Re: Comments on parallel-IO


> One thing that I think needs to be clear is that Parallel I/O will be
> supported for netCDF4 HDF5 files, not for netCDF3 files. ...

Right, and more generally the backward-compatibility interface for
accessing netCDF-3 files won't support any other new features of
netCDF-4 either, including large files, multiple unlimited dimensions,
chunking, etc.  The only purpose of the netCDF-3 interface is to
permit accessing data from existing netCDF datasets.  It will be
invoked automatically when operating on an old netCDF file, by
detecting the file type signature at the beginning of the file on

>  ...  I think there 
> is also an implicit requirement that all APIs and features
> from regular netCDF4 work as expected in parallel. (This is the case
> for HDF5.)

I notice that the Argonne/Northwestern group says about features not
yet implemented in their Parallel-NetCDF software:

 - NC_FILL (no plan to implement)

I'm not sure why they have no plans to implement this, but it may be
that supporting the fill-value semantics of netCDF-3 to permit
detection of unwritten values when reading defeats the performance
gains the parallel interface makes possible.  I think that trade-off
would be OK (no automatic writing of fill-values when writing in
parallel) for the sake of performance.