anyone who has played around with getting performant IO in node will tell you to get really great performance you need to avoid unnecessary memory allocations and memory copying.
Working on pull-file I could get about 1gb/s read on a warm cache, but if I just reused the same buffer over and over I could get 2gb/s!
reusing memory is easy to do in a benchmark, but in an a real system when you read data in order to do something with it. maybe you write it to another file or socket, or maybe you encrypt it? maybe you pass it to a multiplexer which adds framing and then writes that to a socket?