1 /*
2 * Copyright (c) 2016 The WebRTC project authors. All Rights Reserved.
3 *
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
9 */
10
11 #include "modules/congestion_controller/goog_cc/trendline_estimator.h"
12
13 #include <math.h>
14
15 #include <algorithm>
16 #include <string>
17
18 #include "absl/strings/match.h"
19 #include "absl/types/optional.h"
20 #include "modules/remote_bitrate_estimator/include/bwe_defines.h"
21 #include "modules/remote_bitrate_estimator/test/bwe_test_logging.h"
22 #include "rtc_base/checks.h"
23 #include "rtc_base/experiments/struct_parameters_parser.h"
24 #include "rtc_base/logging.h"
25 #include "rtc_base/numerics/safe_minmax.h"
26
27 namespace webrtc {
28
29 namespace {
30
31 // Parameters for linear least squares fit of regression line to noisy data.
32 constexpr double kDefaultTrendlineSmoothingCoeff = 0.9;
33 constexpr double kDefaultTrendlineThresholdGain = 4.0;
34 const char kBweWindowSizeInPacketsExperiment[] =
35 "WebRTC-BweWindowSizeInPackets";
36
ReadTrendlineFilterWindowSize(const WebRtcKeyValueConfig * key_value_config)37 size_t ReadTrendlineFilterWindowSize(
38 const WebRtcKeyValueConfig* key_value_config) {
39 std::string experiment_string =
40 key_value_config->Lookup(kBweWindowSizeInPacketsExperiment);
41 size_t window_size;
42 int parsed_values =
43 sscanf(experiment_string.c_str(), "Enabled-%zu", &window_size);
44 if (parsed_values == 1) {
45 if (window_size > 1)
46 return window_size;
47 RTC_LOG(WARNING) << "Window size must be greater than 1.";
48 }
49 RTC_LOG(LS_WARNING) << "Failed to parse parameters for BweWindowSizeInPackets"
50 " experiment from field trial string. Using default.";
51 return TrendlineEstimatorSettings::kDefaultTrendlineWindowSize;
52 }
53
LinearFitSlope(const std::deque<TrendlineEstimator::PacketTiming> & packets)54 absl::optional<double> LinearFitSlope(
55 const std::deque<TrendlineEstimator::PacketTiming>& packets) {
56 RTC_DCHECK(packets.size() >= 2);
57 // Compute the "center of mass".
58 double sum_x = 0;
59 double sum_y = 0;
60 for (const auto& packet : packets) {
61 sum_x += packet.arrival_time_ms;
62 sum_y += packet.smoothed_delay_ms;
63 }
64 double x_avg = sum_x / packets.size();
65 double y_avg = sum_y / packets.size();
66 // Compute the slope k = \sum (x_i-x_avg)(y_i-y_avg) / \sum (x_i-x_avg)^2
67 double numerator = 0;
68 double denominator = 0;
69 for (const auto& packet : packets) {
70 double x = packet.arrival_time_ms;
71 double y = packet.smoothed_delay_ms;
72 numerator += (x - x_avg) * (y - y_avg);
73 denominator += (x - x_avg) * (x - x_avg);
74 }
75 if (denominator == 0)
76 return absl::nullopt;
77 return numerator / denominator;
78 }
79
ComputeSlopeCap(const std::deque<TrendlineEstimator::PacketTiming> & packets,const TrendlineEstimatorSettings & settings)80 absl::optional<double> ComputeSlopeCap(
81 const std::deque<TrendlineEstimator::PacketTiming>& packets,
82 const TrendlineEstimatorSettings& settings) {
83 RTC_DCHECK(1 <= settings.beginning_packets &&
84 settings.beginning_packets < packets.size());
85 RTC_DCHECK(1 <= settings.end_packets &&
