pnetcdf 0.9.0 versus hdf5 1.4.5

[Rob Ross]

Hi John,

I'm glad to see that you were able to get a comparison done for your 
specific problem.  I'm disappointed that our code didn't perform best, but 
that happens.

Of course we would like to see the test codes and look at what is going 
on!  Remember, we haven't done any performance tuning yet -- we've been 
concentrating on things like the Fortran functionality :).  So this is a 
great opportunity for us in that we have an application code that we know 
can perform better than it is, and it is well-timed in that we seem to 
have filled in most of the missing functionality!  Having both the HDF5 
and PnetCDF versions would be very helpful.

Jianwei will be in touch regarding acquiring the codes for testing 
purposes.

Thanks!

Rob

On Wed, 24 Sep 2003, John Tannahill wrote:

> Rob,
> 
> I am wrapping up my parallel netCDF work.  Timing results look very good
> with respect to serial netCDF (with my test codes).
> 
> I had some parallel I/O folks here translate my parallel netCDF test code
> to parallel HDF5.  It proved pretty difficult to do and equally as hard
> to get the code compiled and linked on at least one machine of interest.
> 
> Well, at long last, I have the parallel netcdf and the parallel hdf5 test
> codes running on NERSC's IBM SP (seaborg).  I was hoping that parallel
> netCDF would perform well against HDF5, but actually the reverse appears
> to be the case.  The HDF5 reads are ~2.2 times faster that pnetcdf and the
> HDF5 writes are ~3.7 times faster!?
> 
> I was hoping to finish my writeup by the end of the week, but could push
> things off a bit if you folks want to take a look at this set of test
> codes and see what you think.  I would be happy to share them with pnetcdf
> developers, but I would not want to make them part of the pnetcdf
> distribution at this point.
> 
> Regards,
> John