[SCore-users] Re: SCore-users digest, Vol 1 #256 - 4 msgs

[Mark Ellerby]

Hi Atsushi (& SCore team),

Thankyou for your reply. I have talked to our key MPI-IO user and he
insists that his code was running perfectly without errors with the
previous version of SCore/MPICH/Redhat 7.2. And the MPI-IO users are
mathemeticians, so they're pretty handy with numbers ;-)
Assuming that they know their code is giving correct results, is there
anyway they can bypass these error messages that they are getting now?
If they can do that, then they can verify the results for themselves.
If it's not possible to do that, is there any other way around this issue
that you can think of? (which allows them to use a shared filesystem)

Best wishes,
Mark


> I am not familiar with MPI-IO at all, but as far as I undestand, 
> MPI-IO on NFS is not recommended. I found MPICH web page 
> <http://www-unix.mcs.anl.gov/mpi/mpich/buglist-tbl.html>, and it says, 
> 
> ---
> Problems with MPI-IO and NFS
> The network file system (NFS) must be configured extremely carefully 
> for MPI-IO (and many other programs) to work correctly. 
> Unfortunately, few systems are so configured, and doing so can 
> adversely impact performance. As a result, programs using files on an 
> NFS system may hang or produce incorrect results. Note that this is, 
> officially, a design feature of NFS; unless the NFS system is 
> configured with no attribute caching, any two processes, accessing 
> the same file, may produce incorrect results. You can use the 
> -file_system=ufs option of configure to build an MPICH that supports 
> only UFS (Unix File System); MPI-IO works correctly with UFS, XFS, 
> PIOFS, HFS, SFS, etc. (more precisely, those file system that 
> correctly implement basic Unix I/O system calls; something that NFS 
> does not do).
> 
> ----
> Atsushi HORI
> Swimmy Software, Inc.



--
Mark Ellerby              ISS, Leeds University
White Rose Grid           issmde at leeds.ac.uk
Support Officer           0113 3435429