Re: [PATCH V2] cpuidle: Change ktime_get() with local_clock()

From: Peter Zijlstra
Date: Fri Apr 22 2016 - 02:42:48 EST


On Thu, Apr 21, 2016 at 09:41:14PM +0200, Rafael J. Wysocki wrote:
> On Thu, Apr 21, 2016 at 10:56 AM, Daniel Lezcano
> <daniel.lezcano@xxxxxxxxxx> wrote:
> > The ktime_get() can have a non negligeable overhead, use local_clock()
> > instead.
> >
> > In order to test the difference between ktime_get() and local_clock(),
> > a quick hack has been added to trigger, via debugfs, 10000 times a
> > call to ktime_get() and local_clock() and measure the elapsed time.
> >
> > Then the average value, the min and max is computed for each call.
> >
> > From userspace, the test above was called 100 times every 2 seconds.
> >
> > So, ktime_get() and local_clock() have been called 1000000 times in
> > total.
> >
> > The results are:
> >
> > ktime_get():
> > ============
> > * average: 101 ns (stddev: 27.4)
> > * maximum: 38313 ns
> > * minimum: 65 ns
> >
> > local_clock():
> > ==============
> > * average: 60 ns (stddev: 9.8)
> > * maximum: 13487 ns
> > * minimum: 46 ns
> >
> > The local_clock() is faster and more stable.
> >
> > Even if it is a drop in the ocean, changing the ktime_get() by the
> > local_clock() allows to save 80ns at idle time (entry + exit). And
> > in some circumstances, especially when there are several CPUs racing
> > for the clock access, we save tens of microseconds.
> >
> > The idle duration resulting from a diff is converted from nanosec to
> > microsec. This could be done with integer division (div 1000) - which is
> > an expensive operation or by 10 bits shifting (div 1024) - which is fast
> > but unprecise.
> >
> > The following table gives some results at the limits.
> >
> > ------------------------------------------
> > | nsec | div(1000) | div(1024) |
> > ------------------------------------------
> > | 1e3 | 1 usec | 976 nsec |
> > ------------------------------------------
> > | 1e6 | 1000 usec | 976 usec |
> > ------------------------------------------
> > | 1e9 | 1000000 usec | 976562 usec |
> > ------------------------------------------
> >
> > There is a linear deviation of 2.34%. This loss of precision is acceptable
> > in the context of the resulting diff which is used for statistics. These
> > ones are processed to guess estimate an approximation of the duration of the
> > next idle period which ends up into an idle state selection. The selection
> > criteria takes into account the next duration based on large intervals,
> > represented by the idle state's target residency.
> >
> > The 2^10 division is enough because the approximation regarding the 1e3
> > division is lost in all the approximations done for the next idle duration
> > computation.
> >
> > Signed-off-by: Daniel Lezcano <daniel.lezcano@xxxxxxxxxx>
>
> Looks good to me.
>
> Peter, are you happy with the changelog now?

Yep, works for me:

Acked-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>