Skip to contentSkip to conversation

sched: Make schedstats a runtime tunable that is disabled by default v4

PATCH In review Mel Gorman 3 participants 7 messages Feb 3, 2016

Subsystems (3)
lore
Share & tools
lore

Mel Gorman

author
#0 lore raw

PATCH 1/1 sched: Make schedstats a runtime tunable that is disabled by default v4

Changelog since v3
o Force enable stats during profiling and latencytop

Changelog since V2
o Print stats that are not related to schedstat
o Reintroduce a static inline for update_stats_dequeue

Changelog since V1
o Introduce schedstat_enabled and address Ingo's feedback
o More schedstat-only paths eliminated, particularly ttwu_stat

schedstats is very useful during debugging and performance tuning but it incurs overhead. As such, even though it can be disabled at build time, it is often enabled as the information is useful. This patch adds a kernel command-line and sysctl tunable to enable or disable schedstats on demand. It is disabled by default as someone who knows they need it can also learn to enable it when necessary.

The benefits are workload-dependent but when it gets down to it, the difference will be whether cache misses are incurred updating the shared stats or not. These measurements were taken from a 48-core 2-socket machine with Xeon(R) E5-2670 v3 cpus although they were also tested on a single socket machine 8-core machine with Intel i7-3770 processors.

netperf-tcp

                           4.5.0-rc1             4.5.0-rc1
                             vanilla          nostats-v3r1
Hmean    64         560.45 (  0.00%)      575.98 (  2.77%)
Hmean    128        766.66 (  0.00%)      795.79 (  3.80%)
Hmean    256        950.51 (  0.00%)      981.50 (  3.26%)
Hmean    1024      1433.25 (  0.00%)     1466.51 (  2.32%)
Hmean    2048      2810.54 (  0.00%)     2879.75 (  2.46%)
Hmean    3312      4618.18 (  0.00%)     4682.09 (  1.38%)
Hmean    4096      5306.42 (  0.00%)     5346.39 (  0.75%)
Hmean    8192     10581.44 (  0.00%)    10698.15 (  1.10%)
Hmean    16384    18857.70 (  0.00%)    18937.61 (  0.42%)

Small gains here, UDP_STREAM showed nothing intresting and neither did the TCP_RR tests. The gains on the 8-core machine were very similar.

tbench4

                                 4.5.0-rc1             4.5.0-rc1
                                   vanilla          nostats-v3r1
Hmean    mb/sec-1         500.85 (  0.00%)      522.43 (  4.31%)
Hmean    mb/sec-2         984.66 (  0.00%)     1018.19 (  3.41%)
Hmean    mb/sec-4        1827.91 (  0.00%)     1847.78 (  1.09%)
Hmean    mb/sec-8        3561.36 (  0.00%)     3611.28 (  1.40%)
Hmean    mb/sec-16       5824.52 (  0.00%)     5929.03 (  1.79%)
Hmean    mb/sec-32      10943.10 (  0.00%)    10802.83 ( -1.28%)
Hmean    mb/sec-64      15950.81 (  0.00%)    16211.31 (  1.63%)
Hmean    mb/sec-128     15302.17 (  0.00%)    15445.11 (  0.93%)
Hmean    mb/sec-256     14866.18 (  0.00%)    15088.73 (  1.50%)
Hmean    mb/sec-512     15223.31 (  0.00%)    15373.69 (  0.99%)
Hmean    mb/sec-1024    14574.25 (  0.00%)    14598.02 (  0.16%)
Hmean    mb/sec-2048    13569.02 (  0.00%)    13733.86 (  1.21%)
Hmean    mb/sec-3072    12865.98 (  0.00%)    13209.23 (  2.67%)

Small gains of 2-4% at low thread counts and otherwise flat. The gains on the 8-core machine were slightly different

tbench4 on 8-core i7-3770 single socket machine

Hmean    mb/sec-1        442.59 (  0.00%)      448.73 (  1.39%)
Hmean    mb/sec-2        796.68 (  0.00%)      794.39 ( -0.29%)
Hmean    mb/sec-4       1322.52 (  0.00%)     1343.66 (  1.60%)
Hmean    mb/sec-8       2611.65 (  0.00%)     2694.86 (  3.19%)
Hmean    mb/sec-16      2537.07 (  0.00%)     2609.34 (  2.85%)
Hmean    mb/sec-32      2506.02 (  0.00%)     2578.18 (  2.88%)
Hmean    mb/sec-64      2511.06 (  0.00%)     2569.16 (  2.31%)
Hmean    mb/sec-128     2313.38 (  0.00%)     2395.50 (  3.55%)
Hmean    mb/sec-256     2110.04 (  0.00%)     2177.45 (  3.19%)
Hmean    mb/sec-512     2072.51 (  0.00%)     2053.97 ( -0.89%)

