Benchmarks

These numbers come from specific machines and should not be read as a general advantage. Each one comes with its method, including the cases where Arca loses.

Requirements

Ryzen 9 5900X, Linux. Everything in this section and the next is reproduced by one script, which downloads the Silesia corpus, pins it by hash and prints the machine and tool versions it ran on:

cargo build --releasenix-shell -p hyperfine zip unzip _7zz --run 'bash bench.sh'

hyperfine -N --warmup 20 'arca list tiny.zip' 'unzip -l tiny.zip' '7zz l tiny.zip'hyperfine -N --warmup 5 'arca list list-6000.zip' 'unzip -l list-6000.zip' '7zz l list-6000.zip'hyperfine -N --prepare 'rm -f out.zip' \  'taskset -c 0,1 arca create out.zip corpus -c deflate -j 1' \  'taskset -c 0,1 arca create out.zip corpus -c deflate -j 2'

Compression

The Silesia corpus, 211.9 MB cut into 120 equal files, pinned to two physical cores.

hyperfine -N --prepare 'rm -f out.zip' \  'taskset -c 0,1 arca create out.zip corpus -c zstd' \  'taskset -c 0,1 arca create out.zip corpus -c deflate' \  'taskset -c 0,1 zip -qr -6 out.zip corpus' \  'taskset -c 0,1 7zz a -tzip -mx5 out.zip corpus'

With Zstandard, Arca was 13.2× faster than zip and 18.1× faster than 7-Zip; 7-Zip still wrote the smallest archive, 1.6% below Arca’s. With Deflate, Arca was 5.7× faster than zip and its archive 1.1% smaller.

Deflate against deflate on Windows 11

16 threads, best of 3.

arca create c1.zip src7z a -tzip -mx5 c2.zip src

With large files Arca compressed 5.7× faster at the cost of 3.7% in size; with many small files 7-Zip was somewhat ahead at the same size. The winner depends on the corpus.

Parallel extraction

Windows 11, 16 threads, 287 MB in 16 text files, best of 3, deleting the output before each pass.

arca create big.zip bigarca extract big.zip -o out -j 1arca extract big.zip -o out7z x -o"out" big7z.zip

Against NanaZip, on an archive big enough to hurt

A 3.13 GB .zip holding 1,513 files that unpack to 6.28 GB, all Deflate, with 13 entries over 100 MB. Windows 11, 16 cores, NanaZip 7.0.1832, and a DRAM-less SATA SSD (Kingston A400).

Decompression alone

Both tools read every entry, check every CRC and write nothing, on one core each.

arca test big.zipNanaZipC t big.zip

The whole job

Writing all 6.28 GB, run in the order A B B A so the disk’s decline can’t favour either side.

arca extract big.zip -o outNanaZipC x -y -o"out" big.zip

The disk sustained 327 MB/s before the four runs and 90 MB/s after them. Arca ran on both the freshest and the most worn disk, and its worst run still beat NanaZip’s best.