lkml.org 
[lkml]   [2012]   [Jun]   [4]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Subject[RFC] [PATCH 0/5] Teach perf tool to profile sleep times (V4)
Date
Do you want to know where your code waits locks for a long time?
Yes! It's for you. This feature helps you to find bottlenecks.

It's not artificial task. Once one of my colleague was investigating a
scalability problem. He pressed sysrq-t some times and tried to merge
call-chains by hand. But perf can do that better.

Here is fourth attempt to add this functionality in the kernel.

Previous attempts:
* [PATCH 0/4] trace: add ability to collect call chains of non-current task.

* [PATCH rh6 0/7] Teach perf tool to collect D-state statistics
This version doesn't work for non-root users.

* "sched/tracing: Add a new tracepoint for sleeptime" by Arun Sharma
It has been committed, but then reverted due to performance reason.

This series resolves a problem with events filtering and non-root users.
Before I tried to use sched_stat* events, but it occurs in a context of
another task. For this case logic of perf_event_context* doesn't work.

This series adds one more event sched:sched_switch_finish, it's sent
when a task starts executing on a cpu. When a task goes out from cpu,
sched:sched_switch is sent.

The time difference between sched_switch and sched_switch_finish is a time
when a task wasn't being executed on a cpu.

For profiling sleep time, events of both types should be collected and
filtered by a defined state (interruptible of uninterruptible sleep).

Then events should be post-processed with help "perf inject". It saves the time
of sched_switch event for each process and when it meets stat_switch_finish,
it sets time difference in sample->period and saves this new sample
in the new perf.data.

./perf record -e sched:sched_switch -g --filter 'prev_state == 1' \
-e sched:sched_switch_finish --filter 'state == 1' -P ~/foo

./perf inject -i ~/perf.data -o perf.data.new -s

A test program:
#include<unistd.h>
#include<time.h>
#include<sys/select.h>

int main()
{
struct timespec ts1;
struct timeval tv1;
int i;
long s;


for (i = 0; i< 2; i++) {
ts1.tv_sec = 0;
ts1.tv_nsec = 10000000;
nanosleep(&ts1, NULL);

tv1.tv_sec = 0;
tv1.tv_usec = 40000;
select(0, NULL, NULL, NULL,&tv1);
if (i % 100 == 0) {
s = time(NULL);
while (time(NULL) - s < 4);
}
}
return 1;
}

Profiling results:
# Overhead Command Shared Object Symbol
# ........ ....... ................. ......................
#
100.00% foo [kernel.kallsyms] [k] finish_task_switch
|
--- finish_task_switch
__schedule
schedule
|
|--79.85%-- schedule_hrtimeout_range_clock
| schedule_hrtimeout_range
| poll_schedule_timeout
| do_select
| core_sys_select
| sys_select
| system_call
| __select
| __libc_start_main
|
--20.15%-- do_nanosleep
hrtimer_nanosleep
sys_nanosleep
system_call
__GI___libc_nanosleep
__libc_start_main

Andrew Vagin (5):
event: add an event sched_switch_finish
sched: send the event sched_switch before perf_event_task_sched_out
sched: save a previous state on task_struct
perf: teach "perf inject" to work with files
perf: teach perf inject to merge sched_switch* events

include/linux/sched.h | 3 +
include/trace/events/sched.h | 29 ++++++++++
kernel/sched/core.c | 3 +-
tools/perf/builtin-inject.c | 119 +++++++++++++++++++++++++++++++++++++++---
4 files changed, 146 insertions(+), 8 deletions(-)



\
 
 \ /
  Last update: 2012-06-04 10:41    [W:0.333 / U:0.008 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site