Avoid premature optimization: worry about performance only after you determine where the bottlenecks are. The netCDF classic file format is mature and implementation is efficient for many uses.

If netCDF access time is a bottleneck, these techniques may help:

• If data will be written once but ready many times, change dimension order for variables, to favor most common access patterns (see ncpdq). Avoid reading one value per disk access!
• For lots of record variables, use record-at-a-time processing rather than variable-at-a-time
• To avoid copying data when the file schema changes:
• Create all needed dimensions, variables, and attributes before writing data, or
• Reserve extra space in the file header for later additions
• nc__enddef() in C
• NF__ENDDEF() or NF90__ENDDEF() for Fortran
• setExtraHeaderBytes() method of NetcdfFileWritable for Java
• With netCDF-3 or -4, all file metadata is read into memory when file is opened, so functions that get metadata ("_inq_" functions) are very fast. However, try to avoid the need for opening many files to get a little data out of each, with can be slow if each file has lots of complex metadata.
• Use an HDF5-based netCDF-4 format to avoid overhead of schema additions

2012 Unidata NetCDF Workshop > Formats and Performance