In constract, this shows a relatively steady 2-3% gain at higher thread counts. Due to the nature of the patch and the type of workload, it's not a surprise that the result will depend on the CPU used.

hackbench-pipes

                         4.5.0-rc1             4.5.0-rc1
                           vanilla          nostats-v3r1
Amean    1        0.0637 (  0.00%)      0.0660 ( -3.59%)
Amean    4        0.1229 (  0.00%)      0.1181 (  3.84%)
Amean    7        0.1921 (  0.00%)      0.1911 (  0.52%)
Amean    12       0.3117 (  0.00%)      0.2923 (  6.23%)
Amean    21       0.4050 (  0.00%)      0.3899 (  3.74%)
Amean    30       0.4586 (  0.00%)      0.4433 (  3.33%)
Amean    48       0.5910 (  0.00%)      0.5694 (  3.65%)
Amean    79       0.8663 (  0.00%)      0.8626 (  0.43%)
Amean    110      1.1543 (  0.00%)      1.1517 (  0.22%)
Amean    141      1.4457 (  0.00%)      1.4290 (  1.16%)
Amean    172      1.7090 (  0.00%)      1.6924 (  0.97%)
Amean    192      1.9126 (  0.00%)      1.9089 (  0.19%)

Some small gains and losses and while the variance data is not included, it's close to the noise. The UMA machine did not show anything particularly different

pipetest

                             4.5.0-rc1             4.5.0-rc1
                               vanilla          nostats-v2r2
Min         Time        4.13 (  0.00%)        3.99 (  3.39%)
1st-qrtle   Time        4.38 (  0.00%)        4.27 (  2.51%)
2nd-qrtle   Time        4.46 (  0.00%)        4.39 (  1.57%)
3rd-qrtle   Time        4.56 (  0.00%)        4.51 (  1.10%)
Max-90%     Time        4.67 (  0.00%)        4.60 (  1.50%)
Max-93%     Time        4.71 (  0.00%)        4.65 (  1.27%)
Max-95%     Time        4.74 (  0.00%)        4.71 (  0.63%)
Max-99%     Time        4.88 (  0.00%)        4.79 (  1.84%)
Max         Time        4.93 (  0.00%)        4.83 (  2.03%)
Mean        Time        4.48 (  0.00%)        4.39 (  1.91%)
Best99%Mean Time        4.47 (  0.00%)        4.39 (  1.91%)
Best95%Mean Time        4.46 (  0.00%)        4.38 (  1.93%)
Best90%Mean Time        4.45 (  0.00%)        4.36 (  1.98%)
Best50%Mean Time        4.36 (  0.00%)        4.25 (  2.49%)
Best10%Mean Time        4.23 (  0.00%)        4.10 (  3.13%)
Best5%Mean  Time        4.19 (  0.00%)        4.06 (  3.20%)
Best1%Mean  Time        4.13 (  0.00%)        4.00 (  3.39%)

Small improvement and similar gains were seen on the UMA machine.

The gain is small but it'll depend on the CPU and the workload whether this patch makes a different. However, it stands to reason that doing less work in the scheduler is a good thing. The downside is that the lack of schedstats and tracepoints will be surprising to experts doing performance analysis until they find the existance of the schedstats= parameter or schedstats sysctl.

---
 Documentation/kernel-parameters.txt |   5 ++
 Documentation/sysctl/kernel.txt     |   8 +++
 include/linux/latencytop.h          |   3 ++
 include/linux/sched.h               |   4 ++
 include/linux/sched/sysctl.h        |   4 ++
 kernel/latencytop.c                 |  14 ++++-
 kernel/profile.c                    |   1 +
 kernel/sched/core.c                 |  70 ++++++++++++++++++++++++-
 kernel/sched/debug.c                | 102 +++++++++++++++++++-----------------
 kernel/sched/fair.c                 |  92 +++++++++++++++++++-------------
 kernel/sched/sched.h                |   1 +
 kernel/sched/stats.h                |   8 +--
 kernel/sysctl.c                     |  13 ++++-
 13 files changed, 235 insertions(+), 90 deletions(-)
Documentation/kernel-parameters.txt +5 -0
index 87d40a72f6a1..846956abfe85 100644
@@ -3523,6 +3523,11 @@ bytes respectively. Such letter suffixes can also be entirely omitted.
 
 	sched_debug	[KNL] Enables verbose scheduler debug messages.
 
