Re: [PATCH V3]hrtimer: Fix a performance regression by disablereprogramming in remove_hrtimer

From: Mike Galbraith
Date: Tue Aug 06 2013 - 03:29:17 EST

On Tue, 2013-07-30 at 11:35 +0200, Peter Zijlstra wrote:

> It would be good if you could do what Thomas suggested and look at which
> timer is actually active during your workload.

Rebuilding regression test trees, some pipe-test results...

I'm missing mwait_idle() rather a lot on Q6600, and at 3.8, E5620 took a
severe NOHZ drubbing from the menu governor.

pipe-test, scheduling cross core

NOTE: nohz is throttled here (patchlet below), as to not eat horrible
microidle cost, see E5620 v3.7.10-nothrottle below.

Q6600
v3.8.13 500.6 KHz 1.000
v3.9.11 422.4 KHz .843
v3.10.4 420.2 KHz .839
v3.11-rc3-4-g36f571e 404.7 KHz .808

Q6600 3.9 regression:
guilty party is 69fb3676 x86 idle: remove mwait_idle() and "idle=mwait" cmdline param
halt sucks, HTH does one activate mwait_idle_with_hints() [processor_idle()] for core2 boxen?

E5620 +write 0 -> /dev/cpu_dma_latency, hold open
v3.7.10 578.5 KHz 1.000 675.4 KHz 1.000
v3.7.10-nothrottle 366.7 KHz .633 395.0 KHz .584
v3.8.13 468.3 KHz .809 690.0 KHz 1.021
v3.8.13 idle=mwait 595.1 KHz 1.028 NA
v3.9.11 462.0 KHz .798 691.1 KHz 1.023
v3.10.4 419.4 KHz .724 570.8 KHz .845
v3.11-rc3-4-g36f571e 400.1 KHz .691 538.5 KHz .797

E5620 3.8 regression:
guilty party: 69a37bea cpuidle: Quickly notice prediction failure for repeat mode


Q6600 (2.4 GHz core2 quad)
v3.11-rc3-4-g36f571e v3.8.13
7.97% [k] reschedule_interrupt 8.63% [k] __schedule
6.27% [k] __schedule 6.07% [k] native_sched_clock
4.74% [k] native_sched_clock 4.96% [k] system_call
4.23% [k] _raw_spin_lock_irqsave 4.30% [k] _raw_spin_lock_irqsave
3.39% [k] system_call 4.06% [k] resched_task
2.89% [k] sched_clock_local 3.44% [k] sched_clock_local
2.79% [k] mutex_lock 3.39% [k] pipe_read
2.57% [k] pipe_read 3.21% [k] mutex_lock
2.55% [k] __switch_to 2.98% [k] read_tsc
2.24% [k] read_tsc 2.87% [k] __switch_to


E5620 (2.4 GHz Westmere quad)
v3.7.10 v3.7.10-nothrottle v3.7.10-nothrottle
8.01% [k] __schedule 25.80% [k] _raw_spin_unlock_irqrestore 21.80% [k] _raw_spin_unlock_irqrestore
4.49% [k] resched_tas 4.64% [k] __hrtimer_start_range_ns - _raw_spin_unlock_irqrestore
3.94% [k] mutex_lock 4.62% [k] timerqueue_add + 37.94% __hrtimer_start_range_ns
3.44% [k] __switch_to 4.54% [k] __schedule 19.69% hrtimer_cancel
3.18% [k] menu_select 2.84% [k] enqueue_hrtimer tick_nohz_restart
3.05% [k] copy_user_generic_string 2.64% [k] resched_task tick_nohz_idle_exit
3.02% [k] task_waking_fair 2.29% [k] _raw_spin_lock_irqsave cpu_idle
2.91% [k] mutex_unlock 2.28% [k] mutex_lock start_secondary
2.82% [k] pipe_read 1.96% [k] __switch_to + 16.05% hrtimer_start_range_ns
2.32% [k] ktime_get_real 1.73% [k] menu_select 15.46% hrtimer_start
tick_nohz_stop_sched_tick
__tick_nohz_idle_enter
tick_nohz_idle_enter
cpu_idle
start_secondary
6.37% hrtimer_try_to_cancel
hrtimer_cancel
tick_nohz_restart
tick_nohz_idle_exit
cpu_idle
start_secondary

