Decompression comparison — 2026-11-09

Benchmark index · All level pages · Published CSV

The published 384 rows cover two decoders on eight identical C-written frames at each source level, recorded as long-20261109 (mzstd 0.2.12, the working tree on 5ecb52f, identified by the source hashes, zstd 1.5.7). mzstd’s median speed relative to C is 1.517× across all 192 dataset/level pairs; mzstd is faster in 151 of them, and in 144 of those its 95% mean timing interval lies wholly below C’s. Empty and tiny inputs have the same weight as large inputs. Per-workload results include cases where mzstd is slower.

The host was loaded by other tenants throughout the run. Sampling was lengthened to 50 samples of 6 s after a 3 s warmup, which keeps the medians stable but leaves a median 95% interval of about 5% per case, against 0.5% in the 2026-10-30 run. Read single cells as ±10%.

Each bar is the median of eight C mean time / decoder mean time ratios, in the same process. The decoder overview lists every source level. Higher is faster; medians do not establish statistical significance.

Empty and tiny frames decode tens of times faster because mzstd needs no preallocated decoding context. Compressible text still trails: all 41 slower pairs are Alice (22 of 24 levels, all but 1 and 20) and cyclic Alice (19 of 24, all but 5, 11, 13, 15 and 16). At level 1 Alice decodes at 1,539 MiB/s against C’s 1,505 MiB/s; cyclic Alice at 8,939 MiB/s against 10,016. Most of that gap is the Huffman literal loop (decoder fast paths). Cyclic Alice’s median fell from 0.965× to 0.937×, well inside the 0.84–1.03× spread its cases show across levels in this run; this run cannot tell that change from the host’s noise.

This suite times only decompress; decompress_to_slice and the streaming sessions are measured by the root decompress benchmark, and lzbench measures whole-archive throughput on silesia.tar.

Measurement contract

Cold one-shot APIs include construction, allocation, decoding and disposal: C reads the frame’s content size (ZSTD_getFrameContentSize) and decodes with ZSTD_decompress into an uninitialized buffer of exactly that size; mzstd runs Decompressor::default().decompress, which returns its own vector. Throughput counts restored bytes. The shared frame sizes belong to the C encoder, not to a decoder. Every case must restore the original bytes through both decoders before timing.

The corpus is empty input, tiny text, Alice, cyclic Alice at 1 MiB, structured binary at 64 KiB, random bytes at 64 KiB and 1 MiB, and repeated a at 1 MiB. The host is an AMD Ryzen 9 7950X KVM guest with two vCPUs, its host shared with other tenants; the decoder suite ran on vCPU 1 after the encoder suite, with nothing else in the guest. Sampling: 50 samples, 3 s warmup and 6 s requested measurement per case (Criterion extends slow cases to collect every sample). Rust 1.99.1, runtime AVX-512 dispatch with the block loops at the AVX2 level; zstd built by zstd-ffi with GCC 15.2 -O3, without its Huffman assembly.

Provenance and limits

The environment record hashes the measured sources and binaries, the size manifest lists every frame, and the raw Criterion samples and estimates are archived. This is one sweep on one KVM guest; frequency, thermal state and scheduling are not controlled. Other tenants loaded the host throughout: a fixed C decode varied by 8-43% between its 10th and 90th percentile wall times during the day, the median case’s 95% interval is 5.2% (90th percentile 9.1%), and C’s own timings moved by a median 8% against the 2026-10-30 run.

Reproduce from the repository root with a fresh baseline name:

cargo bench --manifest-path benchmarks/comparison/Cargo.toml --bench decoders --locked --no-run
cd benchmarks/comparison
taskset -c 1 target/release/deps/decoders-<hash> --bench --save-baseline my-decoders --sample-size 50 --warm-up-time 3 --measurement-time 6 --noplot
cd ../..
python3 benchmarks/comparison/report.py --decoding --baseline my-decoders --csv /tmp/my-decoders.csv
cp benchmarks/comparison/target/criterion-decoders/sizes.csv /tmp/my-decoders-sizes.csv

Use the benchmark guide to regenerate the published pages from the recorded CSV and environment metadata.