Re: [PATCH 2/2] sched/fair: Scale wakeup granularity relative to nr_running

[Peter Zijlstra]

On Wed, Sep 22, 2021 at 04:04:57PM +0100, Mel Gorman wrote:
> On Wed, Sep 22, 2021 at 04:15:27PM +0200, Vincent Guittot wrote:

> The reason I used SCHED_TUNABLESCALING_NONE for the changelog is that
> the exact values depend on the number of CPUs so values are not even
> the same across the range of machines I'm using. sysctl_sched_latency,
> sysctl_sched_min_granularity sysctl_sched_wakeup_granularity are all
> scaled but the ratios remain constant.

It might make sense to reconsider the whole scaling thing FWIW. It used
to be that desktop systems had 'small' number of CPUs (<=8) and servers
had more.

But today it's not uncommon to have 32-64 CPUs in a desktop system. And
even LOG scaling gets us into stupid numbers for the latencies.