v3.8.13 v3.8.13 idle=mwait v3.8.13 (throttled, but menu gov bites.. HARD)
23.16% [k] _raw_spin_unlock_irqrestore 8.35% [k] __schedule - 22.91% [k] _raw_spin_unlock_irqrestore
4.93% [k] __schedule 6.49% [k] __switch_to - _raw_spin_unlock_irqrestore
3.42% [k] resched_task 5.71% [k] resched_task - 47.26% hrtimer_try_to_cancel
3.27% [k] __switch_to 4.64% [k] mutex_lock hrtimer_cancel
3.05% [k] mutex_lock 3.48% [k] copy_user_generic_string menu_hrtimer_cancel
2.32% [k] copy_user_generic_string 3.15% [k] task_waking_fair tick_nohz_idle_exit
2.30% [k] _raw_spin_lock_irqsave 3.13% [k] pipe_read cpu_idle
2.15% [k] pipe_read 2.61% [k] mutex_unlock start_secondary
2.15% [k] task_waking_fair 2.54% [k] finish_task_switch - 40.01% __hrtimer_start_range_ns
2.08% [k] ktime_get 2.29% [k] _raw_spin_lock_irqsave hrtimer_start
1.87% [k] mutex_unlock 1.91% [k] idle_cpu menu_select
1.76% [k] finish_task_switch 1.84% [k] __wake_up_common cpuidle_idle_call
cpu_idle
start_secondary

v3.9.11
18.67% [k] _raw_spin_unlock_irqrestore
4.36% [k] __schedule
3.66% [k] __switch_to
3.13% [k] mutex_lock
2.97% [k] __hrtimer_start_range_ns
2.69% [k] _raw_spin_lock_irqsave
2.38% [k] copy_user_generic_string
2.34% [k] hrtimer_reprogram.isra.32
2.34% [k] task_waking_fair
2.25% [k] ktime_get
2.14% [k] pipe_read
1.98% [k] menu_select

v3.10.4
20.42% [k] _raw_spin_unlock_irqrestore
4.75% [k] __schedule
4.42% [k] reschedule_interrupt <== appears in 3.10, guilty party as yet unknown
3.52% [k] __switch_to
3.27% [k] resched_task
2.64% [k] cpuidle_enter_state
2.63% [k] _raw_spin_lock_irqsave
2.04% [k] copy_user_generic_string
2.00% [k] cpu_idle_loop
1.97% [k] mutex_lock
1.90% [k] ktime_get
1.75% [k] task_waking_fair

v3.11-rc3-4-g36f571e
18.96% [k] _raw_spin_unlock_irqrestore
4.84% [k] __schedule
4.69% [k] reschedule_interrupt
3.75% [k] __switch_to
2.62% [k] _raw_spin_lock_irqsave
2.43% [k] cpuidle_enter_state
2.28% [k] resched_task
2.20% [k] cpu_idle_loop
1.97% [k] copy_user_generic_string
1.88% [k] ktime_get
1.81% [k] task_waking_fair
1.75% [k] mutex_lock

sched: ratelimit nohz

Entering nohz code on every micro-idle is too expensive to bear.

Signed-off-by: Mike Galbraith <efault@xxxxxx>

---
include/linux/sched.h | 5 +++++
kernel/sched/core.c | 5 +++++
kernel/time/tick-sched.c | 2 +-
3 files changed, 11 insertions(+), 1 deletion(-)

--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -235,9 +235,14 @@ extern int runqueue_is_locked(int cpu);
extern void nohz_balance_enter_idle(int cpu);
extern void set_cpu_sd_state_idle(void);
extern int get_nohz_timer_target(void);
+extern int sched_needs_cpu(int cpu);
#else
static inline void nohz_balance_enter_idle(int cpu) { }
static inline void set_cpu_sd_state_idle(void) { }
+static inline int sched_needs_cpu(int cpu)
+{
+ return 0;
+}
#endif

/*
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -650,6 +650,11 @@ static inline bool got_nohz_idle_kick(vo
return false;
}

+int sched_needs_cpu(int cpu)
+{
+ return cpu_rq(cpu)->avg_idle < sysctl_sched_migration_cost;
+}
+
#else /* CONFIG_NO_HZ_COMMON */

static inline bool got_nohz_idle_kick(void)
--- a/kernel/time/tick-sched.c
+++ b/kernel/time/tick-sched.c
@@ -548,7 +548,7 @@ static ktime_t tick_nohz_stop_sched_tick
time_delta = timekeeping_max_deferment();
} while (read_seqretry(&jiffies_lock, seq));

- if (rcu_needs_cpu(cpu, &rcu_delta_jiffies) ||
+ if (sched_needs_cpu(cpu) || rcu_needs_cpu(cpu, &rcu_delta_jiffies) ||
arch_needs_cpu(cpu) || irq_work_needs_cpu()) {
next_jiffies = last_jiffies + 1;
delta_jiffies = 1;


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/