The try_to_wake_up function has an optimization where it can queue
a task for wakeup on its previous CPU, if the task is still in the
middle of going to sleep inside schedule().
Once schedule() re-enables IRQs, the task will be woken up with an
IPI, and placed back on the runqueue.
If we have such a wakeup pending, there is no need to search other
CPUs for runnable tasks. Just skip (or bail out early from) newidle
balancing, and run the just woken up task.
For a memcache like workload test, this reduces total CPU use by
about 2%, proportionally split between user and system time,
and p99 and p95 application response time by 10% on average.
The schedstats run_delay number shows a similar improvement.
Signed-off-by: Rik van Riel <[email protected]>
---
kernel/sched/fair.c | 18 ++++++++++++++++--
1 file changed, 16 insertions(+), 2 deletions(-)
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 69680158963f..fd80175c3b3e 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -10594,6 +10594,14 @@ static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
u64 curr_cost = 0;
update_misfit_status(NULL, this_rq);
+
+ /*
+ * There is a task waiting to run. No need to search for one.
+ * Return 0; the task will be enqueued when switching to idle.
+ */
+ if (this_rq->ttwu_pending)
+ return 0;
+
/*
* We must set idle_stamp _before_ calling idle_balance(), such that we
* measure the duration of idle_balance() as idle time.
@@ -10661,7 +10669,8 @@ static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
* Stop searching for tasks to pull if there are
* now runnable tasks on this rq.
*/
- if (pulled_task || this_rq->nr_running > 0)
+ if (pulled_task || this_rq->nr_running > 0 ||
+ this_rq->ttwu_pending)
break;
}
rcu_read_unlock();
@@ -10688,7 +10697,12 @@ static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
if (this_rq->nr_running != this_rq->cfs.h_nr_running)
pulled_task = -1;
- if (pulled_task)
+ /*
+ * If we are no longer idle, do not let the time spent here pull
+ * down this_rq->avg_idle. That could lead to newidle_balance not
+ * doing enough work, and the CPU actually going idle.
+ */
+ if (pulled_task || this_rq->ttwu_pending)
this_rq->idle_stamp = 0;
rq_repin_lock(this_rq, rf);
--
2.25.4
Hi Rik,
On Tue, 20 Apr 2021 at 18:07, Rik van Riel <[email protected]> wrote:
>
> The try_to_wake_up function has an optimization where it can queue
> a task for wakeup on its previous CPU, if the task is still in the
> middle of going to sleep inside schedule().
>
> Once schedule() re-enables IRQs, the task will be woken up with an
> IPI, and placed back on the runqueue.
>
> If we have such a wakeup pending, there is no need to search other
> CPUs for runnable tasks. Just skip (or bail out early from) newidle
> balancing, and run the just woken up task.
>
> For a memcache like workload test, this reduces total CPU use by
> about 2%, proportionally split between user and system time,
> and p99 and p95 application response time by 10% on average.
> The schedstats run_delay number shows a similar improvement.
>
> Signed-off-by: Rik van Riel <[email protected]>
> ---
> kernel/sched/fair.c | 18 ++++++++++++++++--
> 1 file changed, 16 insertions(+), 2 deletions(-)
>
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 69680158963f..fd80175c3b3e 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -10594,6 +10594,14 @@ static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
> u64 curr_cost = 0;
>
> update_misfit_status(NULL, this_rq);
> +
> + /*
> + * There is a task waiting to run. No need to search for one.
> + * Return 0; the task will be enqueued when switching to idle.
> + */
> + if (this_rq->ttwu_pending)
> + return 0;
> +
> /*
> * We must set idle_stamp _before_ calling idle_balance(), such that we
> * measure the duration of idle_balance() as idle time.
> @@ -10661,7 +10669,8 @@ static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
> * Stop searching for tasks to pull if there are
> * now runnable tasks on this rq.
> */
> - if (pulled_task || this_rq->nr_running > 0)
> + if (pulled_task || this_rq->nr_running > 0 ||
> + this_rq->ttwu_pending)
> break;
> }
> rcu_read_unlock();
> @@ -10688,7 +10697,12 @@ static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
> if (this_rq->nr_running != this_rq->cfs.h_nr_running)
> pulled_task = -1;
>
> - if (pulled_task)
> + /*
> + * If we are no longer idle, do not let the time spent here pull
> + * down this_rq->avg_idle. That could lead to newidle_balance not
> + * doing enough work, and the CPU actually going idle.
> + */
> + if (pulled_task || this_rq->ttwu_pending)
I'm still running some benchmarks to evaluate the impact of your patch
and more especially the line above which clears this_rq->idle_stamp
and skips the time spent in newidle_balance from being accounted for
in avg_idle. I have some results which show some regression because
of this test especially with hackbench.
