RE: [PATCH] Revert "sched/fair: Fix O(nr_cgroups) in the load balancing path"

From: "Doug Smythies" <dsmythies@telus.net>
To: "'Vincent Guittot'" <vincent.guittot@linaro.org>
Cc: "'linux-kernel'" <linux-kernel@vger.kernel.org>,
	"'open list:THERMAL'" <linux-pm@vger.kernel.org>,
	"'Peter Zijlstra'" <peterz@infradead.org>,
	"'Ingo Molnar'" <mingo@kernel.org>,
	"'Linus Torvalds'" <torvalds@linux-foundation.org>,
	"'Thomas Gleixner'" <tglx@linutronix.de>,
	"'Sargun Dhillon'" <sargun@sargun.me>,
	"'Tejun Heo'" <tj@kernel.org>,
	"'Xie XiuQi'" <xiexiuqi@huawei.com>, <xiezhipeng1@huawei.com>,
	"'Srinivas Pandruvada'" <srinivas.pandruvada@linux.intel.com>,
	"'Rik van Riel'" <riel@surriel.com>
Subject: RE: [PATCH] Revert "sched/fair: Fix O(nr_cgroups) in the load balancing path"
Date: Sat, 9 Nov 2019 08:47:09 -0800	[thread overview]
Message-ID: <000001d5971d$57a90c80$06fb2580$@net> (raw)
In-Reply-To: <20191108091834.GA24402@linaro.org>

Hi Vincent,

Thank you for your item 2 patch.

On 2019.11.08 01:19 Vincent Guittot wrote:
...
>> I have to prepare a patch for this part which is item 2
>
> I have finally been able to prepared the patch for item 2. Could you check
> that it also fixes your problem ?
...
> We can still have some spurious call to cpufreq_util_change in 
> update_blocked_average() with this patch but at least the value will be
> up to date in both calls, which was not the case before. If this fix
> Doug's problem, I can prepare an additional one to fix the spurious call
> but I wanted to make sure that this fix the problem first.

Yes, the issue is solved with this patch.
I do wonder if I am seeing the effect of the spurious calls.

Details:

Test 1: an 8000 second trace during system idle:
Maximum duration: 4.00015 seconds. Good.
Typically, there would have been about 300 durations
of over 10 seconds in 8000 seconds.
Number of calls to driver: 103168, which is about 8% more than
the previous experimental solution.
(Should be repeated a few times to verify repeatability, but
not going to.)

Test 2: one 8000 second energy sample, for high accuracy idle power:
3.703 watts which is about +0.7% idle power increase.

Test 3: The load-no-load test with only idle state 1 enabled:
There was never an excessive energy sample for the no load samples.
The test ran for about 8 hours.
Maximum: 9.49 watts
Minimum: 9.13 watts
Recall that with the issue, the max would have been about 14 watts

Kernel: 5.4-rc6 + your items 1 and item 2 patches.
Idle governor = menu, because teo fixes are still pending.
Note: some reference data is from kernel 5.4-rc2, and really
should have been re-done with 5.4-rc6 as the baseline.

... Doug