Re: [PATCH v3 2/7] sched/topology: Define and assign sched_domain flag metadata

[Valentin Schneider]

On 02/07/20 17:37, Quentin Perret wrote:
> On Thursday 02 Jul 2020 at 17:25:41 (+0100), Valentin Schneider wrote:
>> It's actually pretty close to what happens with the LLC domain on SMP -
>> select_idle_sibling() doesn't look outside of it. The wake_affine() stuff
>> might steer the task towards a different LLC, but that's about it for
>> wakeups. We rely on load balancing (fork/exec, newidle, nohz and periodic)
>> to spread this further - and we would here too.
>
> Sure, but on SMP the search space in select_idle_sibling is always
> consistent -- you search within the LLC. With the fix you suggested,
> CPUs 0-3 will search within their LLCs, while CPU4 searches the entire
> system, which creates an imbalanced mess IMO.
>

Yeah, it is a mess.

> For affine wake-ups, you could migrate from CPU4 -> CPU0-3, but CPU0-3
> to CPU4 is not possible

AIU the wake_affine bits, you get to steer the wakeup towards the waking
CPU. So if the task previously ran on CPU0-3, wake_affine can make the
target CPU4 (waking CPU), so it would become a possible candidate.

But as you say, this thing is still an ugly asymmetric mess.

> so this asymmetry is almost guaranteed to
> actively create imbalance. And sure, the periodic load balancer ought to
> fix it, but really wake-up balance and periodic load balance should be
> pushing in the same direction and not fighting against each other.
>
> Anyways, enough bikeshedding for today, I'll try and have look at the
> rest of the series :)
>

Thanks!

> Cheers,
> Quentin