Memory profiling

--profile-memory runs each selected benchmark under memray once, separately from the timing pass, and attaches peak memory, total memory usage, and allocation count metadata to each run.

The capture is scoped to the benchmark’s timing loop: tracking starts at the first for _ in state iteration and stops when the loop ends. Fixture and setup allocations made before the loop are excluded, so the numbers describe the workload — and stay comparable across suites with different setup strategies.

Prerequisites

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

Note: memray is not available on Windows.

Basics

$ mew run --profile-memory
mew · host=laptop cpus=10 …
Benchmark         │  Iters │ Real │ Peak Mem │ Total Alloc
──────────────────────────────────────────────────────────
bench_alloc_list  │ 50,000 │ 8 µs │  3.2 MB  │   5.0 MB

Write a flame graph:

$ mew run --flamegraph alloc.html

This implies --profile-memory, so the bare profile flag is optional. The resulting flame graph is a self-contained HTML page, and can be inspected directly in the browser.

Caveats

  • Memray captures Python-level allocations. Allocations made by C extensions through malloc directly may or may not show up depending on the extension.

  • Like CPU profiling, the memory pass is separate from the timing pass. Don’t expect the memory column to line up with the iteration count from the timing column.

  • Memray’s tracker has measurable overhead. Treat allocations as approximate when they’re already small.

  • The loop-scoped capture applies to the stats columns. The --flamegraph capture wraps whole bodies (one tracker across all selected benchmarks), so setup allocations do appear there — useful when the fixture itself is the thing you’re hunting.