Think of CAF as a built-in version of
MPI_GET/MPI_PUT, but (potentially) without the subroutine call/library
overhead.

It may well be that co-arrays turn out to be useful - but nobody really
knows because
almost nobody has experience with them.

Anyone who has experience with MPI_GET/MPI_PUT and friends should immediately
see the value in CAF.

Walt
.