Skip to content

F
Front-End
Unverified Commit a30c46dc authored by Yegor's avatar Yegor Committed by GitHub
Browse files
Options

Remove outliers in Web benchmarks to reduce noise; add visualization (#54883) 

* web benchmarks: separate outliers to reduce noise; add visualization
parent 20803507
Show whitespace changes
Inline Side-by-side
Showing with 300 additions and 104 deletions
dev/benchmarks/macrobenchmarks/lib/src/web/bench_dynamic_clip_on_static_picture.dart
View file @ a30c46dc
......@@ -36,7 +36,7 @@ class BenchDynamicClipOnStaticPicture extends SceneBuilderRecorder {
// If the scrollable extent is too small, the benchmark may end up
// scrolling the picture out of the clip area entirely, resulting in
// bogus metric vaules.
const double maxScrollExtent = kMaxSampleCount * kScrollDelta;
const double maxScrollExtent = kTotalSampleCount * kScrollDelta;
const double pictureHeight = kRows * kRowHeight;
if (maxScrollExtent > pictureHeight) {
throw Exception(
......
dev/benchmarks/macrobenchmarks/lib/src/web/recorder.dart
View file @ a30c46dc
......@@ -16,28 +16,16 @@ import 'package:flutter/scheduler.dart';
import 'package:flutter/rendering.dart';
import 'package:flutter/widgets.dart';
/// Minimum number of samples collected by a benchmark irrespective of noise
/// levels.
const int kMinSampleCount = 50;
/// Maximum number of samples collected by a benchmark irrespective of noise
/// levels.
/// The number of samples from warm-up iterations.
///
/// If the noise doesn't settle down before we reach the max we'll report noisy
/// results assuming the benchmarks is simply always noisy.
const int kMaxSampleCount = 10 * kMinSampleCount;
/// We warm-up the benchmark prior to measuring to allow JIT and caches to settle.
const int _kWarmUpSampleCount = 200;
/// The number of samples used to extract metrics, such as noise, means,
/// max/min values.
///
/// Keep this constant in sync with the same constant defined in `dev/devicelab/lib/framework/browser.dart`.
const int _kMeasuredSampleCount = 10;
/// The number of samples we use to collect statistics from.
const int _kMeasuredSampleCount = 100;
/// Maximum tolerated noise level.
///
/// A benchmark continues running until a noise level below this threshold is
/// reached.
const double _kNoiseThreshold = 0.05; // 5%
/// The total number of samples collected by a benchmark.
const int kTotalSampleCount = _kWarmUpSampleCount + _kMeasuredSampleCount;
/// Measures the amount of time [action] takes.
Duration timeAction(VoidCallback action) {
......@@ -321,8 +309,7 @@ abstract class WidgetRecorder extends Recorder implements FrameRecorder {
///
/// The widget must create its own animation to drive the benchmark. The
/// animation should continue indefinitely. The benchmark harness will stop
/// pumping frames automatically as soon as the noise levels are sufficiently
/// low.
/// pumping frames automatically.
Widget createWidget();
@override
......@@ -503,54 +490,186 @@ class _WidgetBuildRecorderHostState extends State<_WidgetBuildRecorderHost> {
class Timeseries {
Timeseries(this.name);
/// The label of this timeseries used for debugging and result inspection.
final String name;
/// List of all the values that have been recorded.
///
/// This list has no limit.
final List<num> _allValues = <num>[];
/// List of values that are being used for measurement purposes.
///
/// [average], [standardDeviation] and [noise] are all based on this list, not
/// the [_allValues] list.
final List<num> _measuredValues = <num>[];
final List<double> _allValues = <double>[];
/// The total amount of data collected, including ones that were dropped
/// because of the sample size limit.
int get count => _allValues.length;
/// Computes the average value of the measured values.
double get average => _computeAverage(name, _measuredValues);
/// Computes the standard deviation of the measured values.
double get standardDeviation =>
_computeStandardDeviationForPopulation(name, _measuredValues);
/// Computes noise as a multiple of the [average] value.
