RE: [RFT][PATCH v1] cpuidle: teo: Avoid selecting deepest idle state over-eagerly

From: Doug Smythies
Date: Thu Feb 13 2025 - 23:23:26 EST


Hi Christian,

Thank you for trying to repeat my idle test results.

On 2025.02.13 06:07 Christian Loehle wrote:

> I'm curious, are Doug's numbers reproducible?
> Or could you share the idle state usage numbers? Is that explainable?
> Seems like a lot and it does worry me that I can't reproduce anything
> as drastic.

While I am having some severe repeatability issues with my testing,
not for this test.

Please recall my test conditions because the CPU frequency
scaling governor does matter. I was using "performance".
The power comes from the high amount of time in idle state 1.
I verified the idle state 1 power use by disabling all other idle states.
I also have HWP disabled, but do not know if it matters.
If I use the "powersave" governor (driver is intel_pstate, not
intel_cpufreq) then idle power is < 2 watts.

Anyway, my data:

http://smythies.com/~doug/linux/idle/teo-6.14/idle/perf/

> (Idle numbers aren't really reflective in energy used -> dominated by
> active power.)

Well, it depends on idle time verses active time on the computer.
I also measured the difference in the mains power at 20%, from
43.2 watts to 51.4 watts.

I am about to send a long email with all of my test results.
Since I can not seem to function without making graphs,
it has a lot of links to graphs.