I'm new to this discussion so please excuse me if I cover anything
that has already been worked out.
Regarding Rajeev's proposal to add a boolean for file-interoperability,
I agree with the need for this but I would like to see the proposal
extended. It seems to me that there are several orthogonal aspects of
interoperbility including:
1) interoperability among different MPI implementations
2) interoperability among different platforms
(e.g. IEEE float -vs- Cray float, endianness for ints, ...)
3) interoperability among node configurations
(There are probably several others.)
Given these multiple orthogonal aspects, I think we should extend
the single flag idea (which imposes an all-or-nothing stance) to something
which will allow MPI-IO clients to specify the interoperability desired
along each aspect. This has the disadvantage of being slightly more
complex, but the advantage of increased flexibility.
I feel this could be useful in a number of situations. For instance, suppose
an application is being developed on a single MPP and the the developers are
gradually scaling up the number of nodes on that MPP. Such a scenario
would benefit from type 3 interoperability, but type 2 interoperbility would
presumably add unnecessary overhead (since the work is being done on a single
machine). In such a scenario, I argue that it would be nice to pick and
choose which aspects of interoperability you are willing to pay for.
-terry
-----------------------------------------------------------------------------
terry jones |
lawrence livermore lab |
POB 808 MS: L-61 | "Out of sight, out of mind," when translated into
Livermore, CA 94550 | Russian [by computer], then back again into
email: trj@llnl.gov | English, became "invisible maniac."
voice: (510) 423-9834 |
fax: (510) 423-6961 | -- Arthur Calder-Marshall
-----------------------------------------------------------------------------