[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[netCDF #SKS-197537]: sporadic missing headers in netcdf output

> Hi,
>
> I am running mom4p1, which writes netcdf-output in parallel. Each
> MPI-taks writes its own file.

Howdy Martin!

Do I understand correctly that netCDF installed correctly and passed make
check, and the problems you
are having are in a user program?

Also, does your application work when run on a single processor?

>
> It appears, that in some output file the header information and possibly
> some data blocks are filled with (nul).

This would occur if different processes were updating the netCDF file metadata
at the same time.

You need to be sure that your processes stop fiddling with netCDF metadata
before different processes start
writing to the file. If enddef is called before any data operations are done in
parallel, and if there are no unlimited dimensions in the file, then parallel
writes should work. (This is the same as you could achieve with fopen calls,
which is exactly what happens when you use sequential netCDF.)

If you cannot manage to get this working with sequential netCDF, or if you want
better parallel I/O support, there are two alternatives:

You may want to use pNetCDF, a parallel version of netCDF, not written at
Unidata, but well-supported and known to work well. See
http://www-unix.mcs.anl.gov/parallel-netcdf/. This unfortunately would require
some code rewriting.

You may instead wish to use netcdf-4/HDF5 files. This would mean installing
HDF5-1.8.1 and netCDF-4.0, and creating HDF5 files with the netCDF API. This
would require almost no code changes in your interactions with netCDF, but you
may wish to take advantage of some features, like the ability to switch from
independant to collective operations.

Please let me know if this doesn't help.

Ed



Ticket Details
===================
Ticket ID: SKS-197537
Department: Support netCDF
Priority: Normal
Status: Closed