Re: [PATCH 09/16] sched/fair: Let asymmetric cpu configurations balance at wake-up

[Morten Rasmussen]

On Wed, May 25, 2016 at 06:29:33PM +0800, Wanpeng Li wrote:
> 2016-05-25 17:49 GMT+08:00 Morten Rasmussen <morten.rasmussen@xxxxxxx>:
> > On Wed, May 25, 2016 at 02:57:00PM +0800, Wanpeng Li wrote:
> >> 2016-05-23 18:58 GMT+08:00 Morten Rasmussen <morten.rasmussen@xxxxxxx>:
> >> > Currently, SD_WAKE_AFFINE always takes priority over wakeup balancing if
> >> > SD_BALANCE_WAKE is set on the sched_domains. For asymmetric
> >> > configurations SD_WAKE_AFFINE is only desirable if the waking task's
> >> > compute demand (utilization) is suitable for the cpu capacities
> >> > available within the SD_WAKE_AFFINE sched_domain. If not, let wakeup
> >> > balancing take over (find_idlest_{group, cpu}()).
> >> >
> >> > The assumption is that SD_WAKE_AFFINE is never set for a sched_domain
> >> > containing cpus with different capacities. This is enforced by a
> >> > previous patch based on the SD_ASYM_CPUCAPACITY flag.
> >> >
> >> > Ideally, we shouldn't set 'want_affine' in the first place, but we don't
> >> > know if SD_BALANCE_WAKE is enabled on the sched_domain(s) until we start
> >> > traversing them.
> >> >
> >> > cc: Ingo Molnar <mingo@xxxxxxxxxx>
> >> > cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> >> >
> >> > Signed-off-by: Morten Rasmussen <morten.rasmussen@xxxxxxx>
> >> > ---
> >> >  kernel/sched/fair.c | 28 +++++++++++++++++++++++++++-
> >> >  1 file changed, 27 insertions(+), 1 deletion(-)
> >> >
> >> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> >> > index 564215d..ce44fa7 100644
> >> > --- a/kernel/sched/fair.c
> >> > +++ b/kernel/sched/fair.c
> >> > @@ -114,6 +114,12 @@ unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
> >> >  unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
> >> >  #endif
> >> >
> >> > +/*
> >> > + * The margin used when comparing utilization with cpu capacity:
> >> > + * util * 1024 < capacity * margin
> >> > + */
> >> > +unsigned int capacity_margin = 1280; /* ~20% */
> >> > +
> >> >  static inline void update_load_add(struct load_weight *lw, unsigned long inc)
> >> >  {
> >> >         lw->weight += inc;
> >> > @@ -5293,6 +5299,25 @@ static int cpu_util(int cpu)
> >> >         return (util >= capacity) ? capacity : util;
> >> >  }
> >> >
> >> > +static inline int task_util(struct task_struct *p)
> >> > +{
> >> > +       return p->se.avg.util_avg;
> >> > +}
> >> > +
> >> > +static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
> >> > +{
> >> > +       long delta;
> >> > +       long prev_cap = capacity_of(prev_cpu);
> >> > +
> >> > +       delta = cpu_rq(cpu)->rd->max_cpu_capacity - prev_cap;
> >> > +
> >> > +       /* prev_cpu is fairly close to max, no need to abort wake_affine */
> >> > +       if (delta < prev_cap >> 3)
> >> > +               return 0;
> >> > +
> >> > +       return prev_cap * 1024 < task_util(p) * capacity_margin;
> >> > +}
> >>
> >> If one task util_avg is SCHED_CAPACITY_SCALE and running on x86 box w/
> >> SMT enabled, then each HT has capacity 589, wake_cap() will result in
> >> always not wake affine, right?
> >
> > The idea is that SMT systems would bail out already at the previous
> > condition. We should have max_cpu_capacity == prev_cap == 589, delta
> > should then be zero and make the first condition true and make
> > wake_cap() always return 0 for any system with symmetric capacities
> > regardless of their actual capacity values.
> >
> > Note that this isn't entirely true as I used capacity_of() for prev_cap,
> > if I change that to capacity_orig_of() it should be true.
> >
> > By making the !wake_cap() condition always true for want_affine, we
> > should preserve existing behaviour for SMT/SMP. The only overhead is the
> > capacity delta computation and comparison, which should be cheap.
> >
> > Does that make sense?
> 
> Fair enough, thanks for your explanation.
> 
> >
> > Btw, task util_avg == SCHED_CAPACITY_SCALE should only be possible
> > temporarily, it should decay to util_avg <=
> > capacity_orig_of(task_cpu(p)) over time. That doesn't affect your
> 
> Sorry, I didn't find it will decay to capacity_orig in
> __update_load_avg(), could you elaborate?

I should have checked the code before writing that :-( I thought the
scaling by arch_scale_cpu_capacity() in __update_load_avg() would do
that, but it turns out that the default implementation of
arch_scale_cpu_capacity() doesn't do that when we pass a NULL pointer
for the sched_domain, it would have returned smt_gain/span_weight ==
capacity_orig_of(cpu) otherwise.

Sorry for the confusion, though I'm not sure if it is right to return
SCHED_CAPACITY_SCALE for SMT systems.