Benchmark

To run this benchmark yourself, clone web-infra-dev/bundler-benchmark and then use this command from the root:

./scripts/bench-all.sh ${platform} # `platform` is used to uniquely identify the generated benchmark data.

Benchmark case

all

This project is used for comparing development mode and production mode build performance. It is combined by several sub projects atlaskit-editor, common-libs, common-libs-chunks, rome and esbuild-three, which has a total of 50000 modules. here is the simple description of each sub project:

atlaskit-editor

Basic react application with @atlaskit/editor-core

common-libs

A application that used some common libraries in JavaScript ecosystem.

common-libs-chunks

Same as common-libs, but using async import() syntax to force bundler generate multiple chunks.

esbuild-three

A large JavaScript codebase by duplicating the three.js library 10 times.

rome

The old Rome code base (prior to their Rust rewrite) to approximate a large TypeScript codebase.

Data

Tested on Intel(R) Xeon(R) Platinum 8260 CPU @ 2.40GHz 32Core, 64GB of RAM

mode rspack webpack(with swc) webpack (with babel)
development cold start 3.79s 31.25s 42.61s
production cold start 22.353s 75.054s 160.059s

react-components-10000

10000 react components (each component only have basic skeleton) to test hot module replacement performance of each devServer of bundler.

Data

Tested on Intel(R) Xeon(R) Platinum 8260 CPU @ 2.40GHz 32Core, 64GB of RAM

rspack/root rspack/leaf webpack(with swc)/root webpack (with swc)/leaf webpack (with babel)/root webpack (with babel)/leaf
1 711ms 562ms 2023ms 1534ms 2081ms 1644ms
2 507ms 566ms 1725ms 1593ms 1642ms 1652ms
3 551ms 592ms 1602ms 1454ms 1786ms 1595ms
4 525ms 513ms 1480ms 1587ms 1543ms 1684ms
5 545ms 571ms 1517ms 1470ms 1658ms 1603ms
average 567.8ms 560.8ms 1669.4ms 1527.6ms 1742ms 1635.6ms

Metrics

  • dev-cold-start: the time it takes for the development build to complete without caching.
  • production-build-cold-start: the time it takes for the production build to complete without caching.
  • HMR: the time it takes for the development server rebuild the project when an update is applied to a source file. there are two sub metrics of HMR:
    • leaf: time for updating a react component file that has no dependency
    • root: time for updating a root react component, normally it is named App.tsx or index.tsx

Bundler involved in comparison

  • webpack + ts-loader (or babel-loader in react-components-10000) + terser-webpack-plugin

  • webpack + swc-loader + swc-minifier

  • rspack(using swc as transpiler and swc minifier)

All the configuration in development mode keep as is except the loader which depends on the test target.

In production mode we set devtool to source-map and use loader and minifier depends on the test target, keep other configuration as is.

Methodology

dev-cold-start

Using hyperfine to run each development build command 10 times with warmup and get the average time, finally generate corresponding markdown table.

production-cold-start

Using hyperfine to run each production build command 10 times with warmup and get the average time, finally generate corresponding markdown table.

HMR

The HMR benchmark is relatively complex because it is usually implemented in four stages.

  1. devServer listens for file changes and rebuilds the project (usually a partial build).
  2. devServer sends the built diff to the HMR client over the local network (usually a websocket).
  3. HMR client receives the diff and applies the diff to runtime code of the browser.
  4. the browser rerendering the view.

Sometimes it takes about half time of HMR from second stage to fourth stage.

Some tools' HMR benchmarks use the entire HMR time as the metric time, and since each toolchain may implement each stage of HMR differently, it's fine to compare in this way.

However, our implementation is basically the same as webpack's in phases two through four, unlike webpack, Rspack uses a more fine-grained incremental build algorithm to accelerate phase one, and to better demonstrate the effect of this algorithm, we use the re-build time of both build tools as a metric.