RE: Ask for help on governor

[Andy Tang <andy.tang@nxp.com>]

Hi,

> -----Original Message-----
> From: Doug Smythies [mailto:dsmythies@telus.net]
> Sent: Friday, December 15, 2017 9:31 AM
> To: Andy Tang <andy.tang@nxp.com>; 'Viresh Kumar'
> <viresh.kumar@linaro.org>; 'Stratos Karafotis' <stratosk@semaphore.gr>
> Cc: 'Rafael J. Wysocki' <rafael@kernel.org>; 'Rafael J. Wysocki'
> <rjw@rjwysocki.net>; 'Linux PM' <linux-pm@vger.kernel.org>
> Subject: RE: Ask for help on governor
> 
> On 2017.12.13 18:42 Andy Tang wrote:
> > On 2017.12.14 09:21 Doug wrote:
> >>
> >> diff --git a/drivers/cpufreq/cpufreq_governor.c
> >> b/drivers/cpufreq/cpufreq_governor.c
> >> index 58d4f4e..3493ca7 100644
> >> --- a/drivers/cpufreq/cpufreq_governor.c
> >> +++ b/drivers/cpufreq/cpufreq_governor.c
> >> @@ -222,6 +222,7 @@ unsigned int dbs_update(struct cpufreq_policy
> >> *policy)
> >>                         max_load = load;
> >>         }
> >>
> >> +       idle_periods = 0;
> >
> > With this line added, conservative governor works well.
> >
> > Ondemand governor still can't work well.
> 
> Your original message said:
> 
> On 12-12-17, 02:46, Andy Tang wrote:
> > I run into a question that conservative governor can't work while
> ondemand governor works well.
> 
> So, for my part of it, I never looked at ondemand.
Sorry for the misleading. Actually I observed two issues:
1. conservative governor can't increase cpu frequency even the load is high than 90%
2. CPU load can't be calculated correctly when sampling_rate is 1000 for both conservative and ondemand governor. At first I thought ondemand governor worked well. Actually I did not test it thoroughly. Ondemand governor can work only when cpu load is calculated correct.

> 
> > As I mentioned in earlier email, load varies from different sampling_rate.
> 
> Because that e-mail was html, a lot of people didn't get it, and it also is not in
> the archives. Anyway:
> 
> On 2017.12.12 23:24 Andy wrote:
> 
> > I also found sampling_rate affects the CPU load dramatically.
> >			sample rate
> >			100000	10000	1000
> > actual cpu load	1%		1%	1%
> > cpu report load	1%		20%	40-100%
> >
> > So we can know that when sample_rate is higher than 10000us, cpufreq
> can not calculate the load correctly.
> 
> Can you tell us more details about how you did this test and how you know
> your data is correct?
> In particular, what is your work/sleep frequency for your 1% load.
> In my experience the work/sleep frequency has always been a very
> important consideration with these tests.
> What is the tick rate of your kernel? Myself, I wouldn't expect a 1000 uSec
> sample period to work properly.
My tick rate is 250Hz.  And Yes, cpu load can't be calculated correctly.
We really should set sampling_rate a proper values as default if there are some sample rate don't work properly. We don't want to calculate a workable sampling_rate according to the tick rate manually.
 
That's odd too that you wouldn't expect a 1000 Usec sample period to work properly but set this value as default, even worse, you allow to set a lower value to it.

Is it better if we set a low limit of  sample period? 

Regards,
Andy
 
> 
> I used turbostat to verify my tests, and yes, on demand does start to
> misbehave if the sampling period is short relative to the tick period (jiffy
> time), otherwise it was fine.
> 
> ... Doug
>