Skip to content

COOL_BENCHMARK.md

Latest commit

 

History

History
51 lines (43 loc) · 2.69 KB

COOL_BENCHMARK.md

File metadata and controls

51 lines (43 loc) · 2.69 KB
Raw

Cooler (.cool/.mcool) benchmark results

polars-bio native cooler scans versus the reference cooler package's chunked-pandas approach — the abdenlab/oxbow#180 baseline that native scanning replaces.

  • Dataset: open2c HFF Micro-C test.mcool (Krietenstein et al. 2021, hg38 chr2 + chr17), resolution 10000 → 24,521,334 pixels. SHA-256 a77252c0… (see results/cool_results.json).
  • Machine: macOS arm64, median of 3 interleaved fresh-process runs.
  • Versions: polars-bio 0.34.0-dev (cool branch), polars 1.40.1, cooler 0.10.4, pandas 3.0.3, numpy 2.4.4.
  • Equivalence: benchmarks/verify_cool_equivalence.py passed before the timed runs — identical row counts, count sums, and coordinate checksums for every workload across both implementations.
Workload cooler (chunked pandas) polars-bio t1 t2 t4 t8
stream_count — count all pixels 1.21 s / 784 MB 0.83 s / 232 MB 0.34 s 0.24 s 0.24 s
collect_all — full joined table 1.94 s / 1467 MB 0.93 s / 1839 MB 0.37 s 0.28 s 0.29 s
region — 20 Mb box, 773,355 rows 0.11 s / 256 MB 0.11 s / 276 MB 0.06 s 0.04 s 0.04 s

Numbers are median wall time / peak RSS.

Takeaways

  • polars-bio is faster than or equal to the cooler baseline serially on every workload (1.5x on count, 2.1x on full materialization, parity on the region fetch) and 2.7-6.9x faster at 4 partitions.
  • The provider reads pixel data through a direct-chunk fast path: chunk file addresses are indexed once through libhdf5, then reads are plain file I/O + zlib-rs inflation + byte unshuffling in Rust — the libhdf5 global lock (which previously capped parallel speedups near 1.4x) is not in the data path at all. Every column is validated against a libhdf5 reference read at index time and falls back to ordinary hdf5 reads on any mismatch.
  • Scaling saturates at 4 partitions on this laptop-class machine (memory bandwidth and the final Polars concat), not at 2 as before.
  • Streaming count keeps a flat ~230-480 MB footprint vs cooler's 784 MB chunked pass; region queries now beat cooler's dedicated matrix().fetch while composing with arbitrary Polars expressions.
  • Benchmarking earlier revisions exposed two polars-bio issues, both fixed and included here: typed optimizer literals silently disabling numeric predicate pushdown on the scan path, and lock-bound HDF5 decoding.

Reproduce with:

./setup.sh                                    # downloads + verifies test.mcool
.venv/bin/python run_cool.py --iterations 3   # verify + benchmark + results JSON