I'm running only one concurrent session in the benchmark, because of course if I do more they compete with each other and run up the numbers. The system has 512MB of memory and is lightly loaded; no swapping was going on as the benchmark ran. I believe I ran 50 retrievals; enough to avoid the warning about the sample being too small to be meaningful from ab, anyway.

I was about to go optimize a section of my code that turns out to have almost zero contribution to the total; another example of why one really does need to profile first!