New rtc dump analyzing tool in Python

tools/py_event_log_analyzer/rtp_analyzer.py

+#  Copyright (c) 2016 The WebRTC project authors. All Rights Reserved.
+#
+#  Use of this source code is governed by a BSD-style license
+#  that can be found in the LICENSE file in the root of the source
+#  tree. An additional intellectual property rights grant can be found
+#  in the file PATENTS.  All contributing project authors may
+#  be found in the AUTHORS file in the root of the source tree.
+
+"""Displays statistics and plots graphs from RTC protobuf dump."""
+
+from __future__ import division
+from __future__ import print_function
+
+import collections
+import sys
+import builtins
+import matplotlib.pyplot as plt
+import misc
+import numpy
+import pb_parse
+
+
+
+class RTPStatistics(object):
+  """Has methods for calculating and plotting statistics for RPT

[hlundin-webrtc, 2016/05/25 12:54:23]

RPT -> RTP

+  packets.
+  """
+
+  BANDWIDTH_SMOOTHING_WINDOW_SIZE = 10
+
+  def __init__(self, data_points):
+    """Initializes object with data_points computes simple statistics:

[kwiberg-webrtc, 2016/05/25 12:42:42]

and

[aleloi2, 2016/05/30 14:57:56]

Done.

+    percentages of packets and packet sizes by SSRC.
+
+    Args:
+        data_points: list of pb_parse.DataPoints on which statistics are
+            calculated.
+
+    """
+
+    self.data_points = data_points
+    self.ssrc_frequencies = misc.percent_table([x.ssrc for x in
+                                                self.data_points])
+    self.ssrc_size_table = misc.ssrc_size_table(self.data_points)
+    self.bandwidth_kbps = None
+    self.smooth_bw_kbps = None
+
+  def print_ssrc_info(self, ssrc_id, ssrc):
+    """Prints packet and size statistics for given SSRC.

[peah-webrtc, 2016/05/26 06:44:18]

Maybe "the" or "a" should be added between "for" and "given"?

[aleloi2, 2016/05/30 14:57:56]

Done.

+
+    Args:
+        ssrc_id: textual identifier of SSRC printed beside statistics for it.
+        ssrc: SSRC by which to filter data and display statistics
+
+    Raises:
+         Exception: when different payload types are present in data

[hlundin-webrtc, 2016/05/25 12:54:23]

This will happen, and can be legit. When comfort noise is used, it can either be
codec-internal (such as Opus has) which will use the same payload type as the
regular speech packets, or be codec-external which uses separate payload types
for the speech and noise packets, both under the same SSRC. You will have to
handle this case.

[aleloi2, 2016/05/30 14:57:55]

Now it prints the different payload types instead of raising an exception.

+           for same SSRC
+    """
+    filtered_ssrc = [x for x in self.data_points if x.ssrc == ssrc]
+    payloads = misc.percent_table([x.payload_type for x in filtered_ssrc])
+
+    if len(payloads) == 1:
+      payload_info = "payload type {}".format(*list(payloads))
+    else:
+      raise Exception(
+          "This tool cannot yet handle changes in codec sample rate")
+    print("{} 0X{:X} {}, {:.2f}% packets, {:.2f}% data".format(

[kwiberg-webrtc, 2016/05/25 12:42:42]

Consider using lowercase hexadecimal. It causes less eye bleeding. :-)

[aleloi2, 2016/05/30 14:57:56]

Ok.

+        ssrc_id, ssrc, payload_info, self.ssrc_frequencies[ssrc]*100,
+        self.ssrc_size_table[ssrc]*100))

[kwiberg-webrtc, 2016/05/25 12:42:41]

You still have some binary operators not surrounded by spaces.

[aleloi2, 2016/05/30 14:57:56]

Done.

+    print("  packet sizes:")
+    bin_counts, bin_bounds = numpy.histogram([x.size for x in
+                                              filtered_ssrc], bins=5,
+                                             density=False)
+    bin_proportions = bin_counts / sum(bin_counts)
+    print("\n".join([
+        " {:.1f} - {:.1f}: {:.2f}%".format(bin_bounds[i], bin_bounds[i+1],
+                                           bin_proportions[i]*100)
+        for i in builtins.range(len(bin_proportions))
+    ]))
+
+  def choose_ssrc(self):
+    """Queries user for SSRC."""
+    ssrc_frequencies_lst = list(enumerate(self.ssrc_frequencies))

[kwiberg-webrtc, 2016/05/25 12:42:42]

Eliminate this variable. You don't use it for anything where
self.ssrc_frequencies wouldn't work just as well or better.

[aleloi2, 2016/05/30 14:57:56]

It was used to pair ssrc with an integer counter, which came from the iteration
order of dictionary keys. I checked the documentation to make sure that the
iteration order stays the same for list(d) and enumerate(d).

