On Mon, Jul 17, 2017 at 12:48:38PM -0700, Arjan van de Ven wrote: > On 7/17/2017 12:23 PM, Peter Zijlstra wrote: > > Of course, this all assumes a Gaussian distribution to begin with, if we > > get bimodal (or worse) distributions we can still get it wrong. To fix > > that, we'd need to do something better than what we currently have. > > > > fwiw some time ago I made a chart for predicted vs actual so you can sort > of judge the distribution of things visually > > http://git.fenrus.org/tmp/linux2.png
That shows we get it wrong a lot of times (about 50%, as per the average) and moving the line has benefit. Since for performance you really don't want to pick the deeper idle state, so you want to bias your pick towards a shallower state. Using the CDF approach you can quantify by how much you want it moved.