On large system, the time spent in newidle_balance can be significant
and we can't ignore it just because this_rq->ttwu_pending is set while
looping the domains because without newidle_balance the idle time
would have been large and we end up screwing up the metric
> this_rq->idle_stamp = 0;
>
> rq_repin_lock(this_rq, rf);
> --
> 2.25.4
>
>
The following commit has been merged into the sched/core branch of tip:
Commit-ID: 9c9f520a14670ad59da2f700660f7601ec9e0b07
Gitweb: https://git.kernel.org/tip/9c9f520a14670ad59da2f700660f7601ec9e0b07
Author: Rik van Riel <[email protected]>
AuthorDate: Tue, 20 Apr 2021 12:07:05 -04:00
Committer: Peter Zijlstra <[email protected]>
CommitterDate: Wed, 21 Apr 2021 13:55:43 +02:00
sched,fair: skip newidle_balance if a wakeup is pending
The try_to_wake_up function has an optimization where it can queue
a task for wakeup on its previous CPU, if the task is still in the
middle of going to sleep inside schedule().
Once schedule() re-enables IRQs, the task will be woken up with an
IPI, and placed back on the runqueue.
If we have such a wakeup pending, there is no need to search other
CPUs for runnable tasks. Just skip (or bail out early from) newidle
balancing, and run the just woken up task.
For a memcache like workload test, this reduces total CPU use by
about 2%, proportionally split between user and system time,
and p99 and p95 application response time by 10% on average.
The schedstats run_delay number shows a similar improvement.
Signed-off-by: Rik van Riel <[email protected]>
Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
Link: https://lkml.kernel.org/r/[email protected]
---
kernel/sched/fair.c | 18 ++++++++++++++++--
1 file changed, 16 insertions(+), 2 deletions(-)
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 1d75af1..83cd2bd 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -10592,6 +10592,14 @@ static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
u64 curr_cost = 0;
update_misfit_status(NULL, this_rq);
+
+ /*
+ * There is a task waiting to run. No need to search for one.
+ * Return 0; the task will be enqueued when switching to idle.
+ */
+ if (this_rq->ttwu_pending)
+ return 0;
+
/*
* We must set idle_stamp _before_ calling idle_balance(), such that we
* measure the duration of idle_balance() as idle time.
@@ -10657,7 +10665,8 @@ static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
* Stop searching for tasks to pull if there are
* now runnable tasks on this rq.
*/
- if (pulled_task || this_rq->nr_running > 0)
+ if (pulled_task || this_rq->nr_running > 0 ||
+ this_rq->ttwu_pending)
break;
}
rcu_read_unlock();
@@ -10684,7 +10693,12 @@ out:
if (time_after(this_rq->next_balance, next_balance))
this_rq->next_balance = next_balance;
- if (pulled_task)
+ /*
+ * If we are no longer idle, do not let the time spent here pull
+ * down this_rq->avg_idle. That could lead to newidle_balance not
+ * doing enough work, and the CPU actually going idle.
+ */
+ if (pulled_task || this_rq->ttwu_pending)
this_rq->idle_stamp = 0;
else
nohz_newidle_balance(this_rq);
On Thu, Apr 22, 2021 at 07:36:00AM -0000, tip-bot2 for Rik van Riel wrote:
> @@ -10684,7 +10693,12 @@ out:
> if (time_after(this_rq->next_balance, next_balance))
> this_rq->next_balance = next_balance;
>
> - if (pulled_task)
> + /*
> + * If we are no longer idle, do not let the time spent here pull
> + * down this_rq->avg_idle. That could lead to newidle_balance not
> + * doing enough work, and the CPU actually going idle.
> + */
> + if (pulled_task || this_rq->ttwu_pending)
> this_rq->idle_stamp = 0;
I've un-committed this patch, because vingu was reporting increased idle
time because of this hunk. I had mistakenly assumed that was sorted
with v3, sorry for not keeping better track of things.
(also, now that I look again, please also fix the Subject to have a
capital after the :)
On Wed, 21 Apr 2021 at 19:27, Vincent Guittot
<[email protected]> wrote:
>
> Hi Rik,
>
> On Tue, 20 Apr 2021 at 18:07, Rik van Riel <[email protected]> wrote:
> >
> > The try_to_wake_up function has an optimization where it can queue
> > a task for wakeup on its previous CPU, if the task is still in the
> > middle of going to sleep inside schedule().
> >
> > Once schedule() re-enables IRQs, the task will be woken up with an
> > IPI, and placed back on the runqueue.
