lkml.org 
[lkml]   [2022]   [Apr]   [26]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [PATCH] sched: Take thermal pressure into account when determine rt fits capacity
> > > The best (simplest) way forward IMHO is to introduce a new function
> > >
> > > bool cpu_in_capacity_inversion(int cpu);
> > >
> > > (feel free to pick another name) which will detect the scenario you're in. You
> > > can use this function then in rt_task_fits_capacity()
> > >
> > > diff --git a/kernel/sched/rt.c b/kernel/sched/rt.c
> > > index a32c46889af8..d48811a7e956 100644
> > > --- a/kernel/sched/rt.c
> > > +++ b/kernel/sched/rt.c
> > > @@ -462,6 +462,9 @@ static inline bool rt_task_fits_capacity(struct task_struct *p, int cpu)
> > > if (!static_branch_unlikely(&sched_asym_cpucapacity))
> > > return true;
> > >
> > > + if (cpu_in_capacity_inversion(cpu))
> > > + return false;
> > > +
> > > min_cap = uclamp_eff_value(p, UCLAMP_MIN);
> > > max_cap = uclamp_eff_value(p, UCLAMP_MAX);
> > >
> > > You'll probably need to do something similar in dl_task_fits_capacity().
> > >
> > > This might be a bit aggressive though as we'll steer away all RT tasks from
> > > this CPU (as long as there's another CPU that can fit it). I need to think more
> > > about it. But we could do something like this too
> > >
> > > diff --git a/kernel/sched/rt.c b/kernel/sched/rt.c
> > > index a32c46889af8..f2a34946a7ab 100644
> > > --- a/kernel/sched/rt.c
> > > +++ b/kernel/sched/rt.c
> > > @@ -462,11 +462,14 @@ static inline bool rt_task_fits_capacity(struct task_struct *p, int cpu)
> > > if (!static_branch_unlikely(&sched_asym_cpucapacity))
> > > return true;
> > >
> > > + cpu_cap = capacity_orig_of(cpu);
> > > +
> > > + if (cpu_in_capacity_inversion(cpu))
> >
> > It's a good idea, but as you said, in mainline, the
> > sysctl_sched_uclamp_util_min_rt_default is always 1024,
> > Maybe it's better to add it to the judgment?
>
> I don't think so. If we want to handle finding the next best thing, we need to
> make the search more complex than that. This is no worse than having 2 RT tasks
> waking up at the same time while there's only a single big CPU. One of them
> will end up on a medium or a little and we don't provide better guarantees
> here.

I may have misunderstood your patch before, do you mean this:
1. the cpu has to be inversion, if not, the cpu's capacity is still
the biggest, although the sysctl_sched_uclamp_util_min_rt_default
=1024, it still can put on the cpu.
2. If the cpu is inversion, the thermal pressure should be considered,
at this time, if the sysctl_sched_uclamp_util_min_rt_default is not
1024, make the rt still have chance to select the cpu.
If the sysctl_sched_uclamp_util_min_rt_default is 1024, all of the
cpu actually can not fit the rt, at this time, select cpu without
considering the cap_orig_of(cpu). The worst thing may be that rt
would put on the small core.

I understand right? If so, Perhaps this approach has the least impact
on the current code complexity.

Thanks!
BR
---
xuewen

\
 
 \ /
  Last update: 2022-04-27 03:39    [W:0.148 / U:3.152 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site