Re: [RFC PATCH v2 0/1] cpuidle: teo: Introduce optional util-awareness

From: Lukasz Luba
Date: Tue Oct 11 2022 - 04:22:01 EST


Hi Doug, Kajetan,

On 10/10/22 11:09, Kajetan Puchalski wrote:
On Thu, Oct 06, 2022 at 05:06:58PM -0700, Doug Smythies wrote:

I would suggest a completely new governor for your
only 2 idle states, and of similar power use, scenario.
Incredibly simple:

If CPU utilization >= threshold
select idle state 0
Else
Select idle state 1
Endif

That would be easy to implement at least but sadly just doesn't really
work at all. It would just result in huge amounts of too deep C1 sleeps
which are the main problem for us at the moment. For all intents and
purposes, only ever selecting C0 gives better results than this approach
on our hardware.

As previously mentioned, you are overriding
everything that teo does anyhow.

To an extent yes, this just came from observations of how cpuidle
behaves on arm. We tried an approach with only ever using C0 and it was
all right on power usage, the device would just heat up too much and
lose performance. That's why this approach of using TEO when the core is
completely idle and moving to the lower state when it's not achieves the
best of both worlds.

Note to other readers: I also sent Kajetan an off-list email with more
tests descriptions and web links to results graphs and such.

Thanks Doug for testing this work, we really appreciate that.
Doug maybe you could post these into some public space, so wider
community would also see them. Kajetan has used github to put
a report with testing results containing some graphs/plots.


Looks like the email got lost in my inbox, could be some email filtering
or something like that.. Could you possibly resend and cc my private
email (kajetan.puchalski@xxxxxxx)?
So far I've not directly received any of your emails, they show up on
the mailing list archive but not in my inbox. Could be an IT issue on my
end but might be worth checking your email setup regardless.

I also have some issues with the email inbox. Me apologies Doug, we will
sort it.

Regards,
Lukasz