Blink perf tests are used for micro benchmarking the surface of Blink that is exposed to the Web. They are the counterpart of web_tests/ but for performance coverage.
Each test entry point is a HTML file written using runner.js testing framework. The test file is placed inside a sub folder of blink/perf_tests/ and is started by importing runner.js script into the document:
<script src="../resources/runner.js"></script>
In a nutshell, to measure speed of synchronous code encapsulated in a test run method F, synchronous perf tests exercises this loop:
FOR i = 1 to NUMBER_OF_REPEAT
Start timer
F()
Stop timer
Depending on how fast F runs, one can choose between PerfTestRunner.measureTime or PerfTestRunner.measureRunsPerSecond (very fast). In either case, you create a test object & run by invoking the measure method as follow:
PerfTestRunner.measureTime({ // the "test" object
description: '...',
setup: function () { ... }, // test setup logic, called once before each run
run: function () { ... }, // contains the code to benchmark
iterationCount: 5 // repeat the test 5 times
});
In the case of PerfTestRunner.measureRunsPerSecond, each run invokes test.run multiple times.
Tracing support
When the test is run through Telemetry, you can also collect timing of trace events that happen during each run by specifying tracingCategories & traceEventsToMeasure in the test object. For example:
PerfTestRunner.measureTime({
...
run: foo,
iterationCount: 3,
tracingCategories: 'blink',
traceEventsToMeasure: ['A', 'B'],
});
To illustrate what the framework computes, imaging the test timeline as follow:
Test run times (time duration of each slice is right under it):
-----------[ foo ]-----[ foo ]-----[ foo ]------
u1 u2 u3
---------------[ A ]------------------------------------------------------
v0
-----------------[ A ]--[ A ]---------[ B ]--[A]----------[ B ]--[C]--------
v1 v2 v3 v4 v5 v6
Besides outputting timeseries [u1, u2, u3], telemetry perf test runner will also compute the total CPU times for trace events ‘A’ & ‘B’ per foo() run:
[v0 + v2, v4, 0.0][0.0, v3, v5]Example tracing synchronous tests:
In asynchronous perf test, you define your test scheduler and do your own measurement. For example:
var isDone = false;
var startTime;
function runTest() {
if (startTime) {
PerfTestRunner.measureValueAsync(PerfTestRunner.now() - startTime);
PerfTestRunner.addRunTestEndMarker(); // For tracing metrics
}
if (!isDone) {
PerfTestRunner.addRunTestStartMarker();
startTime = PerfTestRunner.now(); // For tracing metrics
// runTest will be invoked after the async operation finish
runAsyncOperation(runTest);
}
}
PerfTestRunner.startMeasureValuesAsync({
unit: 'ms',
done: function () {
isDone = true;
},
run: function() {
runTest();
},
iterationCount: 6,
});
In the example above, the call PerfTestRunner.measureValueAsync(value) send the metric of a single run to the test runner and also let the runner know that it has finished a single run. Once the number of run reaches iterationCount (6 in the example above), the done callback is invoked, setting the your test state to finished.
Tracing support
Like synchronous perf tests, tracing metrics are only available when you run your tests with Telemetry.
Unlike synchronous perf tests which the test runner framework handles test scheduling and tracing coverage for you, for most asynchronous tests, you need to manually mark when the async test begins (PerfTestRunner.addRunTestStartMarker) and ends (PerfTestRunner.addRunTestEndMarker). Once those are marked, specifying tracingCategories and traceEventsToMeasure will output CPU time metrics of trace events that happen during test runs in the fashion similar to the example of synchronous tracing test above.
Example of tracing asynchronous tests:
color-changes-measure-frame-time.html
simple-blob-measure-async.html
The sub-framework canvas_runner.js is used for tests in the canvas directory. This can measure rasterization and GPU time using requestAnimationFrame (RAF) and contains a callback framework for video.
Normal tests using runTest() work similarly to the asynchronous test above, but crucially wait for RAF after completing a single trial of MEASURE_DRAW_TIMES runs.
RAF tests are triggered by appending the query string raf (case insensitive) to the test's url. These tests wait for RAF to return before making a measurement. This way rasterization and GPU time are included in the measurement.
For example:
The test gpu-bound-shader.html is just measuring CPU, and thus looks extremely fast as the test is just one slow shader.
The url gpu-bound-shader.html?raf will measure rasterization and GPU time as well, thus giving a more realistic measurement of performance.
Running tests directly in browser
Most of Blink Performance tests should be runnable by just open the test file directly in the browser. However, features like tracing metrics & HTML results viewer won't be supported.
Running tests with Telemetry
Assuming your current directory is chromium/src/, you can run tests with:
./tools/perf/run_benchmark run blink_perf [--test-path=<path to your tests>]
For information about all supported options, run:
./tools/perf/run_benchmark run blink_perf --help