20031212: pgf90, nf90_short

Tim,

> To: <address@hidden>
> From: Tim Hoar <address@hidden>
> Subject: pgf90, nf90_short (fwd)
> Organization: NCAR/CGD

The above message contained the following:

> I am creating a netcdf file but do not seem to be getting the right
> _size_ file.
> 
> I have a variable of length 18811 and the netCDF file is
> the same size if I code this thing as a NF90_SHORT or a NF90_FLOAT.
> 
> This is with the pgf90 4.1-2 compiler on poorman (RH 9.0), and the
> 
> /usr/local/netcdf-3.5.1-beta5-pgi-hpf-cc-4.1-2/libnetcdf.a    library.
> 
> :netCDF_library_version = "3.5.0 of Aug 28 2002 15:59:11 $" ;
> 
> 
> if(.not. byteSizesOK()) then
>    print *, "Compiler does not appear to support required kinds of variables."
>    stop
> end if
> 
> Does not error out  ...
> 
> Any ideas?
> 
> 
> call check(NF90_def_var(ncid=ncFileID, name="precip", xtype=NF90_short, &
>               dimids = TimeDimID, varid=PrcpVarID))
> call check(NF90_put_att(ncFileID, PrcpVarID, "long_name",
>                                              "daily total precipitation"))
> call check(NF90_put_att(ncFileID, PrcpVarID, "units", "mm"))
> 
> The (current) length of the (unlimited) TimeDimID is 18811
> 
> Without the precip variable, the file size is:  77064 bytes.
> Coded as a short,                file size is: 152472
> Coded as a float,                file size is: 152472
> Coded as a double,               file size is: 227716
> 
> 77064 + 18811*2 = 114686   ( off by a bunch )
> 77064 + 18811*4 = 152308   ( off by 164 bytes )
> 77064 + 18811*8 = 227552   ( off by 164 bytes )

The "problem" is that the "precip" variable is a 1-D record variable (i.e.,
it uses the unlimited dimension.  When such variables are written to
disk, they are put into a system-independent format by an XDR layer
within the netCDF library.  When this layer writes out a single netCDF
record -- as it does for every element of the "precip" variable -- it
pads the output record to a 4-byte boundary for reasons having to do
with the XDR standard.

If the "precip" variable wasn't a netCDF record variable, then the
"short" form of the resulting file would be smaller than the "float"
form of the file.

In summary, having a single, short, record variable doesn't get you
anything over a single, float, record variable.

> ## Tim Hoar, Associate Scientist              email: address@hidden     ##
> ## Geophysical Statistics Project             phone: 303-497-1708       ##
> ## National Center for Atmospheric Research   FAX  : 303-497-1333       ##
> ## Boulder, CO  80307                    http://www.cgd.ucar.edu/~thoar ##

Regards,
Steve Emmerson