Re: [PATCH RFC 1/2] sched: Minimize the idle cpu selection race window.

[Uladzislau Rezki]

On Thu, Nov 23, 2017 at 02:13:01PM +0100, Mike Galbraith wrote:
> On Thu, 2017-11-23 at 11:52 +0100, Uladzislau Rezki wrote:
> > Hello, Atish, Peter, all.
> > 
> > I have a question about if a task's nr_cpus_allowed is 1.
> > In that scenario we do not call select_task_rq. Therefore
> > even thought a task "p" is placed on idle CPU that CPU
> > will not be marked as claimed for wake-up.
> > 
> > What do you think about adding per_cpu(claim_wakeup, cpu) = 1;
> > to select_task_rq() instead and possibly get rid of them from
> > other places (increases a race window a bit)?
> 
> My thoughts on all of this is that we need less SIS, not more.  Rather
> than trying so hard for the absolute lowest wakeup latency, which
> induces throughput/efficiency robbing bouncing, I think we'd be better
> of considering leaving an already llc affine task where it is if the
> average cycle time is sufficiently low that it will likely hit the CPU
> RSN.
I guess there is misunderstanding here. The main goal is not to cover
pinned case, for sure. I was thinking more about below points:

- Extend a claim_wake_up logic for making an ILB/NO_HZ decision more
  predictable (that is good for mobile workloads). Because as it is
  right now it simply returns a first CPU in a "nohz" mask and if we
  know that CPU has been claimed i think it is worth to go with another
  ILB core, since waking up on a remote CPU + doing nohz_idle_balance
  does not improve wake-up latency and is a miss from ilb point of view.

- Get rid of duplication;

- Be not limited to pinned case. 

If you have any proposal, i would be appreciated if you could
share your specific view.

> Completely ignoring low utilization kernel threads would go a
> long way to getting rid of bouncing userspace (which tends to have a
> meaningful footprint), all over hell and creation.
> 
> You could also periodically send mobile kthreads down the slow path to
> try to keep them the hell away from partially busy CPUs, as well as
> anything else that hasn't run for a while, to keep background cruft
> from continually injecting itself into the middle of a cross core
> cyber-sex.

CPU is not considered idle (in terms of idle_cpu()) until it hits a
first rule checking a condition if current is idle thread or not. If we
hit last check when a claim wake-up is set it means that CPU switches
from idle state to non-idle one and it will happen quite soon depending
on wake-up latency.

Considering a core as not-idle when somebody tends to wake up a task on
it is a good point. If you have any specific example when it is bad,
please share it.

--
Uladzislau Rezki