[mpich2-dev] Fw: 32-bit ROMIO: What is the best solution?

[William Gropp]

One thing that we've talked a little about is providing an option to  
make MPI_Aint and MPI_Offset the same (large) size, even if sizeof 
(void*) < sizeof(MPI_Aint).  Does it make sense to make this the  
option (Aint and Offset the same size) and then make everything else  
match up with that (still as an option, since there are costs to  
making Aints bigger than pointers).

Bill

On Feb 28, 2008, at 4:21 PM, Bob Cernohous wrote:

>
> I'm moving a discussion here from the IBM/Argonne open source DCMF  
> mailing
> list (http://dcmf.anl-external.org/wiki).
>
> Very briefly, we're making MPI_Aint into a 64 bit signed value on
> 32-bit MPICH2/ROMIO platforms.  This will fix problems with files  
> larger
> than 2G (and possibly other 4G virtual address problems).
>
> We'd like to start sending fix patches to the mpich-dev community  
> for comment.
>
> The code base that we're patching is changing under us.  So our  
> patches
> might not apply properly for members of this list.  When we're  
> done, we
> will have a patch for the current release of MPICH2.  For now, we'd  
> like
> to provide these patches to the community for information and  
> discussion.
>
> If you're interested, there might be more technical history in the  
> DCMF
> archives.  Otherwise, start the discussion here.
>
> bobc at us.ibm.com wrote on 02/27/2008 04:19:29 PM:
> >
> > bobc at us.ibm.com wrote on 02/08/2008 02:41:40 PM:
> >
> > >We all realize there are issues with MPI_Aint's in 32 bit  
> implementations.
> > >Using signed 32 bit addresses along with 64 bit offsets can  
> result in
> > >some pretty broken code.  I've reproduced several problems on
> > >BGL/BGP/linux.   The problems are most obvious with romio files  
> > 2G or
> > >virtual addresses > 2G.
> >
> > Earlier this month I posted some comments to this mailing list about
> > 32-bit MPICH/ROMIO/MPI_Aint issues (see above and the archive).
> >
> > We've started to work on the problem.   We're (optionally) making
> > MPI_Aint a 64 bit signed value even when integers/pointers are 32  
> bit
> > values.
> >

William Gropp
Paul and Cynthia Saylor Professor of Computer Science
University of Illinois Urbana-Champaign


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.mcs.anl.gov/mailman/private/mpich2-dev/attachments/20080229/d68b4e52/attachment.htm>