+
+    if len(self.ssrc_frequencies) == 1:
+      chosen_ssrc = self.ssrc_frequencies[0][-1]
+      self.print_ssrc_info("", chosen_ssrc)
+      return chosen_ssrc
+
+    for i, ssrc in enumerate(self.ssrc_frequencies):
+      self.print_ssrc_info(i, ssrc)
+
+    while True:
+      chosen_index = int(builtins.input("choose one> "))
+      if 0 <= chosen_index < len(ssrc_frequencies_lst):
+        return ssrc_frequencies_lst[chosen_index][-1]
+      else:
+        print("Invalid index!")
+
+  def filter_ssrc(self, chosen_ssrc):
+    """Filters and wraps data points.
+
+    Removes data points with `ssrc != chosen_ssrc`. Unwraps sequence
+    numbers and time stamps for the chosen selection.

[hlundin-webrtc, 2016/05/25 12:54:23]

Nit: 'timestamps' is typically written as one word ('sequence numbers' is not).

[aleloi2, 2016/05/30 14:57:56]

Done.

+    """
+    self.data_points = [x for x in self.data_points if x.ssrc ==
+                        chosen_ssrc]
+    unwrapped_sequence_numbers = misc.unwrap([x.sequence_number for x in
+                                              self.data_points],
+                                             2**16-1)
+    for (data_point, sequence_number) in zip(self.data_points,
+                                             unwrapped_sequence_numbers):
+      data_point.sequence_number = sequence_number
+
+    unwrapped_timestamps = misc.unwrap([x.timestamp for x in self.data_points],
+                                       2**32-1)
+
+    for (data_point, timestamp) in zip(self.data_points,
+                                       unwrapped_timestamps):
+      data_point.timestamp = timestamp
+
+  def print_sequence_number_statistics(self):
+    seq_no_set = set(x.sequence_number for x in self.data_points)
+    print("Missing sequence numbers: {} out of {}".format(
+        max(seq_no_set) - min(seq_no_set) + 1 - len(seq_no_set),
+        len(seq_no_set)
+    ))
+    packet_counter = collections.Counter(x.sequence_number for x in
+                                         self.data_points)
+    print("Duplicated packets: {}".format(sum(packet_counter.values())
+                                          - len(packet_counter)
+                                         ))

[kwiberg-webrtc, 2016/05/25 12:42:41]

Unless I'm mistaken, it's simpler and more efficient to compute this number as
len(self.data_points) - len(seq_no_set).

[aleloi2, 2016/05/30 14:57:56]

I agree. Changed.

+    print("Reordered packets: {}".format(
+        misc.count_reordered([x.sequence_number for x in  self.data_points])))
+
+  def print_frequency_duration_statistics(self):

[peah-webrtc, 2016/05/26 06:44:19]

Suggestion:
This is a fairly long method which does many things, I think it may benefit in
terms of readability and encapsulation if it is split into separate methods
calls. 

For instance, the computation of the estimated frequency could be split into a
non-class member method declarable as.

def estimate_frequency(f,data_points):
 ...
 return fs
 

[aleloi2, 2016/05/30 14:57:56]

I agree. I divided it into a frequency related and a duration/bitrate related
part. I didn't do exactly as you said: both new parts are member functions.
There is no strong reason for it, but I think it looks more uniform this way.

+    """Estimates frequency and prints related statistics.
+
+    Guesses the most probable frequency by looking at changes in
+    timestamps (RFC 3550 section 5.1), calculates clock drifts and
+    sending time of packets. Updates `self.data_points` with changes
+    in delay and send time.
+
+    """
+    delta_timestamp = (self.data_points[-1].timestamp -
+                       self.data_points[0].timestamp)
+    delta_arr_timestamp = float((self.data_points[-1].arrival_timestamp_ms -
+                                 self.data_points[0].arrival_timestamp_ms))
+    fs_est = delta_timestamp / delta_arr_timestamp
+
+    fs_vec = [8, 16, 32, 48, 90]  # TODO(aleloi) 90 is a hack for video

[hlundin-webrtc, 2016/05/25 12:54:23]

Not really a "hack"; https://tools.ietf.org/html/rfc3551#section-5.

[peah-webrtc, 2016/05/26 06:44:18]

The python style guide actually has an example where a comment is not terminated
by a period, but I think it would be nice with a terminating period in this
comment.

+    fs = None
+    for f in fs_vec:
+      if abs((fs_est - f)/float(f)) < 0.05:

[kwiberg-webrtc, 2016/05/25 12:42:42]

No need for float, since you import division from __future__.

[aleloi2, 2016/05/30 14:57:56]

OK.

+        fs = f
+
+    print("Estimated frequency: {}".format(fs_est))
+    if fs is None:
+      fs = int(builtins.input(
+          "Frequency could not be guessed. Input frequency> "))
+
+    print("Guessed frequency: {}".format(fs))

[kwiberg-webrtc, 2016/05/25 12:42:41]

Don't print this if the user entered a frequency.

[aleloi2, 2016/05/30 14:57:56]

Done.

