Re: [PATCH] sched: Take thermal pressure into account when determine rt fits capacity

[Lukasz Luba <lukasz.luba@arm.com>]

On 4/19/22 13:51, Vincent Guittot wrote:
> On Tue, 19 Apr 2022 at 14:01, Lukasz Luba <lukasz.luba@arm.com> wrote:
>>
>>
>>
>> On 4/19/22 08:14, Vincent Guittot wrote:
>>> On Sat, 16 Apr 2022 at 04:47, Xuewen Yan <xuewen.yan94@gmail.com> wrote:
>>>>
>>>> Hi Luba  / Dietmar
>>>>
>>>> On Wed, Apr 13, 2022 at 9:26 PM Lukasz Luba <lukasz.luba@arm.com> wrote:
>>>>>
>>>>>
>>>>>
>>>>> On 4/11/22 15:07, Dietmar Eggemann wrote:
>>>>>> On 11/04/2022 10:52, Xuewen Yan wrote:
>>>>>>> HI Dietmar
>>>>>>>
>>>>>>> On Mon, Apr 11, 2022 at 4:21 PM Dietmar Eggemann
>>>>>>> <dietmar.eggemann@arm.com> wrote:
>>>>>>>>
>>>>>>>> On 07/04/2022 07:19, Xuewen Yan wrote:
>>>>>>>>> There are cases when the cpu max capacity might be reduced due to thermal.
>>>>>>>>> Take into the thermal pressure into account when judge whether the rt task
>>>>>>>>> fits the cpu. And when schedutil govnor get cpu util, the thermal pressure
>>>>>>>>> also should be considered.
>>>>>>>>>
>>>>>>>>> Signed-off-by: Xuewen Yan <xuewen.yan@unisoc.com>
>>>>>>>>> ---
>>>>>>>>>     kernel/sched/cpufreq_schedutil.c | 1 +
>>>>>>>>>     kernel/sched/rt.c                | 1 +
>>>>>>>>>     2 files changed, 2 insertions(+)
>>>>>>>>>
>>>>>>>>> diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c
>>>>>>>>> index 3dbf351d12d5..285ad51caf0f 100644
>>>>>>>>> --- a/kernel/sched/cpufreq_schedutil.c
>>>>>>>>> +++ b/kernel/sched/cpufreq_schedutil.c
>>>>>>>>> @@ -159,6 +159,7 @@ static void sugov_get_util(struct sugov_cpu *sg_cpu)
>>>>>>>>>          struct rq *rq = cpu_rq(sg_cpu->cpu);
>>>>>>>>>          unsigned long max = arch_scale_cpu_capacity(sg_cpu->cpu);
>>>>>>>>>
>>>>>>>>> +     max -= arch_scale_thermal_pressure(sg_cpu->cpu);
>>>>>>>>
>>>>>>>> max' = arch_scale_cpu_capacity() - arch_scale_thermal_pressure()
>>>>>>>>
>>>>>>>> For the energy part (A) we use max' in compute_energy() to cap sum_util
>>>>>>>> and max_util at max' and to call em_cpu_energy(..., max_util, sum_util,
>>>>>>>> max'). This was done to match (B)'s `policy->max` capping.
>>>>>>>>
>>>>>>>> For the frequency part (B) we have freq_qos_update_request() in:
>>>>>>>>
>>>>>>>> power_actor_set_power()
>>>>>>>>      ...
>>>>>>>>      cdev->ops->set_cur_state()
>>>>>>>>
>>>>>>>>        cpufreq_set_cur_state()
>>>>>>>>          freq_qos_update_request()      <-- !
>>>>>>>>          arch_update_thermal_pressure()
>>>>>>>>
>>>>>>>> restricting `policy->max` which then clamps `target_freq` in:
>>>>>>>>
>>>>>>>>      cpufreq_update_util()
>>>>>>>>        ...
>>>>>>>>        get_next_freq()
>>>>>>>>          cpufreq_driver_resolve_freq()
>>>>>>>>            __resolve_freq()
>>>>>>>>
>>>>>>>
>>>>>>> Do you mean that the "max" here will not affect the frequency
>>>>>>> conversion, so there is no need to change it?
>>>>>>> But is it better to reflect the influence of thermal here?
>>>>>>
>>>>>> I guess your point is that even though max' has no effect on frequency
>>>>>> since QOS caps policy->max anyway, it is still easier to understand the
>>>>>> dependency between schedutil and EAS/EM when it comes to the use of
>>>>>> thermal pressure.
>>>>>>
>>>>>> I agree. The way how the "hidden" policy->max capping guarantees that
>>>>>> schedutil and EAS/EM are doing the same even when only the latter uses
>>>>>> max' is not obvious.
>>>>>
>>>>> +1 here, IMO we shouldn't rely on hidden stuff. There are two which set
>>>>> the thermal pressure, but one is not setting the freq_qos which causes
>>>>> the update of the 'policy->max'. So the schedutil will send that high
>>>>> frequency but that driver would just ignore and clamp internally. In the
>>>>> end we might argue it still works, but is it clean and visible from the
>>>>> code? Funny thing might start to happen then the driver, which might be
>>>>> the last safety net cannot capture this.
>>>
>>> schedutil reflects what scheduler wants not what HW can do. If you
>>> start to cap the freq with arch_scale_thermal_pressure() in schedutil,
>>
>> s/freq/util ?
> 
> Isn't it the same at the end if you cap util to orig capacity -
> arch_scale_thermal_pressure ?

