Message-ID: <0B3518AC-3F4A-4E89-974A-E7DFD446D7C2@tencent.com>
Date:   Fri, 24 Jul 2020 02:29:24 +0000
From:   benbjiang(蒋彪) <benbjiang@...cent.com>
To:     "Li, Aubrey" <aubrey.li@...ux.intel.com>
CC:     Aubrey Li <aubrey.intel@...il.com>,
        Vineeth Remanan Pillai <vpillai@...italocean.com>,
        Nishanth Aravamudan <naravamudan@...italocean.com>,
        Julien Desfossez <jdesfossez@...italocean.com>,
        Peter Zijlstra <peterz@...radead.org>,
        "Tim Chen" <tim.c.chen@...ux.intel.com>,
        "mingo@...nel.org" <mingo@...nel.org>,
        "tglx@...utronix.de" <tglx@...utronix.de>,
        "pjt@...gle.com" <pjt@...gle.com>,
        "torvalds@...ux-foundation.org" <torvalds@...ux-foundation.org>,
        Aubrey Li <aubrey.li@...el.com>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
        "subhra.mazumdar@...cle.com" <subhra.mazumdar@...cle.com>,
        "fweisbec@...il.com" <fweisbec@...il.com>,
        "keescook@...omium.org" <keescook@...omium.org>,
        "kerrnel@...gle.com" <kerrnel@...gle.com>,
        Phil Auld <pauld@...hat.com>, Aaron Lu <aaron.lwe@...il.com>,
        Valentin Schneider <valentin.schneider@....com>,
        "Mel Gorman" <mgorman@...hsingularity.net>,
        Pawan Gupta <pawan.kumar.gupta@...ux.intel.com>,
        Paolo Bonzini <pbonzini@...hat.com>,
        Joel Fernandes <joelaf@...gle.com>,
        "joel@...lfernandes.org" <joel@...lfernandes.org>,
        "vineethrp@...il.com" <vineethrp@...il.com>,
        "Chen Yu" <yu.c.chen@...el.com>,
        Christian Brauner <christian.brauner@...ntu.com>
Subject: Re: [RFC PATCH 11/16] sched: migration changes for core
 scheduling(Internet mail)

