Date: Mon, 13 May 2024 09:37:07 +0530
From: K Prateek Nayak <kprateek.nayak@....com>
To: Chen Yu <yu.c.chen@...el.com>
Cc: Peter Zijlstra <peterz@...radead.org>, mingo@...hat.com,
 juri.lelli@...hat.com, vincent.guittot@...aro.org, dietmar.eggemann@....com,
 rostedt@...dmis.org, bsegall@...gle.com, mgorman@...e.de,
 bristot@...hat.com, vschneid@...hat.com, linux-kernel@...r.kernel.org,
 wuyun.abel@...edance.com, tglx@...utronix.de, efault@....de,
 tim.c.chen@...el.com, yu.c.chen.y@...il.com
Subject: Re: [RFC][PATCH 10/10] sched/eevdf: Use sched_attr::sched_runtime to
 set request/slice suggestion

Hello Chenyu,

On 5/7/2024 8:45 PM, Chen Yu wrote:
> On 2024-04-05 at 12:28:04 +0200, Peter Zijlstra wrote:
>> Allow applications to directly set a suggested request/slice length using
>> sched_attr::sched_runtime.
>>
>> The implementation clamps the value to: 0.1[ms] <= slice <= 100[ms]
>> which is 1/10 the size of HZ=1000 and 10 times the size of HZ=100.
>>
>> Applications should strive to use their periodic runtime at a high
>> confidence interval (95%+) as the target slice. Using a smaller slice
>> will introduce undue preemptions, while using a larger value will
>> increase latency.
>>
>> Signed-off-by: Peter Zijlstra (Intel) <peterz@...radead.org>
>>
> 
> Is it possible to leverage this task slice to do better task wakeup placement?
> The idea is that, the smaller the slice the wakee has, the less idle CPU it
> should scan. This can reduce wake latency and inhibit costly task migration,
> especially on large systems.
> 
> We did some experiments and got some performance improvements:
> 
> 
> From 9cb806476586d7048fcbd0f66d0101f0dbb8fd2b Mon Sep 17 00:00:00 2001
> From: Chen Yu <yu.c.chen@...el.com>
> Date: Tue, 7 May 2024 22:36:29 +0800
> Subject: [RFC PATCH] sched/eevdf: Use customized slice to reduce wakeup latency
>  and inhibit task migration
> 
> Problem 1:
> The overhead of task migration is high on many-core system. The overhead
> brings performance penalty due to broken cache locality/higher cache-to-cache
> latency.
> 
> Problem 2:
> During wakeup, the time spent on searching for an idle CPU is costly on
> many-core system. Besides, access to other CPU's rq statistics brings
> cace contention:
> 
> available_idle_cpu(cpu) -> idle_cpu(cpu) -> {rq->curr, rq->nr_running}
> 
> Although SIS_UTIL throttles the scan depth based on system utilization,
> there is requirement to further limit the scan depth for specific workload,
> especially for short duration wakee.
> 
> Now we have the interface to customize the request/slice. The smaller the
> slice is, the earlier the task can be picked up, and the lower wakeup latency
> the task expects. Leverage the wakee's slice to further throttle the
> idle CPU scan depth - the shorter slice, the less CPUs to scan.
> 
> Test on 240 CPUs, 2 sockets system. With SNC(sub-numa-cluster) enabled,
> each LLC domain has 60 CPUs. There is noticeable improvement of netperf.
> (With SNC disabled, more improvements should be seen because C2C is higher)
> 
> The global slice is 3 msec(sysctl_sched_base_slice) by default on my ubuntu
> 22.04, and the customized slice is set to 0.1 msec for both netperf and netserver:
> 
> for i in $(seq 1 $job); do
> 	netperf_slice -e 100000 -4 -H 127.0.01 -t TCP_RR -c -C -l 100 &
> done
> 
> case            	load    	baseline(std%)	compare%( std%)
> TCP_RR          	60-threads	 1.00 (  1.60)	 +0.35 (  1.73)
> TCP_RR          	120-threads	 1.00 (  1.34)	 -0.96 (  1.24)
> TCP_RR          	180-threads	 1.00 (  1.59)	+92.20 (  4.24)
> TCP_RR          	240-threads	 1.00 (  9.71)	+43.11 (  2.97)
> 
> Suggested-by: Tim Chen <tim.c.chen@...el.com>
> Signed-off-by: Chen Yu <yu.c.chen@...el.com>
> ---
>  kernel/sched/fair.c     | 23 ++++++++++++++++++++---
>  kernel/sched/features.h |  1 +
>  2 files changed, 21 insertions(+), 3 deletions(-)
> 
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index edc23f6588a3..f269ae7d6e24 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -7368,6 +7368,24 @@ static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd
>  
>  #endif /* CONFIG_SCHED_SMT */
>  
> +/*
> + * Scale the scan number of idle CPUs according to customized
> + * wakee's slice. The smaller the slice is, the earlier the task
> + * wants be picked up, thus the lower wakeup latency the task expects.
> + * The baseline is the global sysctl_sched_base_slice. Task slice
> + * smaller than the global one would shrink the scan number.
> + */
> +static int adjust_idle_scan(struct task_struct *p, int nr)
> +{
> +	if (!sched_feat(SIS_FAST))
> +		return nr;
> +
> +	if (!p->se.custom_slice || p->se.slice >= sysctl_sched_base_slice)
> +		return nr;
> +
> +	return div_u64(nr * p->se.slice, sysctl_sched_base_slice);
> +}
> +
>  /*
>   * Scan the LLC domain for idle CPUs; this is dynamically regulated by
>   * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
> @@ -7384,10 +7402,9 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool
>  	if (sched_feat(SIS_UTIL)) {
>  		sd_share = rcu_dereference(per_cpu(sd_llc_shared, target));
>  		if (sd_share) {
> -			/* because !--nr is the condition to stop scan */> -			nr = READ_ONCE(sd_share->nr_idle_scan) + 1;
> +			nr = adjust_idle_scan(p, READ_ONCE(sd_share->nr_idle_scan));
>  			/* overloaded LLC is unlikely to have idle cpu/core */
> -			if (nr == 1)
> +			if (nr <= 0)

