zqz/platform-external-webrtc
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Avoid precision loss in TrendlineEstimator by passing the arrival time as an int64_t instead of a double.

Browse Source 
BUG=webrtc:6884

Review-Url: https://codereview.webrtc.org/2577463002
Cr-Commit-Position: refs/heads/master@{#15631}
This commit is contained in:
terelius
2016-12-15 06:19:59 -08:00
committed by Commit bot
parent 3168c7a04b
commit c12cbaf9dd
3 changed files with 58 additions and 83 deletions
Download Patch File Download Diff File Expand all files Collapse all files

34
webrtc/modules/congestion_controller/trendline_estimator.cc
View File

@ -13,12 +13,14 @@
#include <algorithm>
#include "webrtc/base/checks.h"
#include "webrtc/base/optional.h"
#include "webrtc/modules/remote_bitrate_estimator/test/bwe_test_logging.h"
namespace webrtc {
namespace {
double LinearFitSlope(const std::list<std::pair<double, double>> points) {
rtc::Optional<double> LinearFitSlope(
const std::list<std::pair<double, double>> points) {
RTC_DCHECK(points.size() >= 2);
// Compute the "center of mass".
double sum_x = 0;
@ -36,7 +38,9 @@ double LinearFitSlope(const std::list<std::pair<double, double>> points) {
numerator += (point.first - x_avg) * (point.second - y_avg);
denominator += (point.first - x_avg) * (point.first - x_avg);
}
return numerator / denominator;
if (denominator == 0)
return rtc::Optional<double>();
return rtc::Optional<double>(numerator / denominator);
}
} // namespace
@ -49,6 +53,7 @@ TrendlineEstimator::TrendlineEstimator(size_t window_size,
smoothing_coef_(smoothing_coef),
threshold_gain_(threshold_gain),
num_of_deltas_(0),
first_arrival_time_ms(-1),
accumulated_delay_(0),
smoothed_delay_(0),
delay_hist_(),
@ -58,30 +63,35 @@ TrendlineEstimator::~TrendlineEstimator() {}
void TrendlineEstimator::Update(double recv_delta_ms,
double send_delta_ms,
double now_ms) {
int64_t arrival_time_ms) {
const double delta_ms = recv_delta_ms - send_delta_ms;
++num_of_deltas_;
if (num_of_deltas_ > kDeltaCounterMax) {
if (num_of_deltas_ > kDeltaCounterMax)
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", now_ms, accumulated_delay_);
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", now_ms, 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(now_ms, smoothed_delay_));
if (delay_hist_.size() > window_size_) {
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();
}
if (delay_hist_.size() == window_size_) {
trendline_ = LinearFitSlope(delay_hist_);
// Only update trendline_ if it is possible to fit a line to the data.
trendline_ = LinearFitSlope(delay_hist_).value_or(trendline_);
}
BWE_TEST_LOGGING_PLOT(1, "trendline_slope", now_ms, trendline_);
BWE_TEST_LOGGING_PLOT(1, "trendline_slope", arrival_time_ms, trendline_);
}
} // namespace webrtc

10
webrtc/modules/congestion_controller/trendline_estimator.h
View File

@ -10,11 +10,13 @@
#ifndef WEBRTC_MODULES_CONGESTION_CONTROLLER_TRENDLINE_ESTIMATOR_H_
#define WEBRTC_MODULES_CONGESTION_CONTROLLER_TRENDLINE_ESTIMATOR_H_
#include <stddef.h>
#include <stdint.h>
#include <list>
#include <utility>
#include "webrtc/base/constructormagic.h"
#include "webrtc/common_types.h"
namespace webrtc {
@ -33,7 +35,9 @@ class TrendlineEstimator {
// Update the estimator with a new sample. The deltas should represent deltas
// between timestamp groups as defined by the InterArrival class.
void Update(double recv_delta_ms, double send_delta_ms, double now_ms);
void Update(double recv_delta_ms,
double send_delta_ms,
int64_t arrival_time_ms);
// Returns the estimated trend k multiplied by some gain.
// 0 < k < 1 -> the delay increases, queues are filling up
@ -51,6 +55,8 @@ class TrendlineEstimator {
const double threshold_gain_;
// Used by the existing threshold.
unsigned int num_of_deltas_;
// Keep the arrival times small by using the change from the first packet.
int64_t first_arrival_time_ms;
// Exponential backoff filtering.
double accumulated_delay_;
double smoothed_delay_;

97
webrtc/modules/congestion_controller/trendline_estimator_unittest.cc
View File

