Intel P-State Is Faster Than ACPI CPUFreq Performance On Linux 4.4

[phoronix]

Phoronix: Intel P-State Is Faster Than ACPI CPUFreq Performance On Linux 4.4

A request recently came in for doing some fresh benchmarks atop the brand new Linux 4.4 kernel while comparing the P-State and CPUFreq CPU scaling drivers and their different scaling governor options...

http://www.phoronix.com/scan.php?page=news_item&px=Linux-4.4-CPUFreq-P-State-Gov

[birdie]

Michael, you are clearly looking at an egregious bug (Intel P-State driver doesn't honor Powersave mode) and you're stating that this driver gives a higher performance. Really?

As for performance mode there's something wrong here again. In this mode both P-State and CPUFreq drivers should set to and keep the CPU at its highest possible performance, yet we see it's not the case. Throttling, maybe? Or new Intel CPUs do not really honor ACPI performance modes.

[milkylainen]

I agree that P-state seems to have improved the last couple of revisions to the point of actually not breaking just about every load you throw at it.
But this is hard to tell from only one machine setup. Intel P-state seems to be more sensitive to what CPU you are actually using according to my own experience. But I could be wrong. I have had a couple of machines where they would have erratic behavior at times using P-state. Never did any real performance measurements besides the "feel" you get when something is seriously wrong.
With 4.4 I shifted my desktop kernels to P-state. Seems to be working fine so far.

[milkylainen]

As for performance mode there's something wrong here again. In this mode both P-State and CPUFreq drivers should set to and keep the CPU at its highest possible performance, yet we see it's not the case. Throttling, maybe? Or new Intel CPUs do not really honor ACPI performance modes.

I think maybe this has something to do with turbo-states of newer cpus. Does cpufreq use turbo freqs? I have not seen working turbo states.
Or rather.. does cpufreq _not_ screw up a cpu that has scaled to turbo by resetting clock to "max" ACPI declaration?

[SuperIce97]

Michael, you are clearly looking at an egregious bug (Intel P-State driver doesn't honor Powersave mode) and you're stating that this driver gives a higher performance. Really?

You're misinterpreting what is meant by "powersave". In ACPI CPUFreq, it means limiting the CPU to minimum frequency at all times. In Intel P-State, it means making the CPU less sensitive to load changes. This will cause it to consume less and be useful for laptops while still generally providing great performance overall. It's just has a generally higher latency in ramping up the frequency and as you can see in a few of the tests and especially in FFMpeg which didn't really ramp up the speed correctly. Latency on powersave was a big issue with text editors as well, but they fixed that some time ago. If you want to reduce your frequency to minimum at all times, just edit the config files in /sys or use the pstate-frequency program (simply "pstate-frequency -Sp1" will set it to minimum speeds).

[hahahafr]

Where is the power consumption for each mode ? Article completely useless without this information.

[Michael]

Where is the power consumption for each mode ? Article completely useless without this information.

Did you read the article?

[Licaon]

"Current default governor for Intel since kernel 3.9 better than older alternatives. News at 11" :-\

[haagch]

or use the pstate-frequency program (simply "pstate-frequency -Sp1" will set it to minimum speeds).

But be careful: Default for pstate-frequency is to disable turbo boost. Makes a big difference on my i7 3632qm: 2.2 GHz vs. 3.2 GHz.
With current pstate-frequency git I'm running with "pstate-frequency -S -p performance"

[GreatEmerald]

Good to know that P-state indeed gives some nice gains over cpufreq. Of course, only the performance settings really matter in this case.