Message-ID: <20240222115557.blnm4uulkxnorrl4@airbuntu>
Date: Thu, 22 Feb 2024 11:55:57 +0000
From: Qais Yousef <qyousef@...alina.io>
To: Pierre Gondois <pierre.gondois@....com>
Cc: "Rafael J. Wysocki" <rafael@...nel.org>,
	Viresh Kumar <viresh.kumar@...aro.org>,
	linux-kernel@...r.kernel.org, linux-pm@...r.kernel.org,
	Ingo Molnar <mingo@...nel.org>,
	Peter Zijlstra <peterz@...radead.org>,
	Vincent Guittot <vincent.guittot@...aro.org>,
	Dietmar Eggemann <dietmar.eggemann@....com>,
	Christian.Loehle@....com
Subject: Re: [PATCH] cpufreq: Change default transition delay to 2ms

On 02/20/24 18:38, Pierre Gondois wrote:
> Hello Qais,
> 
> I added some other remarks,
> 
> On 2/20/24 14:50, Qais Yousef wrote:
> > On 02/14/24 10:19, Pierre Gondois wrote:
> > > Hello,
> > > 
> > > On 2/12/24 16:53, Rafael J. Wysocki wrote:
> > > > On Mon, Feb 5, 2024 at 8:45 AM Viresh Kumar <viresh.kumar@...aro.org> wrote:
> > > > > 
> > > > > On 05-02-24, 02:25, Qais Yousef wrote:
> > > > > > 10ms is too high for today's hardware, even low end ones. This default
> > > > > > end up being used a lot on Arm machines at least. Pine64, mac mini and
> > > > > > pixel 6 all end up with 10ms rate_limit_us when using schedutil, and
> > > > > > it's too high for all of them.
> > > > > > 
> > > > > > Change the default to 2ms which should be 'pessimistic' enough for worst
> > > > > > case scenario, but not too high for platforms with fast DVFS hardware.
> > > > > > 
> > > > > > Signed-off-by: Qais Yousef <qyousef@...alina.io>
> > > > > > ---
> > > > > >    drivers/cpufreq/cpufreq.c | 4 ++--
> > > > > >    1 file changed, 2 insertions(+), 2 deletions(-)
> > > > > > 
> > > > > > diff --git a/drivers/cpufreq/cpufreq.c b/drivers/cpufreq/cpufreq.c
> > > > > > index 44db4f59c4cc..8207f7294cb6 100644
> > > > > > --- a/drivers/cpufreq/cpufreq.c
> > > > > > +++ b/drivers/cpufreq/cpufreq.c
> > > > > > @@ -582,11 +582,11 @@ unsigned int cpufreq_policy_transition_delay_us(struct cpufreq_policy *policy)
> > > > > >                  * for platforms where transition_latency is in milliseconds, it
> > > > > >                  * ends up giving unrealistic values.
> > > > > >                  *
> > > > > > -              * Cap the default transition delay to 10 ms, which seems to be
> > > > > > +              * Cap the default transition delay to 2 ms, which seems to be
> > > > > >                  * a reasonable amount of time after which we should reevaluate
> > > > > >                  * the frequency.
> > > > > >                  */
> > > > > > -             return min(latency * LATENCY_MULTIPLIER, (unsigned int)10000);
> > > > > > +             return min(latency * LATENCY_MULTIPLIER, (unsigned int)(2*MSEC_PER_SEC));
> > > > > 
> > > > > Please add spaces around '*'.
> > > > > 
> > > > > Acked-by: Viresh Kumar <viresh.kumar@...aro.org>
> > > > 
> > > > I've adjusted the whitespace as suggested above and applied the patch
> > > > as 5.9 material.
> > > > 
> > > > Thanks!
> > > > 
> > > 
> > > To add some numbers, on a Juno-r2, with latency measured between the frequency
> > > request on the kernel side and the SCP actually making the frequency update.
> > > 
> > > The SCP is the firmware responsible of making the frequency updates. It receives
> > > the kernel requests and coordinate them/make the actual changes. The SCP also has
> > > a mechanism called 'fast channel' (FC) where the kernel writes the requested
> > > frequency to a memory area shared with the SCP. Every 4ms, the SCP polls/reads
> > > these memory area and make the required modifications.
> > > 
> > > Latency values (in ms)
> > > Workload:
> > > Idle system, during ~30s
> > > +---------------------------------------+
> > > |       |   Without FC  |      With FC  |
> > > +-------+---------------+---------------+
> > > | count |       1663    |        1102   |
> > > | mean  |          2.92 |          2.10 |
> > > | std   |          1.90 |          1.58 |
> > > | min   |          0.21 |          0.00 |
> > > | 25%   |          1.64 |          0.91 |
> > > | 50%   |          2.57 |          1.68 |
> > > | 75%   |          3.66 |          2.97 |
> > > | max   |         14.37 |         13.50 |
> > > +-------+---------------+---------------+
> > > 
> > > Latency values (in ms)
> > > Workload:
> > > One 1% task per CPU, period = 32ms. This allows to wake up the CPU
> > > every 32ms and send more requests/give more work to the SCP. Indeed
> > > the SCP is also responsible of idle state transitions.
> > > Test duration ~=30s.
> > > +---------------------------------------+
> > > |       |   Without FC  |      With FC  |
> > > +-------+---------------+---------------+
> > > | count |       1629    |       1446    |
> > > | mean  |          3.23 |          2.31 |
> > > | std   |          2.40 |          1.73 |
> > > | min   |          0.05 |          0.02 |
> > > | 25%   |          1.91 |          0.98 |
> > > | 50%   |          2.65 |          2.00 |
> > > | 75%   |          3.65 |          3.23 |
> > > | max   |         20.56 |         16.73 |
> > > +-------+---------------+---------------+
> > > 
> > > ---
> 
> 1.
> With this patch, platforms like the Juno which:
> - don't set a `transition_delay_us`
> - have a high `transition_latency` (> 1ms)
> can request freq. changes every 2ms.
> 
> If a platform has a `transition_latency` > 2ms, this means:
>   `transition_latency` > `transition_delay_us`
> I.e. a second freq. requests might be emitted before the first one
> will be completed. On the Juno, this doesn't cause any 'real' issue
> as the SCMI/mailbox mechanism works well, but this doesn't seem
> correct.
> If the util of CPUs is in between OPPs (i.e. freq. changes are often
> required), the Juno:
> - sends a freq. request
> - waits for completion and schedules another task in the meantime
> - upon completion, immediately sends a new freq.
> 
> I think that the following should be respected/checked:
> - `transition_latency` < `transition_delay_us`
> (it might also make sense to have, with K being any factor:)
> - `transition_latency` * K < `transition_delay_us`

