Linux 4.8 Intel P-State vs. CPUFreq Scaling Driver/Governor Benchmarks

[phoronix]

Phoronix: Linux 4.8 Intel P-State vs. CPUFreq Scaling Driver/Governor Benchmarks

Given the underlying work that's been happening in the CPUFreq/scheduler area and the introduce of the new Schedutil CPUFreq governor, I decided to run some fresh performance benchmarks of P-State and CPUFreq with the different governor options when testing from a Linux 4.8 Git kernel atop the current Fedora 25 development packages and using a Core i5 Skylake processor.

http://www.phoronix.com/vr.php?view=23470

[arekm]

Doh, I would expect power usage stats for these.

[davidbepo]

schedutil doesn't seem as a good option for now

[Michael]

Doh, I would expect power usage stats for these.

As said in this article, my WattsUp Pro was busy on another system.

[sdack]

Not only would a power measurement be helpful, but this sort of pure benchmarking is going vertical, meaning, it doesn't run alongside the development and purpose of these schedulers, but it runs its own course and makes it less meaningful. Next to the question "How much power is saved?" should everyone also be asking "Why isn't the cpufreq performance governor the fastest in all tests?". If such questions don't get asked and answered, then nothing will stop people from asking "Are these benchmarks accurate?".

So how come we are seeing these difference and how accurate are the results really? I cannot believe that the ondemand scheduler is beating the performance scheduler.

I use the ondemand governor with modified settings (50% threshold with 10x down sampling factor), to get more performance from a desktop machine, while still saving a good amount of power. But such considerations make little sense with this kind of benchmarking.

[sdack]

- The ondemand scheduler cannot beat the performance scheduler if max CPU frequency is fixed because of overclocking.

- With boost frequencies enabled, the ondemand scheduler can beat the performance scheduler because the performance scheduler never switches to the boost frequency.

- With boost frequencies disabled, the ondemand scheduler on AMD CPUs slows down the system by more % than the ondemand scheduler on Intel CPUs.

- On a notebook with an Intel Bay Trail CPU, the CPU is adjusting its frequency even if the performance scheduler is selected

A German saying I've learned when I was a young electrician said, "Wer viel misst, misst Mist." It was meant as a warning for us students to make an informed decision over what we were trying to measure and not only to measure anything and everything, because when one doesn't know what exactly it is one is measuring then one's measurement has no worth regardless of its precision.

So how much worth do you think these benchmarks are?

[dungeon]

- On a notebook with an Intel Bay Trail CPU, the CPU is adjusting its frequency even if the performance scheduler is selected

Does that even eat any power, i mean Bay Trail is 10 W max... it might save exactly as much as you lose on perf, probably around 2W just there

[MaxToTheMax]

Seems like the "bad" performance of powersave is actually good: that means it's actually saving power by clocking the CPUs down.

[geearf]

Seems like the "bad" performance of powersave is actually good: that means it's actually saving power by clocking the CPUs down.

Not necessarily.
Paraphrasing what one Intel dev wrote a while back, if in powersave the CPU uses 50% of the power used in performance but takes 400% the time to finish its job, it's actually using more power.
To know the actual results here we would need the power data.

[sdack]

Nobody has disproved the numbers in the article yet. ...

You miss the point. The numbers prove little to begin, which is what makes them worthless. You want to use something worthless to prove or disprove something? Or use it to make a random speculation? You don't need Phoronix for this. A pair of dice will do.