+	schedstats=	[KNL,X86] Enable or disable scheduled statistics.
+			Allowed values are enable and disable. This feature
+			incurs a small amount of overhead in the scheduler
+			but is useful for debugging and performance tuning.
+
 	skew_tick=	[KNL] Offset the periodic timer tick per cpu to mitigate
 			xtime_lock contention on larger systems, and/or RCU lock
 			contention on all systems with CONFIG_MAXSMP set.
Documentation/sysctl/kernel.txt +8 -0
index a93b414672a7..be7c3b720adf 100644
@@ -760,6 +760,14 @@ rtsig-nr shows the number of RT signals currently queued.
 
 ==============================================================
 
+schedstats:
+
+Enables/disables scheduler statistics. Enabling this feature
+incurs a small amount of overhead in the scheduler but is
+useful for debugging and performance tuning.
+
+==============================================================
+
 sg-big-buff:
 
 This file shows the size of the generic SCSI (sg) buffer.
include/linux/latencytop.h +3 -0
index e23121f9d82a..59ccab297ae0 100644
@@ -37,6 +37,9 @@ account_scheduler_latency(struct task_struct *task, int usecs, int inter)
 
 void clear_all_latency_tracing(struct task_struct *p);
 
+extern int sysctl_latencytop(struct ctl_table *table, int write,
+			void __user *buffer, size_t *lenp, loff_t *ppos);
+
 #else
 
 static inline void
include/linux/sched.h +4 -0
index a10494a94cc3..a292c4b7e94c 100644
@@ -920,6 +920,10 @@ static inline int sched_info_on(void)
 #endif
 }
 
+#ifdef CONFIG_SCHEDSTATS
+void force_schedstat_enabled(void);
+#endif
+
 enum cpu_idle_type {
 	CPU_IDLE,
 	CPU_NOT_IDLE,
include/linux/sched/sysctl.h +4 -0
index c9e4731cf10b..4f080ab4f2cd 100644
@@ -95,4 +95,8 @@ extern int sysctl_numa_balancing(struct ctl_table *table, int write,
 				 void __user *buffer, size_t *lenp,
 				 loff_t *ppos);
 
+extern int sysctl_schedstats(struct ctl_table *table, int write,
+				 void __user *buffer, size_t *lenp,
+				 loff_t *ppos);
+
 #endif /* _SCHED_SYSCTL_H */
kernel/latencytop.c +13 -1
index a02812743a7e..b5c30d9f46c5 100644
@@ -47,12 +47,12 @@
  * of times)
  */
 
-#include <linux/latencytop.h>
 #include <linux/kallsyms.h>
 #include <linux/seq_file.h>
 #include <linux/notifier.h>
 #include <linux/spinlock.h>
 #include <linux/proc_fs.h>
+#include <linux/latencytop.h>
 #include <linux/export.h>
 #include <linux/sched.h>
 #include <linux/list.h>
@@ -289,4 +289,16 @@ static int __init init_lstats_procfs(void)
 	proc_create("latency_stats", 0644, NULL, &lstats_fops);
 	return 0;
 }
+
+int sysctl_latencytop(struct ctl_table *table, int write,
+			void __user *buffer, size_t *lenp, loff_t *ppos)
+{
+	int err;
+
+	err = proc_dointvec(table, write, buffer, lenp, ppos);
+	if (latencytop_enabled)
+		force_schedstat_enabled();
+
+	return err;
+}
 device_initcall(init_lstats_procfs);
kernel/profile.c +1 -0
index 99513e1160e5..51369697466e 100644
@@ -59,6 +59,7 @@ int profile_setup(char *str)
 
 	if (!strncmp(str, sleepstr, strlen(sleepstr))) {
 #ifdef CONFIG_SCHEDSTATS
+		force_schedstat_enabled();
 		prof_on = SLEEP_PROFILING;
 		if (str[strlen(sleepstr)] == ',')
 			str += strlen(sleepstr) + 1;
kernel/sched/core.c +68 -2

