Re: [PATCH 3/3] sched/rt: fix pushing unfit tasks to a better CPU

From: Qais Yousef
Date: Fri Feb 21 2020 - 06:12:30 EST


On 02/21/20 13:45, Pavan Kondeti wrote:
> On Wed, Feb 19, 2020 at 02:02:44PM +0000, Qais Yousef wrote:
> > On 02/17/20 13:53, Qais Yousef wrote:
> > > On 02/17/20 14:53, Pavan Kondeti wrote:
> > > > I notice a case where tasks would migrate for no reason (happens without this
> > > > patch also). Assuming BIG cores are busy with other RT tasks. Now this RT
> > > > task can go to *any* little CPU. There is no bias towards its previous CPU.
> > > > I don't know if it makes any difference but I see RT task placement is too
> > > > keen on reducing the migrations unless it is absolutely needed.
> > >
> > > In find_lowest_rq() there's a check if the task_cpu(p) is in the lowest_mask
> > > and prefer it if it is.
> > >
> > > But yeah I see it happening too
> > >
> > > https://imgur.com/a/FYqLIko
> > >
> > > Tasks on CPU 0 and 3 swap. Note that my tasks are periodic but the plots don't
> > > show that.
> > >
> > > I shouldn't have changed something to affect this bias. Do you think it's
> > > something I introduced?
> > >
> > > It's something maybe worth digging into though. I'll try to have a look.
> >
> > FWIW, I dug a bit into this and I found out we have a thundering herd issue.
> >
> > Since I just have a set of periodic task that all start together,
> > select_task_rq_rt() ends up selecting the same fitting CPU for all of them
> > (CPU1). The end up all waking up on CPU1, only to get pushed back out
> > again with only one surviving.
> >
> > This reshuffles the task placement ending with some tasks being swapped.
> >
> > I don't think this problem is specific to my change and could happen without
> > it.
> >
> > The problem is caused by the way find_lowest_rq() selects a cpu in the mask
> >
> > 1750 best_cpu = cpumask_first_and(lowest_mask,
> > 1751 sched_domain_span(sd));
> > 1752 if (best_cpu < nr_cpu_ids) {
> > 1753 rcu_read_unlock();
> > 1754 return best_cpu;
> > 1755 }
> >
> > It always returns the first CPU in the mask. Or the mask could only contain
> > a single CPU too. The end result is that we most likely end up herding all the
> > tasks that wake up simultaneously to the same CPU.
> >
> > I'm not sure how to fix this problem yet.
> >
>
> Yes, I have seen this problem too. This is not limited to RT even fair class
> (find_energy_efficient_cpu path) also have the same issue. There is a window
> where we select a CPU for the task and the task being queued there. Because of
> this, we may select the same CPU for two successive waking tasks. Turning off
> TTWU_QUEUE sched feature addresses this up to some extent. At least it would
> solve the cases like multiple tasks getting woken up from an interrupt handler.

Oh, handy. Let me try this out.

I added it to my to-do to investigate it when I have time anyway.

In modern systems where L3 is spanning all CPUs, the migration isn't that
costly, but it'd still be unnecessary wakeup latency that can add up.

Thanks

--
Qais Yousef