CPU profiling

--sample runs each selected benchmark under pyinstrument once, separately from the timing pass, and attaches a summary, including sample count and hottest frame, to each run.

This is in-process sampling: it sees Python frames only. To capture native (C/C++) frames from a compiled extension, use mew profile instead.

Prerequisites

$ uv add 'mew[cpu]'    # or: pip install 'mew[cpu]'

Basics

$ mew run --sample
mew · host=laptop cpus=10 …
Benchmark         │   Iters │   Real │ Samples │ Hottest Frame
─────────────────────────────────────────────────────────────────
bench_sort_quick  │ 100,000 │ 24 µs  │  12,341 │ _quicksort (bench.py:18)

Write a self-contained HTML report:

$ mew run --sample --sample-html cpu.html

This implies --sample, so you can drop one flag:

$ mew run --sample-html cpu.html

Tuning the sampler

Flag	Default	Notes
--sample-interval F	1e-4	Smaller = more samples = higher overhead.
--sample-iterations N	1000	How many times the body runs under the sampler.

A profiling pass is separate from the timing pass. Profiler overhead does not pollute timing numbers, but the profiled iteration count is independent of min_time, so don’t read timings out of the profiling report.

What you’ll see

cpu.html is a single self-contained pyinstrument page with a tree view per benchmark. For the example workloads under benchmarks/bench_cpu.py:

• bench_sort_builtin collapses into a single wide sorted C frame.

• bench_sort_quicksort shows deep recursion into _quicksort.

• bench_sort_bubble shows the flat double-loop hot path.

• bench_fib_naive shows the exponential branching of naive recursion.