Re: [Resend][PATCH] cpufreq: conservative: Decrease frequency faster when the timer deferred

From: Stratos Karafotis <stratosk@semaphore.gr>
To: "Rafael J. Wysocki" <rafael@kernel.org>,
	Viresh Kumar <viresh.kumar@linaro.org>
Cc: "linux-pm@vger.kernel.org" <linux-pm@vger.kernel.org>,
	Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
	"Rafael J. Wysocki" <rjw@rjwysocki.net>
Subject: Re: [Resend][PATCH] cpufreq: conservative: Decrease frequency faster when the timer deferred
Date: Thu, 10 Nov 2016 17:48:32 +0200	[thread overview]
Message-ID: <9420dc99-57ab-389a-dae9-4efae33dd102@semaphore.gr> (raw)
In-Reply-To: <CAJZ5v0hsEd_fVQEnW9ELhauwot8ZdqxWRJJCoF_j599DyrMJ-g@mail.gmail.com>

On 10/11/2016 02:13 πμ, Rafael J. Wysocki wrote:
> On Wed, Nov 9, 2016 at 6:55 AM, Viresh Kumar <viresh.kumar@linaro.org> wrote:
>> On 08-11-16, 21:25, Stratos Karafotis wrote:
>>> But this is the supposed behaviour of conservative governor. We want
>>> the CPU to increase the frequency in steps. The patch just resets
>>> the frequency to a lower frequency in case of idle.
>>>
>>> For argument's sake, let's assume that the governor timer is never
>>> deferred and runs every sampling period even on completely idle CPU.
>>
>> There are no timers now :)
>>
>>> And let's assume, for example, a burst load that runs every 100ms
>>> for 20ms. The default sampling rate is also 20ms.
>>> What would conservative do in case of that burst load? It would
>>> increase the frequency by one freq step after 20ms and then it would
>>> decrease the frequency 4 times by one frequency step. Most probably
>>> on the next burst load, the CPU will run on min frequency.
>>>
>>> I agree that maybe this is not ideal for performance but maybe this is
>>> how we want conservative governor to work (lazily increase and decrease
>>> frequency).
>>
>> Idle periods are already accounted for while calculating system load by legacy
>> governors.
>>
>> And the more and more I think about this, I am inclined towards your patch.
>> Maybe in a bit different form and commit log.
>>
>> If we see how the governors were written initially, there were no deferred
>> timers. And so even if CPUs were idle, we will wake up to adjust the step.
>>
>> Even if we want to make the behavior similar to that, then also we should
>> account of missed sampling periods both while decreasing or increasing
>> frequencies.
>>
>> @Rafael: What do you think ?
> 
> It looks like the issue with the conservative governor is real, but
> I'm a bit concerned about adding things to use by one particular
> governor only to cpufreq_governor.c.

I think the code is minimum and I didn't find a way to do this
calculation in cpufreq_conservative.c. We also use code in
cpufreq_governor.c that it's only specific to ondemand (io_busy).

If you can give me a hint about how to implement this logic in
cpufreq_conservative I would appreciate it.

> Apart from the timer-related terminology that is not applicable any
> more, of course.

I will correct the terminology if the logic is accepted.

Regards,
Stratos