RE: [RFT][PATCH v1] cpuidle: teo: Avoid selecting deepest idle state over-eagerly
From: Doug Smythies
Date: Mon Feb 10 2025 - 10:17:36 EST
On 2025.02.08 03:25 Rafael J. Wysocki wrote:
>On Sat, Feb 8, 2025 at 12:40 AM Doug Smythies <dsmythies@xxxxxxxxx> wrote:
>> On 2025.02.04 12:58 Rafael J. Wysocki wrote:
>>>
>>> From: Rafael J. Wysocki <rafael.j.wysocki@xxxxxxxxx>
... snip ...
>>
>> Test durations were >= 45 minutes each.
>>
>> Kernel 6.14-rc1: Includes cpuidle: teo: Cleanups and very frequent wakeups handling update
>> Average Idle Power: teo governor: 2.199 watts (+25.51%)
>> Average Idle power: menu governor: 1.873 watts (+6.91%)
>
> menu hasn't changed in 6.14-rc1, so this would be variation between
>runs I suppose.
Perhaps the way I presented my data wasn't the best method.
All the % were relative to kernel 6.13 and the teo idle governor.
Relative to kernel 6.13 menu governor that would be: -1.9%,
certainly within the noise floor.
>> Kernel 6.14-rc1-p: Added this patch for teo and "cpuidle: menu: Avoid discarding useful information when processing recent idle intervals" for menu
>> Average Idle Power: teo governor: 9.401 watts (+436.6%)
>> Only 69.61% idle is in the deepest idle state. More typically it would be 98% to 99%.
>
> Ah, not good.
>
> OK, this clearly doesn't go in the right direction.
>
>> 28.6531% idle time is in state 1. More typically it would be 0.03%
>> Average Idle Power: menu governor: 1.820 watts (+3.9%)
>>
>> Kernel 6.13: before "cpuidle: teo: Cleanups and very frequent wakeups handling update"
>> Average Idle Power: teo governor: 1.752 watts (reference: 0.0%)
>> Average Idle power: menu governor: 1.909 watts (+9.0%)
>
> Thanks, I'm just going to drop this patch then.
>
> If you don't mind, I'll have a couple more teo updates for testing.
O.K. I'll watch for the patches.
Plenty to do in the meantime, acquiring baseline data and continuing with the:
"cpuidle: menu: Avoid discarding useful information when processing recent idle intervals"
patch set portion of kernel 6.14-rc1-p