Another good writeup from Scott (time trip: [1], [2], [3, [4]) this time on using NIO vs IO, partly as followup to JD Campbell's Top Java 5 EE Servers Compared.

Scott compares simple and complex cases: in a simple request/response case the overhead of the select() call implied by NIO is noticeable, but in the more complex case where you need to scale, NIO is the clear winner.