Re: [intel-pstate driver regression] processor frequency very high even if in idle

[JÃrg Otte]

2016-04-01 18:46 GMT+02:00 JÃrg Otte <jrg.otte@xxxxxxxxx>:
> 2016-04-01 17:20 GMT+02:00 Doug Smythies <dsmythies@xxxxxxxxx>:
>> On 2016.04.01 05:40 Rafael J. Wysocki wrote:
>>> On Friday, April 01, 2016 11:20:42 AM JÃrg Otte wrote:
>>>> 2016-03-31 17:43 GMT+02:00 Rafael J. Wysocki <rjw@xxxxxxxxxxxxx>:
>>>>> On Thursday, March 31, 2016 05:25:18 PM JÃrg Otte wrote:
>>>>>> 2016-03-31 13:42 GMT+02:00 Rafael J. Wysocki <rjw@xxxxxxxxxxxxx>:
>>>>>>> On Thursday, March 31, 2016 11:05:56 AM JÃrg Otte wrote:
>>
>>>>
>>>> here they are.
>>>>
>>
>>> First of all, the sampling mechanics works as expected
>>> in the failing case, which is the most important thing
>>> I wanted to know.
>>
>> Yes, but that might be part of the problem, as for some CPUs
>> there is never a long duration, and thus the long duration
>> check never kicks in driving the target pstate down.
>>
>>> The core_busy column is clearly suspicious and it
>>> looks like CPUs 2 and 3 never really go idle.
>>
>> This has been observed several times before [1].
>> Due to beat frequencies between desktop type frame rates
>> and such, the worst manifestation of the issue seems to be
>> for 300 Hz kernels, but Ubuntu uses uses 250 Hz.
>>
>> Oh look, JÃrg is using 300 Hz!!
>>
>> $ grep CONFIG_HZ .config_jorg
>> # CONFIG_HZ_PERIODIC is not set
>> # CONFIG_HZ_100 is not set
>> # CONFIG_HZ_250 is not set
>> CONFIG_HZ_300=y
>> # CONFIG_HZ_1000 is not set
>> CONFIG_HZ=300
>>
>
> I use 300Hz because of:
> "250 Hz is a good compromise choice allowing server performance
> while also showing good interactive responsiveness even
> on SMP and NUMA systems. If you are going to be using NTSC video
> or multimedia, selected 300Hz instead." (from KBuild helptext)
>
> -> I often use multimedia so according this text 300 Hz is the better
> choice.
>
>>> I guess we'll need to find out
>>> why they don't go idle to get to the bottom of this, but it firmly falls into
>>> the weird stuff territory already.
>
>> I'm compiling a 300 Hz kernel now, also with "# CONFIG_NO_HZ is not set",
>
> Again from KBuild helptext:
> "CONFIG_NO_HZ:
> This is the old config entry that enables dynticks idle.
> We keep it around for a little while to enforce backward
> compatibility with older config files."
>
> -> NO_HZ outdated.
>
>> but I have never been able to re-create these type of findings before.
>>
>> I have also tried several other things in an attempt re-create JÃrg's
>> Case, so far without success.
>>
>> References:
>> [1] https://bugzilla.kernel.org/show_bug.cgi?id=93521
>> In particular:
>> https://bugzilla.kernel.org/show_bug.cgi?id=93521#c35
>> https://bugzilla.kernel.org/show_bug.cgi?id=93521#c42
>> https://bugzilla.kernel.org/show_bug.cgi?id=93521#c77
>>
>> ... Doug
>
>
> Nevertheless, I'll try setting 250Hz + NO_HZ
>
For me no improvements.
Neither 300->250Hz  nor  NO_HZ_IDLE + NO_HZ

Thanks, JÃrg