[MPI3 Fortran] Request for a straw vote.
crasmussen at newmexicoconsortium.org
Fri Jun 12 12:17:19 CDT 2009
On Jun 12, 2009, at 3:52 AM, Supalov, Alexander wrote:
> Complementing Steve's reply that represents Intel's position, there
> may be two additional aspects to consider, namely:
> 1. We may want to decide what flavor of the new Fortran interface
> we're striving at. There are at least two extremes:
> a. As close as possible to the C interface with all its low level
This is the choice for the MPI standard. There has been some
discussion of this in earlier emails. I'll let Jeff answer if you
would like a recap.
> b. As natural as possible for Fortran with all its modern
> expressive power
I'm going to work on an additional MPI API and reference
implementation that will be outside of the MPI standard (like Boost
MPI is for C++). I'm happy to have additional cooks.
> We may even want to formulate two different interfaces according to
> these extremes, and see what they look like, both in theory and in
> practice. I would even propose a usability review with the
> application programmers. Then we may be better equipped to find a
> golden middle between the extremes, or at least consistently go for
> one of the extremes if it proves superior to the other.
> 2. As soon as the MPI standard allows certain features to be
> expressed in a compiler dependent way, even for a while (the any-
> type-and-shape comes to mind here first and foremost), one is
> exposed to the possible subtle semantical variations between
> different compilers, as well as between the compiler ways and the
> final standard. This may lead to dissatisfied customers who will
> notice that when they change the compiler/MPI implementation, the
> behavior of their applications may change as well.
> Note that even if we try to specify very precisely what we want from
> the Fortran standard, and then, for any reason, the Fortran standard
> committee decides to go a different way, we may be basically left
> with an MPI interface that is difficult to impossible to represent
> faithfully in a Fortran standard compliant manner.
> Hence, it may be advisable to wait until the respective Fortran
> standard syntax and semantics are cast in stone. As we're talking
> about MPI-3 here, we may have some time for this slower, safer
> approach I hope.
Yup, the MPI-3 Fortran interfaces will wait on the Fortran standard to
be finalized so this won't be a problem. The earliest any MPI-3 stuff
could come out would be when the leaves fall in 2010.
More information about the mpiwg-fortran