86 settings.end_packets < packets.size());
87 RTC_DCHECK(settings.beginning_packets + settings.end_packets <=
88 packets.size());
89 TrendlineEstimator::PacketTiming early = packets[0];
90 for (size_t i = 1; i < settings.beginning_packets; ++i) {
91 if (packets[i].raw_delay_ms < early.raw_delay_ms)
92 early = packets[i];
93 }
94 size_t late_start = packets.size() - settings.end_packets;
95 TrendlineEstimator::PacketTiming late = packets[late_start];
96 for (size_t i = late_start + 1; i < packets.size(); ++i) {
97 if (packets[i].raw_delay_ms < late.raw_delay_ms)
98 late = packets[i];
99 }
100 if (late.arrival_time_ms - early.arrival_time_ms < 1) {
101 return absl::nullopt;
102 }
103 return (late.raw_delay_ms - early.raw_delay_ms) /
104 (late.arrival_time_ms - early.arrival_time_ms) +
105 settings.cap_uncertainty;
106 }
107
108 constexpr double kMaxAdaptOffsetMs = 15.0;
109 constexpr double kOverUsingTimeThreshold = 10;
110 constexpr int kMinNumDeltas = 60;
111 constexpr int kDeltaCounterMax = 1000;
112
113 } // namespace
114
115 constexpr char TrendlineEstimatorSettings::kKey[];
116
TrendlineEstimatorSettings(const WebRtcKeyValueConfig * key_value_config)117 TrendlineEstimatorSettings::TrendlineEstimatorSettings(
118 const WebRtcKeyValueConfig* key_value_config) {
119 if (absl::StartsWith(
120 key_value_config->Lookup(kBweWindowSizeInPacketsExperiment),
121 "Enabled")) {
122 window_size = ReadTrendlineFilterWindowSize(key_value_config);
123 }
124 Parser()->Parse(key_value_config->Lookup(TrendlineEstimatorSettings::kKey));
125 if (window_size < 10 || 200 < window_size) {
126 RTC_LOG(LS_WARNING) << "Window size must be between 10 and 200 packets";
127 window_size = kDefaultTrendlineWindowSize;
128 }
129 if (enable_cap) {
130 if (beginning_packets < 1 || end_packets < 1 ||
131 beginning_packets > window_size || end_packets > window_size) {
132 RTC_LOG(LS_WARNING) << "Size of beginning and end must be between 1 and "
133 << window_size;
134 enable_cap = false;
135 beginning_packets = end_packets = 0;
136 cap_uncertainty = 0.0;
137 }
138 if (beginning_packets + end_packets > window_size) {
139 RTC_LOG(LS_WARNING)
140 << "Size of beginning plus end can't exceed the window size";
141 enable_cap = false;
142 beginning_packets = end_packets = 0;
143 cap_uncertainty = 0.0;
144 }
145 if (cap_uncertainty < 0.0 || 0.025 < cap_uncertainty) {
146 RTC_LOG(LS_WARNING) << "Cap uncertainty must be between 0 and 0.025";
147 cap_uncertainty = 0.0;
148 }
149 }
150 }
151
Parser()152 std::unique_ptr<StructParametersParser> TrendlineEstimatorSettings::Parser() {
153 return StructParametersParser::Create("sort", &enable_sort, //
154 "cap", &enable_cap, //
155 "beginning_packets",
156 &beginning_packets, //
157 "end_packets", &end_packets, //
158 "cap_uncertainty", &cap_uncertainty, //
159 "window_size", &window_size);
160 }
161
TrendlineEstimator(const WebRtcKeyValueConfig * key_value_config,NetworkStatePredictor * network_state_predictor)162 TrendlineEstimator::TrendlineEstimator(
163 const WebRtcKeyValueConfig* key_value_config,
164 NetworkStatePredictor* network_state_predictor)
