linux-kernel - Re: [PATCH] sched/cfs: change initial value of runnable

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <c4574b9e-852d-8f04-91cb-0fbae9f89833@applied-asynchrony.com>
Date:   Thu, 25 Jun 2020 12:42:29 +0200
From:   Holger Hoffstätte <holger@...lied-asynchrony.com>
To:     Vincent Guittot <vincent.guittot@...aro.org>
Cc:     Ingo Molnar <mingo@...hat.com>,
        Peter Zijlstra <peterz@...radead.org>,
        Juri Lelli <juri.lelli@...hat.com>,
        Dietmar Eggemann <dietmar.eggemann@....com>,
        Steven Rostedt <rostedt@...dmis.org>,
        Ben Segall <bsegall@...gle.com>, Mel Gorman <mgorman@...e.de>,
        linux-kernel <linux-kernel@...r.kernel.org>,
        kernel test robot <rong.a.chen@...el.com>,
        Valentin Schneider <valentin.schneider@....com>,
        Phil Auld <pauld@...hat.com>, Hillf Danton <hdanton@...a.com>
Subject: Re: [PATCH] sched/cfs: change initial value of runnable_avg

On 2020-06-25 11:56, Vincent Guittot wrote:
> On Thu, 25 Jun 2020 at 11:24, Holger Hoffstätte
> <holger@...lied-asynchrony.com> wrote:
>>
>> On 2020-06-24 17:44, Vincent Guittot wrote:
>>> Some performance regression on reaim benchmark have been raised with
>>>     commit 070f5e860ee2 ("sched/fair: Take into account runnable_avg to classify group")
>>>
>>> The problem comes from the init value of runnable_avg which is initialized
>>> with max value. This can be a problem if the newly forked task is finally
>>> a short task because the group of CPUs is wrongly set to overloaded and
>>> tasks are pulled less agressively.
>>>
>>> Set initial value of runnable_avg equals to util_avg to reflect that there
>>> is no waiting time so far.
>>>
>>> Fixes: 070f5e860ee2 ("sched/fair: Take into account runnable_avg to classify group")
>>> Reported-by: kernel test robot <rong.a.chen@...el.com>
>>> Signed-off-by: Vincent Guittot <vincent.guittot@...aro.org>
>>> ---
>>>    kernel/sched/fair.c | 2 +-
>>>    1 file changed, 1 insertion(+), 1 deletion(-)
>>>
>>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>>> index 0424a0af5f87..45e467bf42fc 100644
>>> --- a/kernel/sched/fair.c
>>> +++ b/kernel/sched/fair.c
>>> @@ -806,7 +806,7 @@ void post_init_entity_util_avg(struct task_struct *p)
>>>                }
>>>        }
>>>
>>> -     sa->runnable_avg = cpu_scale;
>>> +     sa->runnable_avg = sa->util_avg;
>>>
>>>        if (p->sched_class != &fair_sched_class) {
>>>                /*
>>>
>>
>> Something is wrong here. I woke up my machine from suspend-to-RAM this morning
>> and saw that a completely idle machine had a loadavg of ~7. According to my
> 
> Just to make sure: Are you speaking about loadavg that is output by
> /proc/loadavg or load_avg which is the PELT load ?

/proc/loadavg

>> monitoring system this happened to be the loadavg right before I suspended.
>> I've reverted this, rebooted, created a loadavg >0, suspended and after wake up
>> loadavg again correctly ranges between 0 and whatever, as expected.
> 
> I'm not sure to catch why ~7 is bad compared to correctly ranges
> between 0 and whatever. Isn't ~7 part of the whatever ?

After wakeup the _baseline_ for loadavg seemed to be the last value before suspend,
not 0. The 7 then was the base loadavg for a _mostly idle machine_ (just reading
mail etc.), i.e. it never went below said baseline again, no matter the
_actual_ load.

Here's an image: https://imgur.com/a/kd2stqO

Before 02:00 last night the load was ~7 (compiled something), then all processes
were terminated and the machine was suspended. After wakeup the machine was mostly
idle (9am..11am), yet measured loadavg continued with the same value as before.
I didn't notice this right away since my CPU meter on the desktop didn't show any
*actual* activity (because there was none). The spike at ~11am is the revert/reboot.
After that loadavg became normal again, i.e. representative of the actual load,
even after suspend/resume cycles.
I suspend/resume every night and the only thing that changed recently was this
patch, so.. :)

-h