Re: [RESEND PATCH 2/2] sched/fair: Optimize __update_sched_avg()

[Yuyang Du]

On Fri, Mar 31, 2017 at 03:13:55AM +0800, Yuyang Du wrote:
> On Thu, Mar 30, 2017 at 04:14:28PM +0200, Peter Zijlstra wrote:
> > On Thu, Mar 30, 2017 at 02:16:58PM +0200, Peter Zijlstra wrote:
> > > On Thu, Mar 30, 2017 at 04:21:08AM -0700, Paul Turner wrote:
> > 
> > > > > +
> > > > > +       if (unlikely(periods >= LOAD_AVG_MAX_N))
> > > > >                 return LOAD_AVG_MAX;
> > 
> > > > 
> > > > Is this correct in the iterated periods > LOAD_AVG_MAX_N case?
> > > > I don't think the decay above is guaranteed to return these to zero.
> > > 
> > > Ah!
> > > 
> > > Indeed, so decay_load() needs LOAD_AVG_PERIOD * 63 before it truncates
> > > to 0, because every LOAD_AVG_PERIOD we half the value; loose 1 bit; so
> > > 63 of those and we're 0.
> > > 
> > > But __accumulate_sum() OTOH returns LOAD_AVG_MAX after only
> > > LOAD_AVG_MAX_N, which < LOAD_AVG_PERIOD * 63.
> > > 
> > > So yes, combined we exceed LOAD_AVG_MAX, which is bad. Let me think what
> > > to do about that.
> > 
> > 
> > So at the very least it should be decay_load(LOAD_AVG_MAX, 1) (aka
> > LOAD_AVG_MAX - 1024), but that still doesn't account for the !0
> > decay_load() of the first segment.
> > 
> > I'm thinking that we can compute the middle segment, by taking the max
> > value and chopping off the ends, like:
> > 
> > 
> >              p
> >  c2 = 1024 \Sum y^n
> >             n=1
> > 
> >               inf        inf
> >     = 1024 ( \Sum y^n - \Sum y^n - y^0 )
> >               n=0        n=p
>  
> It looks surprisingly kinda works :)
>  
> > +	c2 = LOAD_AVG_MAX - decay_load(LOAD_AVG_MAX, periods) - 1024;
>                             ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> But, I'm not sure               this is what you want (just assume p==0).
> 

Oh, what I meant is when p != 0, actually p>=1.

And thinking about it for a while, it's really what you want, brilliant :)