165 : settings_(key_value_config),
166 smoothing_coef_(kDefaultTrendlineSmoothingCoeff),
167 threshold_gain_(kDefaultTrendlineThresholdGain),
168 num_of_deltas_(0),
169 first_arrival_time_ms_(-1),
170 accumulated_delay_(0),
171 smoothed_delay_(0),
172 delay_hist_(),
173 k_up_(0.0087),
174 k_down_(0.039),
175 overusing_time_threshold_(kOverUsingTimeThreshold),
176 threshold_(12.5),
177 prev_modified_trend_(NAN),
178 last_update_ms_(-1),
179 prev_trend_(0.0),
180 time_over_using_(-1),
181 overuse_counter_(0),
182 hypothesis_(BandwidthUsage::kBwNormal),
183 hypothesis_predicted_(BandwidthUsage::kBwNormal),
184 network_state_predictor_(network_state_predictor) {
185 RTC_LOG(LS_INFO)
186 << "Using Trendline filter for delay change estimation with settings "
187 << settings_.Parser()->Encode() << " and "
188 << (network_state_predictor_ ? "injected" : "no")
189 << " network state predictor";
190 }
191
~TrendlineEstimator()192 TrendlineEstimator::~TrendlineEstimator() {}
193
UpdateTrendline(double recv_delta_ms,double send_delta_ms,int64_t send_time_ms,int64_t arrival_time_ms,size_t packet_size)194 void TrendlineEstimator::UpdateTrendline(double recv_delta_ms,
195 double send_delta_ms,
196 int64_t send_time_ms,
197 int64_t arrival_time_ms,
198 size_t packet_size) {
199 const double delta_ms = recv_delta_ms - send_delta_ms;
200 ++num_of_deltas_;
201 num_of_deltas_ = std::min(num_of_deltas_, kDeltaCounterMax);
202 if (first_arrival_time_ms_ == -1)
203 first_arrival_time_ms_ = arrival_time_ms;
204
205 // Exponential backoff filter.
206 accumulated_delay_ += delta_ms;
207 BWE_TEST_LOGGING_PLOT(1, "accumulated_delay_ms", arrival_time_ms,
208 accumulated_delay_);
209 smoothed_delay_ = smoothing_coef_ * smoothed_delay_ +
210 (1 - smoothing_coef_) * accumulated_delay_;
211 BWE_TEST_LOGGING_PLOT(1, "smoothed_delay_ms", arrival_time_ms,
212 smoothed_delay_);
213
214 // Maintain packet window
215 delay_hist_.emplace_back(
216 static_cast<double>(arrival_time_ms - first_arrival_time_ms_),
217 smoothed_delay_, accumulated_delay_);
218 if (settings_.enable_sort) {
219 for (size_t i = delay_hist_.size() - 1;
220 i > 0 &&
221 delay_hist_[i].arrival_time_ms < delay_hist_[i - 1].arrival_time_ms;
222 --i) {
223 std::swap(delay_hist_[i], delay_hist_[i - 1]);
224 }
225 }
226 if (delay_hist_.size() > settings_.window_size)
227 delay_hist_.pop_front();
228
229 // Simple linear regression.
230 double trend = prev_trend_;
231 if (delay_hist_.size() == settings_.window_size) {
232 // Update trend_ if it is possible to fit a line to the data. The delay
233 // trend can be seen as an estimate of (send_rate - capacity)/capacity.
234 // 0 < trend < 1 -> the delay increases, queues are filling up
235 // trend == 0 -> the delay does not change
236 // trend < 0 -> the delay decreases, queues are being emptied
237 trend = LinearFitSlope(delay_hist_).value_or(trend);
238 if (settings_.enable_cap) {
239 absl::optional<double> cap = ComputeSlopeCap(delay_hist_, settings_);
240 // We only use the cap to filter out overuse detections, not
241 // to detect additional underuses.