94 lines not on this page · raw message

kernel/sched/debug.c +54 -48

135 lines not on this page · raw message

kernel/sched/fair.c +57 -35

161 lines not on this page · raw message

kernel/sched/sched.h +1 -0
index 10f16374df7f..1d583870e1a6 100644
@@ -1022,6 +1022,7 @@ extern struct static_key sched_feat_keys[__SCHED_FEAT_NR];
 #endif /* SCHED_DEBUG && HAVE_JUMP_LABEL */
 
 extern struct static_key_false sched_numa_balancing;
+extern struct static_key_false sched_schedstats;
 
 static inline u64 global_rt_period(void)
 {
kernel/sched/stats.h +5 -3
index b0fbc7632de5..70b3b6a20fb0 100644
@@ -29,9 +29,10 @@ rq_sched_info_dequeued(struct rq *rq, unsigned long long delta)
 	if (rq)
 		rq->rq_sched_info.run_delay += delta;
 }
-# define schedstat_inc(rq, field)	do { (rq)->field++; } while (0)
-# define schedstat_add(rq, field, amt)	do { (rq)->field += (amt); } while (0)
-# define schedstat_set(var, val)	do { var = (val); } while (0)
+# define schedstat_enabled()		static_branch_unlikely(&sched_schedstats)
+# define schedstat_inc(rq, field)	do { if (schedstat_enabled()) { (rq)->field++; } } while (0)
+# define schedstat_add(rq, field, amt)	do { if (schedstat_enabled()) { (rq)->field += (amt); } } while (0)
+# define schedstat_set(var, val)	do { if (schedstat_enabled()) { var = (val); } } while (0)
 #else /* !CONFIG_SCHEDSTATS */
 static inline void
 rq_sched_info_arrive(struct rq *rq, unsigned long long delta)
@@ -42,6 +43,7 @@ rq_sched_info_dequeued(struct rq *rq, unsigned long long delta)
 static inline void
 rq_sched_info_depart(struct rq *rq, unsigned long long delta)
 {}
+# define schedstat_enabled()		0
 # define schedstat_inc(rq, field)	do { } while (0)
 # define schedstat_add(rq, field, amt)	do { } while (0)
 # define schedstat_set(var, val)	do { } while (0)
kernel/sysctl.c +12 -1
index 97715fd9e790..f5102fabef7f 100644
@@ -350,6 +350,17 @@ static struct ctl_table kern_table[] = {
 		.mode		= 0644,
 		.proc_handler	= proc_dointvec,
 	},
+#ifdef CONFIG_SCHEDSTATS
+	{
+		.procname	= "sched_schedstats",
+		.data		= NULL,
+		.maxlen		= sizeof(unsigned int),
+		.mode		= 0644,
+		.proc_handler	= sysctl_schedstats,
+		.extra1		= &zero,
+		.extra2		= &one,
+	},
+#endif /* CONFIG_SCHEDSTATS */
 #endif /* CONFIG_SMP */
 #ifdef CONFIG_NUMA_BALANCING
 	{
@@ -505,7 +516,7 @@ static struct ctl_table kern_table[] = {
 		.data		= &latencytop_enabled,
 		.maxlen		= sizeof(int),
 		.mode		= 0644,
-		.proc_handler	= proc_dointvec,
+		.proc_handler	= sysctl_latencytop,
 	},
 #endif
 #ifdef CONFIG_BLK_DEV_INITRD
Signature · 1 line
2.6.4
#1 lore raw
Quoted 18 lines · Mel Gorman

Changelog since v3
o Force enable stats during profiling and latencytop

Changelog since V2
o Print stats that are not related to schedstat
o Reintroduce a static inline for update_stats_dequeue

Changelog since V1
o Introduce schedstat_enabled and address Ingo's feedback
o More schedstat-only paths eliminated, particularly ttwu_stat

schedstats is very useful during debugging and performance tuning but it incurs overhead. As such, even though it can be disabled at build time, it is often enabled as the information is useful. This patch adds a kernel command-line and sysctl tunable to enable or disable schedstats on demand. It is disabled by default as someone who knows they need it can also learn to enable it when necessary.

Quoted 3 lines

The benefits are workload-dependent but when it gets down to it, the difference will be whether cache misses are incurred updating the shared stats or not. [...]

Hm, which shared stats are those? I think we should really fix those as well:
those shared stats should be percpu collected as well, with no extra cache misses in any scheduler fast path.

