Avoid premature optimization: worry about performance only after you determine where the bottlenecks are. The HDF5 file format is mature and implementation is efficient for many uses.

• A chunk is a fixed-size multidimensional block of data treated as atomic for disk accesses: disk I/O is always in terms of complete chunks. For netCDF-4 variables, chunking should be tailored to make common data access patterns efficient
• Compression (also called "deflation") can save space and either speed up access significantly or slow access to a crawl, depending on details of use, such as chunk cache configuration
• "Endianness" can be specified to avoid byte-swapping during I/O
• Parallel I/O can take advantage of efficiencies of parallel file systems.
• Use of netCDF-4 classic model format can exploit all the performance features above without requiring changes to programs that read the data.

Other netCDF-4 performance features not in netCDF-4 classic model:

• Compound data types: may speed up accessing structured data, because it's stored close together

2012 Unidata NetCDF Workshop > Formats and Performance