> Doing 10 runs of each and averaging the reported duration I wish people would stop doing that. Your program is doing the same work in all benchmarking runs;

> Doing 10 runs of each and averaging the reported duration I wish people would stop doing that. Your program is doing the same work in all benchmarking runs; any variance in the runtime is a result of factors you're probably not interested in measuring, like kernel scheduling nuances or what CPUs your benchmark happened to run on. You should take the min of microbenchmark duration, not the average.