[opensuse-factory] Default CPU performance governor (Intel)
So I’ve noticed that the default CPU performance governor on Intel CPUs is still “powersave”. This causes the CPU to stay in the lowest frequency tier most of the time, and effectively disables the Turbo frequency range for “hot” workloads on at least Haswell and newer CPUs, which is probably not what most people want. It (very) noticeably decreases performance and desktop responsiveness, especially on laptops. Probably also leads to bad benchmark scores for openSUSE on a certain hobby journalist’s website. And it does not even save power! At least not according to my experience with a laptop (ASUS PU551LD, i5-4210U, SSD, replaced display with higher power consumption than the stock model). Idle power consumption with both “powersave” and “performance” governor was exactly the same at slightly less than 5 W according to powertop. No surprise there, really. I do not have a battery plot for actual use (light work and some browsing), but in 4 tries with each governor it actually seems like “powersave” performs worse by half an hour on this hardware. Could something be done about these defaults? Should I spend some time trying to get actually useful data to support my claims? If so, I’d appreciate a little assistance in figuring out the best procedure for this. Benchmarks are easy to screw up, and I’m sure there are people here with more experience in testing such things.
29.03.2018 17:15, Martin Herkt пишет:
So I’ve noticed that the default CPU performance governor on Intel CPUs is still “powersave”. This causes the CPU to stay in the lowest frequency tier most of the time, and effectively disables the Turbo frequency range for “hot” workloads on at least Haswell and newer CPUs, which is probably not what most people want.
It (very) noticeably decreases performance and desktop responsiveness, especially on laptops. Probably also leads to bad benchmark scores for openSUSE on a certain hobby journalist’s website.
And it does not even save power! At least not according to my experience with a laptop (ASUS PU551LD, i5-4210U, SSD, replaced display with higher power
This is Haswell CPU for which intel_pstate is likely used, so "powersave" is misleading. What /sys/devices/system/cpu/cpufreq/*/scaling_driver says?
consumption than the stock model). Idle power consumption with both “powersave” and “performance” governor was exactly the same at slightly less than 5 W according to powertop. No surprise there, really. I do not have a battery plot for actual use (light work and some browsing), but in 4 tries with each governor it actually seems like “powersave” performs worse by half an hour on this hardware.
Could something be done about these defaults? Should I spend some time trying to get actually useful data to support my claims? If so, I’d appreciate a little assistance in figuring out the best procedure for this. Benchmarks are easy to screw up, and I’m sure there are people here with more experience in testing such things.
On Thursday 2018-03-29 16:15, Martin Herkt wrote:
So I’ve noticed that the default CPU performance governor on Intel CPUs is still “powersave”. This causes the CPU to stay in the lowest frequency tier most of the time,
Well, no. The classic "powersave" governor kept the CPU in the lowest frequency *at all times* without exception. What you describe is the "ondemand" governor. Except that... intel_pstate does not even use classic governors. https://plus.google.com/+ArjanvandeVen/posts/dLn9T4ehywL https://lwn.net/Articles/536017/ As to what the strangely named file /sys/devices/system/cpu/cpufreq/policy0/scaling_governor is about, is this: if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) min = max; So setting min_perf_pct and max_perf_pct yourself whilst staying on the "powersave" policy has the same effect. -- To unsubscribe, e-mail: opensuse-factory+unsubscribe@opensuse.org To contact the owner, e-mail: opensuse-factory+owner@opensuse.org
On Thursday, March 29, 2018 6:43:51 PM CEST Jan Engelhardt wrote:
On Thursday 2018-03-29 16:15, Martin Herkt wrote:
So I’ve noticed that the default CPU performance governor on Intel CPUs is still “powersave”. This causes the CPU to stay in the lowest frequency tier most of the time,
Well, no. The classic "powersave" governor kept the CPU in the lowest frequency *at all times* without exception. What you describe is the "ondemand" governor.
Except that... intel_pstate does not even use classic governors.
https://plus.google.com/+ArjanvandeVen/posts/dLn9T4ehywL https://lwn.net/Articles/536017/
As to what the strangely named file /sys/devices/system/cpu/cpufreq/policy0/scaling_governor is about, is this:
if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) min = max;
So setting min_perf_pct and max_perf_pct yourself whilst staying on the "powersave" policy has the same effect.
Ah, right, I was referring to what “cpupower frequency-info” would print, so that’s a little misleading :)
Am 29.03.2018 um 16:15 schrieb Martin Herkt:
So I’ve noticed that the default CPU performance governor on Intel CPUs is still “powersave”.
No, it's not. Just try it on an older CPU and you will find that "ondemand" is the default. Unless you have a CPU where ondemand does not make sense, then "powersave" is the fallback. Which is correct IMHO.
This causes the CPU to stay in the lowest frequency tier most of the time,
That's also wrong.
and effectively disables the Turbo frequency range for “hot” workloads on at least Haswell and newer CPUs, which is probably not what most people want.
Haswell and newer CPUs use the "intel_pstate" driver, which does not really have a concept of "frequencies". Instead it uses "performance states" (P-States) which involve much more than just switching the cpu core frequency. The "performance" or "powersave" governor on this driver merely does hint the driver and CPU on what the "performance bias" should be. The CPU is still using all its available "frequencies" (P-States to phrase it correctly), it's just a bit more conservative in jumping to higher states.
It (very) noticeably decreases performance and desktop responsiveness, especially on laptops.
Show me the numbers. Of real life workloads.
Probably also leads to bad benchmark scores for openSUSE on a certain hobby journalist’s website.
Probably not. Show me the numbers. I have recently rebutted a SUSE Consultants demand (!) for us to switch off power management completely in order for him to be able to support the product with numbers. Power management OFF (on BIOS/Firmware level) vs "performance" governor: => no measurable difference in performance, but 500W idle power consumption vs. 100W idle power consumption "performance" governor vs "powersave" governor: => low 1-figure percent (1-3%) performance impact, still hard to measure, in extreme, non-realistic benchmarks (extremely short, high througput network bursts). Overall power consumption under mixed load was about 10% lower with "powersave" governor compared with "performance" governor (but this is hard to measure as we had to rely on firmware power consumption readings). In the end, we configured the machines with the "performance" governor, but *not* because of real cost/performance calculations, but just because managers want to hear that you have done everything for maximum performance (in the beginning they, supported by this SUSE consultant, demanded that we switch off power management completely in the BIOS!)
And it does not even save power! At least not according to my experience with a laptop (ASUS PU551LD, i5-4210U, SSD, replaced display with higher power consumption than the stock model). Idle power consumption with both “powersave” and “performance” governor was exactly the same at slightly less than 5 W according to powertop.
Of course idle power consumption is the same. As the governor will only influence the "scheduling" of the CPU power state transitions.
No surprise there, really. I do not have a battery plot for actual use (light work and some browsing), but in 4 tries with each governor it actually seems like “powersave” performs worse by half an hour on this hardware.
Could something be done about these defaults? Should I spend some time trying to get actually useful data to support my claims?
Of course. Because changing because it "feels bad" to you is certainly not a good idea. -- Stefan Seyfried "For a successful technology, reality must take precedence over public relations, for nature cannot be fooled." -- Richard Feynman -- To unsubscribe, e-mail: opensuse-factory+unsubscribe@opensuse.org To contact the owner, e-mail: opensuse-factory+owner@opensuse.org
participants (4)
-
Andrei Borzenkov -
Jan Engelhardt -
Martin Herkt -
Stefan Seyfried