Before tackling more serious performance changes, I would like a wider array of benchmarks, each focused on either a realistic scenario or a particular way we could stress the algorithm.