Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1562131 > unrolled thread

[PATCH 2/2] perf,core: use parent avg sample period as child initial period

Back to article view | Back to linux.kernel

This discussion starts older than the indexed window; earlier articles aren't shown. The article labeled Started by below is the oldest one visible, not the original post.

Contents

  [PATCH 2/2] perf,core: use parent avg sample period as child initial period kan.liang@intel.com - 2017-01-18 21:20 +0100
    Re: [PATCH 2/2] perf,core: use parent avg sample period as child  initial period Peter Zijlstra <peterz@infradead.org> - 2017-01-25 16:40 +0100
      RE: [PATCH 2/2] perf,core: use parent avg sample period as child  initial period "Liang, Kan" <kan.liang@intel.com> - 2017-01-29 18:10 +0100

#1562131 — [PATCH 2/2] perf,core: use parent avg sample period as child initial period

From: Kan Liang <kan.liang@intel.com>

perf brings additional overhead when monitoring the task which
frequently generates child task.

When inheriting a event from parent task to child task, the
sample_period of original parent event (parent_event->parent) will be
assigned to child event as its initial period, which is usually the
default sample_period 1. But too many very short period like 1 will
increase overhead and may cause various problems.

avg_time_stamp is introduced to keep the average sample period. Each
child event can use its original parent event's avg period as its initial
sample period, which can reduce the overhead.

The avg_time_stamp doesn't update more than once every tick to avoid the
contention.
For each new child event, the parent event refcount++. Parent will not
go away until all children go away. So it's safe to access its parent.

Here is some data from the overhead test on Broadwell server
  perf record -e $TEST_EVENTS -- ./loop.sh 50000

loop.sh
  start=$(date +%s%N)
  i=0
  while [ "$i" -le "$1" ]
  do
          date > /dev/null
          i=`expr $i + 1`
  done
  end=$(date +%s%N)
  elapsed=`expr $end - $start`

Event#	Original Elapsed time	Elapsed time with patch		delta
1	196,573,192,397		188,480,366,278			-4.12%
2	257,567,753,013		242,256,126,043			-5.94%
3	398,730,726,971		373,882,492,502			-6.23%
4	824,983,761,120		750,906,525,917			-8.98%
5	1,883,411,923,498	1,648,192,098,897		-12.49%

Signed-off-by: Kan Liang <kan.liang@intel.com>
---
 include/linux/perf_event.h |  3 +++
 kernel/events/core.c       | 20 ++++++++++++++++++--
 2 files changed, 21 insertions(+), 2 deletions(-)

diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
index 78ed810..84b0f47 100644
--- a/include/linux/perf_event.h
+++ b/include/linux/perf_event.h
@@ -648,6 +648,9 @@ struct perf_event {
 	struct list_head		child_list;
 	struct perf_event		*parent;
 
+	atomic64_t			avg_sample_period;
+	u64				avg_time_stamp;
+
 	int				oncpu;
 	int				cpu;
 
diff --git a/kernel/events/core.c b/kernel/events/core.c
index 924268c..82a2c0e 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -3237,9 +3237,11 @@ static DEFINE_PER_CPU(u64, perf_throttled_seq);
 
 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
 {
+	struct perf_event *head_event = (event->parent != NULL) ? event->parent : event;
 	struct hw_perf_event *hwc = &event->hw;
 	s64 period, sample_period;
 	s64 delta;
+	u64 now;
 
 	period = perf_calculate_period(event, nsec, count);
 
@@ -3253,6 +3255,15 @@ static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bo
 
 	hwc->sample_period = sample_period;
 
+	now = perf_clock();
+	if ((now - head_event->avg_time_stamp) > TICK_NSEC) {
+		s64 avg_period;
+
+		head_event->avg_time_stamp = now;
+		avg_period = (atomic64_read(&head_event->avg_sample_period) + sample_period) / 2;
+		atomic64_set(&head_event->avg_sample_period, avg_period);
+	}
+
 	if (local64_read(&hwc->period_left) > 8*sample_period) {
 		if (disable)
 			event->pmu->stop(event, PERF_EF_UPDATE);
@@ -9231,8 +9242,13 @@ perf_event_alloc(struct perf_event_attr *attr, int cpu,
 
 	hwc = &event->hw;
 	hwc->sample_period = attr->sample_period;
-	if (attr->freq && attr->sample_freq)
+	if (attr->freq && attr->sample_freq) {
 		hwc->sample_period = 1;
+		if (parent_event)
+			hwc->sample_period = atomic64_read(&parent_event->avg_sample_period);
+		else
+			atomic64_set(&event->avg_sample_period, hwc->sample_period);
+	}
 	hwc->last_period = hwc->sample_period;
 
 	local64_set(&hwc->period_left, hwc->sample_period);
@@ -10464,8 +10480,8 @@ inherit_event(struct perf_event *parent_event,
 		child_event->state = PERF_EVENT_STATE_OFF;
 
 	if (parent_event->attr.freq) {
-		u64 sample_period = parent_event->hw.sample_period;
 		struct hw_perf_event *hwc = &child_event->hw;
+		u64 sample_period = atomic64_read(&parent_event->avg_sample_period);
 
 		hwc->sample_period = sample_period;
 		hwc->last_period   = sample_period;
-- 
2.4.3

[toc] | [next] | [standalone]

#1566688 — Re: [PATCH 2/2] perf,core: use parent avg sample period as child initial period

On Wed, Jan 18, 2017 at 08:21:02AM -0500, kan.liang@intel.com wrote:
> From: Kan Liang <kan.liang@intel.com>
> 
> perf brings additional overhead when monitoring the task which
> frequently generates child task.
> 
> When inheriting a event from parent task to child task, the
> sample_period of original parent event (parent_event->parent) will be
> assigned to child event as its initial period, which is usually the
> default sample_period 1.


Why is that mostly 1? I would expect the parent event's sample_period to
ramp up as well.

[toc] | [prev] | [next] | [standalone]

#1569233 — RE: [PATCH 2/2] perf,core: use parent avg sample period as child initial period

> 
> On Wed, Jan 18, 2017 at 08:21:02AM -0500, kan.liang@intel.com wrote:
> > From: Kan Liang <kan.liang@intel.com>
> >
> > perf brings additional overhead when monitoring the task which
> > frequently generates child task.
> >
> > When inheriting a event from parent task to child task, the
> > sample_period of original parent event (parent_event->parent) will be
> > assigned to child event as its initial period, which is usually the
> > default sample_period 1.
> 
> 
> Why is that mostly 1? I would expect the parent event's sample_period to
> ramp up as well.

The conclusion is based on the observation in the specific test case
mentioned in the description. The sample_period is for the original parent
event (parent_event->parent), not the direct parent.

I did several tests these days, it can be 100% reproduced by this particular
test case in multiplexing.
I think the reason is that the original parent event is scheduled out shortly
and never gets a chance to trigger interrupt in the first round. So its
sample_period doesn't get updated. Then the test case repeatedly
generates child tasks and child events. The new child events will be
scheduled in/out continually. The original parent event never gets a
change to be scheduled again. So the original parent event's
sample_period is kept1.

Thanks,
Kan

[toc] | [prev] | [standalone]

Back to top | Article view | linux.kernel

csiph-web