Thanks,

Ingo

author
#2 lore raw
Quoted 21 lines · Ingo Molnar
Quoted 18 lines · Mel Gorman

Changelog since v3
o Force enable stats during profiling and latencytop

Changelog since V2
o Print stats that are not related to schedstat
o Reintroduce a static inline for update_stats_dequeue

Changelog since V1
o Introduce schedstat_enabled and address Ingo's feedback
o More schedstat-only paths eliminated, particularly ttwu_stat

schedstats is very useful during debugging and performance tuning but it incurs overhead. As such, even though it can be disabled at build time, it is often enabled as the information is useful. This patch adds a kernel command-line and sysctl tunable to enable or disable schedstats on demand. It is disabled by default as someone who knows they need it can also learn to enable it when necessary.

Quoted 5 lines
Quoted 3 lines

The benefits are workload-dependent but when it gets down to it, the difference will be whether cache misses are incurred updating the shared stats or not. [...]

Hm, which shared stats are those?

Extremely poor phrasing on my part. The stats share a cache line and the impact partially depends on whether unrelated stats share a cache line or not during updates.

Quoted 4 lines

I think we should really fix those as well:
those shared stats should be percpu collected as well, with no extra cache misses in any scheduler fast path.

I looked into that but converting those stats to per-cpu counters would incur sizable memory overhead. There are a *lot* of them and the basic structure for the generic percpu-counter is

struct percpu_counter {
        raw_spinlock_t lock;
        s64 count;
#ifdef CONFIG_HOTPLUG_CPU
        struct list_head list;  /* All percpu_counters are on a list */
#endif
        s32 __percpu *counters;
};

That's not taking the associated runtime overhead such as synchronising them. Granted, some specialised implementation could be done for scheduler but it would be massive overkill and maintenance overhead for stats that most users do not even want.

Signature · 2 lines
Mel Gorman
SUSE Labs
#3 lore raw
Quoted 26 lines · Mel Gorman
Quoted 25 lines · Ingo Molnar
Quoted 21 lines · Mel Gorman

Changelog since v3
o Force enable stats during profiling and latencytop

Changelog since V2
o Print stats that are not related to schedstat
o Reintroduce a static inline for update_stats_dequeue

Changelog since V1
o Introduce schedstat_enabled and address Ingo's feedback
o More schedstat-only paths eliminated, particularly ttwu_stat

schedstats is very useful during debugging and performance tuning but it incurs overhead. As such, even though it can be disabled at build time, it is often enabled as the information is useful. This patch adds a kernel command-line and sysctl tunable to enable or disable schedstats on demand. It is disabled by default as someone who knows they need it can also learn to enable it when necessary.

The benefits are workload-dependent but when it gets down to it, the difference will be whether cache misses are incurred updating the shared stats or not. [...]

Quoted 5 lines
Quoted 1 line

Hm, which shared stats are those?

Extremely poor phrasing on my part. The stats share a cache line and the impact partially depends on whether unrelated stats share a cache line or not during updates.

Yes, but the question is, are there true cross-CPU cache-misses? I.e. are there any 'global' (or per node) counters that we keep touching and which keep generating cache-misses?

Quoted 8 lines
Quoted 3 lines

I think we should really fix those as well: those shared stats should be percpu collected as well, with no extra cache misses in any scheduler fast path.

I looked into that but converting those stats to per-cpu counters would incur sizable memory overhead. There are a *lot* of them and the basic structure for the generic percpu-counter is

Quoted 8 lines
struct percpu_counter {
        raw_spinlock_t lock;
        s64 count;
#ifdef CONFIG_HOTPLUG_CPU
        struct list_head list;  /* All percpu_counters are on a list */
#endif
        s32 __percpu *counters;
};

We don't have to reuse percpu_counter().

Quoted 1 line

That's not taking the associated runtime overhead such as synchronising them.

Why do we have to synchronize them in the kernel? User-space can recover them on a percpu basis and add them up if it wishes to. We can update the schedstat utility to handle the more spread out fields as well.

