Performance

The numbers below are the benchmark results committed in benchmarks/. They are one run on one machine, and the files do not record which machine, so read them as orders of magnitude and relative costs, not as a promise about your hardware. Pull requests are not judged against them (see Regression gate).

Rust core (criterion)

benchmarks/latest_benchmarks.json: 7 benchmarks, collected on 2026-02-07 at commit 8135e47. Times are criterion’s mean per iteration with its 95% confidence interval.

Benchmark	What it measures	Mean (ms)	95% CI (ms)	Source
structural_breaks_get_sadf_sm_power	structural_breaks::get_sadf, sm_power model with constant, on 320 log prices, minimum window 20, 5 fixed lags	602.0	600.0–604.0	perf_hotspots.rs
synthetic_ticker_seq_bootstrap_2k_600	sampling::seq_bootstrap: 240 draws from an indicator matrix of 600 labels over 2,000 bars	383.9	360.3–409.8	synthetic_ticker_pipeline.rs
synthetic_ticker_pipeline_end_to_end	EWMA, VaR/ES/CDaR, a 3×3 covariance and portfolio variance, and a 200-draw sequential bootstrap (360 labels over 1,200 bars) on 20,000 prices, in one pass	61.72	60.90–62.48	synthetic_ticker_pipeline.rs
bet_sizing_bet_size_reserve_fit	bet_sizing::bet_size_reserve_full on 500 synthetic events: fits the EF3M mixture (100 runs) and sizes every bet	18.75	17.82–19.81	perf_hotspots.rs
synthetic_ticker_risk_metrics_var_es_cdar	RiskMetrics value at risk, expected shortfall and conditional drawdown at risk (5%) on 119,999 returns	8.01	7.50–8.60	synthetic_ticker_pipeline.rs
synthetic_ticker_ewma_100k	util::fast_ewma::ewma, window 64, on 100,000 prices	4.72	4.69–4.75	synthetic_ticker_pipeline.rs
bet_sizing_bet_size_reserve_reuse_fit	bet_sizing::bet_size_reserve on the same 500 events, reusing already-fitted mixture parameters	0.278	0.274–0.284	perf_hotspots.rs

benchmarks/baseline_benchmarks.json holds the same numbers: the baseline was last refreshed from this run.

Benchmarks that exist in crates/openquant/benches/ but are not in benchmarks/benchmark_manifest.json, so are neither committed here nor gated: hpc_parallel/serial_linear (hpc_parallel_scaling.rs), hpc_parallel/threaded_linear (hpc_parallel_scaling.rs), hpc_parallel/threaded_nested (hpc_parallel_scaling.rs), streaming_hpc/scaling/* (streaming_hpc_scaling.rs). Run them with cargo bench -p openquant --bench <file stem>.

Python data layer

benchmarks/data_processing/latest.json, written by python/benchmarks/benchmark_data_processing.py: 4 symbols × 100,000 rows (400,000 rows), 5 iterations per function. The file records neither a date nor a commit. just py-bench-data regenerates it (with 200,000 rows per symbol, not the size committed here).

Function	Rows	Mean (ms)	p95 (ms)	Rows/s	MB/s
clean_ohlcv	400,000	431.8	430.2	926,317	48.2
data_quality_report	400,000	202.2	205.0	1,977,855	102.8
align_calendar	400,000	487.7	485.1	820,180	49.2
load_ohlcv_csv	400,000	234.6	235.6	1,704,667	120.3
load_ohlcv_parquet	400,000	458.7	468.5	872,087	4.9

There is no committed benchmarks/data_processing/baseline.json, so just py-bench-data-compare has nothing to compare against until one is made, and no CI job runs these benchmarks.

Regression gate

.github/workflows/benchmark-regression.yml runs on pull requests that touch crates/openquant/, the workspace manifests, the toolchain pin or the benchmark scripts. It builds the pull request’s merge base and its head, runs the perf_hotspots and synthetic_ticker_pipeline benches on both on the same runner, and fails if a benchmark’s mean slowed down by more than 35% (scripts/check_bench_thresholds.py), or by more than its entry in benchmarks/threshold_overrides.json. A benchmark that stops producing a result fails the gate too. Comparing base and head back to back removes the machine-to-machine variation that made a gate against the committed baseline fail on unrelated changes. Because the workflow is path-filtered, it is not suitable as a required status check.

Benchmark	Allowed slowdown
structural_breaks_get_sadf_sm_power	35%
bet_sizing_bet_size_reserve_fit	35%
bet_sizing_bet_size_reserve_reuse_fit	55%
synthetic_ticker_ewma_100k	35%
synthetic_ticker_risk_metrics_var_es_cdar	35%
synthetic_ticker_seq_bootstrap_2k_600	35%
synthetic_ticker_pipeline_end_to_end	50%

Reproduce

just bench-all      # cargo bench -p openquant --bench perf_hotspots --bench synthetic_ticker_pipeline
just bench-collect  # target/criterion -> benchmarks/latest_benchmarks.json (manifest only)
just bench-check    # latest vs the committed baseline, with the thresholds CI uses
just py-bench-data  # Python data-layer benchmarks -> benchmarks/data_processing/latest.json
python3 scripts/docs/generate_site_pages.py --write   # re-render this page

Refresh benchmarks/baseline_benchmarks.json only in a pull request that says why; a slowdown and a new baseline in the same change hides the slowdown.