Tasty » Adam Sampson's confidence bookmarks

Publication: Quantifying Performance Changes with Effect Size Confidence Intervals - School of Computing - University of Kent edit / delete

Tech report with more details of the statistics behind Tomas/Richard's approach. In particular, this describes how to do the same thing in either parametric or non-parametric ways, and gives some description of how badly the parametric approach performs when the underlying data isn't normally distributed (not very badly, as it turns out).

to benchmarking confidence effect-size non-parametric performance statistics ... on 26 March 2014

Rigorous Benchmarking in Reasonable Time - Kent Academic Repository edit / delete

Tomas Kalibera and Richard Jones' paper on how to do benchmarking that's actually meaningful -- presenting results as confidence intervals for effect sizes, with techniques to establish i.i.d. results and work out how many repetitions you need to do. Very nice work for a pretty short paper! (I've spent most of today chasing references from this in the interests of understanding the maths behind it...)

to benchmarking compiler confidence effect-size independence java performance reproducibility statistics vm ... on 26 March 2014

Statistically rigorous Java performance evaluation edit / delete

One of the papers that inspired Tomas/Richard's rigorous benchmarking work. This is a much simpler strategy, involving looking for overlapping confidence intervals -- which is statistically pretty dubious, but common in other disciplines...

to benchmarking confidence java performance research statistics ... on 26 March 2014

If we're so different, why do we keep overlapping? When 1 plus 1 doesn't make 2 edit / delete

Why looking for overlapping confidence intervals isn't the right thing to do when comparing two distributions (contrary to some modern benchmarking advice) -- you can have overlapping but also have a significant difference.

to confidence statistics ... on 26 March 2014