Re: [PATCH 1/1] sched/fair: ratelimit update to tg->load_avg

[Mathieu Desnoyers <mathieu.desnoyers@efficios.com>]

On 8/24/23 04:01, Aaron Lu wrote:
> On Wed, Aug 23, 2023 at 10:05:31AM -0400, Mathieu Desnoyers wrote:
>> On 8/23/23 02:08, Aaron Lu wrote:
>>> When using sysbench to benchmark Postgres in a single docker instance
>>> with sysbench's nr_threads set to nr_cpu, it is observed there are times
>>> update_cfs_group() and update_load_avg() shows noticeable overhead on
>>> a 2sockets/112core/224cpu Intel Sapphire Rapids(SPR):
>>>
>>>       13.75%    13.74%  [kernel.vmlinux]           [k] update_cfs_group
>>>       10.63%    10.04%  [kernel.vmlinux]           [k] update_load_avg
>>>
>>> Annotate shows the cycles are mostly spent on accessing tg->load_avg
>>> with update_load_avg() being the write side and update_cfs_group() being
>>> the read side. tg->load_avg is per task group and when different tasks
>>> of the same taskgroup running on different CPUs frequently access
>>> tg->load_avg, it can be heavily contended.
>>>
>>> E.g. when running postgres_sysbench on a 2sockets/112cores/224cpus Intel
>>> Sappire Rapids, during a 5s window, the wakeup number is 14millions and
>>> migration number is 11millions and with each migration, the task's load
>>> will transfer from src cfs_rq to target cfs_rq and each change involves
>>> an update to tg->load_avg. Since the workload can trigger as many wakeups
>>> and migrations, the access(both read and write) to tg->load_avg can be
>>> unbound. As a result, the two mentioned functions showed noticeable
>>> overhead. With netperf/nr_client=nr_cpu/UDP_RR, the problem is worse:
>>> during a 5s window, wakeup number is 21millions and migration number is
>>> 14millions; update_cfs_group() costs ~25% and update_load_avg() costs ~16%.
>>>
>>> Reduce the overhead by limiting updates to tg->load_avg to at most once
>>> per ms. After this change, the cost of accessing tg->load_avg is greatly
>>> reduced and performance improved. Detailed test results below.
>>
>> By applying your patch on top of my patchset at:
>>
>> https://lore.kernel.org/lkml/20230822113133.643238-1-mathieu.desnoyers@efficios.com/
>>
>> The combined hackbench results look very promising:
>>
>> (hackbench -g 32 -f 20 --threads --pipe -l 480000 -s 100)
>> (192 cores AMD EPYC 9654 96-Core Processor (over 2 sockets), with hyperthreading)
>>
>> Baseline:                                       49s
>> With L2-ttwu-queue-skip:                        34s (30% speedup)
>> With L2-ttwu-queue-skip + ratelimit-load-avg:   26s (46% speedup)
>>
>> Feel free to apply my:
>>
>> Reviewed-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
>> Tested-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
> 
> Thanks a lot for running this and reviewing the patch.
> I'll add your number and tag in the changelog when sending a new
> version.

Now that I come to think of it, I have comment: why use 
sched_clock_cpu() rather than just read the jiffies value ? AFAIR, 
sched_clock can be slower than needed when read from a "remote" cpu on 
architectures that have an unsynchronized tsc.

Considering that you only need a time reference more or less accurate at 
the millisecond level, I suspect that jiffies is what you are looking 
for here. This is what the NUMA balance code and rseq mm_cid use to 
execute work every N milliseconds.

