Parallel netcdf & wrf

Robert Latham robl at mcs.anl.gov
Mon Feb 12 16:36:08 CST 2007


On Sat, Feb 10, 2007 at 05:47:26PM -0700, John Michalakes wrote:
> Hi Rob,
> 
> I've isolated one of the problems we're seeing on Blue Gene to a
> very small test case. The script is attached to this email. There is
> a relatively small data file you can test this with at:
> 
>    http://www.mmm.ucar.edu/wrf/WG2/wrfinput_d01
> 
> It is a CDF 1 file (about 7MB), but it still shows the problem we've
> been seeing with the larger CDF 2 files we've been trying to read.
> In a nutshell, the call to set_vara_fileview in mpinetcdf.c around
> line 2401 is not doing the right thing. The file view remains at the
> beginning of the file, even though the offset to the text field
> we're trying to read is something like 26872 bytes in.  So when
> MPI_File_read_all is called in ncmpi_get_vara_all, we get a string
> with "CDF" and then garbage in it instead of the first value of the
> Times variable, "2000-01-24_12:00:00".

Hi John

I just tried this on our BGL here with pnetcdf-1.0.1, and I get the
correct result.  

> This is with the latest version of pnetcdf, the pre-release issued
> last week.

With the latest version, I don't get 'CDF' in the output, but I do get
'AAAAAA', so clearly something is wrong.  

The only changes in the Argonne environment I can think of is that we
have version 10.1 of IBM's XL compilers (xlc and xlf) installed at
Argonne, and our "bluelight" driver is V1R3M1_400_2006-061024

I should be able to figure out what among the changes between 1.0.1
and the 1.0.2 pre-release is the culprit.  I'll keep you posted.

==rob

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




More information about the parallel-netcdf mailing list