From: Jeff Squyres <jsquyres@XXXXXXXXX>
Date: January 18, 2007 10:26:07 PM CST
To: Karl Feind <kaf@XXXXXXX>
Cc: mpi-comments@XXXXXXXXXXXXX,Steven Reinhardt <spr@XXXXXXX>, mpi- 
users@XXXXXXXXXXX
Subject: Re: [MPICH] standardizing 64 bit integer MPI

Greetings Karl!

Just curious: do you have a current customer demand for this?  I.e.,  
are there real-world apps out there that could benefit from Enormous  
counts/sizes?  Or are you mainly trying to look forward and  
anticipate the need?


On Jan 18, 2007, at 11:15 AM, Karl Feind wrote:

>
> Greetings MPI standards enthusiasts:
>
> Memory capacity on large computer systems and clusters conintues to  
> increase.
> As the number of systems with substantially more memory than 2 GB  
> per MPI
> process increases, the demand for a flavor of the MPI library to  
> support MPI
> communication with message sizes or buffer sizes in excess of 2^31  
> elements
> will also increase.
>
> The industry is pretty entrenched in a 32 bit default
> integer size for C "int" and Fortran INTEGER.
> See http://www.unix.org/whitepapers/64bit.html.
>
> Perhaps the MPI Forum should consider adding 64 bit integer variants
> of many of the MPI functions that specify "int" or "INTEGER" arguments
> for the message size/count.  Eg.  MPI_Send and MPI_Recv.
> I envision an interface called MPI_Send64 that specifies the "count"
> argument is "int64" or "INTEGER(KIND=8)".
>
> Do you agree there is a need?  Do you have any ideas about a  
> solution for
> this issue?
>
> Thanks,
> Karl Feind
> SGI


-- 
Jeff Squyres
Server Virtualization Business Unit
Cisco Systems