lkml.org 
[lkml]   [2018]   [Mar]   [26]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRE: [PATCH v3] cpuidle: poll_state: Add time limit to poll_idle()
Date
On 2018.03.25 04:54 Rafael J. Wysocki wrote:
> On Sun, Mar 25, 2018 at 1:28 AM, Doug Smythies <dsmythies@telus.net> wrote:
>> On 2018.03.14 07:04 Rafael J. Wysocki wrote:
>>
>>> If poll_idle() is allowed to spin until need_resched() returns 'true',
>>> it may actually spin for a much longer time than expected by the idle
>>> governor, since set_tsk_need_resched() is not always called by the
>>> timer interrupt handler. If that happens, the CPU may spend much
>>> more time than anticipated in the "polling" state.
>>>
>>> To prevent that from happening, limit the time of the spinning loop
>>> in poll_idle().
>>
>> ...[snip]...
>>
>>> +#define POLL_IDLE_TIME_LIMIT (TICK_NSEC / 16)
>>
>> The other ongoing threads on this aside, potentially, there might
>> be another issue.
>>
>> What if the next available idle state, after 0, has a residency
>> that is greater than TICK_NSEC / 16? Meaning these numbers, for example:
>>
>> /sys/devices/system/cpu/cpu0/cpuidle/state*/residency
>>
>> The suggestion is that upon a timeout exit from idle state 0,
>> the measured_us should maybe be rejected, because the statistics
>> are being biased and it doesn't seem to correct itself.
>
> OK
>
>> Up to 1300% (<- not a typo) extra power consumption has been observed.
>>
>> Supporting experimental data:
>>
>> My processor:
>> /sys/devices/system/cpu/cpu0/cpuidle/state0/residency:0
>> /sys/devices/system/cpu/cpu0/cpuidle/state1/residency:2
>> /sys/devices/system/cpu/cpu0/cpuidle/state2/residency:20
>> /sys/devices/system/cpu/cpu0/cpuidle/state3/residency:211 <<< Important
>> /sys/devices/system/cpu/cpu0/cpuidle/state4/residency:345
>>
>> A 1000 Hz kernel (TICK_NSEC/16) = 62.5 nsec; idle system:
>
> nsec or usec?

Right, uSeconds.

>> Idle state 0 time: Typically 0 uSec.
>> Processor package power: 3.7 watts (steady)
>>
>> Now, disable idle states 1 and 2:
>>
>> Idle state 0 time (all 8 CPUs): ~~ 430 Seconds / minute
>> Processor package power: ~52 watts (1300% more power, 14X)
>
> But that's because you have disabled states 1 and 2, isn't it?

Yes, and perhaps the conclusion here is that we don't care if
the user has disabled intermediate idle states, forcing these
conditions.

>> A 250 Hz kernel (TICK_NSEC/16) = 250 nSec; idle system:
250 uSec.
>>
>> Idle state 0 time: Typically < 1 mSec / minute
>> Processor package power: 3.7 to 3.8 watts
>>
>> Now, disable idle states 1 and 2:
>>
>> Idle state 0 time (all 8 CPUs): Typically 0 to 70 mSecs / minute
>> Processor package power: 3.7 to 3.8 watts
>>
>> A 1000 Hz kernel with:
>>
>> +#define POLL_IDLE_TIME_LIMIT (TICK_NSEC / 4)
>>
>> Note: Just for a test. I am not suggesting this should change.
>>
>> instead. i.e. (TICK_NSEC/4) = 250 nSec.
250 uSec.
>>
>> Idle state 0 time: Typically 0 uSec.
>> Processor package power: 3.7 watts (steady)
>>
>> Now, disable idle states 1 and 2:
>>
>> Idle state 0 time (all 8 CPUs): Typically 0 to 70 mSecs / minute
>> Processor package power: ~3.8 watts
>>
>> Note 1: My example is contrived via disabling idle states, so
>> I don't know if it actually needs to be worried about.
>>
>> Note 2: I do not know if there is some processor where
>> cpuidle/state1/residency is > 62.5 nSec.
62.5 uSec.
> If that's usec, I would be quite surprised if there were any. :-)

O.K.

>> Note 3: I am trying to figure out a way to test rejecting
>> measured_us upon timeout exit, but haven't made much progress.
>
> Rejecting it has side effects too, because it basically means lost information.
>
> Reaching the time limit means that the CPU could have been idle much
> longer, even though we can't say how much. That needs to be recorded
> in the data used for the next-time prediction or that is going to be
> inaccurate too.
>
> Of course, what number to record is a good question. :-)

Maybe it would be O.K. to be aware of this and simply move on.

By the way, I forgot to mention that the above work was done
with kernels based on 4.16-rc6 and only these poll_idle patches.
If I then add the V7.3 idle loop rework patch set, the issue
becomes partially mitigated (151 minute averages):
Idle state 0 time (all 8 CPUs): ~~ 304 Seconds / minute
Processor package power: ~47.7 watts

It'll be tomorrow before I can try the suggestion from the other e-mail.

... Doug


\
 
 \ /
  Last update: 2018-03-26 08:02    [W:0.097 / U:20.904 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site