Benchmarks

Criterion benchmarks for every analysis method on the Ishigami function (d = 3), plus sampling.

Methodology

Criterion 0.5, 100 samples per benchmark, automatic warmup. Statistic: median. Machine: Apple Silicon.

Test function: Ishigami f(x) = \sin(x_1) + 7\sin^2(x_2) + 0.1 x_3^4 \sin(x_1) with x_i \in [-\pi, \pi].

What is timed: Methods marked [fn] evaluate the model function inside the timed loop (sampling + analysis are interleaved in the estimator). All other methods take pre-computed (X, Y) and time the analysis step only.

Analysis

[fn] = benchmark includes Ishigami evaluation inside the timed loop. For these methods, the reported time is analysis + function evaluation. Ishigami is trivial (~3 ns/eval); for expensive models, function evaluation dominates and the analysis overhead shown here becomes negligible.

Morris N is trajectory count r; total evaluations are r \times (d + 1).

FAST N values are odd (required by the algorithm).

Sampling

Saltelli N is base sample size; the matrix has N \times (d + 2) rows. Morris N is trajectory count.

Reproducing

cargo bench --manifest-path crates/salib/Cargo.toml

# Regenerate this document from Criterion results
python benches/generate_benchmark_doc.py