The Dart project benchmark harness is the recommended starting point when building a benchmark for Dart.
By default, the reported runtime in BenchmarkBase
is not for a single call to run()
, but for the average time it takes to call run()
10 times for legacy reasons. The benchmark harness executes a 10-call timing loop repeatedly until 2 seconds have elapsed; the reported result is the average of the runtimes for each loop. This behavior will change in a future major version.
Benchmarks extending BenchmarkBase
can opt into the reporting the average time to call run()
once by overriding the exercise
method:
@override void exercise() => run();
AsyncBenchmarkBase
already reports the average time to call run()
once.
If you are running the same benchmark, on the same machine, running the same OS, the reported run times can be carefully compared across runs. Carefully because there are a variety of factors which could cause error in the run time, for example, the load from other applications running on your machine could alter the result.
Comparing the run time of different benchmarks is not recommended. In other words, don't compare apples with oranges.
BenchmarkBase
class that all new benchmarks should extend
.AsyncBenchmarkBase
for asynchronous benchmarks.1. Add the following to your project's pubspec.yaml
dependencies: benchmark_harness: any
2. Install pub packages
dart pub install
3. Add the following import:
import 'package:benchmark_harness/benchmark_harness.dart';
4. Create a benchmark class which inherits from BenchmarkBase
or AsyncBenchmarkBase
.
Create a dart file in the benchmark/
folder of your package.
// Import BenchmarkBase class. import 'package:benchmark_harness/benchmark_harness.dart'; // Create a new benchmark by extending BenchmarkBase class TemplateBenchmark extends BenchmarkBase { const TemplateBenchmark() : super('Template'); static void main() { const TemplateBenchmark().report(); } // The benchmark code. @override void run() {} // Not measured setup code executed prior to the benchmark runs. @override void setup() {} // Not measured teardown code executed after the benchmark runs. @override void teardown() {} // To opt into the reporting the time per run() instead of per 10 run() calls. //@override //void exercise() => run(); } void main() { // Run TemplateBenchmark TemplateBenchmark.main(); }
Template(RunTime): 0.1568472448997197 us.
This is the average amount of time it takes to run run()
10 times for BenchmarkBase
and once for AsyncBenchmarkBase
.
µs is an abbreviation for microseconds.
bench
commandA convenience command available in package:benchmark_harness
.
If a package depends on benchmark_harness
, invoke the command by running
dart run benchmark_harness:bench
If not, you can use this command by activating it.
dart pub global activate benchmark_harness dart pub global run benchmark_harness:bench
Output from dart run benchmark_harness:bench --help
Runs a dart script in a number of runtimes. Meant to make it easy to run a benchmark executable across runtimes to validate performance impacts. -f, --flavor [aot] Compile and run as a native binary. [jit] Run as-is without compilation, using the just-in-time (JIT) runtime. [js] Compile to JavaScript and run on node. [wasm] Compile to WebAssembly and run on node. --target The target script to compile and run. (defaults to "benchmark/benchmark.dart") -h, --help Print usage information and quit. -v, --verbose Print the full stack trace if an exception is thrown.
Example usage:
dart run benchmark_harness:bench --flavor aot --target example/template.dart AOT - COMPILE /dart_installation/dart-sdk/bin/dart compile exe example/template.dart -o /temp_dir/bench_1747680526905_GtfAeM/out.exe Generated: /temp_dir/bench_1747680526905_GtfAeM/out.exe AOT - RUN /temp_dir/bench_1747680526905_GtfAeM/out.exe Template(RunTime): 0.005620051244379949 us.