No, as I have showed you in the example calculation the 'max_freq'
is always taken as a start then scaled by current 'util/max'.
If the 'max' was 1024 in both cases, then you might claim that
we made a mistake and end up with obviously too low frequency.

That's why I asked you to be precised with your statements in the
description while mentioning variables and signals.

> 
>>
>> To be precised and maybe fix some potential design issues. We are
>> talking here about utilization and set max capacity in function:
>> sugov_get_util()
>> so fields:
>>
>> sugov_cpu::util
>> sugov_cpu::max /* max capacity */
> 
> Yes. With this patch ,util will be lower than current thermal
> mitigation whereas util normally reflects what we need  not what can
> be provided

This is a different requirements: util has to be max capacity and
max capacity has to be original max CPU capacity - for the SchedUtil.
OK, why? What this requirement adds in the design and final values?

> 
>>
>>> you will loose some opportunity to run at higher frequency because
>>> arch_scale_thermal_pressure() is transient and might change just after
>>> using it. This means that you will stay at lower freq after mitigation
>>> stops until a new cpufreq_update_util() happens. ANd I don't vene
>>> mentioned when thermal mitigation is managed by HW at a much higher
>>> frequency than what Linux can handle
>>>
>>> arch_scale_thermal_pressure() must not be used but thermal_load_avg()
>>> like scale_rt_capacity() what Dietmar suggested
>>>
>>
>> First, I would like to see your view to this topic and why you are
>> making such strong statements. I have slightly different view and
>> made dozen of experiments with this thermal pressure in last ~2-3y.
>>
>> The code flow is like this and operates on those fields from above:
>>
>> util, max <--- sugov_get_util()
>> util <--- sugov_iowait_apply()  <--- util, max /* ignore this now */
>>
>> get_next_freq():
>> util <--- map_util_perf() <--- util (1)
>> freq <--- map_util_freq() <--- util, max, max_freq (2)
>>
>>
>> At (1) we add +25% util, at (2) we do the conversion to frequency:
>> freq = max_freq * util / max
>>
>> As you can see with the patch we would still end up with bigger
>> frequency than max_freq (since it can happen: max < util).
>> It's also true currently in mainline, when
>> max=1024 and util=1024+256=1280
>> I would be similar if we cap max capacity:
>> max=800 and util=800+200=1000
> 
> It's not because you end up with a similar value that it's ok. You
> can't use one side to compensate for the other one. 1.25 is there to
> provide spare cycles to a change in the cpu load (load meaning what is
> running on the cpu not load_avg)

It's different. You've made a hard statement that we are going to break 
this frequency selected value, while IMO we aren't. It would behave
the same. I don't compensate anything, the divider ('max') has changed
as well. The patch sets 'sg_cpu->max' not just uses 'max' locally, so
it produces lower 'util' but then might have 'max=1024'. It's not the
case.

> 
>> but then in both cases are multiplied by 'max_freq' in (2)
>>
>> As you can see this is not the situation that you have described, is it?
>> And the transient or non-transient is minor here IMO.
> 
> If max is 512 then util = 640 which is much lower than 1024.

