Messages in this thread | | | Date | Tue, 23 Oct 2018 11:29:37 +0530 | From | Pavan Kondeti <> | Subject | Re: [PATCH v4 2/2] sched/fair: update scale invariance of PELT |
| |
Hi Vincent,
On Fri, Oct 19, 2018 at 06:17:51PM +0200, Vincent Guittot wrote: > > /* > + * The clock_pelt scales the time to reflect the effective amount of > + * computation done during the running delta time but then sync back to > + * clock_task when rq is idle. > + * > + * > + * absolute time | 1| 2| 3| 4| 5| 6| 7| 8| 9|10|11|12|13|14|15|16 > + * @ max capacity ------******---------------******--------------- > + * @ half capacity ------************---------************--------- > + * clock pelt | 1| 2| 3| 4| 7| 8| 9| 10| 11|14|15|16 > + * > + */ > +void update_rq_clock_pelt(struct rq *rq, s64 delta) > +{ > + > + if (is_idle_task(rq->curr)) { > + u32 divider = (LOAD_AVG_MAX - 1024 + rq->cfs.avg.period_contrib) << SCHED_CAPACITY_SHIFT; > + u32 overload = rq->cfs.avg.util_sum + LOAD_AVG_MAX; > + overload += rq->avg_rt.util_sum; > + overload += rq->avg_dl.util_sum; > + > + /* > + * Reflecting some stolen time makes sense only if the idle > + * phase would be present at max capacity. As soon as the > + * utilization of a rq has reached the maximum value, it is > + * considered as an always runnnig rq without idle time to > + * steal. This potential idle time is considered as lost in > + * this case. We keep track of this lost idle time compare to > + * rq's clock_task. > + */ > + if (overload >= divider) > + rq->lost_idle_time += rq_clock_task(rq) - rq->clock_pelt; > +
I am trying to understand this better. I believe we run into this scenario, when the frequency is limited due to thermal/userspace constraints. Lets say frequency is limited to Fmax/2. A 50% task at Fmax, becomes 100% running at Fmax/2. The utilization is built up to 100% after several periods. The clock_pelt runs at 1/2 speed of the clock_task. We are loosing the idle time all along. What happens when the CPU enters idle for a short duration and comes back to run this 100% utilization task?
If the above block is not present i.e lost_idle_time is not tracked, we stretch the idle time (since clock_pelt is synced to clock_task) and the utilization is dropped. Right?
With the above block, we don't stretch the idle time. In fact we don't consider the idle time at all. Because,
idle_time = now - last_time;
idle_time = (rq->clock_pelt - rq->lost_idle_time) - last_time idle_time = (rq->clock_task - rq_clock_task + rq->clock_pelt_old) - last_time idle_time = rq->clock_pelt_old - last_time
The last time is nothing but the last snapshot of the rq->clock_pelt when the task entered sleep due to which CPU entered idle.
Can you please explain the significance of the above block with an example?
> + > + /* The rq is idle, we can sync to clock_task */ > + rq->clock_pelt = rq_clock_task(rq); > + > + > + } else { > + /* > + * When a rq runs at a lower compute capacity, it will need > + * more time to do the same amount of work than at max > + * capacity: either because it takes more time to compute the > + * same amount of work or because taking more time means > + * sharing more often the CPU between entities. > + * In order to be invariant, we scale the delta to reflect how > + * much work has been really done. > + * Running at lower capacity also means running longer to do > + * the same amount of work and this results in stealing some > + * idle time that will disturb the load signal compared to > + * max capacity; This stolen idle time will be automaticcally > + * reflected when the rq will be idle and the clock will be > + * synced with rq_clock_task. > + */ > + > + /* > + * scale the elapsed time to reflect the real amount of > + * computation > + */ > + delta = cap_scale(delta, arch_scale_freq_capacity(cpu_of(rq))); > + delta = cap_scale(delta, arch_scale_cpu_capacity(NULL, cpu_of(rq))); > + > + rq->clock_pelt += delta;
AFAICT, the rq->clock_pelt is used for both utilization and load. So the load also becomes a function of CPU uarch now. Is this intentional?
Thanks, Pavan -- Qualcomm India Private Limited, on behalf of Qualcomm Innovation Center, Inc. Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project.
| |