Skip to content

Benchmarks

pnport 0.1.2 is available. Publication does not establish complete benchmark acceptance or a performance promise. Release-build measurements ran on all four macOS and glibc Linux targets, but complete filesystem/process compatibility remains unverified. Experimental preview and development measurements do not establish stable performance. Windows results remain planned with 0.2.0 support. This page specifies a reproducible comparison method; check installation and known limits before use.

Reproduction protocol

  1. Record the exact pnport version, operating system and architecture, Yarn 4 version, fixture lockfile, command, and host hardware.
  2. Prepare one installed Yarn 4 PnP fixture with fixed dependency bytes. Keep project inputs and child command identical between measurements; prepare dependencies before timed execution.
  3. Measure a cold run with a fresh, empty pnport cache directory. Measure a warm run using the completed cache from that same fixture. Keep the child's incremental or application cache state identical between conditions so both runs perform the same work. Repeat each condition at least five times and report the median and range rather than a single best result.
  4. Record elapsed wall time from pnport start to exit, filesystem operation counts and bytes where the host can observe them, peak process-tree memory, and cache/disk bytes after each run. State the measurement tool, its overhead, and whether the sample includes package materialization or only a reused entry. Distinguish instrumented fixture calls from kernel syscalls or device I/O. For sampled memory, state the interval, possible missed peaks, and whether shared pages can be counted more than once.
  5. Publish the fixture and exact commands with results, and state which host/target actually ran them. Do not infer Windows or Linux performance from a macOS run.

Use --cache-dir to isolate benchmark runs and cache commands to inspect retained entries. Keep the child's own output separate from pnport diagnostics. Release results must include cold and warm behavior; no editor or arbitrary-tool certification follows from one benchmark.