[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <219b8b49-3767-4010-aa68-9e1cf66c2ccb@arm.com>
Date: Fri, 10 May 2024 15:49:46 +0100
From: Luis Machado <luis.machado@....com>
To: Peter Zijlstra <peterz@...radead.org>
Cc: 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, kprateek.nayak@....com,
wuyun.abel@...edance.com, tglx@...utronix.de, efault@....de, nd
<nd@....com>, John Stultz <jstultz@...gle.com>, Hongyan.Xia2@....com
Subject: Re: [RFC][PATCH 08/10] sched/fair: Implement delayed dequeue
On 5/2/24 11:26, Luis Machado wrote:
> On 4/29/24 15:33, Luis Machado wrote:
>> Hi Peter,
>>
>> On 4/26/24 10:32, Peter Zijlstra wrote:
>>> On Thu, Apr 25, 2024 at 01:49:49PM +0200, Peter Zijlstra wrote:
>>>> On Thu, Apr 25, 2024 at 12:42:20PM +0200, Peter Zijlstra wrote:
>>>>
>>>>>> I wonder if the delayed dequeue logic is having an unwanted effect on the calculation of
>>>>>> utilization/load of the runqueue and, as a consequence, we're scheduling things to run on
>>>>>> higher OPP's in the big cores, leading to poor decisions for energy efficiency.
>>>>>
>>>>> Notably util_est_update() gets delayed. Given we don't actually do an
>>>>> enqueue when a delayed task gets woken, it didn't seem to make sense to
>>>>> update that sooner.
>>>>
>>>> The PELT runnable values will be inflated because of delayed dequeue.
>>>> cpu_util() uses those in the @boost case, and as such this can indeed
>>>> affect things.
>>>>
>>>> This can also slightly affect the cgroup case, but since the delay goes
>>>> away as contention goes away, and the cgroup case must already assume
>>>> worst case overlap, this seems limited.
>>>>
>>>> /me goes ponder things moar.
>>>
>>> First order approximation of a fix would be something like the totally
>>> untested below I suppose...
>>
>> I gave this a try on the Pixel 6, and I noticed some improvement (see below), but not
>> enough to bring it back to the original levels.
>>
>> (1) m6.6-stock - Basic EEVDF with wakeup preemption fix (baseline)
>> (2) m6.6-eevdf-complete: m6.6-stock plus this series.
>> (3) m6.6-eevdf-complete-no-delay-dequeue: (2) + NO_DELAY_DEQUEUE
>> (4) m6.6-eevdf-complete-no-delay-dequeue-no-delay-zero: (2) + NO_DELAY_DEQUEUE + NO_DELAY_ZERO
>> (5) m6.6-eevdf-complete-no-delay-zero: (2) + NO_DELAY_ZERO
>> (6) m6.6-eevdf-complete-pelt-fix: (2) + the proposed load_avg update patch.
>>
>> I included (3), (4) and (5) to exercise the impact of disabling the individual
>> scheduler features.
>>
>>
>> Energy use.
>>
>> +------------+------------------------------------------------------+-----------+
>> | cluster | tag | perc_diff |
>> +------------+------------------------------------------------------+-----------+
>> | CPU | m6.6-stock | 0.0% |
>> | CPU-Big | m6.6-stock | 0.0% |
>> | CPU-Little | m6.6-stock | 0.0% |
>> | CPU-Mid | m6.6-stock | 0.0% |
>> | GPU | m6.6-stock | 0.0% |
>> | Total | m6.6-stock | 0.0% |
>> | CPU | m6.6-eevdf-complete | 114.51% |
>> | CPU-Big | m6.6-eevdf-complete | 90.75% |
>> | CPU-Little | m6.6-eevdf-complete | 98.74% |
>> | CPU-Mid | m6.6-eevdf-complete | 213.9% |
>> | GPU | m6.6-eevdf-complete | -7.04% |
>> | Total | m6.6-eevdf-complete | 100.92% |
>> | CPU | m6.6-eevdf-complete-no-delay-dequeue | 117.77% |
>> | CPU-Big | m6.6-eevdf-complete-no-delay-dequeue | 113.79% |
>> | CPU-Little | m6.6-eevdf-complete-no-delay-dequeue | 97.47% |
>> | CPU-Mid | m6.6-eevdf-complete-no-delay-dequeue | 189.0% |
>> | GPU | m6.6-eevdf-complete-no-delay-dequeue | -6.74% |
>> | Total | m6.6-eevdf-complete-no-delay-dequeue | 103.84% |
>> | CPU | m6.6-eevdf-complete-no-delay-dequeue-no-delay-zero | 120.45% |
>> | CPU-Big | m6.6-eevdf-complete-no-delay-dequeue-no-delay-zero | 113.65% |
>> | CPU-Little | m6.6-eevdf-complete-no-delay-dequeue-no-delay-zero | 99.04% |
>> | CPU-Mid | m6.6-eevdf-complete-no-delay-dequeue-no-delay-zero | 201.14% |
>> | GPU | m6.6-eevdf-complete-no-delay-dequeue-no-delay-zero | -5.37% |
>> | Total | m6.6-eevdf-complete-no-delay-dequeue-no-delay-zero | 106.38% |
>> | CPU | m6.6-eevdf-complete-no-delay-zero | 119.05% |
>> | CPU-Big | m6.6-eevdf-complete-no-delay-zero | 107.55% |
>> | CPU-Little | m6.6-eevdf-complete-no-delay-zero | 98.66% |
>> | CPU-Mid | m6.6-eevdf-complete-no-delay-zero | 206.58% |
>> | GPU | m6.6-eevdf-complete-no-delay-zero | -5.25% |
>> | Total | m6.6-eevdf-complete-no-delay-zero | 105.14% |
>> | CPU | m6.6-eevdf-complete-pelt-fix | 105.56% |
>> | CPU-Big | m6.6-eevdf-complete-pelt-fix | 100.45% |
>> | CPU-Little | m6.6-eevdf-complete-pelt-fix | 94.4% |
>> | CPU-Mid | m6.6-eevdf-complete-pelt-fix | 150.94% |
>> | GPU | m6.6-eevdf-complete-pelt-fix | -3.96% |
>> | Total | m6.6-eevdf-complete-pelt-fix | 93.31% |
>> +------------+------------------------------------------------------+-----------+
>>
>> Utilization and load levels.
