linux-kernel.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
* [PATCH v2 0/5] tracing: Extend histogram triggers expression parsing
@ 2021-10-20  1:31 Kalesh Singh
  2021-10-20  1:31 ` [PATCH v2 1/5] tracing: Add support for creating hist trigger variables from literal Kalesh Singh
                   ` (4 more replies)
  0 siblings, 5 replies; 20+ messages in thread
From: Kalesh Singh @ 2021-10-20  1:31 UTC (permalink / raw)
  Cc: surenb, hridya, namhyung, kernel-team, Kalesh Singh,
	Jonathan Corbet, Steven Rostedt, Ingo Molnar, Shuah Khan,
	Masami Hiramatsu, Tom Zanussi, linux-doc, linux-kernel,
	linux-kselftest

Hi all,

The v1 of this series was posted at:
https://lore.kernel.org/r/20210915195306.612966-1-kaleshsingh@google.com/

The cover letter here is mostly identical to that in v1, with a
correction to how the results were obtained.
---

The frequency of the rss_stat trace event is known to be of the same
magnitude as that of the sched_switch event on Android devices. This can
cause flooding of the trace buffer with rss_stat traces leading to a
decreased trace buffer capacity and loss of data.

If it is not necessary to monitor very small changes in rss (as is the
case in Android) then the rss_stat tracepoint can be throttled to only
emit the event once there is a large enough change in the rss size.
The original patch that introduced the rss_stat tracepoint also proposed
a fixed throttling mechanism that only emits the rss_stat event
when the rss size crosses a 512KB boundary. It was concluded that more
generic support for this type of filtering/throttling was need, so that
it can be applied to any trace event. [1]

From the discussion in [1], histogram triggers seemed the most likely
candidate to support this type of throttling. For instance to achieve the
same throttling as was proposed in [1]:

  (1) Create a histogram variable to save the 512KB bucket of the rss size
  (2) Use the onchange handler to generate a synthetic event when the
      rss size bucket changes.

The only missing pieces to support such a hist trigger are:
  (1) Support for setting a hist variable to a specific value -- to set
      the bucket size / granularity.
  (2) Support for division arithmetic operation -- to determine the
      corresponding bucket for an rss size.

This series extends histogram trigger expressions to:
  (1) Allow assigning numeric literals to hist variable (eg. x=1234)
      and using literals directly in expressions (eg. x=size/1234)
  (2) Support division and multiplication in hist expressions.
      (eg. a=$x/$y*z); and
  (3) Fixes expression parsing for non-associative operators: subtraction
      and division. (eg. 8-4-2 should be 2 not 6)

The rss_stat event can then be throttled using histogram triggers as
below:

  # Create a synthetic event to monitor instead of the high frequency
  # rss_stat event
  echo 'rss_stat_throttled unsigned int mm_id; unsigned int curr;
         int member; long size' >> tracing/synthetic_events

  # Create a hist trigger that emits the synthetic rss_stat_throttled
  # event only when the rss size crosses a 512KB boundary.
  echo 'hist:keys=mm_id,member:bucket=size/0x80000:onchange($bucket)
              .rss_stat_throttled(mm_id,curr,member,size)'
        >> events/kmem/rss_stat/trigger

 ------ Test Results ------
Histograms can also be used to evaluate the effectiveness of this
throttling by noting the Total Hits on each trigger:

  echo 'hist:keys=common_pid' >> events/sched/sched_switch/trigger
  echo 'hist:keys=common_pid' >> events/kmem/rss_stat/trigger
  echo 'hist:keys=common_pid'
           >> events/synthetic/rss_stat_throttled/trigger

Allowing the above example (512KB granularity) run for 5 minutes on
an arm64 device with 5.10 kernel:

   sched_switch      : total hits = 147153
   rss_stat          : total hits =  38863
   rss_stat_throttled: total hits =   2409

The synthetic rss_stat_throttled event is ~16x less frequent than the
rss_stat event when using a 512KB granularity.


The results are more pronounced when rss size is changing at a higher
rate in small increments. For instance the following results were obtained
by recording the hits on the above events for a run of Android's
lmkd_unit_test [2], which continually forks processes that map anonymous
memory until there is an oom kill:

   sched_switch      : total hits =  148832
   rss_stat          : total hits = 4754802
   rss_stat_throttled: total hits =   96214

In this stress test, the synthetic rss_stat_throttled event is ~50x less
frequent than the rss_stat event when using a 512KB granularity.

[1] https://lore.kernel.org/lkml/20190903200905.198642-1-joel@joelfernandes.org/
[2] https://cs.android.com/android/platform/superproject/+/master:system/memory/lmkd/tests/lmkd_test.cpp


Kalesh Singh (5):
  tracing: Add support for creating hist trigger variables from literal
  tracing: Add division and multiplication support for hist triggers
  tracing: Fix operator precedence for hist triggers expression
  tracing/selftests: Add tests for hist trigger expression parsing
  tracing/histogram: Document expression arithmetic and constants

 Documentation/trace/histogram.rst             |  14 +
 kernel/trace/trace_events_hist.c              | 318 +++++++++++++++---
 .../testing/selftests/ftrace/test.d/functions |   4 +-
 .../trigger/trigger-hist-expressions.tc       |  74 ++++
 4 files changed, 358 insertions(+), 52 deletions(-)
 create mode 100644 tools/testing/selftests/ftrace/test.d/trigger/trigger-hist-expressions.tc


base-commit: d9abdee5fd5abffd0e763e52fbfa3116de167822
-- 
2.33.0.1079.g6e70778dc9-goog


^ permalink raw reply	[flat|nested] 20+ messages in thread

end of thread, other threads:[~2021-10-20 16:22 UTC | newest]

Thread overview: 20+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2021-10-20  1:31 [PATCH v2 0/5] tracing: Extend histogram triggers expression parsing Kalesh Singh
2021-10-20  1:31 ` [PATCH v2 1/5] tracing: Add support for creating hist trigger variables from literal Kalesh Singh
2021-10-20 15:32   ` Steven Rostedt
2021-10-20 15:55     ` Kalesh Singh
2021-10-20 15:59       ` Steven Rostedt
2021-10-20  1:31 ` [PATCH v2 2/5] tracing: Add division and multiplication support for hist triggers Kalesh Singh
2021-10-20  2:27   ` Steven Rostedt
2021-10-20 14:54     ` Kalesh Singh
2021-10-20 15:13       ` Steven Rostedt
2021-10-20 15:23         ` Kalesh Singh
2021-10-20  1:31 ` [PATCH v2 3/5] tracing: Fix operator precedence for hist triggers expression Kalesh Singh
2021-10-20 14:28   ` Steven Rostedt
2021-10-20 15:06     ` Kalesh Singh
2021-10-20 15:13       ` Steven Rostedt
2021-10-20 15:48   ` Steven Rostedt
2021-10-20 16:11     ` Kalesh Singh
2021-10-20 16:19       ` Steven Rostedt
2021-10-20 16:22     ` Kalesh Singh
2021-10-20  1:31 ` [PATCH v2 4/5] tracing/selftests: Add tests for hist trigger expression parsing Kalesh Singh
2021-10-20  1:31 ` [PATCH v2 5/5] tracing/histogram: Document expression arithmetic and constants Kalesh Singh

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).