242 if (trend >= 0 && cap.has_value() && trend > cap.value()) {
243 trend = cap.value();
244 }
245 }
246 }
247 BWE_TEST_LOGGING_PLOT(1, "trendline_slope", arrival_time_ms, trend);
248
249 Detect(trend, send_delta_ms, arrival_time_ms);
250 }
251
Update(double recv_delta_ms,double send_delta_ms,int64_t send_time_ms,int64_t arrival_time_ms,size_t packet_size,bool calculated_deltas)252 void TrendlineEstimator::Update(double recv_delta_ms,
253 double send_delta_ms,
254 int64_t send_time_ms,
255 int64_t arrival_time_ms,
256 size_t packet_size,
257 bool calculated_deltas) {
258 if (calculated_deltas) {
259 UpdateTrendline(recv_delta_ms, send_delta_ms, send_time_ms, arrival_time_ms,
260 packet_size);
261 }
262 if (network_state_predictor_) {
263 hypothesis_predicted_ = network_state_predictor_->Update(
264 send_time_ms, arrival_time_ms, hypothesis_);
265 }
266 }
267
State() const268 BandwidthUsage TrendlineEstimator::State() const {
269 return network_state_predictor_ ? hypothesis_predicted_ : hypothesis_;
270 }
271
Detect(double trend,double ts_delta,int64_t now_ms)272 void TrendlineEstimator::Detect(double trend, double ts_delta, int64_t now_ms) {
273 if (num_of_deltas_ < 2) {
274 hypothesis_ = BandwidthUsage::kBwNormal;
275 return;
276 }
277 const double modified_trend =
278 std::min(num_of_deltas_, kMinNumDeltas) * trend * threshold_gain_;
279 prev_modified_trend_ = modified_trend;
280 BWE_TEST_LOGGING_PLOT(1, "T", now_ms, modified_trend);
281 BWE_TEST_LOGGING_PLOT(1, "threshold", now_ms, threshold_);
282 if (modified_trend > threshold_) {
283 if (time_over_using_ == -1) {
284 // Initialize the timer. Assume that we've been
285 // over-using half of the time since the previous
286 // sample.
287 time_over_using_ = ts_delta / 2;
288 } else {
289 // Increment timer
290 time_over_using_ += ts_delta;
291 }
292 overuse_counter_++;
293 if (time_over_using_ > overusing_time_threshold_ && overuse_counter_ > 1) {
294 if (trend >= prev_trend_) {
295 time_over_using_ = 0;
296 overuse_counter_ = 0;
297 hypothesis_ = BandwidthUsage::kBwOverusing;
298 }
299 }
300 } else if (modified_trend < -threshold_) {
301 time_over_using_ = -1;
302 overuse_counter_ = 0;
303 hypothesis_ = BandwidthUsage::kBwUnderusing;
304 } else {
305 time_over_using_ = -1;
306 overuse_counter_ = 0;
307 hypothesis_ = BandwidthUsage::kBwNormal;
308 }
309 prev_trend_ = trend;
310 UpdateThreshold(modified_trend, now_ms);
311 }
312
UpdateThreshold(double modified_trend,int64_t now_ms)313 void TrendlineEstimator::UpdateThreshold(double modified_trend,
314 int64_t now_ms) {
315 if (last_update_ms_ == -1)
316 last_update_ms_ = now_ms;
317
318 if (fabs(modified_trend) > threshold_ + kMaxAdaptOffsetMs) {
319 // Avoid adapting the threshold to big latency spikes, caused e.g.,
320 // by a sudden capacity drop.
321 last_update_ms_ = now_ms;
322 return;
323 }
324
325 const double k = fabs(modified_trend) < threshold_ ? k_down_ : k_up_;
326 const int64_t kMaxTimeDeltaMs = 100;
327 int64_t time_delta_ms = std::min(now_ms - last_update_ms_, kMaxTimeDeltaMs);
328 threshold_ += k * (fabs(modified_trend) - threshold_) * time_delta_ms;
329 threshold_ = rtc::SafeClamp(threshold_, 6.f, 600.f);
330 last_update_ms_ = now_ms;
331 }
332
333 } // namespace webrtc
334