I was wondering if this would preserve the current behavior with
SIS_FAST toggled off? Since the implementation below still does a
"--nr <= 0" , wouldn't it effectively visit one CPU less overall now?

Have you tried something similar to the below hunk?

	/* because !--nr is the condition to stop scan */
	nr = adjust_idle_scan(p, READ_ONCE(sd_share->nr_idle_scan)) + 1;
	if (nr == 1)
		return -1;

I agree with Mike that looking at slice to limit scan-depth seems odd.
My experience with netperf is that the workload cares more about the
server-client being co-located on the closest cache domain and by
limiting scan-depth using slice, this is indirectly achieved since all
the wakeups carry the WF_SYNc flag.

P.S. have you tried using the slice in __select_idle_cpu()? Similar to
sched_idle_cpu() check, perhaps an additional sched_preempt_short_cpu()
which compares rq->curr->se.slice with the waking task's slice and
returs that cpu if SIS_SHORT can help run the workload quicker? Note:
This will not work if the SIS scan itself is the largest overhead in the
wakeup cycle and not the task placement itself. Previously during
SIS_UTIL testing, to measure the overheads of scan vs placement, we
would do a full scan but return the result that SIS_UTIL would have
returned to determine the overhead of the search itself.

>  				return -1;
>  		}
>  	}
> diff --git a/kernel/sched/features.h b/kernel/sched/features.h
> index 143f55df890b..176324236018 100644
> --- a/kernel/sched/features.h
> +++ b/kernel/sched/features.h
> @@ -50,6 +50,7 @@ SCHED_FEAT(TTWU_QUEUE, true)
>   * When doing wakeups, attempt to limit superfluous scans of the LLC domain.
>   */
>  SCHED_FEAT(SIS_UTIL, true)
> +SCHED_FEAT(SIS_FAST, true)
>  
>  /*
>   * Issue a WARN when we do multiple update_rq_clock() calls

--
Thanks and Regards,
Prateek

Powered by blists - more mailing lists