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

[darkfires]

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.

I agree completely. I don't know why he posted the damn article without the power figures, it's all meaningless. I prefer quality over quantity, otherwise I just wasted 3-4 mins of my time for nothing, again.

For what it's worth, I have to constantly disable intel_pstate on newer Xeon servers running centos 7 because it's totally broken, setting performance mode doesn't lock the cpufreq to max, it's all over the place, in fact it didn't matter what it was set to it does it's own thing anyway. I have NO clue who thought it worked well enough to make it the default.

[droste]

schedutil doesn't seem as a good option for now

My own experience is that overall the performance is the same with schedutil and as soon as there's real work going on it goes to max freq. but the power consumption on normal use ("measured" in heat -> noise of the fan) is waaaaay better with this one. Ondemand is switching either way to fast or way to slow to the max freq. while schedutil seems to take more steps in between (or doesn't go to max freq.) without slowing down everything too much. It is of course slower, because it is running some tasks with lower freq., but it does not feel like it.

Regarding cpufreq powersave:
The result is what is to be expected. This governor is the opposite of the performance govenor. Performance = max freq. all the time, powersave = min freq. all the time.
Both of these governors actually do nothing but set a fixed freq.

[jabl]

For what it's worth, I have to constantly disable intel_pstate on newer Xeon servers running centos 7 because it's totally broken, setting performance mode doesn't lock the cpufreq to max, it's all over the place, in fact it didn't matter what it was set to it does it's own thing anyway. I have NO clue who thought it worked well enough to make it the default.

intel_pstate with the performance governor is not supposed to lock the frequency to the max. With intel_pstate, both the performance and the powersave governor are more like the cpufreq ondemand governor, they just differ in how aggressively they ramp the frequency up and down.

[karolherbst]

- 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.

this is so wrong, it ain't funny anymore.

[stqn]

Here intel_pstate in powersave mode is (more or less) stuck to the turbo speed in idle, like the performance mode, anyway.

[sdack]

If you happen to have an Intel CPU you can make a script to measure CPU frequency every 10 milliseconds, run the benchmarks and post the CPU frequency charts to this forum.

You do it. I am looking forward to your data.

[starshipeleven]

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

When you are on battery and with shitty cooling systems, yes it matters. On a desktop, not so much.

[Inopia]

Michael It is very confusing when the options on graphs change order between each benchmark. Also these benchmarks don't make too much sense without metering the power usage.

[Michael]

Michael It is very confusing when the options on graphs change order between each benchmark. Also these benchmarks don't make too much sense without metering the power usage.

Patches welcome if you have a better idea for graph presentation...

As mentioned in the article, these tests were just some tests I did for raw performance as my lone power meter was busy on another system. Power measurements will come in other tests later on.

[sdack]

Patches welcome if you have a better idea for graph presentation...

As mentioned in the article, these tests were just some tests I did for raw performance as my lone power meter was busy on another system. Power measurements will come in other tests later on.

--- graphs.old
+++ graphs.new

- blue
+ /* Added extra colours to the bars! */
+ red
+ green
+ yellow
+ orange
+ blue
+ magenta
+ cyan

Here is a patch.