How to narrow down perf.data to a time sub interva

I use linux perf (perf_events) to produce a perf.data file with timestamps.

How can I generate a report of all the events in a sub interval of time [i-start, i-end]?

Can I maybe narrow down perf.data to a perf_subinterv.data file with only events in [i-start, i-end]?

I need to do this to analyze short intervals (2s - 6s) of poor performance every 5mins or so.

回答1:

I still cannot produce a perf_subinterv.data, but I can narrow down the perf trace in textual representation. Then, for further analysis, I can for instance generate a flame graph.

Command to narrow to time interval at 19:29:43 with 3.47s duration:

perf script -i ./perf_2016-03-23_192924.468036489.data | awk -v perfdata=./perf_2016-03-23_192924.468036489.data -v interval_start=19:29:43 -v duration=3.47 -f perf_script_cut_interval.awk > perf_2016-03-23_192924.468036489_INTERV_19:29:43.txt

Generation of flame graph:

stackcollapse-perf.pl perf_2016-03-23_192924.468036489_INTERV_19:29:43.txt | flamegraph.pl > perf_2016-03-23_192924.468036489_INTERV_19:29:43.svg

gawk script:

#
# Consumes output of 'perf script profile.data' and filters events from a given
# time interval
#
# input variables:
#
# perfdata:
#
#      File with profiling data. Name must be perf_<date>_<time>.data, where
#      <time> has the format <hh><mm><secs>, e.g. perf_2016-03-23_140426.002147215.data
#     
# interval_start:
#
#      Start time of the interval with format <hh><mm><secs>, e.g. 19:29:43.890735
#
# duration:
#
#      length of the interval
#

BEGIN {
  print("processing", perfdata) > "/dev/stderr"
  # parse timestamp of perf rec start
  match(perfdata, /.*perf_.*_(..)(..)(.+)\.data/, ts_perf_rec)
  # parse interval start
  match(interval_start, /(..):(..):(.+)/ , ts_interval)
  hh=1
  mm=2
  ss=3
  printf("ts_perf_rec            = %02d:%02d:%05f\n", ts_perf_rec[hh], ts_perf_rec[mm], ts_perf_rec[ss]) > "/dev/stderr"
  printf("ts_interval            = %02d:%02d:%05f\n", ts_interval[hh], ts_interval[mm], ts_interval[ss]) > "/dev/stderr"
  # current line belongs to header
  in_header = 1
  # current line belongs to selected interval
  in_interval = 0

  # first timestamp in perf.data
  first_ts = -1
  FS="[ :]"
}

# find end of header
/^[^#]/ {
  if (in_header) {
    in_header = 0
  }
}

# find timestamps
# example line: java 15950 515784.682786: cycles: 
/^.+ [0-9]+ [0-9]+\.[0-9]+:/ {
  cur_ts = $3 + 0.0
  if (first_ts == -1) {
    # translate ts_interval to profile data timestamps by identifying the first
    # timestamp with ts_perf_rec
    first_ts = cur_ts

    # delta_recstart_intervalstart is the time difference from the first
    # profiling event to the filter interval
    delta_recstart_intervalstart[ss] = ts_interval[ss] - ts_perf_rec[ss]
    if (delta_recstart_intervalstart[ss] < 0) {
      delta_recstart_intervalstart[ss] += 60
      delta_recstart_intervalstart[mm] = -1
    }
    delta_recstart_intervalstart[mm] += ts_interval[mm] - ts_perf_rec[mm]
    if (delta_recstart_intervalstart[mm] < 0) {
      delta_recstart_intervalstart[mm] += 60
      delta_recstart_intervalstart[hh] = -1
    }
    delta_recstart_intervalstart[hh] += ts_interval[hh] - ts_perf_rec[hh]

    # beginning and end of the interval in profiling timestamps
    interval_begin_s = delta_recstart_intervalstart[hh] * 3600 + delta_recstart_intervalstart[mm] * 60 + delta_recstart_intervalstart[ss] + first_ts
    interval_end_s = interval_begin_s + duration

    printf("ts_perf_rec                  = %02d:%02d:%05f\n", ts_perf_rec[hh], ts_perf_rec[mm], ts_perf_rec[ss]) > "/dev/stderr"
    printf("first_ts                     = %f\n", first_ts) > "/dev/stderr"
    printf("ts_interval                  = %02d:%02d:%05f\n", ts_interval[hh], ts_interval[mm], ts_interval[ss]) > "/dev/stderr"
    printf("delta_recstart_intervalstart = %02d:%02d:%05f\n",
           delta_recstart_intervalstart[hh], delta_recstart_intervalstart[mm], delta_recstart_intervalstart[ss]) > "/dev/stderr"
    printf("duration                     = %f\n", duration) > "/dev/stderr"
    printf("interval_begin_s             = %05f\n", interval_begin_s) > "/dev/stderr"
    printf("interval_end_s               = %05f\n", interval_end_s) > "/dev/stderr"
  }

  in_interval = ((cur_ts >= interval_begin_s) && (cur_ts < interval_end_s))
}

# print every line that belongs to the header or the selected time interval
in_interval || in_header {
  print $0
}

回答2:

Most perf tools, including perf report, support filtering by time:

--time::
  Only analyze samples within given time window: <start>,<stop>. Times
  have the format seconds.microseconds. If start is not given (i.e., time
  string is ',x.y') then analysis starts at the beginning of the file. If
  stop time is not given (i.e, time string is 'x.y,') then analysis goes
  to end of file.

For more details see man perf-report.

This is present since version 4.10 (Feb 2017). If you run an older kernel you can try to build the userspace tools part of perf yourself. In more recent versions it is possible to specify time percent and multiple time ranges.