Re: [PATCH V2 0/7] CPU cooling device new strategies

From: Daniel Thompson
Date: Thu Mar 08 2018 - 07:04:01 EST


On Wed, Mar 07, 2018 at 07:57:17PM +0100, Daniel Lezcano wrote:
> >> The preliminary benchmarks show the following changes:
> >>
> >> On the hikey6220, dhrystone shows a throughtput increase of 40% for an
> >> increase of the latency of 16% while sysbench shows a latency increase
> >> of 5%.
> >
> > I don't follow these numbers. Throughput increase while injecting idle?
> > compared to what? percentages of what? Please be more specific to better
> > describer your work..
>
> The dhrystone throughput is based on the virtual timer, when we are
> running, it is at max opp, so the throughput increases. But regarding
> the real time, it takes obviously more time to achieve as we are
> artificially inserting idle cycles. With the cpufreq governor, we run at
> a lower opp, so the throughput is less for dhrystone but it takes less
> time to achieve.
>
> Percentages are comparing cpufreq vs cpuidle cooling devices. I will
> take care of presenting the results in a more clear way in the next version.

I think we should also note that the current hikey settings for cpufreq
are very badly tuned for this platform. It has a single temp threshold
and it jumps from max freq to min freq.

IIRC Leo's work on Hikey thermals correctly it would be much better if
it used the power-allocator thermal governor or if if copied some of
the Samsung 32-bit platform by configuring the step governor with a
graduated with a slightly lower threshold that moves two stops back in
the OPP table (which is still fairly high clock speed... but it
thermally sustainable).


Daniel.


> >> Initially, the first version provided also the cpuidle + cpufreq combo
> >> cooling device but regarding the latest comments, there is a misfit with
> >> how the cpufreq cooling device and suspend/resume/cpu hotplug/module
> >> loading|unloading behave together while the combo cooling device was
> >> designed assuming the cpufreq cooling device was always there. This
> >> dynamic is investigated and the combo cooling device will be posted
> >> separetely after this series gets merged.
> >
> > Yeah, this is one of the confusing parts. Could you please
> > remind us of the limitations here? Why can't we enable CPUfreq
> > on higher trip points and CPUidle on lower trip points, for example?
>
> Sorry, I'm not getting the question. We don't want to enable cpuidle or
> cpufreq at certain point but combine the cooling effect of both in order
> to get the best tradeoff power / performance.
>
> Let me give an example with a simple SoC - one core.
>
> Let's say we have 4 OPPs and a core-sleep idle state. Respectively, the
> OPPs consume 100mW, 500mW, 2W, 4W. Now the CPU is in an intensive work
> running at the highest OPP, thus consuming 4W. The temperature increases
> and reaches 75ÂC which is the mitigation point and where the sustainable
> power is 1.7W.
>
> - With the cpufreq cooling device, we can't have 4W, so we go back and
> forth between 2W and 500mW.
>
> - With the cpuidle cooling device, we are at the highest OPP (there is
> no cpufreq driver) and we insert 47.5% of idle duration
>
> - With the combo cooling device, we compute the round-up OPP (here 2W)
> and we insert idle cycles for the remaining power to reach the
> sustainable power, so 15%.
>
> With the combo, we increase the performances for the same requested
> power. There is no yet the state2power callbacks but we expect the
> combination of dropping the static leakage and the higher OPP to give
> better results in terms of performance and mitigation on energy eager
> CPUs like the recent big ARM cpus with the IPA governor.
>
> Going straight to the point of your question above, we can see the
> cpufreq cooling device and the cpuidle cooling device have to
> collaborate. If we unregister the cpufreq device, we have to do the math
> for the power again in the combo cooling device. It is not a problem by
> itself but needs an extra reflexion in the current code.
>
> > Specially from a system design point of view, the system engineer
> > typically would benefit of idle injections to achieve overall
> > average CPU frequencies in a more granular fashion, for example,
> > achieving performance vs. cooling between available real
> > frequencies, avoiding real switches.
> >
> > Also, there is a major design question here. After Linaro's attempt
> > to send a cpufreq+cpuidle cooling device(s), there was an attempt
> > to generalize and extend intel powerclamp driver.
>
> I'm not aware of such attempt.
>
> > Do you mind
> > explaining why refactoring intel powerclamp is not possible? Basic
> > idea is the same, no?
>
> Basically the idea is the same: run synchronized idle thread and call
> play_idle(). That is all. Putting apart the intel_powerclamp is very x86
> centric and contains a plethora of code not fitting our purpose, it
> increases the idle duration while we are increasing the number of idle
> cycles but keep the idle duration constant in order to have a control on
> the latency for the user interactivity. If you compare the idle
> injection threads codes (powerclamp and cpuidle cooling device), you
> will also notice they are very different in terms of implementation.
>
> The combo cooling device collaborates with the cpufreq cooling device
> and reuses the DT binding, and finally it uses the power information
> provided in the DT. The idle injection is a brick to the combo cooling
> device.
>
> Initially I thought we should refactor the intel_powerclamp but it
> appears the combo cooling device reuses the cpufreq and cpuidle cooling
> device, making sense to have them all in a single file and evolving to a
> single cooling device with different strategies.
>
>
>
> >> Daniel Lezcano (7):
> >> thermal/drivers/cpu_cooling: Fixup the header and copyright
> >> thermal/drivers/cpu_cooling: Add Software Package Data Exchange (SPDX)
> >> thermal/drivers/cpu_cooling: Remove pointless field
> >> thermal/drivers/Kconfig: Convert the CPU cooling device to a choice
> >> thermal/drivers/cpu_cooling: Add idle cooling device documentation
> >> thermal/drivers/cpu_cooling: Introduce the cpu idle cooling driver
> >> cpuidle/drivers/cpuidle-arm: Register the cooling device
> >>
> >> Documentation/thermal/cpu-idle-cooling.txt | 165 ++++++++++
> >> drivers/cpuidle/cpuidle-arm.c | 5 +
> >> drivers/thermal/Kconfig | 30 +-
> >> drivers/thermal/cpu_cooling.c | 480 +++++++++++++++++++++++++++--
> >> include/linux/cpu_cooling.h | 15 +-
> >> 5 files changed, 668 insertions(+), 27 deletions(-)
> >> create mode 100644 Documentation/thermal/cpu-idle-cooling.txt
> >>
> >> --
> >> 2.7.4
> >>
>
>
> --
> <http://www.linaro.org/> Linaro.org â Open source software for ARM SoCs
>
> Follow Linaro: <http://www.facebook.com/pages/Linaro> Facebook |
> <http://twitter.com/#!/linaroorg> Twitter |
> <http://www.linaro.org/linaro-blog/> Blog
>