@ -15,100 +15,59 @@
namespace webrtc {
namespace {
constexpr size_t kWindowSize = 15;
constexpr size_t kWindowSize = 20;
constexpr double kSmoothing = 0.0;
constexpr double kGain = 1;
constexpr int64_t kAvgTimeBetweenPackets = 10;
constexpr size_t kPacketCount = 2 * kWindowSize + 1;
} // namespace
TEST(TrendlineEstimator, PerfectLineSlopeOneHalf) {
void TestEstimator(double slope, double jitter_stddev, double tolerance) {
TrendlineEstimator estimator(kWindowSize, kSmoothing, kGain);
Random rand(0x1234567);
double now_ms = rand.Rand<double>() * 10000;
for (size_t i = 1; i < 2 * kWindowSize; i++) {
double send_delta = rand.Rand<double>() * 2 * kAvgTimeBetweenPackets;
double recv_delta = 2 * send_delta;
now_ms += recv_delta;
estimator.Update(recv_delta, send_delta, now_ms);
Random random(0x1234567);
int64_t send_times[kPacketCount];
int64_t recv_times[kPacketCount];
int64_t send_start_time = random.Rand(1000000);
int64_t recv_start_time = random.Rand(1000000);
for (size_t i = 0; i < kPacketCount; ++i) {
send_times[i] = send_start_time + i * kAvgTimeBetweenPackets;
double latency = i * kAvgTimeBetweenPackets / (1 - slope);
double jitter = random.Gaussian(0, jitter_stddev);
recv_times[i] = recv_start_time + latency + jitter;
}
for (size_t i = 1; i < kPacketCount; ++i) {
double recv_delta = recv_times[i] - recv_times[i - 1];
double send_delta = send_times[i] - send_times[i - 1];
estimator.Update(recv_delta, send_delta, recv_times[i]);
if (i < kWindowSize)
EXPECT_NEAR(estimator.trendline_slope(), 0, 0.001);
else
EXPECT_NEAR(estimator.trendline_slope(), 0.5, 0.001);
EXPECT_NEAR(estimator.trendline_slope(), slope, tolerance);
}
}
TEST(TrendlineEstimator, PerfectLineSlopeOneHalf) {
TestEstimator(0.5, 0, 0.001);
}
TEST(TrendlineEstimator, PerfectLineSlopeMinusOne) {
TrendlineEstimator estimator(kWindowSize, kSmoothing, kGain);
Random rand(0x1234567);
double now_ms = rand.Rand<double>() * 10000;
for (size_t i = 1; i < 2 * kWindowSize; i++) {
double send_delta = rand.Rand<double>() * 2 * kAvgTimeBetweenPackets;
double recv_delta = 0.5 * send_delta;
now_ms += recv_delta;
estimator.Update(recv_delta, send_delta, now_ms);
if (i < kWindowSize)
EXPECT_NEAR(estimator.trendline_slope(), 0, 0.001);
else
EXPECT_NEAR(estimator.trendline_slope(), -1, 0.001);
}
TestEstimator(-1, 0, 0.001);
}
TEST(TrendlineEstimator, PerfectLineSlopeZero) {
TrendlineEstimator estimator(kWindowSize, kSmoothing, kGain);
Random rand(0x1234567);
double now_ms = rand.Rand<double>() * 10000;
for (size_t i = 1; i < 2 * kWindowSize; i++) {
double send_delta = rand.Rand<double>() * 2 * kAvgTimeBetweenPackets;
double recv_delta = send_delta;
now_ms += recv_delta;
estimator.Update(recv_delta, send_delta, now_ms);
EXPECT_NEAR(estimator.trendline_slope(), 0, 0.001);
}
TestEstimator(0, 0, 0.001);
}
TEST(TrendlineEstimator, JitteryLineSlopeOneHalf) {
TrendlineEstimator estimator(kWindowSize, kSmoothing, kGain);
Random rand(0x1234567);
double now_ms = rand.Rand<double>() * 10000;
for (size_t i = 1; i < 2 * kWindowSize; i++) {
double send_delta = rand.Rand<double>() * 2 * kAvgTimeBetweenPackets;
double recv_delta = 2 * send_delta + rand.Gaussian(0, send_delta / 3);
now_ms += recv_delta;
estimator.Update(recv_delta, send_delta, now_ms);
if (i < kWindowSize)
EXPECT_NEAR(estimator.trendline_slope(), 0, 0.001);
else
EXPECT_NEAR(estimator.trendline_slope(), 0.5, 0.1);
}
TestEstimator(0.5, kAvgTimeBetweenPackets / 3.0, 0.01);
}
TEST(TrendlineEstimator, JitteryLineSlopeMinusOne) {
TrendlineEstimator estimator(kWindowSize, kSmoothing, kGain);
Random rand(0x1234567);
double now_ms = rand.Rand<double>() * 10000;
for (size_t i = 1; i < 2 * kWindowSize; i++) {
double send_delta = rand.Rand<double>() * 2 * kAvgTimeBetweenPackets;
double recv_delta = 0.5 * send_delta + rand.Gaussian(0, send_delta / 25);
now_ms += recv_delta;
estimator.Update(recv_delta, send_delta, now_ms);
if (i < kWindowSize)
EXPECT_NEAR(estimator.trendline_slope(), 0, 0.001);
else
EXPECT_NEAR(estimator.trendline_slope(), -1, 0.1);
}
TestEstimator(-1, kAvgTimeBetweenPackets / 3.0, 0.075);
}
TEST(TrendlineEstimator, JitteryLineSlopeZero) {
TrendlineEstimator estimator(kWindowSize, kSmoothing, kGain);
Random rand(0x1234567);
double now_ms = rand.Rand<double>() * 10000;
for (size_t i = 1; i < 2 * kWindowSize; i++) {
double send_delta = rand.Rand<double>() * 2 * kAvgTimeBetweenPackets;
double recv_delta = send_delta + rand.Gaussian(0, send_delta / 8);
now_ms += recv_delta;
estimator.Update(recv_delta, send_delta, now_ms);
EXPECT_NEAR(estimator.trendline_slope(), 0, 0.1);
}
TestEstimator(0, kAvgTimeBetweenPackets / 3.0, 0.02);
}
} // namespace webrtc