The other thing that caught my eye was an illustration of how on modern processors, optimization can be extremely nonobvious. The example was a subproblem I've solved before: Given a value x between 0 and 1 and an array of ascending values in the same range, figure out which pair of values x falls between. I used a binary search. It turns out that on the x86, binary search is predicted poorly by the processor's branch prediction logic, while in a linear search every compare but the last one is predicted correctly. A mispredicted branch is very expensive. So, for a small number of bins a linear search is 30-50% faster (!), and the problem size has to reach 64 bins before the improved asymptotic efficiency of the binary search actually beats the linear search in real life. My problem had at most a few tens of bins, and the search was deep in an inner loop, so this would have been useful information to have known.