On Thu, 11 Jun 2020 at 12:24, Qais Yousef <qais.yousef@xxxxxxx> wrote:
>
> On 06/09/20 19:10, Vincent Guittot wrote:
> > On Mon, 8 Jun 2020 at 14:31, Qais Yousef <qais.yousef@xxxxxxx> wrote:
> > >
> > > On 06/04/20 14:14, Vincent Guittot wrote:
> > >
> > > [...]
> > >
> > > > I have tried your patch and I don't see any difference compared to
> > > > previous tests. Let me give you more details of my setup:
> > > > I create 3 levels of cgroups and usually run the tests in the 4 levels
> > > > (which includes root). The result above are for the root level
> > > >
> > > > But I see a difference at other levels:
> > > >
> > > > root level 1 level 2 level 3
> > > >
> > > > /w patch uclamp disable 50097 46615 43806 41078
> > > > tip uclamp enable 48706(-2.78%) 45583(-2.21%) 42851(-2.18%)
> > > > 40313(-1.86%)
> > > > /w patch uclamp enable 48882(-2.43%) 45774(-1.80%) 43108(-1.59%)
> > > > 40667(-1.00%)
> > > >
> > > > Whereas tip with uclamp stays around 2% behind tip without uclamp, the
> > > > diff of uclamp with your patch tends to decrease when we increase the
> > > > number of level
> > >
> > > So I did try to dig more into this, but I think it's either not a good
> > > reproducer or what we're observing here is uArch level latencies caused by the
> > > new code that seem to produce a bigger knock on effect than what they really
> > > are.
> > >
> > > First, CONFIG_FAIR_GROUP_SCHED is 'expensive', for some definition of
> > > expensive..
> >
> > yes, enabling CONFIG_FAIR_GROUP_SCHED adds an overhead
> >
> > >
> > > *** uclamp disabled/fair group enabled ***
> > >
> > > # Executed 50000 pipe operations between two threads
> > >
> > > Total time: 0.958 [sec]
> > >
> > > 19.177100 usecs/op
> > > 52145 ops/sec
> > >
> > > *** uclamp disabled/fair group disabled ***
> > >
> > > # Executed 50000 pipe operations between two threads
> > > Total time: 0.808 [sec]
> > >
> > > 16.176200 usecs/op
> > > 61819 ops/sec
> > >
> > > So there's a 15.6% drop in ops/sec when enabling this option. I think it's good
> > > to look at the absolutely number of usecs/op, Fair group adds around
> > > 3 usecs/op.
> > >
> > > I dropped FAIR_GROUP_SCHED from my config to eliminate this overhead and focus
> > > on solely on uclamp overhead.
> >
> > Have you checked that both tests run at the root level ?
>
> I haven't actively moved tasks to cgroups. As I said that snippet was
> particularly bad and I didn't see that level of nesting in every call.
>
> > Your function-graph log below shows several calls to
> > update_cfs_group() which means that your trace below has not been made
> > at root level but most probably at the 3rd level and I wonder if you
> > used the same setup for running the benchmark above. This could
> > explain such huge difference because I don't have such difference on
> > my platform but more around 2%
>
> What promoted me to look at this is when you reported that even without uclamp
> the nested cgroup showed a drop at each level. I was just trying to understand
> how both affect the hot path in hope to understand the root cause of uclamp
> overhead.
>
> >
> > For uclamp disable/fair group enable/ function graph enable : 47994ops/sec
> > For uclamp disable/fair group disable/ function graph enable : 49107ops/sec
> >
> > >
> > > With uclamp enabled but no fair group I get
> > >
> > > *** uclamp enabled/fair group disabled ***
> > >
> > > # Executed 50000 pipe operations between two threads
> > > Total time: 0.856 [sec]
> > >
> > > 17.125740 usecs/op
> > > 58391 ops/sec
> > >
> > > The drop is 5.5% in ops/sec. Or 1 usecs/op.
> > >
> > > I don't know what's the expectation here. 1 us could be a lot, but I don't
> > > think we expect the new code to take more than few 100s of ns anyway. If you
> > > add potential caching effects, reaching 1 us wouldn't be that hard.
> > >
> > > Note that in my runs I chose performance governor and use `taskset 0x2` to
> >
> > You might want to set 2 CPUs in your cpumask instead of 1 in order to
> > have 1 CPU for each thread
>
> I did try that but it didn't seem to change the number. I think the 2 tasks
> interleave so running in 2 CPUs doesn't change the result. But to ease ftrace
> capture, it's easier to monitor a single cpu.
>
> >
> > > force running on a big core to make sure the runs are repeatable.
> >
> > I also use performance governor but don't pinned tasks because I use smp.
>
> Is your arm platform SMP?
Yes, all my tests are done on the Arm64 octo core smp system
>
> >
> > >
> > > On Juno-r2 I managed to scrap most of the 1 us with the below patch. It seems
> > > there was weird branching behavior that affects the I$ in my case. It'd be good
> > > to try it out to see if it makes a difference for you.
> >
> > The perf are slightly worse on my setup:
> > For uclamp enable/fair group disable/ function graph enable : 48413ops/sec
> > with patch below : 47804os/sec
>
> I am not sure if the new code could just introduce worse cache performance
> in a platform dependent way. The evidences I have so far point in this
> direction.
>
> >
> > >
> > > The I$ effect is my best educated guess. Perf doesn't catch this path and
> > > I couldn't convince it to look at cache and branch misses between 2 specific
> > > points.
> > >
> > > Other subtle code shuffling did have weird effect on the result too. One worthy
> > > one is making uclamp_rq_dec() noinline gains back ~400 ns. Making
> > > uclamp_rq_inc() noinline *too* cancels this gain out :-/
> > >
> > >
> > > diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> > > index 0464569f26a7..0835ee20a3c7 100644
> > > --- a/kernel/sched/core.c
> > > +++ b/kernel/sched/core.c
> > > @@ -1071,13 +1071,11 @@ static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
> > >
> > > static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
> > > {
> > > - enum uclamp_id clamp_id;
> > > -
> > > if (unlikely(!p->sched_class->uclamp_enabled))
> > > return;
> > >
> > > - for_each_clamp_id(clamp_id)
> > > - uclamp_rq_inc_id(rq, p, clamp_id);
> > > + uclamp_rq_inc_id(rq, p, UCLAMP_MIN);
> > > + uclamp_rq_inc_id(rq, p, UCLAMP_MAX);
> > >
> > > /* Reset clamp idle holding when there is one RUNNABLE task */
> > > if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
> > > @@ -1086,13 +1084,11 @@ static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
> > >
> > > static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
> > > {
> > > - enum uclamp_id clamp_id;
> > > -
> > > if (unlikely(!p->sched_class->uclamp_enabled))
> > > return;
> > >
> > > - for_each_clamp_id(clamp_id)
> > > - uclamp_rq_dec_id(rq, p, clamp_id);
> > > + uclamp_rq_dec_id(rq, p, UCLAMP_MIN);
> > > + uclamp_rq_dec_id(rq, p, UCLAMP_MAX);
> > > }
> > >
> > > static inline void
> > >
> > >
> > > FWIW I fail to see activate/deactivate_task in perf record. They don't show up
> > > on the list which means this micro benchmark doesn't stress them as Mel's test
> > > does.
> >
> > Strange because I have been able to trace them.
>
> On your arm platform? I can certainly see them on x86.
yes on my arm platform
>
> Thanks
>
> --
> Qais Yousef
