Messages in this thread | | | From | Vincent Guittot <> | Date | Fri, 27 Mar 2015 10:05:29 +0100 | Subject | Re: [PATCH v10 04/11] sched: Make sched entity usage tracking scale-invariant |
| |
On 27 March 2015 at 09:17, Vincent Guittot <vincent.guittot@linaro.org> wrote: > On 26 March 2015 at 18:38, Morten Rasmussen <morten.rasmussen@arm.com> wrote: >> On Wed, Mar 25, 2015 at 06:08:42PM +0000, Vincent Guittot wrote: >>> On 25 March 2015 at 18:33, Peter Zijlstra <peterz@infradead.org> wrote: >>> > On Tue, Mar 24, 2015 at 11:00:57AM +0100, Vincent Guittot wrote: >>> >> On 23 March 2015 at 14:19, Peter Zijlstra <peterz@infradead.org> wrote: >>> >> > On Fri, Feb 27, 2015 at 04:54:07PM +0100, Vincent Guittot wrote: >>> >> > >>> >> >> + unsigned long scale_freq = arch_scale_freq_capacity(NULL, cpu); >>> >> > >>> >> >> + sa->running_avg_sum += delta_w * scale_freq >>> >> >> + >> SCHED_CAPACITY_SHIFT; >>> >> > >>> >> > so the only thing that could be improved is somehow making this >>> >> > multiplication go away when the arch doesn't implement the function. >>> >> > >>> >> > But I'm not sure how to do that without #ifdef. >>> >> > >>> >> > Maybe a little something like so then... that should make the compiler >>> >> > get rid of those multiplications unless the arch needs them. >>> >> >>> >> yes, it removes useless multiplication when not used by an arch. >>> >> It also adds a constraint on the arch side which have to define >>> >> arch_scale_freq_capacity like below: >>> >> >>> >> #define arch_scale_freq_capacity xxx_arch_scale_freq_capacity >>> >> with xxx_arch_scale_freq_capacity an architecture specific function >>> > >>> > Yeah, but it not being weak should make that a compile time warn/fail, >>> > which should be pretty easy to deal with. >>> > >>> >> If it sounds acceptable i can update the patch with your proposal ? >>> > >>> > I'll stick it to the end, I just wanted to float to patch to see if >>> > people had better solutions. >>> >>> ok. all other methods that i have tried, was removing the optimization >>> when default arch_scale_freq_capacity was used >> >> Another potential solution is to stay with weak functions but move the >> multiplication and shift into the arch_scale_*() functions by passing >> the value we want to scale into the arch_scale_*() function. That way we >> can completely avoid multiplication and shift in the default case (no >> arch_scale*() implementations, which is better than what we have today. > > the sched_rt_avg_update only uses the mul with > arch_scale_freq_capacity because the shift by SCHED_CAPACITY_SHIFT has > been factorized in scale_rt_capacity
when arch_scale_freq_capacity is not defined by an arch, the mul is optimized in a lsl 10
> >> >> The only downside is that for frequency invariance we need three >> arch_scale_freq_capacity() calls instead of two.
| |