Re: [RFC PATCH v2 1/2] sched/fair: Introduce short duration task check

[Chen Yu <yu.c.chen@intel.com>]

Hi Peter,
thanks for your review,
On 2022-11-03 at 13:44:12 +0100, Peter Zijlstra wrote:
> On Sun, Oct 23, 2022 at 11:32:31PM +0800, Chen Yu wrote:
> 
> > +	if (sched_feat(SIS_SHORT) && !prev->on_rq) {
> > +		/*
> > +		 * sum_exec_runtime has been updated in update_curr()
> > +		 * because we reach here via dequeue.
> > +		 */
> > +		this_dur_avg = se->sum_exec_runtime - se->prev_sum_runtime_vol;
> > +		/*
> > +		 * Record the accumulated runtime when task voluntarily
> > +		 * switches out. End of old duration period, a new period
> > +		 * starts.
> > +		 */
> > +		se->prev_sum_runtime_vol = se->sum_exec_runtime;
> > +
> > +		last_dur_avg = se->dur_avg;
> > +		delta = this_dur_avg - last_dur_avg;
> > +		/* consider large change to avoid frequent update */
> > +		if (abs(delta) >= sysctl_sched_min_granularity) {
> > +			/*
> > +			 * If it is the first time the task starts to
> > +			 * record dur_avg, discard the initial value 0.
> > +			 * Otherwise, calculate the EWMA.
> > +			 */
> > +			if (unlikely(!this_dur_avg))
> > +				se->dur_avg = this_dur_avg;
> > +			else
> > +				update_avg(&se->dur_avg, this_dur_avg);
> > +		}
> > +	}
> 
> This seems excessively convoluted; what's wrong with something like:
> 
> 	if (sched_feat(SIS_SHORT) && !prev->on_rq) {
> 		u64 this_dur = se->sum_exec_runtime - se->prev_sum_exec_runtime_vol;
> 		se->prev_sum_exec_runtime_vol = se->sum_exec_runtime;
> 
> 		update_avg(&se->dur_avg, this_dur);
> 	}
> 
> All that other stuff just makes it unreadable and probably slower.
OK, I'll try your suggestion in next version.

thanks,
Chenyu