Upgrade trendline estimator to improve low bandwidth conditions (#1055)

* Uprade trendline estimator to improve low bandwidth conditions

* Missing constant change

worker/deps/libwebrtc/libwebrtc/modules/congestion_controller/goog_cc/delay_based_bwe.cc

@@ -61,7 +61,7 @@ DelayBasedBwe::DelayBasedBwe(const WebRtcKeyValueConfig* key_value_config,
       network_state_predictor_(network_state_predictor),
       inter_arrival_(),
       delay_detector_(
-          new TrendlineEstimator(key_value_config_, network_state_predictor_)),
+          new TrendlineEstimator(network_state_predictor_)),
       last_seen_packet_(Timestamp::MinusInfinity()),
       uma_recorded_(false),
       rate_control_(key_value_config, /*send_side=*/true),
@@ -127,7 +127,7 @@ void DelayBasedBwe::IncomingPacketFeedback(const PacketResult& packet_feedback,
         new InterArrival((kTimestampGroupLengthMs << kInterArrivalShift) / 1000,
                          kTimestampToMs, true));
     delay_detector_.reset(
-        new TrendlineEstimator(key_value_config_, network_state_predictor_));
+        new TrendlineEstimator(network_state_predictor_));
   }
   last_seen_packet_ = at_time;
 

worker/deps/libwebrtc/libwebrtc/modules/congestion_controller/goog_cc/trendline_estimator.cc

@@ -13,96 +13,92 @@
 
 #include "modules/congestion_controller/goog_cc/trendline_estimator.h"
 
-#include "modules/remote_bitrate_estimator/include/bwe_defines.h"
-#include "rtc_base/numerics/safe_minmax.h"
-
-#include "Logger.hpp"
-
-#include <absl/types/optional.h>
 #include <math.h>
+
 #include <algorithm>
 #include <string>
 
+#include "absl/strings/match.h"
+#include "absl/types/optional.h"
+#include "api/network_state_predictor.h"
+#include "rtc_base/numerics/safe_minmax.h"
+#include "api/transport/webrtc_key_value_config.h"
+
+#include "Logger.hpp"
+
 namespace webrtc {
 
 namespace {
 
 // Parameters for linear least squares fit of regression line to noisy data.
-constexpr size_t kDefaultTrendlineWindowSize = 20;
-constexpr double kDefaultTrendlineSmoothingCoeff = 0.6;
+constexpr double kDefaultTrendlineSmoothingCoeff = 0.9;
 constexpr double kDefaultTrendlineThresholdGain = 4.0;
-const char kBweWindowSizeInPacketsExperiment[] =
-    "WebRTC-BweWindowSizeInPackets";
-
-size_t ReadTrendlineFilterWindowSize(
-    const WebRtcKeyValueConfig* key_value_config) {
-  std::string experiment_string =
-      key_value_config->Lookup(kBweWindowSizeInPacketsExperiment);
-  size_t window_size;
-  int parsed_values =
-      sscanf(experiment_string.c_str(), "Enabled-%zu", &window_size);
-  if (parsed_values == 1) {
-    if (window_size > 1)
-      return window_size;
-    MS_WARN_DEV("window size must be greater than 1");
-  }
-  MS_WARN_DEV(
-    "failed to parse parameters for BweWindowSizeInPackets"
-    " experiment from field trial string, using default");
-  return kDefaultTrendlineWindowSize;
-}
 