Thanks,

Ingo

author
#4 lore raw
Quoted 31 lines · Ingo Molnar
Quoted 28 lines · Mel Gorman
Quoted 26 lines · Ingo Molnar
Quoted 18 lines · Mel Gorman

Changelog since v3
o Force enable stats during profiling and latencytop

Changelog since V2
o Print stats that are not related to schedstat
o Reintroduce a static inline for update_stats_dequeue

Changelog since V1
o Introduce schedstat_enabled and address Ingo's feedback
o More schedstat-only paths eliminated, particularly ttwu_stat

schedstats is very useful during debugging and performance tuning but it incurs overhead. As such, even though it can be disabled at build time, it is often enabled as the information is useful. This patch adds a kernel command-line and sysctl tunable to enable or disable schedstats on demand. It is disabled by default as someone who knows they need it can also learn to enable it when necessary.

Quoted 3 lines

The benefits are workload-dependent but when it gets down to it, the difference will be whether cache misses are incurred updating the shared stats or not. [...]

Hm, which shared stats are those?

Quoted 8 lines
Quoted 3 lines

Extremely poor phrasing on my part. The stats share a cache line and the impact partially depends on whether unrelated stats share a cache line or not during updates.

Yes, but the question is, are there true cross-CPU cache-misses? I.e. are there any 'global' (or per node) counters that we keep touching and which keep generating cache-misses?

I haven't specifically identified them as I consider the calculations for some of them to be expensive in their own right even without accounting for cache misses. Moving to per-cpu counters would not eliminate all cache misses as a stat updated on one CPU for a task that is woken on a separate CPU is still going to trigger a cache miss. Even if such counters were identified and moved to separate cache lines, the calculation overhead would remain.

Quoted 17 lines
Quoted 8 lines
Quoted 3 lines

I think we should really fix those as well: those shared stats should be percpu collected as well, with no extra cache misses in any scheduler fast path.

I looked into that but converting those stats to per-cpu counters would incur sizable memory overhead. There are a *lot* of them and the basic structure for the generic percpu-counter is

Quoted 8 lines
struct percpu_counter {
        raw_spinlock_t lock;
        s64 count;
#ifdef CONFIG_HOTPLUG_CPU
        struct list_head list;  /* All percpu_counters are on a list */
#endif
        s32 __percpu *counters;
};
Quoted 2 lines

We don't have to reuse percpu_counter().

No, but rolling a specialised solution for a debugging feature is overkill and the calculation overhead would remain. It's specialised code with very little upside.

The main gain from the patch is that the calculation overhead is avoided. Avoid any potential cache miss is a bonus.

Quoted 3 lines
Quoted 1 line

That's not taking the associated runtime overhead such as synchronising them.

Why do we have to synchronize them in the kernel?

Because some simply require it or are not suitable for moving to per-cpu counters at all. sleep_start is an obvious one as it can wake on another CPU.

Quoted 4 lines

? User-space can recover them on a
percpu basis and add them up if it wishes to. We can update the schedstat utility to handle the more spread out fields as well.

Any user of /proc/pid/sched would also need updating, including latencytop and all of them will need to be able to handle CPU hotplug or else deal with the output from all possible CPUs instead of the currently online ones.

Signature · 2 lines
Mel Gorman
SUSE Labs
author
#5 lore raw
Quoted 5 lines · Mel Gorman
Quoted 4 lines

Yes, but the question is, are there true cross-CPU cache-misses? I.e. are there any 'global' (or per node) counters that we keep touching and which keep generating cache-misses?

Quoted 7 lines

I haven't specifically identified them as I consider the calculations for some of them to be expensive in their own right even without accounting for cache misses. Moving to per-cpu counters would not eliminate all cache misses as a stat updated on one CPU for a task that is woken on a separate CPU is still going to trigger a cache miss. Even if such counters were identified and moved to separate cache lines, the calculation overhead would remain.

I looked closer with perf stat to see if there was a good case for reducing cache misses using per-cpu counters.

Workload was hackbench with pipes and twice as many processes as there are CPUs to generate a reasonable amount of scheduler activity.