Hi,

> On Jul 24, 2020, at 10:05 AM, Li, Aubrey <aubrey.li@...ux.intel.com> wrote:
> 
> On 2020/7/24 9:26, benbjiang(蒋彪) wrote:
>> Hi,
>> 
>>> On Jul 24, 2020, at 7:43 AM, Aubrey Li <aubrey.intel@...il.com> wrote:
>>> 
>>> On Thu, Jul 23, 2020 at 4:28 PM benbjiang(蒋彪) <benbjiang@...cent.com> wrote:
>>>> 
>>>> Hi,
>>>> 
>>>>> On Jul 23, 2020, at 4:06 PM, Li, Aubrey <aubrey.li@...ux.intel.com> wrote:
>>>>> 
>>>>> On 2020/7/23 15:47, benbjiang(蒋彪) wrote:
>>>>>> Hi,
>>>>>> 
>>>>>>> On Jul 23, 2020, at 1:39 PM, Li, Aubrey <aubrey.li@...ux.intel.com> wrote:
>>>>>>> 
>>>>>>> On 2020/7/23 12:23, benbjiang(蒋彪) wrote:
>>>>>>>> Hi,
>>>>>>>>> On Jul 23, 2020, at 11:35 AM, Li, Aubrey <aubrey.li@...ux.intel.com> wrote:
>>>>>>>>> 
>>>>>>>>> On 2020/7/23 10:42, benbjiang(蒋彪) wrote:
>>>>>>>>>> Hi,
>>>>>>>>>> 
>>>>>>>>>>> On Jul 23, 2020, at 9:57 AM, Li, Aubrey <aubrey.li@...ux.intel.com> wrote:
>>>>>>>>>>> 
>>>>>>>>>>> On 2020/7/22 22:32, benbjiang(蒋彪) wrote:
>>>>>>>>>>>> Hi,
>>>>>>>>>>>> 
>>>>>>>>>>>>> On Jul 22, 2020, at 8:13 PM, Li, Aubrey <aubrey.li@...ux.intel.com> wrote:
>>>>>>>>>>>>> 
>>>>>>>>>>>>> On 2020/7/22 16:54, benbjiang(蒋彪) wrote:
>>>>>>>>>>>>>> Hi, Aubrey,
>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>> On Jul 1, 2020, at 5:32 AM, Vineeth Remanan Pillai <vpillai@...italocean.com> wrote:
>>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>> From: Aubrey Li <aubrey.li@...el.com>
>>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>> - Don't migrate if there is a cookie mismatch
>>>>>>>>>>>>>>> Load balance tries to move task from busiest CPU to the
>>>>>>>>>>>>>>> destination CPU. When core scheduling is enabled, if the
>>>>>>>>>>>>>>> task's cookie does not match with the destination CPU's
>>>>>>>>>>>>>>> core cookie, this task will be skipped by this CPU. This
>>>>>>>>>>>>>>> mitigates the forced idle time on the destination CPU.
>>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>> - Select cookie matched idle CPU
>>>>>>>>>>>>>>> In the fast path of task wakeup, select the first cookie matched
>>>>>>>>>>>>>>> idle CPU instead of the first idle CPU.
>>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>> - Find cookie matched idlest CPU
>>>>>>>>>>>>>>> In the slow path of task wakeup, find the idlest CPU whose core
>>>>>>>>>>>>>>> cookie matches with task's cookie
>>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>> - Don't migrate task if cookie not match
>>>>>>>>>>>>>>> For the NUMA load balance, don't migrate task to the CPU whose
>>>>>>>>>>>>>>> core cookie does not match with task's cookie
>>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>> Signed-off-by: Aubrey Li <aubrey.li@...ux.intel.com>
>>>>>>>>>>>>>>> Signed-off-by: Tim Chen <tim.c.chen@...ux.intel.com>
>>>>>>>>>>>>>>> Signed-off-by: Vineeth Remanan Pillai <vpillai@...italocean.com>
>>>>>>>>>>>>>>> ---
>>>>>>>>>>>>>>> kernel/sched/fair.c  | 64 ++++++++++++++++++++++++++++++++++++++++----
>>>>>>>>>>>>>>> kernel/sched/sched.h | 29 ++++++++++++++++++++
>>>>>>>>>>>>>>> 2 files changed, 88 insertions(+), 5 deletions(-)
>>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>>>>>>>>>>>>>>> index d16939766361..33dc4bf01817 100644
>>>>>>>>>>>>>>> --- a/kernel/sched/fair.c
>>>>>>>>>>>>>>> +++ b/kernel/sched/fair.c
>>>>>>>>>>>>>>> @@ -2051,6 +2051,15 @@ static void task_numa_find_cpu(struct task_numa_env *env,
>>>>>>>>>>>>>>>           if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
>>>>>>>>>>>>>>>                   continue;
>>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>> +#ifdef CONFIG_SCHED_CORE
>>>>>>>>>>>>>>> +           /*
>>>>>>>>>>>>>>> +            * Skip this cpu if source task's cookie does not match
>>>>>>>>>>>>>>> +            * with CPU's core cookie.
>>>>>>>>>>>>>>> +            */
>>>>>>>>>>>>>>> +           if (!sched_core_cookie_match(cpu_rq(cpu), env->p))
>>>>>>>>>>>>>>> +                   continue;
>>>>>>>>>>>>>>> +#endif
>>>>>>>>>>>>>>> +
>>>>>>>>>>>>>>>           env->dst_cpu = cpu;
>>>>>>>>>>>>>>>           if (task_numa_compare(env, taskimp, groupimp, maymove))
>>>>>>>>>>>>>>>                   break;