 absl::optional<double> LinearFitSlope(
-    const std::deque<std::pair<double, double>>& points) {
-  //RTC_DCHECK(points.size() >= 2);
+    const std::deque<TrendlineEstimator::PacketTiming>& packets) {
+  // RTC_DCHECK(packets.size() >= 2);
   // Compute the "center of mass".
   double sum_x = 0;
   double sum_y = 0;
-  for (const auto& point : points) {
-    sum_x += point.first;
-    sum_y += point.second;
+  for (const auto& packet : packets) {
+    sum_x += packet.arrival_time_ms;
+    sum_y += packet.smoothed_delay_ms;
   }
-  double x_avg = sum_x / points.size();
-  double y_avg = sum_y / points.size();
+  double x_avg = sum_x / packets.size();
+  double y_avg = sum_y / packets.size();
   // Compute the slope k = \sum (x_i-x_avg)(y_i-y_avg) / \sum (x_i-x_avg)^2
   double numerator = 0;
   double denominator = 0;
-  for (const auto& point : points) {
-    numerator += (point.first - x_avg) * (point.second - y_avg);
-    denominator += (point.first - x_avg) * (point.first - x_avg);
+  for (const auto& packet : packets) {
+    double x = packet.arrival_time_ms;
+    double y = packet.smoothed_delay_ms;
+    numerator += (x - x_avg) * (y - y_avg);
+    denominator += (x - x_avg) * (x - x_avg);
   }
   if (denominator == 0)
     return absl::nullopt;
   return numerator / denominator;
 }
 
+absl::optional<double> ComputeSlopeCap(
+    const std::deque<TrendlineEstimator::PacketTiming>& packets,
+    const TrendlineEstimatorSettings& settings) {
+  // RTC_DCHECK(1 <= settings.beginning_packets &&
+  //            settings.beginning_packets < packets.size());
+  // RTC_DCHECK(1 <= settings.end_packets &&
+  //            settings.end_packets < packets.size());
+  // RTC_DCHECK(settings.beginning_packets + settings.end_packets <=
+  //            packets.size());
+  TrendlineEstimator::PacketTiming early = packets[0];
+  for (size_t i = 1; i < settings.beginning_packets; ++i) {
+    if (packets[i].raw_delay_ms < early.raw_delay_ms)
+      early = packets[i];
+  }
+  size_t late_start = packets.size() - settings.end_packets;
+  TrendlineEstimator::PacketTiming late = packets[late_start];
+  for (size_t i = late_start + 1; i < packets.size(); ++i) {
+    if (packets[i].raw_delay_ms < late.raw_delay_ms)
+      late = packets[i];
+  }
+  if (late.arrival_time_ms - early.arrival_time_ms < 1) {
+    return absl::nullopt;
+  }
+  return (late.raw_delay_ms - early.raw_delay_ms) /
+             (late.arrival_time_ms - early.arrival_time_ms) +
+         settings.cap_uncertainty;
+}
+
 constexpr double kMaxAdaptOffsetMs = 15.0;
-constexpr double kOverUsingTimeThreshold = 30;
+constexpr double kOverUsingTimeThreshold = 10;
 constexpr int kMinNumDeltas = 60;
 constexpr int kDeltaCounterMax = 1000;
 
 }  // namespace
 
 TrendlineEstimator::TrendlineEstimator(
-    const WebRtcKeyValueConfig* key_value_config,
     NetworkStatePredictor* network_state_predictor)
-    : TrendlineEstimator(
-          key_value_config->Lookup(kBweWindowSizeInPacketsExperiment)
-                      .find("Enabled") == 0
-              ? ReadTrendlineFilterWindowSize(key_value_config)
-              : kDefaultTrendlineWindowSize,
-          kDefaultTrendlineSmoothingCoeff,
-          kDefaultTrendlineThresholdGain,
-          network_state_predictor) {}
-
-TrendlineEstimator::TrendlineEstimator(
-    size_t window_size,
-    double smoothing_coef,
-    double threshold_gain,
-    NetworkStatePredictor* network_state_predictor)
-    : window_size_(window_size),
-      smoothing_coef_(smoothing_coef),
-      threshold_gain_(threshold_gain),
+    : smoothing_coef_(kDefaultTrendlineSmoothingCoeff),
+      threshold_gain_(kDefaultTrendlineThresholdGain),
       num_of_deltas_(0),
       first_arrival_time_ms_(-1),
       accumulated_delay_(0),
@@ -120,63 +116,75 @@ TrendlineEstimator::TrendlineEstimator(
       hypothesis_(BandwidthUsage::kBwNormal),
       hypothesis_predicted_(BandwidthUsage::kBwNormal),
       network_state_predictor_(network_state_predictor) {
-  MS_DEBUG_DEV(
-    "using Trendline filter for delay change estimation with window size: %zu",
-    window_size_);
 }
 
