Running benchmarks

The benchmark module measures end-to-end performance of the Meta processing pipeline, from CSV input to triplestore upload.

Usage

uv run python -m oc_meta.run.benchmark -c <CONFIG> [options]

Parameters

Parameter	Default	Description
-c, --config	Required	Path to benchmark config YAML
--sizes	None	Generate N synthetic records. Multiple values for scalability analysis
--runs	1	Execute benchmark multiple times for statistical analysis
--seed	42	Random seed for reproducible data
--fresh-data	False	Generate new data for each run
--no-cleanup	False	Skip database reset after benchmark
--update-scenario	False	Test graph diff performance (preload partial, then complete data)
--preload-high-authors	None	Preload BR with N authors before benchmark

Examples

Single run with 100 synthetic records:

uv run python -m oc_meta.run.benchmark -c benchmark_config.yaml --sizes 100

Statistical analysis with 5 runs:

uv run python -m oc_meta.run.benchmark -c benchmark_config.yaml --sizes 100 --runs 5

Scalability analysis across multiple sizes:

uv run python -m oc_meta.run.benchmark -c benchmark_config.yaml --sizes 10 50 100 500 --runs 3

Update scenario (tests graph diff when updating existing entities):

uv run python -m oc_meta.run.benchmark -c benchmark_config.yaml --sizes 100 --update-scenario

High-author stress test (simulates ATLAS paper with 2869 authors):

uv run python -m oc_meta.run.benchmark -c benchmark_config.yaml --preload-high-authors 2869

Output

Reports are saved in oc_meta/run/benchmark/reports/:

• benchmark_<size>.json - raw timing data and statistics

• benchmark_<size>.png - phase breakdown and throughput charts

Metrics collected

• Total duration

• Throughput (records/sec)

• Per-phase timing: curation (collect IDs, clean, merge), RDF creation, storage

• Memory usage per phase

• 95% confidence intervals (with multiple runs)

• Outlier detection