• Choosing good chunksizes depends on the access patterns of your data. Are you trying to optimize writing, reading, or both? What are the access patterns at I/O bottlenecks?
• Choose chunksizes so that the subsets of data you are accessing fit into a chunk. That is, the chunks should be as large, or larger than, the subsets you are reading/writing.
• The cache size must also be adjusted for good performance. The cache must be large enough to hold at least one chunk.
• Setting a larger cache (for example, big enough to hold tens or hundreds of chunks) will pay off only if the access patterns support it.
• On today's high-performance systems, large amounts of RAM are available (both to the user and as internal hardware caching.) This suggests that chunks and caches should be large, and programs should take large sips of data.

2009 Unidata NetCDF Workshop for Developers and Data Providers > Chunking and Deflating Data with NetCDF-4