Re: [PATCH] sched/fair: Don't increase sd->balance_interval on newidle balance

From: Valentin Schneider
Date: Wed Sep 26 2018 - 09:16:13 EST


On 26/09/18 11:33, Vincent Guittot wrote:
> On Wed, 26 Sep 2018 at 11:35, Valentin Schneider
> [...]
>>> Can you give us details about the use case that you care about ?
>>>
>>
>> It's the same as I presented last week - devlib (some python target communication
>
> ok. you mean at linaro connect
>

Yeah, sorry.

>> library I use) has some phase where it spawns at lot of tasks at once to do
>> some setup (busybox, shutils, bash...). Some of those tasks are pinned to a
>> particular CPU, and that can lead to failed load_balance() - and to make things
>> worse, there's a lot of idle_balance() in there.
>>
>> Eventually when I start running my actual workload a few ~100ms later, it's
>> impacted by that balance_interval increase.
>>
>> Admittedly that's a specific use-case, but I don't think this quick increase
>> is something that was intended.
>
> Yes, this really sounds like a specific use-case. Unluckily you find a
> way to reach max interval quite easily/every time with your test
> set-up but keep in mind that this can also happen in real system life
> and without using the newly idle path.
> So if it's a problem to have a interval at max value for your unitary
> test, it probably means that it's a problem for the system and the max
> value is too high
>

Limiting the max value for those tests is actually a good point, and I think
I'll give it a shot. However...

> Taking advantage of all load_balance event to update the interval
> makes sense to me. It seems that you care about a short and regular
> balance interval more that minimizing overhead of load balancing.
> At the opposite, i'm sure that you don't complain if newly idle load
> balance resets the interval to min value and overwrite what the
> periodic load balance set up previously :-)
>

...My concern is more about increasing balance_interval faster than we should.
The proposed "fix" is to prevent any balance_interval increase when going
through idle_balance(), but Patrick (added in cc) suggested offline that
we could simply limit the rate at which we do these increases, so that they
match what we do in rebalance_domains().

We'd still increase balance_interval on failed newidle load_balance(), but
we wouldn't increase from min to max in e.g. 50ms. Would that work better
for you?

>> [...]