Skip to content

performance-benchmarks.adoc

Latest commit

 

History

History
30 lines (22 loc) · 1.88 KB

performance-benchmarks.adoc

File metadata and controls

30 lines (22 loc) · 1.88 KB

Performance Benchmarks

Chronicle Software

Current Baseline

  • The published README table reports throughput (GB/s) and bootstrap latency (ns) per algorithm. MetroHash and wyHash entries are currently marked TODO pending refreshed measurements (README.adoc:53-66).

  • No automated benchmarking harness ships with the repository today. Bench figures are captured offline and copied into the documentation.

Measurement Protocol

  • Build the library in release mode (mvn -q verify) before running any benchmarks to ensure all classes are compiled and tests have passed.

  • Pin tests to an isolated core and disable extraneous system load; past measurements targeted Intel Core i7-4870HQ @ 2.50 GHz, and new numbers should quote the exact CPU model and frequency.

  • Use a harness that exercises each LongHashFunction via the public API (for example, hashBytes(byte[])) over representative buffer sizes. Warm up the JVM until results stabilise before sampling throughput.

  • Record:

    • Speed (GB/s): sustained throughput while hashing large buffers in steady state.

    • Bootstrap (ns): per-call overhead measured on very small inputs (e.g., empty or sub-cache-line data).

  • Capture both native and non-native endianness where practical; conversion overhead on big-endian systems should be noted if it materially deviates from little-endian results.

Reporting Guidelines

  • Update README.adoc with the latest benchmark table, noting the hardware, JVM version, and any JVM flags used during the run.

  • Flag missing data as TODO (as done for MetroHash and wyHash) rather than leaving stale values in place.

  • When an optimisation changes algorithm behaviour or memory access, include before/after data in the pull request description so reviewers can confirm the impact.

  • Archive raw benchmark logs alongside the pull request or link to them from the decision log to ease future regression analysis.