You know jack squat about statistics. That's basically how I summarize your statement. You don't have a mean, median, mode, or standard deviation in your "10% 40%". You don't understand confounding and want to test the performance of poll vs. epoll on file descriptors by testing an entire server. You think science is to use available data, when it's nothing of the sort, and in fact the way you remove confounding is to remove data.

Basically, you're a FUD slinger who's got a beef and doesn't know enough to hold his own in a real discussion about measurement and analysis.