Re: [RFC PATCH] sched/fair: Skip idle CPU search on busy system

[Ingo Molnar]

* Shrikanth Hegde <sshegde@xxxxxxxxxxxxxxxxxx> wrote:

> When the system is fully busy, there will not be any idle CPU's.
> In that case, load_balance will be called mainly with CPU_NOT_IDLE
> type. In should_we_balance its currently checking for an idle CPU if
> one exist. When system is 100% busy, there will not be an idle CPU and
> these idle_cpu checks can be skipped. This would avoid fetching those rq
> structures.
> 
> This is a minor optimization for a specific case of 100% utilization.
> 
> .....
> Coming to the current implementation. It is a very basic approach to the
> issue. It may not be the best/perfect way to this.  It works only in
> case of CONFIG_NO_HZ_COMMON. nohz.nr_cpus is a global info available which
> tracks idle CPU's. AFAIU there isn't any other. If there is such info, we
> can use that instead. nohz.nr_cpus is atomic, which might be costly too.
> 
> Alternative way would be to add a new attribute to sched_domain and update
> it in cpu idle entry/exit path per CPU. Advantage is, check can be per
> env->sd instead of global. Slightly complicated, but maybe better. there
> could other advantage at wake up to limit the scan etc.
> 
> Your feedback would really help. Does this optimization makes sense?
> 
> Signed-off-by: Shrikanth Hegde <sshegde@xxxxxxxxxxxxxxxxxx>
> ---
>  kernel/sched/fair.c | 6 ++++++
>  1 file changed, 6 insertions(+)
> 
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 373ff5f55884..903d59b5290c 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -10713,6 +10713,12 @@ static int should_we_balance(struct lb_env *env)
>  		return 1;
>  	}
> 
> +#ifdef CONFIG_NO_HZ_COMMON
> +	/* If the system is fully busy, its better to skip the idle checks */
> +	if (env->idle == CPU_NOT_IDLE && atomic_read(&nohz.nr_cpus) == 0)
> +		return group_balance_cpu(sg) == env->dst_cpu;
> +#endif

Not a big fan of coupling NOHZ to a scheduler optimization in this fashion, 
and not a big fan of the nohz.nr_cpus global cacheline either.

I think it should be done unconditionally, via the scheduler topology tree:

 - We should probably slow-propagate "permanently busy" status of a CPU
   down the topology tree, ie.:

     - mark a domain fully-busy with a delay & batching, probably driven
       by the busy-tick only,

     - while marking a domain idle instantly & propagating this up the
       domain tree only if necessary. The propagation can stop if it
       finds a non-busy domain, so usually it won't reach the root domain.

 - This approach ensures there's no real overhead problem in the domain 
   tree: think of hundreds of CPUs all accessing the nohz.nr_cpus global 
   variable... I bet it's a measurable problem already on large systems.

 - The "atomic_read(&nohz.nr_cpus) == 0" condition in your patch is simply
   the busy-flag checked at the root domain: a readonly global cacheline
   that never gets modified on a permanently busy system.

Thanks,

	Ingo