• Pack floating-point data into a narrower (and lower precision) numeric type, for example pack 32-bit floats into 16-bit shorts or 8-bit bytes using scale_factor and add_offset attributes. The NCO utilities ncpack or ncpdq will pack variables into smaller types.
• Store whole files compressed. This has the drawback that the whole files must be uncompressed to access even a small subset of the data.
• The netCDF Java library uncompresses files with ".Z", ".zip", ".gzip", ".gz", or ".bz2" extensions and caches the expanded file for subsequent reads, keeping cache size within a specified limit. This works well in read-only applications like servers.
• Use sparse data structures for sparse grids with lots of missing data, for example store only array indices and valid values. Other sparse formats such as CSR (Compressed Sparse Row) may be useful.

NetCDF-4 note:  provides per-variable compression. If you only access a small amount of data in a large file, only a small part of the data gets uncompressed.

2010 Unidata NetCDF Workshop > Formats and Performance