 TrendlineEstimator::~TrendlineEstimator() {}
 
+void TrendlineEstimator::UpdateTrendline(double recv_delta_ms,
+                                         double send_delta_ms,
+                                         int64_t send_time_ms,
+                                         int64_t arrival_time_ms) {
+  const double delta_ms = recv_delta_ms - send_delta_ms;
+  ++num_of_deltas_;
+  num_of_deltas_ = std::min(num_of_deltas_, kDeltaCounterMax);
+  if (first_arrival_time_ms_ == -1)
+    first_arrival_time_ms_ = arrival_time_ms;
+
+  // Exponential backoff filter.
+  accumulated_delay_ += delta_ms;
+  // BWE_TEST_LOGGING_PLOT(1, "accumulated_delay_ms", arrival_time_ms,
+  //                       accumulated_delay_);
+  smoothed_delay_ = smoothing_coef_ * smoothed_delay_ +
+                    (1 - smoothing_coef_) * accumulated_delay_;
+  // BWE_TEST_LOGGING_PLOT(1, "smoothed_delay_ms", arrival_time_ms,
+  //                       smoothed_delay_);
+
+  // Maintain packet window
+  delay_hist_.emplace_back(
+      static_cast<double>(arrival_time_ms - first_arrival_time_ms_),
+      smoothed_delay_, accumulated_delay_);
+  if (settings_.enable_sort) {
+    for (size_t i = delay_hist_.size() - 1;
+         i > 0 &&
+         delay_hist_[i].arrival_time_ms < delay_hist_[i - 1].arrival_time_ms;
+         --i) {
+      std::swap(delay_hist_[i], delay_hist_[i - 1]);
+    }
+  }
+  if (delay_hist_.size() > settings_.window_size)
+    delay_hist_.pop_front();
+
+  // Simple linear regression.
+  double trend = prev_trend_;
+  if (delay_hist_.size() == settings_.window_size) {
+    // Update trend_ if it is possible to fit a line to the data. The delay
+    // trend can be seen as an estimate of (send_rate - capacity)/capacity.
+    // 0 < trend < 1   ->  the delay increases, queues are filling up
+    //   trend == 0    ->  the delay does not change
+    //   trend < 0     ->  the delay decreases, queues are being emptied
+    trend = LinearFitSlope(delay_hist_).value_or(trend);
+    if (settings_.enable_cap) {
+      absl::optional<double> cap = ComputeSlopeCap(delay_hist_, settings_);
+      // We only use the cap to filter out overuse detections, not
+      // to detect additional underuses.
+      if (trend >= 0 && cap.has_value() && trend > cap.value()) {
+        trend = cap.value();
+      }
+    }
+  }
+  // BWE_TEST_LOGGING_PLOT(1, "trendline_slope", arrival_time_ms, trend);
+
+  Detect(trend, send_delta_ms, arrival_time_ms);
+}
+
 void TrendlineEstimator::Update(double recv_delta_ms,
                                 double send_delta_ms,
                                 int64_t send_time_ms,
                                 int64_t arrival_time_ms,
                                 bool calculated_deltas) {
   if (calculated_deltas) {
-    const double delta_ms = recv_delta_ms - send_delta_ms;
-    ++num_of_deltas_;
-    num_of_deltas_ = std::min(num_of_deltas_, kDeltaCounterMax);
-    if (first_arrival_time_ms_ == -1)
-      first_arrival_time_ms_ = arrival_time_ms;
-
-    // Exponential backoff filter.
-    accumulated_delay_ += delta_ms;