>>>>>>>>>>>>>>> @@ -5963,11 +5972,17 @@ find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this
>>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>>   /* Traverse only the allowed CPUs */
>>>>>>>>>>>>>>>   for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
>>>>>>>>>>>>>>> +           struct rq *rq = cpu_rq(i);
>>>>>>>>>>>>>>> +
>>>>>>>>>>>>>>> +#ifdef CONFIG_SCHED_CORE
>>>>>>>>>>>>>>> +           if (!sched_core_cookie_match(rq, p))
>>>>>>>>>>>>>>> +                   continue;
>>>>>>>>>>>>>>> +#endif
>>>>>>>>>>>>>>> +
>>>>>>>>>>>>>>>           if (sched_idle_cpu(i))
>>>>>>>>>>>>>>>                   return i;
>>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>>           if (available_idle_cpu(i)) {
>>>>>>>>>>>>>>> -                   struct rq *rq = cpu_rq(i);
>>>>>>>>>>>>>>>                   struct cpuidle_state *idle = idle_get_state(rq);
>>>>>>>>>>>>>>>                   if (idle && idle->exit_latency < min_exit_latency) {
>>>>>>>>>>>>>>>                           /*
>>>>>>>>>>>>>>> @@ -6224,8 +6239,18 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int t
>>>>>>>>>>>>>>>   for_each_cpu_wrap(cpu, cpus, target) {
>>>>>>>>>>>>>>>           if (!--nr)
>>>>>>>>>>>>>>>                   return -1;
>>>>>>>>>>>>>>> -           if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
>>>>>>>>>>>>>>> -                   break;
>>>>>>>>>>>>>>> +
>>>>>>>>>>>>>>> +           if (available_idle_cpu(cpu) || sched_idle_cpu(cpu)) {
>>>>>>>>>>>>>>> +#ifdef CONFIG_SCHED_CORE
>>>>>>>>>>>>>>> +                   /*
>>>>>>>>>>>>>>> +                    * If Core Scheduling is enabled, select this cpu
>>>>>>>>>>>>>>> +                    * only if the process cookie matches core cookie.
>>>>>>>>>>>>>>> +                    */
>>>>>>>>>>>>>>> +                   if (sched_core_enabled(cpu_rq(cpu)) &&
>>>>>>>>>>>>>>> +                       p->core_cookie == cpu_rq(cpu)->core->core_cookie)
>>>>>>>>>>>>>> Why not also add similar logic in select_idle_smt to reduce forced-idle? :)
>>>>>>>>>>>>> We hit select_idle_smt after we scaned the entire LLC domain for idle cores
>>>>>>>>>>>>> and idle cpus and failed,so IMHO, an idle smt is probably a good choice under
>>>>>>>>>>>>> this scenario.
>>>>>>>>>>>> 
>>>>>>>>>>>> AFAIC, selecting idle sibling with unmatched cookie will cause unnecessary fored-idle, unfairness and latency, compared to choosing *target* cpu.
>>>>>>>>>>> Choosing target cpu could increase the runnable task number on the target runqueue, this
>>>>>>>>>>> could trigger busiest->nr_running > 1 logic and makes the idle sibling trying to pull but
>>>>>>>>>>> not success(due to cookie not match). Putting task to the idle sibling is relatively stable IMHO.
>>>>>>>>>> 
>>>>>>>>>> I’m afraid that *unsuccessful* pullings between smts would not result in unstableness, because
>>>>>>>>>> the load-balance always do periodicly , and unsuccess means nothing happen.
>>>>>>>>> unsuccess pulling means more unnecessary overhead in load balance.
>>>>>>>>> 
>>>>>>>>>> On the contrary, unmatched sibling tasks running concurrently could bring forced-idle to each other repeatedly,
>>>>>>>>>> Which is more unstable, and more costly when pick_next_task for all siblings.
>>>>>>>>> Not worse than two tasks ping-pong on the same target run queue I guess, and better if
>>>>>>>>> - task1(cookie A) is running on the target, and task2(cookie B) in the runqueue,
>>>>>>>>> - task3(cookie B) coming
>>>>>>>>> 
>>>>>>>>> If task3 chooses target's sibling, it could have a chance to run concurrently with task2.
>>>>>>>>> But if task3 chooses target, it will wait for next pulling luck of load balancer
>>>>>>>> That’s more interesting. :)
>>>>>>>> Distributing different cookie tasks onto different cpus(or cpusets) could be the *ideal stable status* we want, as I understood.
>>>>>>>> Different cookie tasks running on sibling smts could hurt performance, and that should be avoided with best effort.
>>>>>>> We already tried to avoid when we scan idle cores and idle cpus in llc domain.
>>>>>> 
>>>>>> I’m afraid that’s not enough either, :)
>>>>>> 1. Scanning Idle cpus is not a full scan, there is limit according to scan cost.
