Zarr Benchmarks

This repository contains benchmarks of Zarr V3 implementations.

Note

Contributions are welcomed for additional benchmarks, more implementations, or otherwise cleaning up this repository.

Also consider restarting development of the official zarr benchmark repository: https://github.com/zarr-developers/zarr-benchmark

Implementations Benchmarked

• zarrs/zarrs via zarrs/zarrs_tools
  • Read executable: zarrs_benchmark_read_sync
  • Round trip executable: zarrs_reencode
• google/tensorstore
• zarr-developers/zarr-python
  • With and without the ZarrsCodecPipeline from zarrs/zarrs-python
  • With and without dask

Implementation versions are listed in the benchmark charts.

Warning

Python benchmarks are subject to the overheads of Python and may not be using an optimal API/parameters.

Please open a PR if you can improve these benchmarks.

make Targets

• pydeps: install python dependencies (recommended to activate a venv first)
• zarrs_tools: install zarrs_tools (set CARGO_HOME to override the installation dir)
• generate_data: generate benchmark data
• benchmark_read_all: run read all benchmark
• benchmark_read_chunks: run chunk-by-chunk benchmark
• benchmark_roundtrip: run roundtrip benchmark
• benchmark_all: run all benchmarks

Benchmark Data

All datasets are $1024x2048x2048$ uint16 arrays.

Name	Chunk / Shard Shape	Inner Chunk Shape	Compression	Size
Uncompressed	$256^3$		None	8.00 GB
Compressed	$256^3$		blosclz 9 + bitshuffling	659 MB
Compressed + Sharded	$256^3$	$32^3$	blosclz 9 + bitshuffling	1.20 GB

Benchmark System

• AMD Ryzen 5900X
• 64GB DDR4 3600MHz (16-19-19-39)
• 2TB Samsung 990 Pro
• Arch Linux (in Windows 11 WSL2, swap disabled, 32GB available memory)

Round Trip Benchmark

This benchmark measures time and peak memory usage to "round trip" a dataset (potentially chunk-by-chunk).

• The disk cache is cleared between each measurement
• These are best of 3 measurements

Table of raw measurements (benchmarks_roundtrip.md)

Standalone

Dask

Read Chunk-By-Chunk Benchmark

This benchmark measures the the minimum time and peak memory usage to read a dataset chunk-by-chunk into memory.

• The disk cache is cleared between each measurement
• These are best of 1 measurements

Table of raw measurements (benchmarks_read_chunks.md)

Standalone

Note

zarr-python benchmarks with sharding are not visible in this plot

Dask

Read All Benchmark

This benchmark measures the minimum time and and peak memory usage to read an entire dataset into memory.

• The disk cache is cleared between each measurement
• These are best of 3 measurements

Table of raw measurements (benchmarks_read_all.md)

Standalone

Dask