On 2018.12.13 04:42 Paul Menzel wrote:
> On 12/13/18 11:39, Rafael J. Wysocki wrote:
>> On Thu, Dec 13, 2018 at 10:54 AM Paul Menzel <pmen...@molgen.mpg.de> wrote:
>>> On 12/13/18 00:06, Doug Smythies wrote:
>>>> On 2018.12.12 13:40 Paul Menzel wrote:
>>>>
>>>>> Using *powersave* as P-state selection algorithm, on an idle system
>>>>
>>>> Define "idle system".
>>>> If your computer is running a GUI, or is even a server without a GUI
>>>> but with many services running, then "idle" really isn't.
>>>> Below is from my test server, with many services disabled, so
>>>> "idle" really is quite "idle"
>>>>
>>>> doug@s15:~/temp$ sudo turbostat --Summary --quiet --show
>>>> Busy%,Bzy_MHz,PkgTmp,PkgWatt --interval 15
>>>> Busy% Bzy_MHz PkgTmp PkgWatt
>>>> 0.01 1608 27 3.71
>>>> 0.01 1619 27 3.71
>>>> 0.01 1600 28 3.71
>>>> 0.01 1600 28 3.70
>>>>
>>>> Note that p state 16 (1600 MHz) is the minimum for my older i7-2600k
>>>> processor.
>>>
>>> The thing is, on an Intel Kaby Lake laptop with Ubuntu 18.10 and GNOME
>>> running, it goes down to the lowest listed frequency.
>
> Checking the numbers again, I was mistaken. The lowest possible frequency
> of the Intel Kaby Lake i7-7500U in that laptop is 400 MHz, and it is
> going down to 600 MHz. Busy% from turbostat is 0.3 to 0.4.
>
>> Kaby Lake has hardware-managed P-states (HWP) which is a different mechanism.
>
> Isn’t HWP also available for the 6th generation?
>
> $ dmesg | grep intel_pstate
> [ 2.092456] intel_pstate: Intel P-state driver initializing
> [ 2.094820] intel_pstate: HWP enabled
Yes, you have HWP. I do not have HWP. I don't think HWP or not is
particularly relevant to this discussion.
>>>>> Shouldn’t it go down until 800 MHz?
>>>>
>>>> We would need some actual busy information, turbostat is the
>>>> recommended tool, to know for sure.
>>>
>>> Here you go.
>>>
>>> ```
>>> tools/power/x86/turbostat> sudo ./turbostat --Summary --quiet --show
>>> Busy%,Bzy_MHz,PkgTmp,PkgWatt --interval 15
>>> Busy% Bzy_MHz PkgTmp PkgWatt
>>> 3.59 1167 31 1.68
>>> 3.21 903 31 1.34
>>> 3.21 906 31 1.34
>>> 3.27 901 31 1.35
>>> 8.23 2715 30 2.32 ← stopping GDM (systemctl stop gdm)
>>> 2.95 915 30 1.18
>>> 2.91 906 30 1.18
>>> 2.92 903 30 1.17
>>> 2.90 900 29 1.17
>>> 2.89 903 29 1.18
>>> 2.91 903 30 1.18
>>> 2.89 903 29 1.18
>>> 2.89 900 29 1.18
>>> 2.90 903 30 1.18
>>> 2.90 903 29 1.17
>>> 2.90 903 29 1.17
>>> 2.90 900 29 1.16
>>> 2.90 903 29 1.14
>>> 2.90 903 28 1.11
>>> 2.90 903 29 1.10
>>> 2.91 900 29 1.16
>>> 2.91 903 29 1.14
>>> 2.90 903 29 1.12
>>> 2.90 903 29 1.16
>>> 2.90 900 28 1.17
>>> 2.92 903 29 1.16
>>> 2.90 903 29 1.16
>>> 2.90 903 29 1.16
>>> ```
>>>
>>> 800 MHz should be enough to keep GDM running, shouldn’t it?
>>
>> Well, depending.
>>
>>> Otherwise only SSH was running.
>>
>> There obviously is something that causes it to stay at 900 MHz.
Agreed. Clearly some more stuff was using some CPU.
The resulting requested CPU frequency is also highly dependent on
the frequency of the load. Example:
doug@s15:~/csv-plot$ sudo turbostat --Summary --quiet --show
Busy%,Bzy_MHz,PkgTmp,PkgWatt --interval 60
Busy% Bzy_MHz PkgTmp PkgWatt
0.08 2567 27 3.77 <<< Note the high active CPU frequency with
virtually no load.
0.01 1600 26 3.69
0.01 1600 27 3.70
0.83 1600 30 4.18
1.80 1600 30 4.59 <<< This load is at 435 Hertz work / sleep
frequency
1.80 1600 32 4.59
0.74 3033 27 4.49 <<< This load is at 1 Hertz work / sleep frequency
0.76 3380 27 4.63
0.76 3384 28 4.62
0.38 3356 28 4.16
0.01 1615 28 3.71
0.01 1609 27 3.70
>
> It’s not obvious to me, but you have more experience. It’d expect
> to at least one core(?) to go down to 800 MHz and not all to stay
> at 900 MHz.
There is only PLL (Phase Locked Loop). The highest requested vote
for CPU frequency wins. CPUs in deep idle do not get a vote.
>> Please check max_perf_pct, min_perf_pct and num_pstates under
>> /sys/devices/system/cpu/intel_pstate/ .
>
> /sys/devices/system/cpu/intel_pstate> cat max_perf_pct min_perf_pct
> num_pstates
> 100
> 20
> 33
>
>> Also cpuinfo_max_freq, cpuinfo_min_freq, scaling_max_freq,
>> scaling_min_freq under /sys/devices/system/cpu/cpufreq/policy0/ .
>
> /sys/devices/system/cpu/cpufreq/policy0> cat cpuinfo_{min,max}_freq
> scaling_{min,max}_freq
> 800000
> 4000000
> 800000
> 4000000
>
>> However, please note that Busy% of 3 isn't particularly low.
>
> Indeed. On the laptop it is around 0.3 to 0.4 even with GNOME
> running.
>
> So, to check if everything is working, I boot into initramfs
> and check the numbers there?
>
Myself, I think there is nothing wrong here. If you want
to pursue further, then my suggestion would be to disable
HWP and use trace and the intel_pstate_tracer.py utility
to fully analyze your "idle" system load and your systems
response to it.
... Doug
