Intel's P-State Driver Does Seem To Be In Bad Shape On Linux 4.6~4.7

A few days ago when delivering benchmarks of the new CPUFreq "Schedutil" governor in Linux 4.7 the P-State comparison results on this Git kernel looked particularly terrible. I've since done some P-State tests on the same system using the Linux 4.5 and 4.6 kernels that further point towards a regression having taken place.

At least for this Intel Xeon E5-268W v3 system, the P-State CPU frequency scaling driver has become a lot worse since Linux 4.5 while the recent CPUFreq Linux 4.7 results show no issues there -- thus making it look like it's a problem with P-State on recent kernels itself as opposed to a regression elsewhere.

I ran 4.5, 4.6, and 4.7 kernel benchmarks with the P-State powersave and performance governors on this same system.

4.7 is surely much slower than 4.6~4.5 while the performance governor showed it regressing in 4.6 rather than 4.7.

The performance is surely a lot slower for this Xeon E5 system on Linux 4.6~4.7 with P-State. Again, CPUFreq on Linux 4.7 delivers the more optimal results -- in line with P-State from Linux 4.5.


You can dig through more of these P-State Linux 4.5/4.6/4.7 results via this OpenBenchmarking.org result file. I don't have any more information at this time nor have I had a chance yet to do similar tests yet on other Intel hardware to see how isolated or not this apparent P-State regression is besides on this Xeon E5 system. If there's enough interest, I'll fire up Phoromatic and tell it to auto-bisect this performance regression.

Stay tuned.