pNetCDF hanging writing output from WRF

[John Michalakes]

Using a first-cut implementation of pNetCDF in WRF, I'm attempting to run
and write history from 4 MPI processes on an IBM Power5+ system, using GPFS.
The code is calling NFMPI_PUT_VARA_<type>_ALL routines for each output
field, and it seems to work well until it gets to writing the 221st of 235
fields at which point two processes (0 and 1) hang on a call to
MPI_File_set_view down in set_vara_fileview (mpinetcdf.c).  The other two
processes finish the sequence of writes to the WRF output file and go to the
next phase of the computation.  The code will stay hung this way until
killed.

The fields being output are a mix of integer and real, and a mix of
dimensions (3, 2, 1, and 0 dimensional).

I've checked the sequence of calls to the pNetCDF layer and they match up
field-for-field on all four processes and they all match up.

Almost certainly this is a goof on my part, but any pointers where to look
next will be much appreciated.

Thanks,

John

-----------------------------------------------------------
John Michalakes, Software Engineer        michalak at ucar.edu
NCAR, MMM Division                   voice: +1 303 497 8199
3450 Mitchell Lane                     fax: +1 303 497 8181
Boulder, Colorado 80301 U.S.A.        cell: +1 720 209 2320
        http://www.mmm.ucar.edu/individual/michalakes
-----------------------------------------------------------