Recommendation:

We recommend that the dimensions in grid structure datasets be ordered in a manner that facilitates readability for the anticipated end users.

Recommendation Details: The order of the dimensions in a group structure should be carefully considered, since this can have significant impact on the ease with which an end user can read the data in the group structure. While tools such as NCO's ncpdq can re-order dataset dimensions, thereby permitting dataset designers to test the effects of their ordering choices against common access patterns, re-ordering large datasets can be time consuming and is best avoided.

Therefore, producers should design products in which the initial dimension order makes it easy for the most typical end users to read. This order is not necessarily the same as what may be simplest to produce.

The Aura File Format Guidelines suggests that the dimensions in HDF-EOS files be ordered based on how rapidly the dimensions change, with the most rapidly changing dimension located first in Fortran arrays, e.g., (XDim,YDim,time) , and last in C arrays, e.g., (time,YDim,XDim) . Adhering to this method when using either the Fortran or C API results in identical ordering of values and efficient storage orders on disk in the HDF or netCDF product files. The COARDS conventions suggest that dimensions in addition to space and time be added "to the left" of the space and time dimensions.