I think MeowChow's would be nearly as fast, except that for reasons of brevity it performs this crazy map operation on every base-pair triplet substitution.

On this train of thought, is there such as thing as a Benchmark-type routine that will test performance on a variety of data sizes? So many times people benchmark a variety of routines with only one set of data, which has the result of being a 1-dimensional test where there are actually 2 independent variables (function and data set size).

In line with Big-O Notation, is it possible and/or has someone written a Benchmark-type module which would estimate what kind of O(f(n)) function would best represent how the algorithm in question scales? Certainly not trivial by any means, but not impossible either.