Received-SPF: softfail (google.com: domain of transitioning linux-kernel-owner@vger.kernel.org does not designate 23.128.96.19 as permitted sender) client-ip=23.128.96.19;
From:   Valentin Schneider <vschneid@redhat.com>
To:     Tianchen Ding <dtcccc@linux.alibaba.com>,
        Mel Gorman <mgorman@suse.de>
Cc:     Ingo Molnar <mingo@redhat.com>,
        Peter Zijlstra <peterz@infradead.org>,
        Juri Lelli <juri.lelli@redhat.com>,
        Vincent Guittot <vincent.guittot@linaro.org>,
        Dietmar Eggemann <dietmar.eggemann@arm.com>,
        Steven Rostedt <rostedt@goodmis.org>,
        Ben Segall <bsegall@google.com>,
        Daniel Bristot de Oliveira <bristot@redhat.com>,
        linux-kernel@vger.kernel.org
Subject: Re: [PATCH v2] sched: Queue task on wakelist in the same llc if the
 wakee cpu is idle
In-Reply-To: <a5d04748-b34b-3b92-fb1d-bf85c2019cc3@linux.alibaba.com>
References: <20220527090544.527411-1-dtcccc@linux.alibaba.com>
 <xhsmhleuj7zve.mognet@vschneid.remote.csb>
 <1d0eb8f4-e474-86a9-751a-7c2e1788df85@linux.alibaba.com>
 <xhsmhilpl9azq.mognet@vschneid.remote.csb> <20220531135532.GA3332@suse.de>
 <xhsmhfskp8zlg.mognet@vschneid.remote.csb>
 <a5d04748-b34b-3b92-fb1d-bf85c2019cc3@linux.alibaba.com>
Date:   Wed, 01 Jun 2022 11:58:48 +0100
Message-ID: <xhsmhczfs8xav.mognet@vschneid.remote.csb>
MIME-Version: 1.0
Content-Type: text/plain
Precedence: bulk

On 01/06/22 13:54, Tianchen Ding wrote:
> On 2022/5/31 23:56, Valentin Schneider wrote:
>
>> Thanks!
>> 
>> So I'm thinking we could first make that into
>> 
>> 	if ((wake_flags & WF_ON_CPU) && !cpu_rq(cpu)->nr_running)
>> 
>> Then building on this, we can generalize using the wakelist to any remote
>> idle CPU (which on paper isn't as much as a clear win as just WF_ON_CPU,
>> depending on how deeply idle the CPU is...)
>> 
>> We need the cpu != this_cpu check, as that's currently served by the
>> WF_ON_CPU check (AFAIU we can only observe p->on_cpu in there for remote
>> tasks).
>> 
>> ---
>> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
>> index 66c4e5922fe1..60038743f2f1 100644
>> --- a/kernel/sched/core.c
>> +++ b/kernel/sched/core.c
>> @@ -3830,13 +3830,20 @@ static inline bool ttwu_queue_cond(int cpu, int wake_flags)
>>   	if (!cpus_share_cache(smp_processor_id(), cpu))
>>   		return true;
>>   
>> +	if (cpu == smp_processor_id())
>> +		return false;
>> +
>>   	/*
>>   	 * If the task is descheduling and the only running task on the
>>   	 * CPU then use the wakelist to offload the task activation to
>>   	 * the soon-to-be-idle CPU as the current CPU is likely busy.
>>   	 * nr_running is checked to avoid unnecessary task stacking.
>> +	 *
>> +	 * Note that we can only get here with (wakee) p->on_rq=0,
>> +	 * p->on_cpu can be whatever, we've done the dequeue, so
>> +	 * the wakee has been accounted out of ->nr_running
>>   	 */
>> -	if ((wake_flags & WF_ON_CPU) && cpu_rq(cpu)->nr_running <= 1)
>> +	if (!cpu_rq(cpu)->nr_running)
>>   		return true;
>>   
>>   	return false;
>
> Hi Valentin. I've done a simple unixbench test (Pipe-based Context 
> Switching) on my x86 machine with full threads (104).
>
>               old            patch1           patch1+patch2
> score       7825.4     7500(more)-8000          9061.6
>
> patch1: use !cpu_rq(cpu)->nr_running instead of cpu_rq(cpu)->nr_running <= 1
> patch2: ignore WF_ON_CPU check
>
> The score of patch1 is not stable. I've tested for many times and the 
> score is floating between about 7500-8000 (more at 7500).
>
> patch1 means more strict limit on using wakelist. But it may cause 
> performance regression.
>
> It seems that, using wakelist properly can help improve wakeup 
> performance, but using it too much may cause more IPIs. It's a trade-off 
> about how strict the ttwu_queue_cond() is.
>
> Anyhow, I think patch2 should be a pure improvement. What's your idea?

Thanks for separately testing these two.

I take it the results for patch1 are noticeably more swingy than the
baseline? (FWIW boxplots are usually a nice way to summarize those sort of
results).

WF_ON_CPU && nr_running == 1 means the wakee is scheduling out *and* there
is another task queued, I'm guessing that's relatively common in your
unixbench scenario...

Either way, I think we want to keep the two changes separate for the sake
of testing and bisecting.