///
/// This value can be multiplied by 100.0 to get noise as a percentage of
/// the average.
/// Extracts useful statistics out of this timeseries.
///
/// If [average] is zero, treats the result as perfect score, returns zero.
double get noise => average > 0.0 ? standardDeviation / average : 0.0;
/// See [TimeseriesStats] for more details.
TimeseriesStats computeStats() {
// The first few values we simply discard and never look at. They're from the warm-up phase.
final List<double> warmUpValues = _allValues.sublist(0, _allValues.length - _kMeasuredSampleCount);
// Values we analyze.
final List<double> candidateValues = _allValues.sublist(_allValues.length - _kMeasuredSampleCount);
// The average that includes outliers.
final double dirtyAverage = _computeAverage(name, candidateValues);
// The standard deviation that includes outliers.
final double dirtyStandardDeviation = _computeStandardDeviationForPopulation(name, candidateValues);
// Any value that's higher than this is considered an outlier.
final double outlierCutOff = dirtyAverage + dirtyStandardDeviation;
// Candidates with outliers removed.
final Iterable<double> cleanValues = candidateValues.where((double value) => value <= outlierCutOff);
// Outlier candidates.
final Iterable<double> outliers = candidateValues.where((double value) => value > outlierCutOff);
// Final statistics.
final double cleanAverage = _computeAverage(name, cleanValues);
final double outlierAverage = _computeAverage(name, outliers);
final double standardDeviation = _computeStandardDeviationForPopulation(name, cleanValues);
final double noise = cleanAverage > 0.0 ? standardDeviation / cleanAverage : 0.0;
final List<AnnotatedSample> annotatedValues = <AnnotatedSample>[
for (final double warmUpValue in warmUpValues)
AnnotatedSample(
magnitude: warmUpValue,
isOutlier: warmUpValue > outlierCutOff,
isWarmUpValue: true,
),
for (final double candidate in candidateValues)
AnnotatedSample(
magnitude: candidate,
isOutlier: candidate > outlierCutOff,
isWarmUpValue: false,
),
];
return TimeseriesStats(
name: name,
average: cleanAverage,
outlierCutOff: outlierCutOff,
outlierAverage: outlierAverage,
standardDeviation: standardDeviation,
noise: noise,
cleanSampleCount: cleanValues.length,
outlierSampleCount: outliers.length,
samples: annotatedValues,
);
}
/// Adds a value to this timeseries.
void add(num value) {
void add(double value) {
if (value < 0.0) {
throw StateError(
'Timeseries $name: negative metric values are not supported. Got: $value',
);
}
_measuredValues.add(value);
_allValues.add(value);
// Don't let the [_measuredValues] list grow beyond [_kMeasuredSampleCount].
if (_measuredValues.length > _kMeasuredSampleCount) {
_measuredValues.removeAt(0);
}
}
/// Various statistics about a [Timeseries].
///
/// See the docs on the individual fields for more details.
@sealed
class TimeseriesStats {
const TimeseriesStats({
@required this.name,
@required this.average,
@required this.outlierCutOff,
@required this.outlierAverage,
@required this.standardDeviation,
@required this.noise,
@required this.cleanSampleCount,
@required this.outlierSampleCount,
@required this.samples,
});
/// The label used to refer to the corresponding timeseries.
final String name;
/// The average value of the measured samples without outliers.
final double average;
/// The standard deviation in the measured samples without outliers.
final double standardDeviation;
/// The noise as a multiple of the [average] value takes from clean samples.
///
/// This value can be multiplied by 100.0 to get noise as a percentage of
/// the average.
///
/// If [average] is zero, treats the result as perfect score, returns zero.
final double noise;
/// The maximum value a sample can have without being considered an outlier.
///
/// See [Timeseries.computeStats] for details on how this value is computed.
final double outlierCutOff;
/// The average of outlier samples.