What scenario is this?
Is 1024 the utilization that we might have from the CPU rq?
What is the original CPU capacity, 1024?

> 
>>
>> Secondly, you have mentioned the mitigation in HW and issue between
>> instantaneous vs. PELT-one thermal pressure information. This is
>> something that I'm stretching my head for long. I'm trying to
>> develop this for new Arm FW thermal. You have mentioned:
>> 'thermal mitigation is managed by HW at a much higher
>> frequency than what Linux can handle' - I would be also more
>> precised here: HW or FW? How often the HW can change max freq or
>> how often FW can change that? If we don't have those numbers
>> than statement: 'a much higher' doesn't help in solving this
> 
> By much higher means that Linux can't react fast enough and should not
> try to sync because it's a lost game

As I said, 'much higher' is not a number to base a design on it.
We need real numbers from real platforms. Currently we have two
places where the thermal pressure is set:
1) cpufreq_cooling.c [1]
2) Qcom driver [2]
(we might have 3rd soon for Arm SCMI+FW)

For the 2nd I would like to see numbers. For the 1st one when
kernel thermal is used (which supports higher number of platforms
comparing to Qcom driver) as it's by design kernel tries to control
thermal, so changes are not that frequent.

As for now, I can see in experiments the 1st is suffering long decay
delays and also corner cases with long idle CPUs.

> 
>> problem that Xuewen (and others) faces. IMO it's not technical
>> argument for blocking the patch and incremental development.
>>
>> It's about timing, when we talk about thermal pressure signals and
>> those two information. For the PELT-one there are also two use cases:
>> raising time and decay time (where we're actually increasing the
>> visible capacity of the CPU). The decay period is quite long,
>> e.g.
>> Thermal pressure of 140 is removed, signal should converge to 0 from 140
>> in 130ms (90% decayed),
>> in 230ms (fully decayed).
>> The default kernel code allows to slow down the decay period, which is
>> a derivative from current global PELT default setting.
>> We can slow it down, but we cannot make it to react faster (BTW I made
>> such change to compare experiments). It's not always good to have
>> such long delays.
>>
>> For asymmetric CPUs that I was describing and also Xuewen, where mid
>> core might be faster than big, we need this information in RT class.
>> Android is such system, so the situation is real (DL is not used there).
>> You have questioned this that:
>> 'arch_scale_thermal_pressure() must not be used'
>> I wouldn't be so sure for the RT change.
>> Are you sure about that? Do you have experiments for it? I would
>> like to see them. I have run dozen of experiments and measurements
>> for this thermal pressure information on a few platforms. How
>> they behave on those task placements and what are the thermal
>> signal decay delay impacts. I'm still not sure which one is
>> best and thus not proposed any changes. But I'll refactor my
>> test code and send a patch with trace event for the new
>> topology_update_thermal_pressure(), which then allows to compare
>> those two designs and nasty timings issues. We would than see
>> how often (if 'much higher' is true) platforms set this value.
>> Currently, in mainline there are two clients which set this
>> value.
>>
>> I've been investigating this PELT signal ~1-2 year ago and found
>> an issue when it's actually updated with delays for the long idle CPU.
>> When one CPU was running fast and thermal throttling kicked in, while
>> the other was idle, the idle one didn't have recent thermal information,
>> but could be picked as a candidate because visible capacity was ~max
>> possible - which is wrong because they both share the clock.
>> Check the function others_have_blocked() and the design around it.
>>
>> That's why I'm a bit more careful with statements that one type of
>> information is better that other.
>>
>> Furthermore, check the code in rt_task_fits_capacity(), there is no
>> PELT signal from the RT task. There is only uclamp_eff_value() from
>> task 'p', which is not PELT information. So all involved variables
>> are not PELT, why you recommend the PELT thermal pressure here?
>>
>> As I said, this patch for the RT class is an incremental step into the
>> right direction.
>>
>>

You haven't answered my questions, which are about technical details of
your recommendations and statements.

I'm trying to help Xuewen to solve his/her issues with the RT class
incrementally. I don't want to push him/her into a deep dark water
of PELT signals, to what variable compare them, corner cases when they
are (or not) updated or completely not implemented. I'm not even sure
if those extra complexities make sense for the RT/DL (since they
make some difference on big.mid.little specific platforms but not for
the rest).