Skip to content

performance.md

Latest commit

 

History

History
49 lines (34 loc) · 1.45 KB

performance.md

File metadata and controls

49 lines (34 loc) · 1.45 KB

Performance

ferryx treats performance as a compatibility dimension.

Measured Dimensions

  • call overhead (Rust function -> generated Python surface)
  • zero-copy throughput
  • NumPy interop overhead
  • async invocation latency
  • serialization/deserialization throughput
  • memory overhead under repeated calls
  • wheel import startup time
  • API generation speed

Methodology

  • Bench harnesses live in benchmarks/.
  • Each benchmark run writes JSON artifacts under evaluation/results/.
  • CI stores historical benchmark artifacts for trend analysis.
  • Criterion benches run with cargo bench --manifest-path crates/ferryx-bench/Cargo.toml.
  • Markdown reports generated with evaluation/scripts/generate_report.py.

Comparison Matrix Targets

Bench suites include compatibility runners for:

  • PyO3
  • cffi
  • ctypes
  • pybind11

ferryx does not publish synthetic numbers. All published metrics must point to reproducible scripts and raw outputs.

Regression Budget Policy

  • Parser + emitter generation speed: no >10% regression without explicit approval.
  • Runtime call path: no >5% regression without explicit approval.
  • Memory footprint: no >10% regression for baseline scenarios.

Benchmark Reproducibility

Use:

cargo run -p cargo-ferryx -- build --input examples/tensor/src/lib.rs --out-dir target/ferryx-bench --package ferryx_tensor
cargo run --manifest-path crates/ferryx-bench/Cargo.toml -- --suite all --output verification/benchmarks/latest.json