///
/// This value can be used to judge how badly we jank, when we jank.
///
/// Another useful metrics is the difference between [outlierAverage] and
/// [average]. The smaller the value the more predictable is the performance
/// of the corresponding benchmark.
final double outlierAverage;
/// The number of measured samples after outlier are removed.
final int cleanSampleCount;
/// The number of outliers.
final int outlierSampleCount;
/// All collected samples, annotated with statistical information.
///
/// See [AnnotatedSample] for more details.
final List<AnnotatedSample> samples;
@override
String toString() {
final StringBuffer buffer = StringBuffer();
buffer.writeln(
'$name: (samples: $cleanSampleCount clean/$outlierSampleCount outliers/'
'${cleanSampleCount + outlierSampleCount} measured/'
'${samples.length} total)');
buffer.writeln(' | average: $average μs');
buffer.writeln(' | outlier average: $outlierAverage μs');
buffer.writeln(' | noise: ${_ratioToPercent(noise)}');
return buffer.toString();
}
}
/// Annotates a single measurement with statistical information.
@sealed
class AnnotatedSample {
const AnnotatedSample({
@required this.magnitude,
@required this.isOutlier,
@required this.isWarmUpValue,
});
/// The non-negative raw result of the measurement.
final double magnitude;
/// Whether this sample was considered an outlier.
final bool isOutlier;
/// Whether this sample was taken during the warm-up phase.
///
/// If this value is `true`, this sample does not participate in
/// statistical computations. However, the sample would still be
/// shown in the visualization of results so that the benchmark
/// can be inspected manually to make sure there's a predictable
/// warm-up regression slope.
final bool isWarmUpValue;
}
/// Base class for a profile collected from running a benchmark.
class Profile {
Profile({@required this.name}) : assert(name != null);
......@@ -572,7 +691,7 @@ class Profile {
}
void addDataPoint(String key, Duration duration) {
scoreData.putIfAbsent(key, () => Timeseries(key)).add(duration.inMicroseconds);
scoreData.putIfAbsent(key, () => Timeseries(key)).add(duration.inMicroseconds.toDouble());
}
/// Decides whether the data collected so far is sufficient to stop, or
......@@ -584,56 +703,15 @@ class Profile {
/// method will return true (asking the benchmark to continue collecting
/// data).
bool shouldContinue() {
// If we haven't recorded anything yet, we don't wanna stop now.
// If there are no `Timeseries` in the `scoreData`, then we haven't
// recorded anything yet. Don't stop.
if (scoreData.isEmpty) {
return true;
}
// Accumulates all the messages to be printed when the final decision is to
// stop collecting data.
final StringBuffer buffer = StringBuffer();
final Iterable<bool> shouldContinueList = scoreData.keys.map((String key) {
final Timeseries timeseries = scoreData[key];
// Collect enough data points before considering to stop.
if (timeseries.count < kMinSampleCount) {
return true;
}
// Is it still too noisy?
if (timeseries.noise > _kNoiseThreshold) {
// If the timeseries has enough data, stop it, even if it's noisy under
// the assumption that this benchmark is always noisy and there's nothing
// we can do about it.
if (timeseries.count > kMaxSampleCount) {
buffer.writeln(
'WARNING: Noise of benchmark "$name.$key" did not converge below '
'${_ratioToPercent(_kNoiseThreshold)}. Stopping because it reached the '
'maximum number of samples $kMaxSampleCount. Noise level is '
'${_ratioToPercent(timeseries.noise)}.',
);
return false;
} else {
return true;
}
}
buffer.writeln(
'SUCCESS: Benchmark "$name.$key" converged below ${_ratioToPercent(_kNoiseThreshold)}. '
'Noise level is ${_ratioToPercent(timeseries.noise)}.',
);
return false;
});
// If any of the score data needs to continue to be collected, we should
// return true.
final bool finalDecision =
shouldContinueList.any((bool element) => element);
if (!finalDecision) {
print(buffer.toString());
}
return finalDecision;
// We have recorded something, but do we have enough samples? If every
// timeseries has collected enough samples, stop the benchmark.
