> Albert points out the desire to be able to read files written with
> MPI/IO on sequential machines, or read files written on sequential
> machines via MPI-IO.
>
> Unfortunately even the canonical representation cannot make any
> promises about either of these operations.
>
> The reason for this omission is that there are many native file
> formats (particularly for files written from Fortran), and therefore
> MPI cannot be both canonical and inter-operable.
>
> The simplest solution which is guaranteed to work is very likely going
> to remain "read on one process using MPI-IO, and write using the
> normal mechanism".
Unfortunately, I have to agree with this - except for one thing.
The reason that Fortran is more of a problem is PRECISELY because
almost all current workstations run Unix or Unix-like systems, which
use a byte-stream representation for files, and use one or more
bytes to represent newline. And this is the C model.
On systems that are based on a record model (such as MVS, CMS, VMS,
and many older ones), exactly the converse is true. There is usually
only one representation for Fortran, but one per compiler for C. Yes,
I know that recent versions of VMS have added a revolting kludge for C.
And there are probably still some specialised systems that can support
neither model in a "natural" way, though I am not sure how many will
be appropriate for MPI.
Nick Maclaren,
University of Cambridge Computer Laboratory,
New Museums Site, Pembroke Street, Cambridge CB2 3QG, England.
Email: nmm1@cam.ac.uk
Tel.: +44 1223 334761 Fax: +44 1223 334679