-    // BWE_TEST_LOGGING_PLOT(1, "accumulated_delay_ms", arrival_time_ms,
-                          // accumulated_delay_);
-    // smoothed_delay_ = smoothing_coef_ * smoothed_delay_ +
-                      // (1 - smoothing_coef_) * accumulated_delay_;
-    // MS_NOTE: Apply WEMA to the current delta_ms. Don't consider the
-    // accumulated delay. Tests show it behaves more robustly upon delta peaks.
-    smoothed_delay_ = smoothing_coef_ * delta_ms +
-                      (1 - smoothing_coef_) * smoothed_delay_;
-    // BWE_TEST_LOGGING_PLOT(1, "smoothed_delay_ms", arrival_time_ms,
-                          // smoothed_delay_);
-
-    // Simple linear regression.
-    delay_hist_.push_back(std::make_pair(
-        static_cast<double>(arrival_time_ms - first_arrival_time_ms_),
-        smoothed_delay_));
-    if (delay_hist_.size() > window_size_)
-      delay_hist_.pop_front();
-    double trend = prev_trend_;
-    if (delay_hist_.size() == window_size_) {
-      // Update trend_ if it is possible to fit a line to the data. The delay
-      // trend can be seen as an estimate of (send_rate - capacity)/capacity.
-      // 0 < trend < 1   ->  the delay increases, queues are filling up
-      //   trend == 0    ->  the delay does not change
-      //   trend < 0     ->  the delay decreases, queues are being emptied
-      trend = LinearFitSlope(delay_hist_).value_or(trend);
-    }
-
-    // BWE_TEST_LOGGING_PLOT(1, "trendline_slope", arrival_time_ms, trend);
-
-    MS_DEBUG_DEV("trend:%f, send_delta_ms:%f, recv_delta_ms:%f, delta_ms:%f arrival_time_ms:%" PRIi64 ", accumulated_delay_:%f, smoothed_delay_:%f", trend, send_delta_ms, recv_delta_ms, delta_ms, arrival_time_ms, accumulated_delay_, smoothed_delay_);
-    Detect(trend, send_delta_ms, arrival_time_ms);
-  }
-  else {
-    MS_DEBUG_DEV("no calculated deltas");
+    UpdateTrendline(recv_delta_ms, send_delta_ms, send_time_ms, arrival_time_ms);
   }
-
   if (network_state_predictor_) {
     hypothesis_predicted_ = network_state_predictor_->Update(
         send_time_ms, arrival_time_ms, hypothesis_);
@@ -212,15 +220,14 @@ void TrendlineEstimator::Detect(double trend, double ts_delta, int64_t now_ms) {
       if (trend >= prev_trend_) {
         time_over_using_ = 0;
         overuse_counter_ = 0;
+        hypothesis_ = BandwidthUsage::kBwOverusing;
         MS_DEBUG_DEV("hypothesis_: BandwidthUsage::kBwOverusing");
 
 #if MS_LOG_DEV_LEVEL == 3
         for (auto& kv : delay_hist_) {
           MS_DEBUG_DEV("arrival_time_ms - first_arrival_time_ms_:%f, smoothed_delay_:%f", kv.first, kv.second);
         }
 #endif
-
-        hypothesis_ = BandwidthUsage::kBwOverusing;
       }
     }
   } else if (modified_trend < -threshold_) {
@@ -231,8 +238,8 @@ void TrendlineEstimator::Detect(double trend, double ts_delta, int64_t now_ms) {
   } else {
     time_over_using_ = -1;
     overuse_counter_ = 0;
-    MS_DEBUG_DEV("---- BandwidthUsage::kBwNormal ---");
     hypothesis_ = BandwidthUsage::kBwNormal;
+    MS_DEBUG_DEV("---- BandwidthUsage::kBwNormal ---");
   }
   prev_trend_ = trend;
   UpdateThreshold(modified_trend, now_ms);