Interface ResolvedBenchOptions

The number of times that a task should run if even the time option is finished.

Function to get the current timestamp in milliseconds.

Setup function to run before each benchmark task (cycle)

Whether to subtract an estimated timestamp provider call overhead from each raw latency sample.

Each sample is measured as t1 - t0 around a single call to the task function, so every raw sample is inflated by approximately one timestamp provider call cost C. When this option is true, an estimate Ĉ is computed once at construction time via calibrateTimerOverhead, and max(0, raw_sample - Ĉ) is used in place of each non-overridden sample before statistics are computed.

Statistics after correction. All fields of Statistics are derived from the clamped corrected samples, not from the raw distribution. With M denoting the raw-sample mean:

• Clean-shift regime (X >> Ĉ). The clamp max(0, …) rarely triggers, so the correction acts as a translation by Ĉ. Location statistics (mean, min, max, all percentiles) decrease by Ĉ; absolute-unit dispersion (vr, sd, sem, moe, mad, aad) is essentially unchanged. Because rme = moe / mean, it inflates by the deterministic factor M / (M − Ĉ) whenever Ĉ > 0.
• Sub-overhead regime (X ≈ Ĉ). A non-trivial fraction of samples clamp to 0, biasing the corrected mean upward, contracting vr/sd/sem/moe/aad, and compounding the M / (M − Ĉ) factor in rme. Once the cumulative mass of raw samples at or below Ĉ reaches a given quantile, that percentile collapses to 0; in particular p50 collapses once at least half of the raw samples satisfy raw_sample ≤ Ĉ, which then forces mad and aad toward 0. Prefer overriddenDuration for sub-overhead measurements.

Three observable consequences of the clamp.

1. latency.min may be exactly 0 even when no zero-duration sample was actually observed.
2. The throughput estimator substitutes 1000 / latency.mean (or 0 when mean === 0) for every clamped sample.
3. detectTimerSaturation criterion 'zero-dominated' cannot distinguish clamped samples from genuine zero-duration timer reads, so a 'warning' event may be dispatched in the sub-overhead regime even when the timer itself is not saturated.

Caveat — concurrency: "task". The overhead is calibrated once at construction time with sequential timer calls. Setting both options causes the constructor (and run()) to throw, since the sequentially-calibrated estimate would not reflect the per-iteration timer call cost under concurrent execution.

Caveat — overriddenDuration. Samples returned by the task function via overriddenDuration are intentional user values and are never modified by the correction. They are also excluded from Task.detectedResolution and from timer-saturation detection.

On runtimes with a coarse timer (resolution >= 1 ms), the calibration returns 0 and this option becomes a no-op.

Teardown function to run after each benchmark task (cycle).

Throws if a task fails.

Time needed for running a benchmark task in milliseconds.

Warmup benchmark.

Warmup iterations.

Warmup time in milliseconds.