gstreamer/subprojects/gst-devtools/tracer/gsttr-tsplot.py

288 lines
10 KiB
Python

#!/usr/bin/env python3
'''
Plot buffer pts and events in relation to the wall clock. The plots can be used
to spot anomalies, such as processing gaps.
How to run:
1) generate a log
GST_DEBUG="GST_TRACER:7" GST_TRACERS=stats GST_DEBUG_FILE=trace.log <application>
2) generate the images
python3 gsttr-tsplot.py trace.log <outdir>
eog <outdir>/*.png
'''
# TODO:
# - improve event plot
# - ideally each event is a vertical line
# http://stackoverflow.com/questions/35105672/vertical-lines-from-data-in-file-in-time-series-plot-using-gnuplot
# - this won't work well if the event is e.g. 'qos'
# - we could sort them by event type and separate them by double new-lines,
# we'd then use 'index <x>' to plot them in different colors with
# - buffer-pts should be ahead of clock time of the pipeline
# - we don't have the clock ts in the log though
import logging
import os
from subprocess import Popen, PIPE, DEVNULL
from string import Template
from tracer.analysis_runner import AnalysisRunner
from tracer.analyzer import Analyzer
from tracer.parser import Parser
from tracer.structure import Structure
logging.basicConfig(level=logging.WARNING)
logger = logging.getLogger('gsttr-tsplot')
_HANDLED_CLASSES = ('buffer', 'event', 'new-pad', 'new-element')
_GST_BUFFER_FLAG_DISCONT = (1 << 6)
_PLOT_SCRIPT_HEAD = Template(
'''
set term pngcairo truecolor size $width,$height font "Helvetica,14"
set style line 1 lc rgb '#8b1a0e' pt 1 ps 1 lt 1 lw 1 # --- red
set style line 2 lc rgb '#5e9c36' pt 6 ps 1 lt 1 lw 1 # --- green
set style line 100 lc rgb '#999999' lt 0 lw 1
set grid back ls 100
set key font ",10"
set label font ",10"
set tics font ",10"
set xlabel font ",10"
set ylabel font ",10"
''')
_PLOT_SCRIPT_BODY = Template(
'''
set output '$png_file_name'
set multiplot layout 3,1 title "$title\\n$subtitle"
set xlabel ""
set xrange [*:*] writeback
set xtics format ""
set ylabel "Buffer Time (sec.msec)" offset 1,0
set yrange [*:*]
set ytics
plot '$buf_file_name' using 1:2 with linespoints ls 1 notitle
set xrange restore
set ylabel "Duration (sec.msec)" offset 1,0
plot '$buf_file_name' using 1:3 with linespoints ls 1title "cycle", \
'' using 1:4 with linespoints ls 2 title "duration"
set xrange restore
set xtics format "%g" scale .5 offset 0,.5
set xlabel "Clock Time (sec.msec)" offset 0,1
set ylabel "Events" offset 1,0
set yrange [$ypos_max:10]
set ytics format ""
plot '$ev_file_name' using 1:4:3:(0) with vectors heads size screen 0.008,90 ls 1 notitle, \
'' using 2:4 with points ls 1 notitle, \
'' using 2:4:5 with labels font ',7' offset char 0,-0.5 notitle
unset multiplot
''')
class TsPlot(Analyzer):
'''Generate a timestamp plots from a tracer log.
These show the buffer pts on the y-axis and the wall-clock time the buffer
was produced on the x-axis. This helps to spot timing issues, such as
stalled elements.
'''
def __init__(self, outdir, show_ghost_pads, size):
super(TsPlot, self).__init__()
self.outdir = outdir
self.show_ghost_pads = show_ghost_pads
self.params = {
'width': size[0],
'height': size[1],
}
self.buf_files = {}
self.buf_cts = {}
self.ev_files = {}
self.element_names = {}
self.element_info = {}
self.pad_names = {}
self.pad_info = {}
self.ev_labels = {}
self.ev_data = {}
self.ev_ypos = {}
def _get_data_file(self, files, key, name_template):
data_file = files.get(key)
if not data_file:
pad_name = self.pad_names.get(key)
if pad_name:
file_name = name_template % (self.outdir, key, pad_name)
data_file = open(file_name, 'w')
files[key] = data_file
return data_file
def _log_event_data(self, pad_file, ix):
data = self.ev_data.get(ix)
if not data:
return
line = self.ev_labels[ix]
ct = data['ct']
x1 = data['first-ts']
# TODO: scale 'y' according to max-y of buf or do a multiplot
y = (1 + data['ypos']) * -10
if ct == 1:
pad_file.write('%f %f %f %f "%s"\n' % (x1, x1, 0.0, y, line))
else:
x2 = data['last-ts']
xd = (x2 - x1)
xm = x1 + xd / 2
pad_file.write('%f %f %f %f "%s (%d)"\n' % (x1, xm, xd, y, line, ct))
def _log_event(self, s):
# build a [ts, event-name] data file
ix = int(s.values['pad-ix'])
