Writing large files from pnetcdf on IBM

[Robert Latham]

On Tue, Nov 28, 2006 at 05:50:00AM -0700, John Michalakes wrote:
> I'm running into problems creating and writing large files (> 2GB)
> from Fortran on an IBM Power5 (bluevista.ucar.edu).  When the output
> file is less than 2GB it writes okay.  When output is greater than
> 2GB, ncdump reports that the file is not a NetCDF file and od shows
> that a large section at the beginning of the file
> is all zeros.

What happens if you use the 'ncmpidump' tool that comes with
parallel-netcdf?  (I'm curious if the problem is with the on-disk file
or with the dumping tool)
 
> I'm creating the file:
> 
>   stat = NFMPI_CREATE(Comm, FileName, IOR(NF_CLOBBER, NF_64BIT_OFFSET),
> MPI_INFO_NULL, DH%NCID)
> 
> The value of stat is zero.
> 
> Code is compiled with OBJECT_MODE set to 64, and I'm pretty sure all the
> relevant offset types are 64-bit. I'm using the version of parallel netcdf in
> bluevista:/contrib/netcdf .  The file pnetcdf.h contains the string:
> 
>   /* "$Id: pnetcdf.h,v 1.20 2005/10/22 16:31:31 jianwli Exp $" */
> 
> Any ideas?

One thing to check is the type of netcdf file you are creating.  You
are already passing in the appropriate flag.  I'd like to take a peek
at the first few bytes of the datafile header.  One of these
approaches should do the trick:

$ xxd my_netcdf_file.nc | head   
$ od -x my_netcdf_file.nc | head 
$ hexdump my_netcdf_file.nc | head

Let us know what that gives you.

If you have access to the 'nccconfig.h' header, you should see 
'#define SIZEOF_OFF_T 8' in there as well as 
'#define SIZEOF_MPI_OFFSET 8'

Thanks for the report.  We should be able to get this figured out
shortly.
==rob

-- 
Rob Latham
Mathematics and Computer Science Division    A215 0178 EA2D B059 8CDF
Argonne National Lab, IL USA                 B29D F333 664A 4280 315B