netcdf on the Cray, once more..

• To: netcdfgroup@xxxxxxxxxxxxxxxx
• Subject: netcdf on the Cray, once more..
• From: Harvey Davies <hld@xxxxxxxxxxxx>
• Date: Thu, 7 Jul 1994 16:35:24 +1200 (EST)

You can avoid the overhead of conversion to/from ieee floating point by
using one of the integer types.  Of course you need to use attributes
add_offset and scale_factor to scale the values.

I did a timing test to see how much time this saved.  In each case the
input array was dimensioned 12 x 56 x 64 while the output array was
56 x 64.  The job calculated the mean over the 1st dimension.
The same data was used for both tests.

With type float for both input and output the cpu (user) time was 0.59 sec.
With type short for both input and output the cpu (user) time was 0.35 sec.
Thus using shorts saved 41% of the time needed for floats!

Harvey Davies,                              Home: +61 3 772 5199
CSIRO Division of Atmospheric Research,     Work: +61 3 586 7574
Private Bag No. 1, Mordialloc,               Fax: +61 3 586 7600
Victoria 3195,  Australia                 E-mail: hld@xxxxxxxxxxxx