pad_file = self._get_data_file(self.ev_files, ix, '%s/ev_%d_%s.dat')
if not pad_file:
return
# convert timestamps to seconds
x = int(s.values['ts']) / 1e9
# some events fire often, labeling each would be unreadable
# so we aggregate a series of events of the same type
line = s.values['name']
if line == self.ev_labels.get(ix):
# count lines and track last ts
data = self.ev_data[ix]
data['ct'] += 1
data['last-ts'] = x
else:
self._log_event_data(pad_file, ix)
# start new data, assign a -y coord by event type
if ix not in self.ev_ypos:
ypos = {}
self.ev_ypos[ix] = ypos
else:
ypos = self.ev_ypos[ix]
if line in ypos:
y = ypos[line]
else:
y = len(ypos)
ypos[line] = y
self.ev_labels[ix] = line
self.ev_data[ix] = {
'ct': 1,
'first-ts': x,
'ypos': y,
}
def _log_buffer(self, s):
if not int(s.values['have-buffer-pts']):
return
# build a [ts, buffer-pts] data file
ix = int(s.values['pad-ix'])
pad_file = self._get_data_file(self.buf_files, ix, '%s/buf_%d_%s.dat')
if not pad_file:
return
flags = int(s.values['buffer-flags'])
if flags & _GST_BUFFER_FLAG_DISCONT:
pad_file.write('\n')
# convert timestamps to e.g. seconds
cts = int(s.values['ts']) / 1e9
pts = int(s.values['buffer-pts']) / 1e9
dur = int(s.values['buffer-duration']) / 1e9
if ix not in self.buf_cts:
dcts = 0
else:
dcts = cts - self.buf_cts[ix]
self.buf_cts[ix] = cts
pad_file.write('%f %f %f %f\n' % (cts, pts, dcts, dur))
def handle_tracer_entry(self, event):
if event[Parser.F_FUNCTION]:
return
msg = event[Parser.F_MESSAGE]
p = msg.find(',')
if p == -1:
return
entry_name = msg[:p]
if entry_name not in _HANDLED_CLASSES:
return
try:
s = Structure(msg)
except ValueError:
logger.warning("failed to parse: '%s'", msg)
return
if entry_name == 'new-element':
ix = int(s.values['ix'])
self.element_names[ix] = s.values['name']
self.element_info[ix] = 'Element Type: %s' % s.values['type']
elif entry_name == 'new-pad':
pad_type = s.values['type']
if self.show_ghost_pads or pad_type not in ['GstGhostPad', 'GstProxyPad']:
parent_ix = int(s.values['parent-ix'])
parent_name = self.element_names.get(parent_ix, '')
ix = int(s.values['ix'])
self.pad_names[ix] = '%s.%s' % (parent_name, s.values['name'])
self.pad_info[ix] = '(%s, Pad Type: %s)' % (
self.element_info.get(parent_ix, ''), pad_type)
elif entry_name == 'event':
self._log_event(s)
else: # 'buffer'
self._log_buffer(s)
def report(self):
for ix, pad_file in self.ev_files.items():
self._log_event_data(pad_file, ix)
pad_file.close()
script = _PLOT_SCRIPT_HEAD.substitute(self.params)
for ix, pad_file in self.buf_files.items():
pad_file.close()
name = self.pad_names[ix]
buf_file_name = '%s/buf_%d_%s.dat' % (self.outdir, ix, name)
ev_file_name = '%s/ev_%d_%s.dat' % (self.outdir, ix, name)
png_file_name = '%s/%d_%s.png' % (self.outdir, ix, name)
sub_title = self.pad_info[ix]
ypos_max = (2 + len(self.ev_ypos[ix])) * -10
script += _PLOT_SCRIPT_BODY.substitute(self.params, title=name,
subtitle=sub_title, buf_file_name=buf_file_name,
ev_file_name=ev_file_name, png_file_name=png_file_name,
ypos_max=ypos_max)
# plot PNGs
p = Popen(['gnuplot'], stdout=DEVNULL, stdin=PIPE)
p.communicate(input=script.encode('utf-8'))
# cleanup
for ix, pad_file in self.buf_files.items():
name = self.pad_names[ix]
buf_file_name = '%s/buf_%d_%s.dat' % (self.outdir, ix, name)
os.unlink(buf_file_name)
for ix, pad_file in self.ev_files.items():
name = self.pad_names[ix]
ev_file_name = '%s/ev_%d_%s.dat' % (self.outdir, ix, name)
os.unlink(ev_file_name)
if __name__ == '__main__':
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('file', nargs='?', default='debug.log')
parser.add_argument('outdir', nargs='?', default='tsplot')
parser.add_argument('-g', '--ghost-pads', action='store_true',
help='also plot data for ghost-pads')
parser.add_argument('-s', '--size', action='store', default='1600x600',
help='graph size as WxH')
args = parser.parse_args()
os.makedirs(args.outdir, exist_ok=True)
size = [int(s) for s in args.size.split('x')]
with Parser(args.file) as log:
tsplot = TsPlot(args.outdir, args.ghost_pads, size)
runner = AnalysisRunner(log)
runner.add_analyzer(tsplot)
runner.run()
tsplot.report()