>>>>>> 2. That's only trying at the *core/cpu* level, *SMT* level should be considered too.
>>>>>> 
>>>>>>> 
>>>>>>>> For above case, selecting idle sibling cpu can improve the concurrency indeed, but it decrease the imbalance for load-balancer.
>>>>>>>> In that case, load-balancer could not notice the imbalance, and would do nothing to improve the unmatched situation.
>>>>>>>> On the contrary, choosing the *target* cpu could enhance the imbalance, and load-balancer could try to pull unmatched task away,
>>>>>>> Pulling away to where needs another bunch of elaboration.
>>>>>> 
>>>>>> Still with the SMT2+3tasks case,
>>>>>> if *idle sibling* chosen,
>>>>>> Smt1’s load = task1+task2, smt2’s load = task3. Task3 will run intermittently because of forced-idle,
>>>>>> so smt2’s real load could low enough, that it could not be pulled away forever. That’s indeed a stable state,
>>>>>> but with performance at a discount.
>>>>>> 
>>>>>> If *target sibling* chose,
>>>>>> Smt1’s load = task1+task2+task3, smt2’s load=0. It’s a obvious imbalance, and load-balancer will pick a task to pull,
>>>>>> 1. If task1(cookie A) picked, that’s done for good.
>>>>>> 2. If task2(cookie B) or task3(cookie B) picked, that’s ok too, the rest task(cookie B) could be pulled away at next balance(maybe need to improve the pulling to tend to pull matched task more aggressively).
>>>>>> And then, we may reach a more stable state *globally* without performance discount.
>>>>> 
>>>>> I'm not sure what you mean pulled away,
>>>> I mean pulled away by other cpus, may be triggered by idle balance or periodic balance on other cpus.
>>>> 
>>>>> - if you mean pulled away from this core, cookieA in idle sibling case can be
>>>>> pulled away too.
>>>> Yep, cookieA(task1) in idle sibling case could be pulled away, but
>>>> cookieB(task3) on the smt2 could never get the chance being pulled
>>>> away(unless being waken up).
>>>> If cookieA(task1) failed being pulled(cookieB(task2) on smt1 may be pulled,
>>>> 50% chance), cookieA(task1) and cookieB(task3) would reach the stable state
>>>> with performance discount.
>>>> 
>>> If you meant pulled away from this core, I didn't see how two cases are
>>> different either. For example, when task2(cookieB) runs on SMT1, task3
>>> cookieb can be pulled to SMT2. and when task1(cookieA) switch onto SMT1,
>>> task2(cookieB) can be pulled away by other cpus, too.
>> That’s the case only if SMT2’s pulling happens when task2(cookieB) is running
>> on SMT1, which depends on,
>> 1. Smt2 not entering tickless or nohz_balancer_kick picks smt2 before other
>> cpu’s pulling, may be unlikely. :)
>> 2. Task1(cookieA) is not running on SMT1.
>> otherwise it would be the case I described above.  
>> 
>> Besides, for other cases, like smt2+2task(CookieA+CookieB), picking *target*
>> cpu instead of *idle sibling* could be more helpful to reach the global stable
>> status(distribute different cookies onto different cores). 
>> If the task number of two cookies has a big difference, then distributing
> different cookies onto different cores leads to a big imbalance, that state may
> be stable but not an optimal state, I guess that's why core run queue does
> not refuse different cookies onto its rb tree.
That’s the overcommit case, in which case distributing is not possible, and would
fallback to the *local stable status*(as what current implementation does) too.

If the total task number is not overcommit, distributing could work. :)

> 
> I think I understand your concern but IMHO I'm not convinced adding cookie match
> in idle SMT selection is a best choice, if you have some performance data of your
> workload, that will be very helpful to understand the case.
Got it, I’ll try later.  

> 
> If distributing different cookies onto different cores is a hard requirement from
> your side, you are welcome to submit a patch to see others opinion.

Thanks for your patience. If possible, could you please loop me in for future
discussions about core-scheduling. 

Thx.
Regard,
Jiang

> 
> Thanks,
> -Aubrey

Powered by blists - more mailing lists