>>
>> +---------+------------------------------------------------------+----------+-----------+
>> | cluster | tag | variable | perc_diff |
>> +---------+------------------------------------------------------+----------+-----------+
>> | little | m6.6-stock | load | 0.0% |
>> | little | m6.6-stock | util | 0.0% |
>> | little | m6.6-eevdf-complete | load | 29.56% |
>> | little | m6.6-eevdf-complete | util | 55.4% |
>> | little | m6.6-eevdf-complete-no-delay-dequeue | load | 42.89% |
>> | little | m6.6-eevdf-complete-no-delay-dequeue | util | 69.47% |
>> | little | m6.6-eevdf-complete-no-delay-dequeue-no-delay-zero | load | 51.05% |
>> | little | m6.6-eevdf-complete-no-delay-dequeue-no-delay-zero | util | 76.55% |
>> | little | m6.6-eevdf-complete-no-delay-zero | load | 34.51% |
>> | little | m6.6-eevdf-complete-no-delay-zero | util | 72.53% |
>> | little | m6.6-eevdf-complete-pelt-fix | load | 29.96% |
>> | little | m6.6-eevdf-complete-pelt-fix | util | 59.82% |
>> | mid | m6.6-stock | load | 0.0% |
>> | mid | m6.6-stock | util | 0.0% |
>> | mid | m6.6-eevdf-complete | load | 29.37% |
>> | mid | m6.6-eevdf-complete | util | 75.22% |
>> | mid | m6.6-eevdf-complete-no-delay-dequeue | load | 36.4% |
>> | mid | m6.6-eevdf-complete-no-delay-dequeue | util | 80.28% |
>> | mid | m6.6-eevdf-complete-no-delay-dequeue-no-delay-zero | load | 30.35% |
>> | mid | m6.6-eevdf-complete-no-delay-dequeue-no-delay-zero | util | 90.2% |
>> | mid | m6.6-eevdf-complete-no-delay-zero | load | 37.83% |
>> | mid | m6.6-eevdf-complete-no-delay-zero | util | 93.79% |
>> | mid | m6.6-eevdf-complete-pelt-fix | load | 33.57% |
>> | mid | m6.6-eevdf-complete-pelt-fix | util | 67.83% |
>> | big | m6.6-stock | load | 0.0% |
>> | big | m6.6-stock | util | 0.0% |
>> | big | m6.6-eevdf-complete | load | 97.39% |
>> | big | m6.6-eevdf-complete | util | 12.63% |
>> | big | m6.6-eevdf-complete-no-delay-dequeue | load | 139.69% |
>> | big | m6.6-eevdf-complete-no-delay-dequeue | util | 22.58% |
>> | big | m6.6-eevdf-complete-no-delay-dequeue-no-delay-zero | load | 125.36% |
>> | big | m6.6-eevdf-complete-no-delay-dequeue-no-delay-zero | util | 23.15% |
>> | big | m6.6-eevdf-complete-no-delay-zero | load | 128.56% |
>> | big | m6.6-eevdf-complete-no-delay-zero | util | 25.03% |
>> | big | m6.6-eevdf-complete-pelt-fix | load | 130.73% |
>> | big | m6.6-eevdf-complete-pelt-fix | util | 17.52% |
>> +---------+------------------------------------------------------+----------+-----------+
>
> Going through the code, my understanding is that the util_est functions seem to be getting
> called correctly, and in the right order. That is, we first util_est_enqueue, then util_est_dequeue
> and finally util_est_update. So the stats *should* be correct.
>
> On dequeuing (dequeue_task_fair), we immediately call util_est_dequeue, even for the case of
> a DEQUEUE_DELAYED task, since we're no longer going to run the dequeue_delayed task for now, even
> though it is still in the rq.
>
> We delay the util_est_update of dequeue_delayed tasks until a later time in dequeue_entities.
>
> Eventually the dequeue_delayed task will have its lag zeroed when it becomes eligible again,
> (requeue_delayed_entity) while still being in the rq. It will then get dequeued/enqueued (requeued),
> and marked as a non-dequeue-delayed task.
>
> Next time we attempt to enqueue such a task (enqueue_task_fair), it will skip the ENQUEUE_DELAYED
> block and call util_est_enqueue.
>
> Still, something seems to be signalling that util/load is high, and causing migration to the big cores.
>
> Maybe we're not decaying the util/load properly at some point, and inflated numbers start to happen.
>
> I'll continue investigating.
>
Just a quick update on this. While investigating this behavior, I
spotted very high loadavg values on an idle system. For instance:
load average: 4733.84, 4721.24, 4680.33
I wonder if someone else also spotted this.
These values keep increasing slowly but steadily. When the system
is under load they increase a bit more rapidly. Makes me wonder
if we're missing decrementing nr_uninterruptible in some path,
since that is what seems to be causing the loadavg to be off.
Powered by blists - more mailing lists