[MPI3 Fortran] Argument data sizes

Aleksandar Donev donev1 at llnl.gov
Thu Sep 18 12:19:30 CDT 2008


Hello,

My 0.1 cents. To summarize:
The Fortran interface should use default integers. The implementation should 
be allowed to abort with an "out of range" error.

> I don't follow Cray hardware -- do you actually have machines that  
> have more than 32GB per core?
This question should use the future tense. Aren't we designing MPI-3? Surely 
32GB is at most 5 years away from being common?

That said, I am with Jeff :-) And Bill.

Firstly, I am with Jeff that we should simply follow the MPI-3 base standard. 
Let them argue over what it says (C int, intptr_t, MPI_Int, or whatever), and 
we simply translate it into Fortran. But, clarify for me please. What does 
MPI 2.1(2?) say about this: Does it say that counts are limited to at most 
what a C int can accomodate, or does it simply say that in the C interface 
and we are simply trying to ensure that we don't make the Fortran 
implementation any more demanding than simply slapping a wrapper on top of 
the C routine?

I am also with Bill that we should not bowl to the C binding and make using it 
in Fortran difficult. That is, if C is free to *translate* the MPI standard 
by saying a count is a default C integer (int), then why isn't Fortran free 
to say those same counts are default integers in Fortran. Seems natural to 
me. The MPI *standard* itself can say that an implementation may not support 
counts bigger than a certain internal limit and it is allowed to abort with a 
well-defined error. This is not specifically for the Fortran binding, but any 
binding.

An important point here is that the C standard does not say an int is 32 bits. 
So if there is a machine where people really use 32GB arrays, the C compiler 
will adjust to that as well and the internal limit in the MPI implementation 
and the associated error message will simply go away.

Best,
Aleks

-- 
Aleksandar Donev, Ph.D.
Lawrence Postdoctoral Fellow @ Lawrence Livermore National Laboratory
High Performance Computational Materials Science and Chemistry
E-mail: donev1 at llnl.gov
Phone: (925) 424-6816  Fax: (925) 423-0785
Address: P.O.Box 808, L-367, Livermore, CA 94551-9900
Web: http://cherrypit.princeton.edu/donev




More information about the mpiwg-fortran mailing list