Re: [PATCH v4] sched/fair: Consider cpu affinity when allowing NUMA imbalance in find_idlest_group

From: K Prateek Nayak <kprateek.nayak@amd.com>
To: Mel Gorman <mgorman@techsingularity.net>
Cc: peterz@infradead.org, aubrey.li@linux.intel.com, efault@gmx.de,
	gautham.shenoy@amd.com, linux-kernel@vger.kernel.org,
	mingo@kernel.org, song.bao.hua@hisilicon.com,
	srikar@linux.vnet.ibm.com, valentin.schneider@arm.com,
	vincent.guittot@linaro.org
Subject: Re: [PATCH v4] sched/fair: Consider cpu affinity when allowing NUMA imbalance in find_idlest_group
Date: Tue, 22 Feb 2022 15:09:27 +0530	[thread overview]
Message-ID: <fad35024-32c6-21bb-17b2-9fd7e6c781f3@amd.com> (raw)
In-Reply-To: <20220222084535.GB4423@techsingularity.net>

Hello Mel,

On 2/22/2022 2:15 PM, Mel Gorman wrote:
> [..snip..]
>> Following are the results from testing:
>>
>> - Un-affined runs:
>>   Command: stress-ng -t 30s --exec <Worker>
>>
>>   Kernel versions:
>>   - balance-wake - This patch
>>   - branch - This patch + Mel's suggested branch
>>   - branch-unlikely - This patch + Mel's suggested branch + unlikely
>>
>>   Result format: Amean in ns [Co-eff of Var] (% Improvement)
>>
>>   Workers balance-wake            	  branch          		  branch-unlikely
>>   1       18613.20 [0.01] (0.00 pct)      18348.00 [0.04] (1.42 pct)      18299.20 [0.02] (1.69 pct)
>>   2       18634.40 [0.03] (0.00 pct)      18163.80 [0.04] (2.53 pct)      19037.80 [0.05] (-2.16 pct)
>>   4       20997.40 [0.02] (0.00 pct)      20980.80 [0.02] (0.08 pct)      21527.40 [0.02] (-2.52 pct)
>>   8       20890.20 [0.01] (0.00 pct)      19714.60 [0.07] (5.63 pct)      20021.40 [0.05] (4.16 pct)
>>   16      21200.20 [0.02] (0.00 pct)      20564.40 [0.00] (3.00 pct)      20676.00 [0.01] (2.47 pct)
>>   32      21301.80 [0.02] (0.00 pct)      20767.40 [0.02] (2.51 pct)      20945.00 [0.01] (1.67 pct)
>>   64      22772.40 [0.01] (0.00 pct)      22505.00 [0.01] (1.17 pct)      22629.40 [0.00] (0.63 pct)
>>   128     25843.00 [0.01] (0.00 pct)      25124.80 [0.00] (2.78 pct)      25377.40 [0.00] (1.80 pct)
>>   256     18691.00 [0.02] (0.00 pct)      19086.40 [0.05] (-2.12 pct)     18013.00 [0.04] (3.63 pct)
>>   512     19658.40 [0.03] (0.00 pct)      19568.80 [0.01] (0.46 pct)      18972.00 [0.02] (3.49 pct)
>>   1024    19126.80 [0.04] (0.00 pct)      18762.80 [0.02] (1.90 pct)      18878.20 [0.04] (1.30 pct)
>>
> Co-eff of variance looks low but for the lower counts before the machine
> is saturated (>=256?) it does not look like it helps and if anything,
> it hurts.  A branch mispredict profile might reveal more but I doubt
> it's worth the effort at this point.
The positive percentage here represents improvement i.e., the time
between the events sched_process_fork and sched_wakeup_new has come
down in most cases after adding the branch.
Same is applicable for results below.
>> - Affined runs:
>>   Command: taskset -c 0-254 stress-ng -t 30s --exec <Worker>
>>
>>   Kernel versions:
>>   - balance-wake-affine - This patch + affined run
>>   - branch-affine - This patch + Mel's suggested branch + affined run
>>   - branch-unlikely-affine - This patch + Mel's suggested branch + unlikely + affined run
>>
>>   Result format: Amean in ns [Co-eff of Var] (% Improvement)
>>
>>   Workers balance-wake-affine             branch-affine           	  branch-unlikely-affine
>>   1       18515.00 [0.01] (0.00 pct)      18538.00 [0.02] (-0.12 pct)     18568.40 [0.01] (-0.29 pct)
>>   2       17882.80 [0.01] (0.00 pct)      19627.80 [0.09] (-9.76 pct)     18790.40 [0.01] (-5.08 pct)
>>   4       21204.20 [0.01] (0.00 pct)      21410.60 [0.04] (-0.97 pct)     21715.20 [0.03] (-2.41 pct)
>>   8       20840.20 [0.01] (0.00 pct)      19684.60 [0.07] (5.55 pct)      21074.20 [0.02] (-1.12 pct)
>>   16      21115.20 [0.02] (0.00 pct)      20823.00 [0.01] (1.38 pct)      20719.80 [0.00] (1.87 pct)
>>   32      21159.00 [0.02] (0.00 pct)      21371.20 [0.01] (-1.00 pct)     21253.20 [0.01] (-0.45 pct)
>>   64      22768.20 [0.01] (0.00 pct)      22816.80 [0.00] (-0.21 pct)     22662.00 [0.00] (0.47 pct)
>>   128     25671.80 [0.00] (0.00 pct)      25528.20 [0.00] (0.56 pct)      25404.00 [0.00] (1.04 pct)
>>   256     27209.00 [0.01] (0.00 pct)      26751.00 [0.01] (1.68 pct)      26733.20 [0.00] (1.75 pct)
>>   512     20241.00 [0.03] (0.00 pct)      19378.60 [0.03] (4.26 pct)      19671.40 [0.00] (2.81 pct)
>>   1024    19380.80 [0.05] (0.00 pct)      18940.40 [0.02] (2.27 pct)      19071.80 [0.00] (1.59 pct)
> Same here, the cpumask check obviously hurts but it does not look like
> the unlikely helps.
I agree. unlikely doesn't show consistent results.
>> With or without the unlikely, adding the check before doing the
>> cpumask operation benefits most cases of un-affined tasks.
>>
> I think repost the patch with the num_online_cpus check added in. Yes,
> it hurts a bit for the pure fork case when the cpus_ptr is contrained by
> a scheduler policy but at least it makes sense.
I'll post the V5 soon with the check as you suggested.
--
Thanks and Regards,
Prateek