Re: [Patch v9 7/8] sched/fair: Enable tuning of decay period

[Thara Gopinath]

On 02/04/2020 03:39 AM, Dietmar Eggemann wrote:
> On 03/02/2020 16:55, Peter Zijlstra wrote:
>> On Mon, Feb 03, 2020 at 07:07:57AM -0500, Thara Gopinath wrote:
>>> On 01/28/2020 06:56 PM, Randy Dunlap wrote:
>>>> Hi,
>>>>
>>>> On 1/28/20 2:36 PM, Thara Gopinath wrote:
>>>>> diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt
>>>>> index e35b28e..be4147b 100644
>>>>> --- a/Documentation/admin-guide/kernel-parameters.txt
>>>>> +++ b/Documentation/admin-guide/kernel-parameters.txt
>>>>> @@ -4376,6 +4376,11 @@
>>>>>  			incurs a small amount of overhead in the scheduler
>>>>>  			but is useful for debugging and performance tuning.
>>>>>  
>>>>> +	sched_thermal_decay_shift=
>>>>> +			[KNL, SMP] Set decay shift for thermal pressure signal.
>>>>> +			Format: integer between 0 and 10
>>>>> +			Default is 0.
>>>>> +
>>>>
>>>> That tells an admin [or any reader] almost nothing about this kernel parameter
>>>> or what it does.  And nothing about what unit the value is in.
>>>> Does the value 0 disable this feature?
>>>
>>> Thanks for the review. 0 does not disable "thermal pressure" feature. 0
>>> means the default decay period for averaging PELT signals (which is
>>> usually 32 but configurable) will also be applied for thermal pressure
>>> signal. A shift will shift the default decay period.
>>>
>>> You are right. It needs more explanation here. I will fix it and send v10.
>>
>> Or just send an update for this patch? I'm thinking most of this is
>> looking good.
> 
> I do agree. IMHO, there are just two little things outstanding:
> 
> (1) arch_scale_thermal_pressure() instead  of
>     arch_cpu_thermal_pressure() in v8 4/7

The "scale_" part was discussed in v6. Ionela had suggested that having
"scale" is not suited for this function because "thermal pressure" is
not exactly scaled but subtracted. I actually agree with that.

https://lore.kernel.org/lkml/20191223175005.GA31446@xxxxxxx/

Having said that if everyone feel the same about naming of this
function, I can change it one last time.

> 
> (2) guarding of thermal pressure code in Arm's arch_topology driver  w/
>     CONFIG_HAVE_SCHED_THERMAL_PRESSURE plus disabling it by default for
>     Arm64.
It was enabled by default as per your suggestion in v9.

The patch can be dropped.

I don't understand the need to guard arch_topology with
CONFIG_HAVE_SCHED_THERMAL_PRESSURE. CONFIG_HAVE_SCHED_THERMAL_PRESSURE
is for scheduler to enable/disable averaging of thermal pressure. We
wanted to separate updating and retrieving of instantaneous thermal
pressure from scheduler. Guarding it with
CONFIG_HAVE_SCHED_THERMAL_PRESSURE is to me equivalent to putting back
this whole code in the scheduler framework. I am against it. I also do
not see other arch_ functions guarded similarly.

> 


-- 
Warm Regards
Thara