Thanks,

Mathieu

> 
> Regards,
> Aaron
>   
>>>
>>> ==============================
>>> postgres_sysbench on SPR:
>>> 25%
>>> base:   42382±19.8%
>>> patch:  50174±9.5%  (noise)
>>>
>>> 50%
>>> base:   67626±1.3%
>>> patch:  67365±3.1%  (noise)
>>>
>>> 75%
>>> base:   100216±1.2%
>>> patch:  112470±0.1% +12.2%
>>>
>>> 100%
>>> base:    93671±0.4%
>>> patch:  113563±0.2% +21.2%
>>>
>>> ==============================
>>> hackbench on ICL:
>>> group=1
>>> base:    114912±5.2%
>>> patch:   117857±2.5%  (noise)
>>>
>>> group=4
>>> base:    359902±1.6%
>>> patch:   361685±2.7%  (noise)
>>>
>>> group=8
>>> base:    461070±0.8%
>>> patch:   491713±0.3% +6.6%
>>>
>>> group=16
>>> base:    309032±5.0%
>>> patch:   378337±1.3% +22.4%
>>>
>>> =============================
>>> hackbench on SPR:
>>> group=1
>>> base:    100768±2.9%
>>> patch:   103134±2.9%  (noise)
>>>
>>> group=4
>>> base:    413830±12.5%
>>> patch:   378660±16.6% (noise)
>>>
>>> group=8
>>> base:    436124±0.6%
>>> patch:   490787±3.2% +12.5%
>>>
>>> group=16
>>> base:    457730±3.2%
>>> patch:   680452±1.3% +48.8%
>>>
>>> ============================
>>> netperf/udp_rr on ICL
>>> 25%
>>> base:    114413±0.1%
>>> patch:   115111±0.0% +0.6%
>>>
>>> 50%
>>> base:    86803±0.5%
>>> patch:   86611±0.0%  (noise)
>>>
>>> 75%
>>> base:    35959±5.3%
>>> patch:   49801±0.6% +38.5%
>>>
>>> 100%
>>> base:    61951±6.4%
>>> patch:   70224±0.8% +13.4%
>>>
>>> ===========================
>>> netperf/udp_rr on SPR
>>> 25%
>>> base:   104954±1.3%
>>> patch:  107312±2.8%  (noise)
>>>
>>> 50%
>>> base:    55394±4.6%
>>> patch:   54940±7.4%  (noise)
>>>
>>> 75%
>>> base:    13779±3.1%
>>> patch:   36105±1.1% +162%
>>>
>>> 100%
>>> base:     9703±3.7%
>>> patch:   28011±0.2% +189%
>>>
>>> ==============================================
>>> netperf/tcp_stream on ICL (all in noise range)
>>> 25%
>>> base:    43092±0.1%
>>> patch:   42891±0.5%
>>>
>>> 50%
>>> base:    19278±14.9%
>>> patch:   22369±7.2%
>>>
>>> 75%
>>> base:    16822±3.0%
>>> patch:   17086±2.3%
>>>
>>> 100%
>>> base:    18216±0.6%
>>> patch:   18078±2.9%
>>>
>>> ===============================================
>>> netperf/tcp_stream on SPR (all in noise range)
>>> 25%
>>> base:    34491±0.3%
>>> patch:   34886±0.5%
>>>
>>> 50%
>>> base:    19278±14.9%
>>> patch:   22369±7.2%
>>>
>>> 75%
>>> base:    16822±3.0%
>>> patch:   17086±2.3%
>>>
>>> 100%
>>> base:    18216±0.6%
>>> patch:   18078±2.9%
>>>
>>> Reported-by: Nitin Tekchandani <nitin.tekchandani@intel.com>
>>> Suggested-by: Vincent Guittot <vincent.guittot@linaro.org>
>>> Signed-off-by: Aaron Lu <aaron.lu@intel.com>
>>> Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
>>> ---
>>>    kernel/sched/fair.c  | 13 ++++++++++++-
>>>    kernel/sched/sched.h |  1 +
>>>    2 files changed, 13 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>>> index c28206499a3d..a5462d1fcc48 100644
>>> --- a/kernel/sched/fair.c
>>> +++ b/kernel/sched/fair.c
>>> @@ -3664,7 +3664,8 @@ static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
>>>     */
>>>    static inline void update_tg_load_avg(struct cfs_rq *cfs_rq)
>>>    {
>>> -	long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
>>> +	long delta;
>>> +	u64 now;
>>>    	/*
>>>    	 * No need to update load_avg for root_task_group as it is not used.
>>> @@ -3672,9 +3673,19 @@ static inline void update_tg_load_avg(struct cfs_rq *cfs_rq)
>>>    	if (cfs_rq->tg == &root_task_group)
>>>    		return;
>>> +	/*
>>> +	 * For migration heavy workload, access to tg->load_avg can be
>>> +	 * unbound. Limit the update rate to at most once per ms.
>>> +	 */
>>> +	now = sched_clock_cpu(cpu_of(rq_of(cfs_rq)));
>>> +	if (now - cfs_rq->last_update_tg_load_avg < NSEC_PER_MSEC)
>>> +		return;
>>> +
>>> +	delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
>>>    	if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
>>>    		atomic_long_add(delta, &cfs_rq->tg->load_avg);
>>>    		cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
>>> +		cfs_rq->last_update_tg_load_avg = now;
>>>    	}
>>>    }
>>> diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
>>> index 6a8b7b9ed089..52ee7027def9 100644
>>> --- a/kernel/sched/sched.h
>>> +++ b/kernel/sched/sched.h
>>> @@ -593,6 +593,7 @@ struct cfs_rq {
>>>    	} removed;
>>>    #ifdef CONFIG_FAIR_GROUP_SCHED
>>> +	u64			last_update_tg_load_avg;
>>>    	unsigned long		tg_load_avg_contrib;
>>>    	long			propagate;
>>>    	long			prop_runnable_sum;
>>
>> -- 
>> Mathieu Desnoyers
>> EfficiOS Inc.
>> https://www.efficios.com
>>

-- 
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com