Hi Tim,
Le mardi 23 mars 2021 à 14:37:59 (-0700), Tim Chen a écrit :
>
>
> On 1/29/21 9:27 AM, Vincent Guittot wrote:
> >
> > The patch below moves the update of the blocked load of CPUs outside
> > newidle_balance().
>
> On a well known database workload, we also saw a lot of overhead to do
> update_blocked_averages
> in newidle_balance(). So changes to reduce this overhead is much welcomed.
>
> Turning on cgroup induces 9% throughput degradation on a 2 socket 40 cores
> per socket Icelake system.
>
> A big part of the overhead in our database workload comes from updating
> blocked averages in newidle_balance, caused by I/O threads making
> some CPUs go in and out of idle frequently in the following code path:
>
> ----__blkdev_direct_IO_simple
> |
> |----io_schedule_timeout
> | |
> | ----schedule_timeout
> | |
> | ----schedule
> | |
> | ----__schedule
> | |
> | ----pick_next_task_fair
> | |
> |
> ----newidle_balance
> | |
>
>
> ----update_blocked_averages
>
> We found update_blocked_averages() now consumed most CPU time, eating up 2%
> of the CPU cycles once cgroup
> gets turned on.
>
> I hacked up Joe's original patch to rate limit the update of blocked
> averages called from newidle_balance(). The 9% throughput degradation
> reduced to
> 5.4%. We'll be testing Vincent's change to see if it can give
> similar performance improvement.
>
> Though in our test environment, sysctl_sched_migration_cost was kept
> much lower (25000) compared to the default (500000), to encourage migrations
> to idle cpu
> and reduce latency. We got quite a lot of calls to update_blocked_averages
> directly
> and then try to load_balance in newidle_balance instead of relegating
> the responsibility to idle load balancer. (See code snippet in
> newidle_balance below)
>
>
> if (this_rq->avg_idle < sysctl_sched_migration_cost ||
> <-----sched_migration_cost check
> !READ_ONCE(this_rq->rd->overload)) {
>
> rcu_read_lock();
> sd = rcu_dereference_check_sched_domain(this_rq->sd);
> if (sd)
> update_next_balance(sd, &next_balance);
> rcu_read_unlock();
>
> goto out; <--- invoke idle load balancer
> }
>
> raw_spin_unlock(&this_rq->lock);
>
> update_blocked_averages(this_cpu);
>
> .... followed by load balance code ---
>
> So the update_blocked_averages offload to idle_load_balancer in Vincent's
> patch is less
> effective in this case with small sched_migration_cost.
>
> Looking at the code a bit more, we don't actually load balance every time in
> this code path
> unless our avg_idle time exceeds some threshold. Doing
> update_blocked_averages immediately
IIUC your problem, we call update_blocked_averages() but because of:
if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
update_next_balance(sd, &next_balance);
break;
}
the for_each_domain loop stops even before running load_balance on the 1st
sched domain level which means that update_blocked_averages() was called
unnecessarily.
And this is even more true with a small sysctl_sched_migration_cost which
allows newly
idle LB for very small this_rq->avg_idle. We could wonder why you set such a
low value
for sysctl_sched_migration_cost which is lower than the max_newidle_lb_cost of
the
smallest domain but that's probably because of task_hot().
if avg_idle is lower than the sd->max_newidle_lb_cost of the 1st sched_domain,
we should
skip spin_unlock/lock and for_each_domain() loop entirely
Maybe something like below:
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 76e33a70d575..08933e0d87ed 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -10723,17 +10723,21 @@ static int newidle_balance(struct rq *this_rq, struct
rq_flags *rf)
*/
rq_unpin_lock(this_rq, rf);
+ rcu_read_lock();
+ sd = rcu_dereference_check_sched_domain(this_rq->sd);
+
if (this_rq->avg_idle < sysctl_sched_migration_cost ||
- !READ_ONCE(this_rq->rd->overload)) {
+ !READ_ONCE(this_rq->rd->overload) ||
+ (sd && this_rq->avg_idle < sd->max_newidle_lb_cost)) {
- rcu_read_lock();
- sd = rcu_dereference_check_sched_domain(this_rq->sd);
if (sd)
update_next_balance(sd, &next_balance);
rcu_read_unlock();
goto out;
}
+ rcu_read_unlock();
+
raw_spin_unlock(&this_rq->lock);
> is only needed if we do call load_balance(). If we don't do any load balance
> in the code path,
> we can let the idle load balancer update the blocked averages lazily.
>
> Something like the following perhaps on top of Vincent's patch? We haven't
> really tested
> this change yet but want to see if this change makes sense to you.
>
> Tim
>
> Signed-off-by: Tim Chen <tim.c.c...@linux.intel.com>
> ---
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 63950d80fd0b..b93f5f52658a 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -10591,6 +10591,7 @@ static int newidle_balance(struct rq *this_rq, struct
> rq_flags *rf)
> struct sched_domain *sd;
> int pulled_task = 0;
> u64 curr_cost = 0;
> + bool updated_blocked_avg = false;
>
> update_misfit_status(NULL, this_rq);
> /*
> @@ -10627,7 +10628,6 @@ static int newidle_balance(struct rq *this_rq, struct
> rq_flags *rf)
>
> raw_spin_unlock(&this_rq->lock);
>
> - update_blocked_averages(this_cpu);
> rcu_read_lock();
> for_each_domain(this_cpu, sd) {
> int continue_balancing = 1;
> @@ -10639,6 +10639,11 @@ static int newidle_balance(struct rq *this_rq,
> struct rq_flags *rf)
> }
>
> if (sd->flags & SD_BALANCE_NEWIDLE) {
> + if (!updated_blocked_avg) {
> + update_blocked_averages(this_cpu);
> + updated_blocked_avg = true;
> + }
> +
> t0 = sched_clock_cpu(this_cpu);
>
> pulled_task = load_balance(this_cpu, this_rq,
>
>