+
+    for f in self.data_points:
+      f.real_send_time_ms = (f.timestamp -
+                             self.data_points[0].timestamp) / fs
+      f.delay = f.arrival_timestamp_ms - f.real_send_time_ms
+
+    min_delay = min(f.delay for f in self.data_points)
+
+    for f in self.data_points:

[hlundin-webrtc, 2016/05/25 12:54:23]

Oh. Python cannot do "vector +/- scalar" without explicitly looping?

[kwiberg-webrtc, 2016/05/25 13:00:31]

No. list + list returns a new list that's a concatenation of the two argument
lists, and list + number is simply disallowed.

[ivoc, 2016/05/25 13:17:37]

It could do it if it was a numpy array instead of a list, but in this case it's
actually a list of objects, so I don't think that would work in matlab either.

[kwiberg-webrtc, 2016/05/25 13:29:04]

Right, I'd forgotten about numpy, because I never use it---but if you're coming
from Matlab, that's probably the first thing you should look at once you're past
the initial Python tutorial.

+      f.absdelay = f.delay - min_delay
+
+    stream_duration_sender = self.data_points[-1].real_send_time_ms / 1000
+    print("Stream duration at sender: {:.1f} seconds".format(
+        stream_duration_sender
+    ))
+
+    arrival_timestamps_ms = [pt.arrival_timestamp_ms for pt in
+                             self.data_points]
+    stream_duration_receiver = (max(arrival_timestamps_ms) -
+                                min(arrival_timestamps_ms)) / 1000
+    print("Stream duration at receiver: {:.1f} seconds".format(
+        stream_duration_receiver
+    ))
+
+    print("Clock drift: {:.2f}%".format(
+        100* (stream_duration_receiver / stream_duration_sender - 1)
+    ))
+
+    print("Send average bitrate: {:.2f} kbps".format(
+        sum(x.size for x
+            in self.data_points) * 8 / stream_duration_sender / 1000))
+
+    print("Receive average bitrate: {:.2f} kbps".format(
+        sum(x.size
+            for x in self.data_points) * 8 / stream_duration_receiver /
+        1000))
+
+  def remove_reordered(self):
+    last = self.data_points[0]
+    data_points_ordered = [last]
+    for x in self.data_points[1:]:
+      if x.sequence_number > last.sequence_number and (x.real_send_time_ms >
+                                     last.real_send_time_ms):
+        data_points_ordered.append(x)
+        last = x
+    self.data_points = data_points_ordered
+
+  def compute_bandwidth(self):
+    """Computes bandwidth averaged over several consecutive packets.
+
+    The number of consecutive packets used in the average is
+    BANDWIDTH_SMOOTHING_WINDOW_SIZE. Averaging is done with
+    numpy.correlate.
+    """
+    self.bandwidth_kbps = []
+    for i in range(len(self.data_points)-1):
+      self.bandwidth_kbps.append(
+          self.data_points[i].size*8 / (self.data_points[i+1].real_send_time_ms
+                                        - self.data_points[i].real_send_time_ms)
+      )
+    correlate_filter = (numpy.ones(
+        RTPStatistics.BANDWIDTH_SMOOTHING_WINDOW_SIZE) /
+                        RTPStatistics.BANDWIDTH_SMOOTHING_WINDOW_SIZE)
+    self.smooth_bw_kbps = numpy.correlate(self.bandwidth_kbps, correlate_filter)
+
+  def plot_statistics(self):
+    """Plots changes in delay and average bandwidth."""
+    plt.figure(1)
+    plt.plot([f.real_send_time_ms/1000 for f in self.data_points],
+             [f.absdelay for f in self.data_points])
+    plt.xlabel("Send time [s]")
+    plt.ylabel("Relative transport delay [ms]")
+
+    plt.figure(2)
+    plt.plot([f.real_send_time_ms / 1000 for f in
+              self.data_points][:len(self.smooth_bw_kbps)],
+             self.smooth_bw_kbps[:len(self.data_points)])
+    plt.xlabel("Send time [s]")
+    plt.ylabel("Bandwidth [kbps]")
+
+    plt.show()
+
+
+def main():
+
+  if len(sys.argv) < 2:
+    print("Usage: python rtp_analyzer.py <filename of rtc event log>")
+    sys.exit(0)
+
+  data_points = pb_parse.parse_protobuf(sys.argv[1])
+  rtp_stats = RTPStatistics(data_points)
+  chosen_ssrc = rtp_stats.choose_ssrc()
+  print("Chosen SSRC: 0X{:X}".format(chosen_ssrc))
+
+  rtp_stats.filter_ssrc(chosen_ssrc)
+  print("Statistics:")
+  rtp_stats.print_sequence_number_statistics()
+  rtp_stats.print_frequency_duration_statistics()
+  rtp_stats.remove_reordered()
+  rtp_stats.compute_bandwidth()
+  rtp_stats.plot_statistics()
+
+if __name__ == "__main__":
+  main()