>>Taking an average of wildly varying microbenchmark scores doesn't seem particularly useful.<< Correct. Which is why, unlike Rohit Joshi, the benchmarks game

>>Taking an average of wildly varying microbenchmark scores doesn't seem particularly useful.<< Correct. Which is why, unlike Rohit Joshi, the benchmarks game simply shows a broad median; and when a summary measure is required to order the box plots, uses the geometric mean: http://benchmarksgame.alioth.debian.org/u64q/which-programs-are-fastest.html http://benchmarksgame.alioth.debian.org/u64q/which-programs-... (Note the reminders -- "Please don't obsess over tiny differences in median values from such a small number of examples." -- "Please don't obsess about which programming language implementation is shown 10th and which is shown 11th. You can see that the order would be different if it was based on the median scores instead of the geometric mean scores.") >>as though any of its microbenchmarks measure anything other than "is this language capable of binding to libpcre" or "is this language capable of binding to libgmp"<< Incorrect. Everyone can see that there is only 1 regex task and only 1 arbitrary precision arithmetic task -- and they sensibly allow library use.