README.md

Event Loop Benchmark

This benchmark compares:

• stdlib asyncio
• uvloop
• the Rust prototype in rsloop

Run it from the repository root so Python resolves the editable Rust package cleanly:

uv run --with maturin maturin develop --release
uv run --with uvloop python benchmarks/compare_event_loops.py

Benchmark runner: benchmarks/compare_event_loops.py

The Rust prototype should be installed in release mode before benchmarking. Using the default debug build will heavily skew the comparison against asyncio and uvloop.

Useful quick run:

uv run --with maturin maturin develop --release
uv run --with uvloop python benchmarks/compare_event_loops.py \
  --warmups 0 \
  --repeat 3 \
  --callbacks 50000 \
  --tasks 10000 \
  --tcp-roundtrips 1000 \
  --payload-size 512

To launch an unmeasured Tracy session for each rsloop workload before the measured runs, add a label directory:

uv run --with maturin maturin develop --release --features profiler
uv run --with uvloop python benchmarks/compare_event_loops.py \
  --loops rsloop \
  --workloads callbacks,tasks,tcp_streams \
  --profile-rsloop-dir benchmarks/profiles

No files are written by Tracy. The directory argument is only used to derive a human-readable label for the unmeasured profiling pass before the warmup and measured runs. Open the Tracy desktop profiler and connect while that pass is running.

The runner executes each loop/workload in a fresh subprocess and reports:

• median time
• best time
• operations per second
• relative slowdown versus the fastest loop for that workload

Current workloads:

• callbacks: chained call_soon() dispatch
• tasks: many tiny asyncio.sleep(0) tasks
• tcp_streams: local asyncio.start_server() / asyncio.open_connection() echo round trips

By default, tcp_streams uses rsloop's native fast streams. If you want all three loops to go through the stdlib asyncio streams layer instead, pass:

uv run --with uvloop python benchmarks/compare_event_loops.py --no-rsloop-fast-streams