> >
> > If we have such a wakeup pending, there is no need to search other
> > CPUs for runnable tasks. Just skip (or bail out early from) newidle
> > balancing, and run the just woken up task.
> >
> > For a memcache like workload test, this reduces total CPU use by
> > about 2%, proportionally split between user and system time,
> > and p99 and p95 application response time by 10% on average.
> > The schedstats run_delay number shows a similar improvement.
> >
> > Signed-off-by: Rik van Riel <[email protected]>
> > ---
> > kernel/sched/fair.c | 18 ++++++++++++++++--
> > 1 file changed, 16 insertions(+), 2 deletions(-)
> >
> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > index 69680158963f..fd80175c3b3e 100644
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -10594,6 +10594,14 @@ static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
> > u64 curr_cost = 0;
> >
> > update_misfit_status(NULL, this_rq);
> > +
> > + /*
> > + * There is a task waiting to run. No need to search for one.
> > + * Return 0; the task will be enqueued when switching to idle.
> > + */
> > + if (this_rq->ttwu_pending)
> > + return 0;
> > +
> > /*
> > * We must set idle_stamp _before_ calling idle_balance(), such that we
> > * measure the duration of idle_balance() as idle time.
> > @@ -10661,7 +10669,8 @@ static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
> > * Stop searching for tasks to pull if there are
> > * now runnable tasks on this rq.
> > */
> > - if (pulled_task || this_rq->nr_running > 0)
> > + if (pulled_task || this_rq->nr_running > 0 ||
> > + this_rq->ttwu_pending)
> > break;
> > }
> > rcu_read_unlock();
> > @@ -10688,7 +10697,12 @@ static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
> > if (this_rq->nr_running != this_rq->cfs.h_nr_running)
> > pulled_task = -1;
> >
> > - if (pulled_task)
> > + /*
> > + * If we are no longer idle, do not let the time spent here pull
> > + * down this_rq->avg_idle. That could lead to newidle_balance not
> > + * doing enough work, and the CPU actually going idle.
> > + */
> > + if (pulled_task || this_rq->ttwu_pending)
>
> I'm still running some benchmarks to evaluate the impact of your patch
> and more especially the line above which clears this_rq->idle_stamp
> and skips the time spent in newidle_balance from being accounted for
> in avg_idle. I have some results which show some regression because
> of this test especially with hackbench.
> On large system, the time spent in newidle_balance can be significant
> and we can't ignore it just because this_rq->ttwu_pending is set while
> looping the domains because without newidle_balance the idle time
> would have been large and we end up screwing up the metric
I confirmed that the line above generate hackbench regression on my
large arm64 system (2 * 112 CPUs)
I'm testing hackbench with various number of group : 1, 2, 4, 16, 32,
64, 128, 256 but I have only put the 2 results which significantly
regress. The other ones are in the +/-1% variation range
hackbench -g $group
group v5.12-rc8+tip w/ this patch w/ this patch without
the line above
64 2.862(+/- 9%) 2.952(+/-11%) -3% 2.807(+/- 7%) +2%
128 3.334(+/-10%) 3.561-+/-13%) -7% 3.181(+/- 6%) +4%
>
> > this_rq->idle_stamp = 0;
> >
> > rq_repin_lock(this_rq, rf);
> > --
> > 2.25.4
> >
> >
On Thu, 2021-04-22 at 10:37 +0200, Vincent Guittot wrote:
> On Wed, 21 Apr 2021 at 19:27, Vincent Guittot
> <[email protected]> wrote:
> >
> > > - if (pulled_task)
> > > + /*
> > > + * If we are no longer idle, do not let the time spent
> > > here pull
> > > + * down this_rq->avg_idle. That could lead to
> > > newidle_balance not
> > > + * doing enough work, and the CPU actually going idle.
> > > + */
> > > + if (pulled_task || this_rq->ttwu_pending)
> >
> I confirmed that the line above generate hackbench regression on my
> large arm64 system (2 * 112 CPUs)
> I'm testing hackbench with various number of group : 1, 2, 4, 16, 32,
> 64, 128, 256 but I have only put the 2 results which significantly
> regress. The other ones are in the +/-1% variation range
>
> hackbench -g $group
>
> group v5.12-rc8+tip w/ this patch w/ this patch
> without
> the line above
> 64 2.862(+/- 9%) 2.952(+/-11%) -3% 2.807(+/- 7%) +2%
> 128 3.334(+/-10%) 3.561-+/-13%) -7% 3.181(+/- 6%) +4%
OK, I guess this part of the patch needs additional work.
I'll send a v4 with just the first two changes.
Thank you for running those tests.
--
All Rights Reversed.