netcdf-4 superseding PNetCDF?

[Rob Latham]

On Thu, Jul 21, 2011 at 05:18:15PM -0400, John Urbanic wrote:
> Our initial investigations into which method to use to parallelize the IO in
> the MFIX community code have suggested that, on the strictly coding side,
> either netcdf-4 parallel routines or PNetCDF would be sufficient.  However,
> the code maintenance issue seems to greatly favor netcdf-4 as it seems the
> PNetCDF community is declining in activity.  This reflects my experience
> with forums and the age of support documents.  I have the impression that
> this trend is accelerating.

Well, I guess you're right that we don't make as many commits to the
svn repository as we used to.    I'd rather like to think that pnetcdf
is a mature project.

> However, as you can guess, we lack real experience or knowledge in this
> domain, so I would welcome some better informed opinions.

I think this page still sums up my thinking on the matter:

http://trac.mcs.anl.gov/projects/parallel-netcdf/wiki/PnetcdfAndNetcdf4

If you want the flat-out highest parallel I/O performance, I'm pretty
sure you'll be better served with parallel-netcdf.  We've got some
pretty clever optimizations, particularly with write-combining of
non-blocking operations and our support for MPI datatypes.  

==rob

-- 
Rob Latham
Mathematics and Computer Science Division
Argonne National Lab, IL USA