Add documentation regarding how Criterion does its measurements

[paulolieuthier]

This is really important for people writing papers about Haskell who use Criterion to perform measurements. I'm helping a team that is using a modified version of Criterion to measure the energy consuption of benchmarks (we plan to file a pull request when it's done).

It's mainly to explain the statistical reasoning behind the runBenchmark function from Criterion.Measurement.

[rrnewton]

That sounds great. If you have suggestions or a PR, they would be welcome. We're trying to make a new wave of changes to this library, so now is a good time.

At the very least we could link some of the good blog posts that explain the linear regression methodology.

One thing I would like to clarify to users is the difference between taking the "Mean" and taking the "OLS regression". I have found that people are often not clear on that and it's not self explanatory.

Likewise the "y" and "iters" labels are hard to remember and not super accessible.

[Blaisorblade]

This is really important for people writing papers about Haskell who use Criterion to perform measurements.

+💯

One thing I would like to clarify to users is the difference between taking the "Mean" and taking the "OLS regression". I have found that people are often not clear on that and it's not self explanatory.

Also + 💯 . I have some clue about linear regression, but that makes sense if I'm plotting a dependent variable against an independent variable—which are the variables here? The dependent one should be runtime, but what's the independent variable?

Never got a clue until today. #122 suggests it's the number of iterations in a "single" measurement, something I've never seen in benchmarking frameworks. (Also, the paper cited in #122 seems to question the approach).