Benchmarks
Two numbers matter, and both are measured on the machine you run them on rather than quoted from a spec sheet.
Cold start
bun scripts/bench.ts
Times process start to the page-load-finished callback, in a release binary,
against a bare wry+tao binary that does the same work: one event loop, one
window, one webview, one page. Seven samples each, reported as a median.
The script also reports if a sample is far above the median, which means the machine was busy rather than the shell being slow. It exits non-zero if the overhead exceeds +10%, so CI can gate on it.
Async dispatch
bun scripts/bench.ts --dispatch
Measures the three execution models against each other:
| Path | Model | Cost per call |
|---|---|---|
command | UI thread | a function call |
async_command | thread per call | a thread spawn and join |
async_fn | task on a shared runtime | a task submit |
The interesting comparison is the last two, because that is the decision an app
author actually makes. async_command spawns a thread per call, which is fine for
a handful of concurrent calls and wasteful for a thousand. async_fn submits a
task, which costs a task rather than a stack.
The benchmark exists to keep that claim honest: it prints the per-call cost of each model at several concurrency levels, so "async_fn scales better" is a number rather than an opinion. It also fails if the pooled path is ever slower than the thread-per-call path at the concurrency level where it is supposed to win, which would mean the runtime is misconfigured.
What is not benchmarked
- Startup on Windows and macOS. The binary is built and tested there, but the cold-start number in the README is Linux, and says so.
- Render performance. That is the webview's job, and measuring it here would measure the platform rather than Voussoir.
- Binary size over time.
verify.tschecks the size bound; it is a gate, not a trend.