Re: [RFC PATCH] sched/fair: Interleave cfs bandwidth timers for improved single thread performance at low utilization

[Benjamin Segall]

shrikanth hegde <sshegde@xxxxxxxxxxxxxxxxxx> writes:

> On 2/20/23 11:08 PM, Peter Zijlstra wrote:
>> On Tue, Feb 14, 2023 at 08:54:09PM +0530, shrikanth hegde wrote:
>> 
>>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>>> index ff4dbbae3b10..7b69c329e05d 100644
>>> --- a/kernel/sched/fair.c
>>> +++ b/kernel/sched/fair.c
>>> @@ -5939,14 +5939,25 @@ static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
>>>
>>>  void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
>>>  {
>>> -	lockdep_assert_held(&cfs_b->lock);
>>> +	struct hrtimer *period_timer = &cfs_b->period_timer;
>>> +	s64 incr = ktime_to_ns(cfs_b->period) / 10;
>>> +	ktime_t delta;
>>> +	u64 orun = 1;
>>>
>>> +	lockdep_assert_held(&cfs_b->lock);
>>>  	if (cfs_b->period_active)
>>>  		return;
>>>
>>>  	cfs_b->period_active = 1;
>>> -	hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
>>> -	hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
>>> +	delta = ktime_sub(period_timer->base->get_time(),
>>> +			hrtimer_get_expires(period_timer));
>>> +	if (unlikely(delta >= cfs_b->period)) {
>>> +		orun = ktime_divns(delta, incr);
>>> +		hrtimer_add_expires_ns(period_timer, incr * orun);
>>> +	}
>>> +
>>> +	hrtimer_forward_now(period_timer, cfs_b->period);
>>> +	hrtimer_start_expires(period_timer, HRTIMER_MODE_ABS_PINNED);
>>>  }
>> 
>> What kind of mad hackery is this? Why can't you do the sane thing and
>> initialize the timer at !0 in init_cfs_bandwidth(), then any of the
>> forwards will stay in period -- as they should.
>> 
>> Please, go re-read Thomas's email.
>
> Thank you Peter for taking a look and review.
> we can scrap this implementation and switch to the one you suggested.
> All we need is to initialize the offset. 
>
> Only reason was the way i had implemented. This implementation couldn't be
> fit into init_cfs_bandwidth as timers would align if the cgroups are 
> created together and continue to align forever. 
>
>> 
>> *completely* untested...
>> 
>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>> index 7c46485d65d7..4d6ea76096dc 100644
>> --- a/kernel/sched/fair.c
>> +++ b/kernel/sched/fair.c
>> @@ -5915,6 +5915,7 @@ void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
>> 
>>  	INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
>>  	hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
>> +	cfs_b->period_timer.node.expires = get_random_u32_below(cfs_b->period);
>
> This approach/implementation is better as the random function provides uniform  
> distribution. Had to modify this a bit to make it work.  Along with setting     
> setting node.expires, we need to set _softexpires as well. Which is what        
> hrtimer_set_expires does.
>
> Here are the similar numbers again.
> 8 Core system with SMT=8. Total of 64 CPU                                       
> Workload: stress-ng --cpu=32 --cpu-ops=50000                                    
>                                                                                 
>            6.2-rc6                     |   with patch                           
> 8Core   1CG    power    2CG     power  |  1CG    power  2CG    power           
>         27.5    80.6    40      90     |  27.3    82    32.3    104             
>         27.5    81      40.2    91     |  27.5    81    38.7     96             
>         27.7    80      40.1    89     |  27.6    80    29.7    115             
>         27.7    80.1    40.3    94     |  27.6    80    31.5    105   
>
> Will collect some more benchmarks numbers w.r.t to performance.
>
>
>>  	cfs_b->period_timer.function = sched_cfs_period_timer;
>>  	hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
>>  	cfs_b->slack_timer.function = sched_cfs_slack_timer;
>
> This below patch worked. 
> Does the below patch look okay? shall i send the [PATCH V1] with this
> change?

Yeah, this design makes way more sense.

>
> Question. 
> Should we skip this adding the offset for root_task_group?

The value should never come up, so it's just a question of if it's fine
to call get_random_* in early contexts, which I don't know offhand.

>
>
> ---
>  kernel/sched/fair.c | 3 +++
>  1 file changed, 3 insertions(+)
>
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index ff4dbbae3b10..6448533178af 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -5923,6 +5923,9 @@ void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
>  	INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
>  	hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
>  	cfs_b->period_timer.function = sched_cfs_period_timer;
> +	/* Add a random offset so that timers interleave */
> +	hrtimer_set_expires(&cfs_b->period_timer, get_random_u32_below(cfs_b->period));
> +
>  	hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
>  	cfs_b->slack_timer.function = sched_cfs_slack_timer;
>  	cfs_b->slack_started = false;