Song,

On Tue, 20 Nov 2018, Muchun Song wrote:
> John Stultz <john.stu...@linaro.org> 于2018年11月20日周二 上午2:16写道:
> >
> > On Mon, Nov 19, 2018 at 6:10 AM, Muchun Song <smuc...@gmail.com> wrote:
> > > The elements of the heads array are a linked list of timer events that
> > > expire at the current time. And it can contain up to LVL_DEPTH levels
> > > and the lower the level represents the smaller the time granularity.
> > >
> > > Now the result is that the function, which will be called when the timer
> > > expires, in the higher-level is called first than the lower-level 
> > > function.
> > > I think it might be better to call the lower-level timer function first
> > > than the higher-level function. Because the lower-level has the smaller
> > > granularity and delay has less impact on higher-level. So fix it.
> >
> > Interesting.
> >
> > Do you have any specific examples of where this was helpful?  Maybe
> > data on how much this helped the case your concerned about?
> >
> 
>                      heads(with HZ > 100)
> +-----+-----+-----+-----+-----+-----+-----+-----+-----+
> |  0  |  1  |  2  |  3  |  4  |  5  |  6  |  7  |  8  |
> +--+--+--+--+--+--+-----+-----+-----+-----+-----+-----+
>    |     |     |
>    |     |     |
>    |     |     +--->timer4--->timer5
>    |     +--->timer1--->timer2--->timer3
>    +--->timer0
> 
> If we have 6 timers that expire at the current time. And the heads array
> layout as shown above. The collect_expired_timers() will return 3. If timer0
> belong to the first wheel level(level 0), timer1-3 belong to level 2 and
> timer4-5 belong to level 5.
> 
> Follow the current code logic, the timer0 function is called until the
> function call of timer1-5 is completed. So the delay of timer0 is the time
> spent by other timer function calls. If we can call the timer function in
> the following order, this should be more friendly to lower-level timers.
> 
>         timer0->timer1->->timer2->->timer3->->timer4->->timer5
> 
> Although not friendly to higher-level timers, higher-level has larger
> granularity. Therefore the delay has less impact on higher-level.

Well yes, that's clear. But is it a problem in practice and if so, what is
the measurable benefit.

Thanks,

        tglx

Reply via email to