Re: [RFC][PATCH 1/2] cpufreq: Add special-purpose fast-switching callback for drivers

[Rafael J. Wysocki]

On Wed, Dec 2, 2020 at 4:59 PM Doug Smythies <dsmythies@xxxxxxxxx> wrote:
>
> Hi Rafael,
>
> On 2020.11.30 10:37 Rafael J. Wysocki wrote:
>
> > First off, some cpufreq drivers (eg. intel_pstate) can pass hints
> > beyond the current target frequency to the hardware and there are no
> > provisions for doing that in the cpufreq framework.  In particular,
> > today the driver has to assume that it should allow the frequency to
>
> Forgot the important "not":

Right, thanks for noticing that!

> today the driver has to assume that it should allow not the frequency to
>
> > fall below the one requested by the governor (or the required capacity
> > may not be provided) which may not be the case and which may lead to
> > excessive energy usage in some scenarios.
> >
> > Second, the hints passed by these drivers to the hardware neeed not
>
> s/neeed/need

Yup, thanks!

> ...
>
> O.K. this is good.
>
> The problem with my basic CPU frequency verses load test with the
> schedutil governor is that it is always so oscillatory it is pretty
> much not possible to conclude anything. So I re-worked the test
> to look at Processor Package Power load instead.
>
> In a previous e-mail [1] I had reported the power differences
> for one periodic load at one frequency, as a (apparently cherry picked)
> example. Quoted:
>
> > schedutil governor:
> > acpi-cpufreq: good
> > intel_cpufreq hwp: bad    <<<<< Now good, with this patch set.

OK, great!

> > intel_cpufreq no hwp: good
> > ...
> > periodic workflow at 347 hertz.
> > ~36% load at 4.60 GHz (where hwp operates)
> > ~55% load at 3.2 GHz (where no hwp operates)
> >
> > intel_cpufreq hwp: 9.6 processor package watts. 45.8 watts on the mains to the computer.
> > intel_cpufreq no hwp: ~6 processor package watts. ~41 watts on the mains to the computer. (noisy)
>
> So this time, I only have power/energy data, and a relatively easy way to compress all 12,000
> samples into some concise summary is to simply find the average power for the entire experiment:
>
> Legend:
> hwp: Kernel 5.10-rc6, HWP enabled; intel_cpufreq; schedutil (always)
> rjw: Kernel 5.10-rc6 + this patch set, HWP enabled; intel_cpu-freq; schedutil
> no-hwp: Kernel 5.10-rc6, HWP disabled; intel_cpu-freq; schedutil
> acpi-cpufreq: Kernel 5.10-rc6, HWP disabled; acpi-cpufreq; schedutil
>
> load work/sleep frequency: 73 Hertz:
> hwp: Average: 12.00822 watts
> rjw: Average: 10.18089 watts
> no-hwp: Average: 10.21947 watts
> acpi-cpufreq: Average:  9.06585 watts
>
> load work/sleep frequency: 113 Hertz:
>
> hwp: Average: 12.01056
> rjw: Average: 10.12303
> no-hwp: Average: 10.08228
> acpi-cpufreq: Average:  9.02215
>
> load work/sleep frequency: 211 Hertz:
>
> hwp: Average: 12.16067
> rjw: Average: 10.24413
> no-hwp: Average: 10.12463
> acpi-cpufreq: Average:  9.19175
>
> load work/sleep frequency: 347 Hertz:
>
> hwp: Average: 12.34169
> rjw: Average: 10.79980
> no-hwp: Average: 10.57296
> acpi-cpufreq: Average:  9.84709
>
> load work/sleep frequency: 401 Hertz:
>
> hwp: Average: 12.42562
> rjw: Average: 11.12465
> no-hwp: Average: 11.24203
> acpi-cpufreq: Average: 10.78670
>
> [1] https://marc.info/?l=linux-pm&m=159769839401767&w=2
>
> My tests results graphs:
> Note: I have to code the web site, or I get hammered by bots.
> Note: it is .com only because it was less expensive than .org
> 73 Hertz:
> Double u double u double u dot smythies dot .com/~doug/linux/s18/hwp/k510-rc6/su73/
> 113 Hertz:
> Double u double u double u dot smythies dot .com/~doug/linux/s18/hwp/k510-rc6/su113/
> 211 Hertz:
> Double u double u double u dot smythies dot .com/~doug/linux/s18/hwp/k510-rc6/su211/
> 347 Hertz:
> Double u double u double u dot smythies dot .com/~doug/linux/s18/hwp/k510-rc6/su347/
> 401 Hertz:
> Double u double u double u dot smythies dot .com/~doug/linux/s18/hwp/k510-rc6/su401/

Thanks for the data, this is encouraging!