Decompression comparison — 2026-10-30

Benchmark index · Quality pages · CSV

The complete 384-case run dec-2026-10-30-amd compares Google C Brotli 1.2.0, the local mbrotli 0.5.19 checkout, Rust brotli 9.0.0 and Burli 0.3.3. All four restore identical C-produced streams at source qualities q0–q11, generic mode and window 22. SIMD Brotli re-exports the Rust Brotli decoder project and is omitted.

Across 96 corpus/quality pairs, mbrotli’s median speed is 2.588× C and 1.006× Burli. Its mean latency is lower than C in 80/96 cases and lower than Burli in 53/96. Its upper 95% mean-time bound is below the peer’s lower bound in 80/96 C comparisons and 50/96 Burli comparisons. These are sweep counts, not universal leads.

Final measurements

Library columns are medians of eight per-dataset C-time/library-time ratios. The last column is the median of direct Burli-time/mbrotli-time ratios; it does not divide C-relative medians. Overall medians use all 96 direct ratios. Empty and tiny inputs have the same weight as larger inputs. Higher ratios are faster.

Measurement contract

One process ran at a time on logical CPU 1 of an AMD Ryzen 9 7950X Linux VDS with two KVM vCPUs. Encoding and decoding sweeps ran serially; no builds, tests, fuzzers or other benchmarks ran concurrently. Rust 1.99.1 and GCC 15.2 used ordinary release defaults, with runtime backend selection. Criterion used 20 samples, 1 s warmup and 3 s requested measurement per case.

The eight corpora are empty, tiny text, Alice, cyclic Alice at 1 MiB, structured 64 KiB binary, random bytes at 64 KiB/1 MiB and repeated a at 1 MiB. C prepares 96 streams once; all 384 decoder cases must restore the original bytes before timing. Quality belongs to that source encoder, not to a decoder. Throughput counts restored bytes, and compressed sizes are properties of the shared inputs.

Cold iterations include construction, allocation, decoding and disposal. C gets the known output capacity and a freshly initialized buffer; Rust Vec APIs grow their output. Rust Brotli includes native 4 KiB I/O adaptation. Burli decodes every source quality. No decoder retains state between iterations.

Provenance and limits

The environment record identifies the source hashes, lockfile, measured executable, commands and CSV. The matching size manifest records the shared compressed/restored lengths. This is one sweep on a virtual host with fixed implementation order. Affinity does not control clocks, thermals, allocator effects or host scheduling; sampling intervals do not cover every environmental bias. The measured cold APIs do not cover retained state, streaming, dictionaries or streams from other producers. mbrotli remains 3–8% slower than C on Alice at source qualities other than q10 (14% at q8) and 11–24% slower on tiny text at q5–q9, where fixed setup cost dominates a 44-byte output.

Reproduce

cargo bench --manifest-path benchmarks/comparison/Cargo.toml --bench decoders --locked --no-run
taskset -c 1 cargo bench --manifest-path benchmarks/comparison/Cargo.toml --bench decoders --locked -- --sample-size 20 --warm-up-time 1 --measurement-time 3 --noplot --save-baseline dec-2026-10-30-amd
python3 benchmarks/comparison/report.py --decoding --baseline dec-2026-10-30-amd --csv docs/benchmarks/decoder-comparison.csv

Use a fresh baseline for a new run and archive its size manifest immediately. The measured run used the frozen executable recorded in the environment file. Benchmarking covers chart generation and separate API measurements.