Steady-State Warmup

Overview

By default Sailfish runs a fixed number of warmup iterations (NumWarmupIterations, default 3) before it starts measuring. For code whose performance shifts while the runtime tiers up (tiered JIT compilation and on-stack replacement), a fixed count can start measuring too early — capturing the tail of compilation — or waste time warming up longer than needed.

Steady-state warmup (opt-in) instead warms up until the per-iteration timing stops trending and stabilizes, bounded by a floor and a cap.

How it works

Each warmup iteration is timed locally (it is not recorded into your measured samples) and fed to a detector. The detector looks at a sliding window of recent warmup durations and declares steady state when both:

• The trend has flattened — the median of the most recent half of the window is within ~5% of the median of the prior half (timings have stopped falling). Medians are used so a single cold-start spike doesn't dominate.
• Dispersion is low — the coefficient of variation across the window is within ~15%.

Warmup runs at least NumWarmupIterations (the floor) and at most MaxWarmupIterations (the cap, default 50), stopping as soon as steady state is detected. The detector needs a full window (6 samples) before it can decide, so that is the effective minimum.

Usage

[Sailfish(
    UseSteadyStateWarmup = true,  // opt in
    NumWarmupIterations = 4,      // floor: always warm up at least this many
    MaxWarmupIterations = 50,     // cap: never warm up more than this
    SampleSize = 100)]
public class MyBenchmarks
{
    [SailfishMethod]
    public void Work() { /* ... */ }
}

Configured per class on the [Sailfish] attribute (like OperationsPerInvoke). Defaults: UseSteadyStateWarmup = false (fixed-count warmup), MaxWarmupIterations = 50.

Notes

• Opt-in and backward compatible. With the default (false), warmup is the existing fixed NumWarmupIterations loop — unchanged.
• Bounded. A noisy method whose timing never settles will warm up to MaxWarmupIterations and then proceed; it won't loop forever.
• The detector thresholds (window 6, ~5% drift, ~15% CV) are tuned internally and not currently exposed.
• This shapes warmup only; it's independent of Adaptive Sampling (which decides how many measured samples to collect) and composes with it.