return !scoreData.keys.every((String key) => scoreData[key].count >= kTotalSampleCount);
}
/// Returns a JSON representation of the profile that will be sent to the
......@@ -647,9 +725,12 @@ class Profile {
for (final String key in scoreData.keys) {
scoreKeys.add('$key.average');
scoreKeys.add('$key.outlierAverage');
final Timeseries timeseries = scoreData[key];
json['$key.average'] = timeseries.average;
json['$key.noise'] = timeseries.noise;
final TimeseriesStats stats = timeseries.computeStats();
json['$key.average'] = stats.average;
json['$key.outlierAverage'] = stats.outlierAverage;
json['$key.noise'] = stats.noise;
}
json.addAll(extraData);
......@@ -663,9 +744,8 @@ class Profile {
buffer.writeln('name: $name');
for (final String key in scoreData.keys) {
final Timeseries timeseries = scoreData[key];
buffer.writeln('$key: (samples=${timeseries.count})');
buffer.writeln(' | average: ${timeseries.average} μs');
buffer.writeln(' | noise: ${_ratioToPercent(timeseries.noise)}');
final TimeseriesStats stats = timeseries.computeStats();
buffer.writeln(stats.toString());
}
for (final String key in extraData.keys) {
final dynamic value = extraData[key];
......@@ -683,12 +763,12 @@ class Profile {
}
/// Computes the arithmetic mean (or average) of given [values].
double _computeAverage(String label, Iterable<num> values) {
double _computeAverage(String label, Iterable<double> values) {
if (values.isEmpty) {
throw StateError('$label: attempted to compute an average of an empty value list.');
}
final num sum = values.reduce((num a, num b) => a + b);
final double sum = values.reduce((double a, double b) => a + b);
return sum / values.length;
}
......@@ -699,14 +779,14 @@ double _computeAverage(String label, Iterable<num> values) {
/// See also:
///
/// * https://en.wikipedia.org/wiki/Standard_deviation
double _computeStandardDeviationForPopulation(String label, Iterable<num> population) {
double _computeStandardDeviationForPopulation(String label, Iterable<double> population) {
if (population.isEmpty) {
throw StateError('$label: attempted to compute the standard deviation of empty population.');
}
final double mean = _computeAverage(label, population);
final double sumOfSquaredDeltas = population.fold<double>(
0.0,
(double previous, num value) => previous += math.pow(value - mean, 2),
(double previous, double value) => previous += math.pow(value - mean, 2),
);
return math.sqrt(sumOfSquaredDeltas / population.length);
}
......
dev/benchmarks/macrobenchmarks/lib/web_benchmarks.dart
View file @ a30c46dc
......@@ -5,6 +5,7 @@
import 'dart:async';
import 'dart:convert' show json;
import 'dart:html' as html;
import 'dart:math' as math;
import 'package:macrobenchmarks/src/web/bench_text_layout.dart';
import 'package:macrobenchmarks/src/web/bench_text_out_of_picture_bounds.dart';
......@@ -83,6 +84,7 @@ Future<void> _runBenchmark(String benchmarkName) async {
if (!_client.isInManualMode) {
await _client.sendProfileData(profile);
} else {
_printResultsToScreen(profile);
print(profile);
}
} catch (error, stackTrace) {
......@@ -121,6 +123,120 @@ void _fallbackToManual(String error) {
}
}
/// Visualizes results on the Web page for manual inspection.
void _printResultsToScreen(Profile profile) {
html.document.body.remove();
html.document.body = html.BodyElement();
html.document.body.appendHtml('<h2>${profile.name}</h2>');
profile.scoreData.forEach((String scoreKey, Timeseries timeseries) {
html.document.body.appendHtml('<h2>$scoreKey</h2>');
html.document.body.appendHtml('<pre>${timeseries.computeStats()}</pre>');
html.document.body.append(TimeseriesVisualization(timeseries).render());
});
}
/// Draws timeseries data and statistics on a canvas.