Makes sense. How about this? I am not sure it is better to multiply with
a factor if the platform is already slow. Even the current 1000 multiply factor
is high but this is a territory I am not ready to step into yet.

diff --git a/drivers/cpufreq/cpufreq.c b/drivers/cpufreq/cpufreq.c
index 66cef33c4ec7..68a5ba24a5e0 100644
--- a/drivers/cpufreq/cpufreq.c
+++ b/drivers/cpufreq/cpufreq.c
@@ -576,6 +576,15 @@ unsigned int cpufreq_policy_transition_delay_us(struct cpufreq_policy *policy)

        latency = policy->cpuinfo.transition_latency / NSEC_PER_USEC;
        if (latency) {
+               unsigned int max_delay_us = 2 * MSEC_PER_SEC;;
+
+               /*
+                * If the platform already has high transition_latency, use it
+                * as-is.
+                */
+               if (latency > max_delay_us)
+                       return latency;
+
                /*
                 * For platforms that can change the frequency very fast (< 10
                 * us), the above formula gives a decent transition delay. But
@@ -586,7 +595,7 @@ unsigned int cpufreq_policy_transition_delay_us(struct cpufreq_policy *policy)
                 * a reasonable amount of time after which we should reevaluate
                 * the frequency.
                 */
-               return min(latency * LATENCY_MULTIPLIER, (unsigned int)(2 * MSEC_PER_SEC));
+               return min(latency * LATENCY_MULTIPLIER, max_delay_us);
        }

        return LATENCY_MULTIPLIER;

> 
> 
> 2.
> There are references to the 10ms values at other places in the code:
> 
> include/linux/cpufreq.h
>  * ondemand governor will work on any processor with transition latency <= 10ms,

Not sure this one needs updating. Especially with the change above which means
10ms could theoretically happen. But if there are suggestions happy to take
them.

> 
> drivers/cpufreq/cpufreq.c
>  * For platforms that can change the frequency very fast (< 10
>  * us), the above formula gives a decent transition delay. But
> -> the 10us value matches 10ms = 10us * LATENCY_MULTIPLIER

I can't find this one.

> 
> Documentation/admin-guide/pm/cpufreq.rst
>  Typically, it is set to values of the order of 10000 (10 ms).  Its
>  default value is equal to the value of ``cpuinfo_transition_latency``

I am not sure about this one. It refers to cpuinfo_transition_latency not the
delay and uses a formula to calculate it based on that.

Seems the paragraph needs updating in general to reflect other changes?

> 
> 
> 3.
> There seems to be a dependency of the conservative/ondemand governors
> over the the value returned by `cpufreq_policy_transition_delay_us()`:
> 
> drivers/cpufreq/cpufreq_governor.c
>   dbs_data->sampling_rate = max_t(unsigned int,
>     CPUFREQ_DBS_MIN_SAMPLING_INTERVAL,            // = 2 * tick period = 8ms
>     cpufreq_policy_transition_delay_us(policy));  // [1]: val <= 2ms
> 
> [1]
> if `transition_latency` is not set and `transition_delay_us` is,
> which is the case for the Juno.
> 
> The `sampling_rate` is, FYIU, the period used to evaluate the ratio
> of the idle/busy time, and if necessary increase/decrease the freq.
> 
> This patch will likely reduce this sampling rate from 10ms -> 8ms
> (if `cpufreq_policy_transition_delay_us()`` now returns 2ms for some
> platforms). This is not much, but just wanted to note it.

I don't think this is a problem as tick being 1ms is common and
transition_delay_us is not 10ms on all platforms. On my amd system it is 1ms
and on another intel i5 it is 5ms. So it should have already been coping with
various combination.


Thanks!

--
Qais Yousef

Powered by blists - more mailing lists