Re: [PATCH v2 6/8] sched/idle: Move busy_cpu accounting to idle callback

From: Aubrey Li <aubrey.li@linux.intel.com>
To: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Cc: Ingo Molnar <mingo@kernel.org>,
	Peter Zijlstra <peterz@infradead.org>,
	LKML <linux-kernel@vger.kernel.org>,
	Mel Gorman <mgorman@techsingularity.net>,
	Rik van Riel <riel@surriel.com>,
	Thomas Gleixner <tglx@linutronix.de>,
	Valentin Schneider <valentin.schneider@arm.com>,
	Vincent Guittot <vincent.guittot@linaro.org>,
	Dietmar Eggemann <dietmar.eggemann@arm.com>,
	Gautham R Shenoy <ego@linux.vnet.ibm.com>,
	Parth Shah <parth@linux.ibm.com>
Subject: Re: [PATCH v2 6/8] sched/idle: Move busy_cpu accounting to idle callback
Date: Wed, 19 May 2021 17:43:55 +0800	[thread overview]
Message-ID: <32b98350-35e4-7475-2d19-9101f50ecc63@linux.intel.com> (raw)
In-Reply-To: <20210518071843.GZ2633526@linux.vnet.ibm.com>

On 5/18/21 3:18 PM, Srikar Dronamraju wrote:

>>>> This is v3. It looks like hackbench gets better. And netperf still has
>>>> some notable changes under 2 x overcommit cases.
>>>>
>>>
>>> Thanks Aubrey for the results. netperf (2X) case does seem to regress.
>>> I was actually expecting the results to get better with overcommit.
>>> Can you confirm if this was just v3 or with v3 + set_next_idle_core
>>> disabled?
>>
>> Do you mean set_idle_cores(not set_next_idle_core) actually? Gautham's patch
>> changed "this" to "target" in set_idle_cores, and I removed it to apply
>> v3-2-8-sched-fair-Maintain-the-identity-of-idle-core.patch for tip/sched/core
>> commit-id 915a2bc3c6b7.
> 
> Thats correct,
> 
> In the 3rd patch, I had introduced set_next_idle_core
> which is suppose to set idle_cores in the LLC.
> What I suspected was is this one is causing issues in your 48 CPU LLC.
> 
> I am expecting set_next_idle_core to be spending much time in your scenario.
> I was planning for something like the below on top of my patch.
> With this we dont look for an idle-core if we already know that we dont find one.
> But in the mean while I had asked if you could have dropped the call to
> set_next_idle_core.
> 

+	if (atomic_read(&sd->shared->nr_busy_cpus) * 2 >=  per_cpu(sd_llc_size, target))
+		goto out;

Does this has side effect if waker and wakee are coalesced on a portion of cores?
Also, is 2 a SMT2 assumption?

I did a quick testing on this, it looks like the regression of netperf 2x cases are 
improved indeed, but hackbench two mid-load cases get worse.

process-sockets 	group-2 	 1.00 (  5.32)	-18.40 (  7.32)
threads-sockets 	group-2 	 1.00 (  5.44)	-20.44 (  4.60)

Thanks,
-Aubrey