zena:bench

Passed to each benchmark function. By default the full invocation is timed; call start()/stop() (possibly repeatedly) to measure only the region of interest and exclude setup/teardown.

TODO: replace with a record type when records get optional fields.

Best-effort host description recorded with each run. Fields are null / negative when the source (/proc, /etc/hostname) is not readable — e.g. when the runner's preopens don't include the host root. Load averages matter most: a loaded machine widens CIs, and recording load explains why after the fact.

A variant's name and collected samples (milliseconds), for analyzing measurements gathered outside BenchRunner — e.g. by a host-side orchestrator round-robining across separate binaries or processes.

Summary of one benchmark variant's sample distribution. Times are in milliseconds. The confidence interval is a 95% t-interval of the mean.

95% confidence interval of (a.mean - b.mean), Welch's construction. Positive values mean a is slower than b. The pct bounds are the same interval expressed as a percentage of b's mean, so "a vs b" conclusions read as "a is faster/slower than b by X%–Y%".

Resolution thresholds, as in Tachometer: a pairwise difference is resolved once its CI straddles no horizon boundary. The default single horizon of 0% resolves when the CI excludes zero (a definite faster/slower) — but a CI lying entirely inside [-1%, +1%] is also a conclusion ("differs by less than 1%") if you configure a 1% horizon.

Monotonic time in milliseconds (nanosecond-resolution WASI clock).

Wall-clock milliseconds since the Unix epoch, for stamping results.

Builds a full BenchReport from externally collected samples: the same statistics, resolution, machine-info, and JSON pipeline as BenchRunner.run(), minus the sampling loop. durationMs is the caller's total sampling time (0.0 if unknown).

Formats a non-huge f64 with a fixed number of decimals (no exponent).

Renders the Tachometer-style terminal report.

Serializes a report for recording and later cross-run comparison. The schema is versioned; consumers must reject versions they don't know.

Times testFunc once and prints the elapsed milliseconds, after one untimed warm-up call. filter selects by substring; an empty filter runs everything.

The one-shot counterpart to BenchRunner, which samples repeatedly and reports a distribution. Use this when a single number read off stdout is the whole point, as the benchmark programs under test-files/ do.

Returns a sorted copy of the samples. Insertion sort: sample arrays are at most a few thousand entries and typically nearly sorted regions dominate.

Two-sided 95% critical value of Student's t distribution for the given (possibly fractional, from Welch–Satterthwaite) degrees of freedom.

Summarizes a sample distribution. With fewer than two samples the spread statistics are zero and the CI collapses to the mean.

Parses horizon specs: "0%", "1.5%" (relative to the compared-to mean) or "0.5ms" / "2" (absolute milliseconds). Values are symmetric: "1%" tests both +1% and -1%.

True when the difference CI straddles no horizon boundary, i.e. more sampling is not expected to change the conclusion.