Benchmark results

Two kinds of runs, both on an AMD Ryzen 9 7950X (Linux KVM guest with two vCPUs, its host shared with other tenants): the cold library comparison of mzstd 0.2.12 against zstd 1.5.7, recorded on 2026-11-09, and lzbench on silesia.tar. Every report identifies its measured sources, sampling settings and limitations. Other tenants loaded the host during both: the medians are stable, single cells are good to about ±10%.

Compression

Each speed bar is the median of eight per-dataset ratios to C, timed in the same process. Higher speed and lower size are better. Every mzstd frame is byte-identical to zstd’s, so the size ratio is 1× at every level. The median over all 192 dataset/level pairs is 1.248×, with lower mean latency in 174 pairs (2026-10-30: 1.256×, 179 pairs). The per-dataset medians agree with that run to within a few percent: Alice 1.008×, cyclic Alice 1.088×, structured binary 1.241×, repeated bytes 2.437×. The 18 slower pairs are Alice at ten levels, cyclic Alice at four and random data at four, between 0.84× and 0.998×. Report · exact results and confidence bounds.

Decompression

Both decoders restore the same C-written frames; level is the source encoder’s setting. The median over all 192 pairs is 1.517× (2026-10-30: 1.502×), faster in 151 pairs: tens of times faster on empty and tiny frames, 3.2× on random 64 KiB, 1.53× on repeated bytes and 1.45× on structured binary, but 0.948× on Alice and 0.937× on cyclic Alice, the datasets of all 41 slower pairs. Report · exact results.

lzbench

lzbench master a6fa8aa (2.4.1), gcc 15.2 with -march=native, silesia.tar as one frame on one vCPU, the fastest of three rounds: mzstd 0.2.12 writes byte-identical output to zstd 1.5.7 at every level, compresses 6.6-17.5% faster at levels 15, 18 and 22, at zstd’s speed at level 3 and 2.4-6.6% slower at the other levels, and decompresses 4.7-7.2% slower at every level. Against 0.2.11 in the same rounds, relative to zstd, decompression gained 0.8-5 points at levels 1 to 22 and level 3 compression went from 0.973 to 1.011. Compression at level 2 and above moves by 3-21% between rounds on this host. Command, environment, raw output of every round.

How to interpret the results

The cold suites time construction, allocation, the call and disposal. Empty and tiny inputs have equal weight with large inputs; medians describe this corpus, not a production workload. C decoding knows the output capacity from the frame header. The lowest mean is not necessarily a statistically significant lead, and on this run’s loaded host a single cell’s ratio can move by 10% between sweeps.

The datasets, generators and seeds are mbrotli’s, so the two projects’ comparisons describe one corpus.

Reproduce

Use benchmarking and profiling for validation, fresh timings, the main branch’s series, the Silesia sweep and the report-generation commands.