Re: Plain DFS (no voltage scaling)

[Viresh Kumar <viresh.kumar@linaro.org>]

On 03-02-16, 19:19, Mason wrote:
> On 03/02/2016 17:13, Viresh Kumar wrote:
> > On 03-02-16, 16:07, Mason wrote:
> >> On 02/02/2016 22:11, Mason wrote:
> >>
> >>> I plan to enable the on-demand governor on the tango platform:
> >>> https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/tree/arch/arm/boot/dts/tango4-smp8758.dtsi
> >>>
> >>> I found the cpufreq-dt binding doc:
> >>>
> >>> https://www.kernel.org/doc/Documentation/devicetree/bindings/cpufreq/cpufreq-dt.txt
> >>> https://www.kernel.org/doc/Documentation/devicetree/bindings/opp/opp.txt
> >>>
> >>> Something is not clear to me:
> >>>
> >>> If my platform cannot scale the voltage, what information
> >>> should I put in the voltage part of the DT?
> >>
> >> Someone pointed out that tweaking the frequency without tweaking
> >> the voltage might be counter-productive.
> >>
> >> I measured the power consumption of the entire board (at the power outlet)
> >> for 3 CPU frequencies (all other things being equal, I hope).
> >>
> >> idle @ 111 MHz = 4.6 W
> >> idle @ 333 MHz = 4.6 W
> >> idle @ 999 MHz = 4.6 W
> >>
> >> load @ 111 MHz = 5.0 W
> >> load @ 333 MHz = 5.7 W
> >> load @ 999 MHz = 7.7 W
> >>
> >> When idle, the kernel calls WFI, which "turns off" most of the CPU
> >> (clock gating?) such that the actual frequency does not matter.
> >>
> >> At full load (I use cpuburn to jog as many FUs simultaneously as
> >> possible) it looks like each additional MHz requires ~3 mW.
> >>
> >> So it would appear that an on-demand governor might not help to
> >> save power.

IOW, what's so special about how on-demand will behave in your case ?
If it doesn't work properly, then that's the same for everybody, isn't
it ?

> > Why do you say so ?
> 
> Here's my (possibly incorrect) reasoning:
> 
> If the CPU is idle, kernel calls WFI and frequency apparently
> doesn't matter.

Right.

> If the CPU has work to do, the on-demand governor will bump the
> frequency to the max (I think).

Not always, but it will choose the best frequency based on the load.

> I don't think the CPU spends a lot of time in the intermediate
> frequencies (neither max nor min).

It does, it was fixed somewhere in v3.* .

> But maybe my logic is flawed?

Yes.

> Also, some users (of our older kernel) reported problems were
> they considered the on-demand governor was "too slow" to ramp
> the frequency up. (But I don't know if they'd played with the
> configuration knobs.)

Shouldn't be the case anymore, or maybe those guys are more in favor
of the interactive governor :)

-- 
viresh