News for hackers. Stuff that matters.
Re:What is BPF and why is it taking over Linux Performance A... by static_noise on Sunday August 21, 2016 @06:02AM (#12330233)
Replace the average (mean) with the median and you will have better results in almost all cases.

When the data is good the result ist virtually identical. When there are outliers, the median just ignores them unless there lots of them. When the data is that bad, that the median doesn't work anymore you need to up your statistics quite a few notches.

Re:What is BPF and why is it taking over Linux Performance A... by gnufx on Sunday August 21, 2016 @02:44PM (#12332141)
Replacing a single number with a different one hardly helps. The distribution matters, specifically the outliers in performance measurements.
Re:What is BPF and why is it taking over Linux Performance A... by SFJulie on Sunday August 21, 2016 @12:37PM (#12331501)
Well the median of measure that are only aggregated by average (RRD, carbon ...) is ... biased.

Hence the reason I made my own backend for making very high precision measurement using a percentile filter (10%, 50%, 90% and average for the fun).

It was fun to watch linux «monotonic» clock not being monotonic.

Re:What is BPF and why is it taking over Linux Performance A... by omginternets on Sunday August 21, 2016 @01:12PM (#12331678)
In psychophysics we either use the median for reaction times (Gamma distribution) or we normalize with something like the logarithm.

I think some of the techniques might be applicable to system performance analysis, so I submit the following introductory publication, in case any are interested [0].

[0] http://opensiuc.lib.siu.edu/cgi/viewcontent.cgi?article=1077...