Initial Benchmarks Of Schedutil Performance On Linux 5.7 Show Room Still For Improvement

[phoronix]

Phoronix: Initial Benchmarks Of Schedutil Performance On Linux 5.7 Show Room Still For Improvement

With Linux 5.7 the kernel is preparing to use the Schedutil governor more often on Intel systems. That change affects the CPUfreq default as well as the Intel P-State driver when in passive mode. While Schedutil holds a lot of hope, at least on Linux 5.7 with the testing I've done thus far the results show the raw performance slipping while testing on more platforms is forthcoming.

http://www.phoronix.com/vr.php?view=29095

[Mario Junior]

So, what is the purpose of Schedutil? I don't see any benefits. 🤔🤷🏻♂️

[kcrudup]

So, what is the purpose of Schedutil? I don't see any benefits

I was about to ask that, too, and was heading for my kernel tree to put it back to "powersave", then I'd read to the end:

When looking at the CPU package power consumption, using Schedutil had a lower power draw than even powersave while often leading to better results than powersave. Schedutil led to 15% on average lower CPU power consumption than the performance governor or 8% for this particular CPU setup.

So instead I'd modified my "tlp" config to use "schedutil" only when on battery and "powersave" when on AC.

[polarathene]

It's great that we got the power savings and temperature results at the end, missing the perf/watt ratio that other articles in the past have provided though, although that may provide more useful insights at higher granularity where we see notable differences on some benchmarks(eg compared to performance, how much watt difference for the added performance).

Thanks Michael for the article!

So, what is the purpose of Schedutil? I don't see any benefits. 🤔🤷🏻♂️

I think it was meant to be power efficient like powersave but able to put out better performance on demand while keeping temperatures in check? For intel it's a bit different since powersave doesn't keep the frequency locked. On a desktop you'd generally just want performance and be done with it.

On my laptop I use powersave. Performance can naturally be faster but the fans kick in much sooner/frequently and rev up like jet engine sounds, plus I'm generally ok with slower performance on the laptop if the battery life will be much better. It seems you'll get more battery life via schedutil without too much of a drawback to powersave most of the time(remember these are benchmark results on particular workloads under stress, not likely to represent a regular user experience with the extremes).

[smitty3268]

I think it was meant to be power efficient like powersave but able to put out better performance on demand while keeping temperatures in check?

I believe theoretically it's meant to be the best of both worlds. The highest performance + better power usage. Obviously in practice it's not there yet.

[polarathene]

I believe theoretically it's meant to be the best of both worlds. The highest performance + better power usage. Obviously in practice it's not there yet.

I wouldn't expect highest performance at the power cost of powersave. Higher performance draws more power and raises the temps needing the fan to come on. Needs to strike a balance somewhere.. but I don't see how it'd compete at both extremes? Short term high performance bursts perhaps?

[smitty3268]

I wouldn't expect highest performance at the power cost of powersave. Higher performance draws more power and raises the temps needing the fan to come on. Needs to strike a balance somewhere.. but I don't see how it'd compete at both extremes? Short term high performance bursts perhaps?

It's meant to behave something like ondemand. That is, it provides high performance under use, but scales down when not.

The primary difference is that it's driven by information from the scheduler (CFS) as its setting up tasks to run rather than by setting timers that poll the current load every so often, and it should be smarter about being able to ramp up to full speed much faster. The hope is that it will be fast enough at changing frequencies that it can match performance, and completely outclass ondemand/conservative while also reducing frequencies fast enough that you get low power use compared to those 2 options.

I'm not really sure how it uses less power than powersave does here, but I know the kernel devs have said in the past that they hope schedutil could replace all the existing governors so I assume it's expected to compete with it powerwise too - or else it will have some kind of flag to configure it to prefer power over performance.

[polarathene]

I'm not really sure how it uses less power than powersave does here, but I know the kernel devs have said in the past that they hope schedutil could replace all the existing governors so I assume it's expected to compete with it powerwise too - or else it will have some kind of flag to configure it to prefer power over performance.

Note that these benchmarks are showing comparison to Intel Performance and Powersave profiles, these are not the same as non-pstate options with the same names. Eg Powersave isn't locked at the lowest frequency, it can ramp up but on my laptop, the fans may come on under load but never as loud as with performance mode, frequency does get ramped up when that is happening but it's not performing at it's peak, there is still limits.

So I imagine Schedutil is similar to that as the benchmark results are showing except for a few cases, which as you point out is probably because pstate is driven by internal data Intel uses to monitor load, while Schedutil is being driven/influenced from scheduling data from the kernel.

[aphysically]

Does this system have HWP? It looks like schedutil only became the default for systems without HWP.

It would be interesting to see this benchmark with HWP disabled as well to compare (intel_pstate=no_hwp)

I also seem to not even get schedutil as an option on my pstate driver so I'm probably still confused about the changes here.

Is it that it's not even possible to run schedutil with pstate without switching to passive mode?

[Caturba]

Is it that it's not even possible to run schedutil with pstate without switching to passive mode?

Yes, it is true. When intel_pstate is active, it takes over cpufreq and you can't use any governor, not only schedutil. Performance vs powersave choices when intel_pstate is active are in fact intel_pstate's algorithms and have nothing to do with cpufreq. See: https://01.org/linuxgraphics/gfx-doc...el_pstate.html

For me personally the benefit of abandoning intel_pstate is that the latter is not controllable. When it is active performance vs powersave is the only knob you have. Everything else has been decided by Intel. On my system, for example, it seems like it limits the package temperature to 65°C, though nobody asked it to. As the result I can't achieve in Linux the performance available on the same machine in Windows. Well, unless I disable intel_pstate or set it to passive of course.