(j3.2006) [MPI3 Fortran] Feedback from Fortran J3 meeting

[Bill Long]

Van Snyder wrote:
>
> If we were to provide these hooks, and encourage encapsulating MPI, PVM,
>   

I think we can limit the discussion to MPI.  As far as I know there is 
only one remaining major program that uses PVM. :)

> etc... transactions under I/O, the problems we are currently pondering
> concerning code motion would be subsumed under ASYNCHRONOUS.  We
> wouldn't have to worry whether SYNC MEMORY interacts with ASYNCHRONOUS
> or VOLATILE or BLORTZ.
>   

Besides the brain cramp resulting from trying to shoehorn MPI data 
transfer routines into the I/O infrastructure,  I think a key part of 
the big picture is being missed here.  We are mainly talking about 
providing an improved interface for the millions of lines of existing 
MPI code.  There is a high resistance level for significant 
modifications to the user's source code. Especially if that code is 
currently working correctly.  We can create interface block widgets all 
we want, since that could all go into a  module, and a lot of the codes 
already have "use mpi" statements.  We might even get people to add an 
asynchronous attribute to the declarations of a buffer variable, since 
this is only relevant in the cases where nonblocking transfers are being 
used.  I don't think "encourage encapsulating MPI transactions under 
I/O" has any reasonable change of getting traction with MPI users.  If 
people are going to do that much modification to their code, they would 
be better off converting it to coarrays.

Cheers,
Bill

  


-- 
Bill Long                                   longb at cray.com
Fortran Technical Support    &              voice: 651-605-9024
Bioinformatics Software Development         fax:   651-605-9142
Cray Inc., 1340 Mendota Heights Rd., Mendota Heights, MN, 55120