class TimeseriesVisualization {
TimeseriesVisualization(this._timeseries) {
_stats = _timeseries.computeStats();
_canvas = html.CanvasElement();
_screenWidth = html.window.screen.width;
_canvas.width = _screenWidth;
_canvas.height = (_kCanvasHeight * html.window.devicePixelRatio).round();
_canvas.style
..width = '100%'
..height = '${_kCanvasHeight}px'
..outline = '1px solid green';
_ctx = _canvas.context2D;
// The amount of vertical space available on the chart. Because some
// outliers can be huge they can dwarf all the useful values. So we
// limit it to 1.5 x the biggest non-outlier.
_maxValueChartRange = 1.5 * _stats.samples
.where((AnnotatedSample sample) => !sample.isOutlier)
.map<double>((AnnotatedSample sample) => sample.magnitude)
.fold<double>(0, math.max);
}
static const double _kCanvasHeight = 200;
final Timeseries _timeseries;
TimeseriesStats _stats;
html.CanvasElement _canvas;
html.CanvasRenderingContext2D _ctx;
int _screenWidth;
// Used to normalize benchmark values to chart height.
double _maxValueChartRange;
/// Converts a sample value to vertical canvas coordinates.
///
/// This does not work for horizontal coordinates.
double _normalized(double value) {
return _kCanvasHeight * value / _maxValueChartRange;
}
/// A utility for drawing lines.
void drawLine(num x1, num y1, num x2, num y2) {
_ctx.beginPath();
_ctx.moveTo(x1, y1);
_ctx.lineTo(x2, y2);
_ctx.stroke();
}
/// Renders the timeseries into a `<canvas>` and returns the canvas element.
html.CanvasElement render() {
_ctx.translate(0, _kCanvasHeight * html.window.devicePixelRatio);
_ctx.scale(1, -html.window.devicePixelRatio);
final double barWidth = _screenWidth / _stats.samples.length;
double xOffset = 0;
for (int i = 0; i < _stats.samples.length; i++) {
final AnnotatedSample sample = _stats.samples[i];
if (sample.isWarmUpValue) {
// Put gray background behing warm-up samples.
_ctx.fillStyle = 'rgba(200,200,200,1)';
_ctx.fillRect(xOffset, 0, barWidth, _normalized(_maxValueChartRange));
}
if (sample.magnitude > _maxValueChartRange) {
// The sample value is so big it doesn't fit on the chart. Paint it purple.
_ctx.fillStyle = 'rgba(100,50,100,0.8)';
} else if (sample.isOutlier) {
// The sample is an outlier, color it light red.
_ctx.fillStyle = 'rgba(255,50,50,0.6)';
} else {
// A non-outlier sample, color it light blue.
_ctx.fillStyle = 'rgba(50,50,255,0.6)';
}
_ctx.fillRect(xOffset, 0, barWidth - 1, _normalized(sample.magnitude));
xOffset += barWidth;
}
// Draw a horizontal solid line corresponding to the average.
_ctx.lineWidth = 1;
drawLine(0, _normalized(_stats.average), _screenWidth, _normalized(_stats.average));
// Draw a horizontal dashed line corresponding to the outlier cut off.
_ctx.setLineDash(<num>[5, 5]);
drawLine(0, _normalized(_stats.outlierCutOff), _screenWidth, _normalized(_stats.outlierCutOff));
// Draw a light red band that shows the noise (1 stddev in each direction).
_ctx.fillStyle = 'rgba(255,50,50,0.3)';
_ctx.fillRect(
0,
_normalized(_stats.average * (1 - _stats.noise)),
_screenWidth,
_normalized(2 * _stats.average * _stats.noise),
);
return _canvas;
}
}
/// Implements the client REST API for the local benchmark server.
///
/// The local server is optional. If it is not available the benchmark UI must
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment

HIAST - GitLab