Kernel 4.5-rc2 vanilla
Performance counter stats for './hackbench -pipe 96 process 1000' (5 runs):

      54355.194747      task-clock (msec)         #   35.825 CPUs utilized            ( +-  0.72% )  (100.00%)
         6,654,707      context-switches          #    0.122 M/sec                    ( +-  1.56% )  (100.00%)
           376,624      cpu-migrations            #    0.007 M/sec                    ( +-  3.43% )  (100.00%)
           128,533      page-faults               #    0.002 M/sec                    ( +-  1.80% )  (100.00%)
   111,173,775,559      cycles                    #    2.045 GHz                      ( +-  0.76% )  (52.55%)
   <not supported>      stalled-cycles-frontend
   <not supported>      stalled-cycles-backend
    87,243,428,243      instructions              #    0.78  insns per cycle          ( +-  0.38% )  (63.74%)
    17,067,078,003      branches                  #  313.992 M/sec                    ( +-  0.39% )  (61.79%)
        65,864,607      branch-misses             #    0.39% of all branches          ( +-  2.10% )  (61.51%)
    26,873,984,605      L1-dcache-loads           #  494.414 M/sec                    ( +-  0.45% )  (33.08%)
     1,531,628,468      L1-dcache-load-misses     #    5.70% of all L1-dcache hits    ( +-  1.14% )  (31.65%)
       410,990,209      LLC-loads                 #    7.561 M/sec                    ( +-  1.08% )  (31.38%)
        38,279,473      LLC-load-misses           #    9.31% of all LL-cache hits     ( +-  6.82% )  (42.35%)

1.517251315 seconds time elapsed ( +- 1.55% )

Note that the actual cache miss ratio is quite low and indicates that there is potentially little to gain from using per-cpu counters.

Kernel 4.5-rc2 plus patch that disables schedstats by default

Performance counter stats for './hackbench -pipe 96 process 1000' (5 runs):

      51904.139186      task-clock (msec)         #   35.322 CPUs utilized            ( +-  2.07% )  (100.00%)
         5,958,009      context-switches          #    0.115 M/sec                    ( +-  5.90% )  (100.00%)
           327,235      cpu-migrations            #    0.006 M/sec                    ( +-  8.24% )  (100.00%)
           130,063      page-faults               #    0.003 M/sec                    ( +-  1.10% )  (100.00%)
   104,926,877,727      cycles                    #    2.022 GHz                      ( +-  2.12% )  (52.08%)
   <not supported>      stalled-cycles-frontend  
   <not supported>      stalled-cycles-backend   
    83,768,167,895      instructions              #    0.80  insns per cycle          ( +-  1.25% )  (63.49%)
    16,379,438,730      branches                  #  315.571 M/sec                    ( +-  1.47% )  (61.99%)
        59,841,332      branch-misses             #    0.37% of all branches          ( +-  4.60% )  (61.68%)
    25,749,569,276      L1-dcache-loads           #  496.099 M/sec                    ( +-  1.37% )  (34.08%)
     1,385,090,233      L1-dcache-load-misses     #    5.38% of all L1-dcache hits    ( +-  3.40% )  (31.88%)
       358,531,172      LLC-loads                 #    6.908 M/sec                    ( +-  4.65% )  (31.04%)
        33,476,691      LLC-load-misses           #    9.34% of all LL-cache hits     ( +-  4.95% )  (41.71%)

1.469447783 seconds time elapsed ( +- 2.23% )

Now, note that there is a reduction in cache misses but it's not a major percentage and the miss ratio is only dropped slightly in comparison to having stats enabled.

While a perf report shows there is a drop in cache references in functions like ttwu_stat and [en|de]queue_entity but it's a small percentage overall. The same is true for the cycle count. The overall percentage is small but the patch eliminates them.

Based on the low level of cache misses, I see no value to using per-cpu counters as an alternative.

Signature · 2 lines
Mel Gorman
SUSE Labs
Reviewed-by on v1 1/1
#6 lore raw
In reply to Mel Gorman
Quoted 25 lines

Changelog since v3
o Force enable stats during profiling and latencytop

Changelog since V2
o Print stats that are not related to schedstat
o Reintroduce a static inline for update_stats_dequeue

Changelog since V1
o Introduce schedstat_enabled and address Ingo's feedback
o More schedstat-only paths eliminated, particularly ttwu_stat

schedstats is very useful during debugging and performance tuning but it incurs overhead. As such, even though it can be disabled at build time, it is often enabled as the information is useful. This patch adds a kernel command-line and sysctl tunable to enable or disable schedstats on demand. It is disabled by default as someone who knows they need it can also learn to enable it when necessary.

The benefits are workload-dependent but when it gets down to it, the difference will be whether cache misses are incurred updating the shared stats or not. These measurements were taken from a 48-core 2-socket machine with Xeon(R) E5-2670 v3 cpus although they were also tested on a single socket machine 8-core machine with Intel i7-3770 processors.

Quoted 2 lines
Signature · 2 lines
Thanks and Regards
Srikar Dronamraju