StrataZip is an archiver built for one thing: the smallest file you can get back
exactly. Every block of your data is tried against several coders, the one that wins
on that block is kept, and every byte is checked on the way out.
$ stratazip -l 9 -c prices.csv prices.stz
$ stratazip -d prices.stz prices.csv
Why StrataZip
- It picks the coder per block, by measurement. Text, code and tables go to a
context-mixing model with forty contexts, among them one that knows where each byte
stands in its line and field; 16-bit images to a picture coder that finds the width
and byte order itself; 16-bit audio to a waveform coder; DNA to a two-bit coder that
learns both strands. Nothing is guessed from a file name.
- The smallest archives of the tools we tested. On the Silesia corpus
level 9 is 7.4% under zpaq's strongest mode and level 10 12.7% under, the smallest of every
tool on all twelve files. On market data, 26.7% under zpaq; on sensor logs, 8.9%.
- It sees inside ZIP, JAR, DOCX, gzip and PDF files. Their deflate streams are
unwrapped and compressed as what they hold, wherever the bundled zlib can rebuild them
bit for bit on extraction; the rest are kept as they are.
- Your data, verified. A CRC-32 on every block and a SHA-256 on every file in a
tree. A damaged archive is refused, never silently mis-decoded, and optional
Reed-Solomon recovery records (
-r 5 for 5%) can repair it.
- Directory trees, incrementally. Archive a whole tree with permissions, times and
links; identical files are stored once. Give it last night's archive (
--from) and
files that have not changed are not even read again.
- One file out of a big archive (
--get) decodes only the blocks that hold it, and
--info shows what an archive is made of without decoding it (--json for tools).
- Safe writes. Output goes to a temporary name and is renamed into place only when
it is complete.
Benchmarks
One thread per tool, every tool on the same machine, and every result decompressed and
compared byte for byte (FLAC's on their audio samples). The Silesia numbers are reproducible with
bench/run.sh on public data. docs/BENCHMARKS.md
has the full tables, the machine, and the caveats.
The research compressors cmix and paq8px publish smaller results on these corpora; they
were not run here.
Structured data, genomes and text
| Data | Size | xz -9e | zstd -19 | zpaq -m5 | StrataZip -l 9 | StrataZip -l 10 |
|---|
| Bitcoin price history, 2017 to 2025, CSV | 2,575,631 | 153,772 | 200,502 | 113,063 | 82,925 | 82,818 |
| Bike&Safe sensor logs (the dataset as a tar) | 47,544,320 | 15,109,452 | 15,663,947 | 11,540,076 | 10,516,775 | 10,222,204 |
| E. coli genome (Canterbury large corpus) | 4,638,690 | 1,186,180 | 1,148,133 | 1,114,234 | 1,099,133 | 1,096,355 |
| enwik8, 100 MB of Wikipedia | 100,000,000 | 24,831,648 | 26,954,633 | 19,625,015 | 19,953,846 | 18,700,345 |
On enwik8 zpaq -m5 beats level 9 by 1.6%; level 10, which sees the file in two blocks
rather than twenty-four, is 4.7% under it.
Audio: smaller than FLAC
| Recording (16-bit PCM) | Size | FLAC -8 | StrataZip | |
|---|
| Field recordings, ten ESC-50 clips | 4,410,078 | 1,880,800 | 1,784,415 | −5.1% |
| Pop song, 30 s (decoded from MP3) | 5,292,162 | 3,613,378 | 3,400,260 | −5.9% |
| Pop song, 30 s (decoded from MP3) | 5,292,184 | 3,095,882 | 2,874,306 | −7.2% |
The audio coder runs at every level; these are level 4. FLAC decodes much faster, so
StrataZip is for keeping audio, not for playing it.
The Silesia corpus: general data
The standard corpus of mixed real data (12 files, 212 MB).
| Tool | Size (MB) | Ratio | Compress (MB/s) | Decompress (MB/s) |
|---|
| gzip -9 | 67.63 | 3.134 | 9.45 | 133.11 |
| StrataZip -l 4 | 56.18 | 3.773 | 2.84 | 60.88 |
| bzip2 -9 | 54.51 | 3.888 | 11.38 | 26.77 |
| zstd -19 | 53.02 | 3.997 | 1.97 | 483.12 |
| StrataZip -l 6 | 52.73 | 4.019 | 0.75 | 62.31 |
| zstd --ultra -22 --long=27 | 52.52 | 4.035 | 1.55 | 431.27 |
| brotli -q 11 -w 24 | 49.56 | 4.276 | 0.43 | 251.15 |
| 7z -mx=9 | 48.76 | 4.347 | 1.91 | 62.29 |
| xz -9e | 48.46 | 4.374 | 1.45 | 61.21 |
| bzip3 | 46.93 | 4.516 | 9.17 | 8.99 |
| zpaq -m5 | 39.11 | 5.419 | 0.39 | 0.38 |
| StrataZip -l 9 | 36.23 | 5.850 | 0.08 | 0.10 |
| StrataZip -l 10 | 34.16 | 6.205 |
Levels 9 and 10 are the smallest, and the slowest: zpaq -m5 compresses at 0.39 MB/s
here, StrataZip's level 9 at 0.08, on one thread. Two of the files (mr, x-ray) are
16-bit medical images, which StrataZip codes with its picture coder at every level, so
even level 4 makes them smaller than any other tool does. Per-file results are
in docs/BENCHMARKS.md.
Choosing a level
| Level | Use it for | Speed |
|---|
-l 1 to -l 4 (default 4) | everyday archives, fast both ways | level 4 decompresses Silesia at about 61 MB/s on one thread |
-l 5, -l 6 | smaller, still fast to open: the parse is optimised, priced in bits | slow to compress, fast to decompress |
-l 7 to -l 9 | the smallest files at a practical pace: context mixing | about as slow to decompress as to compress |
-l 10, -e | the last bits: 64 MiB blocks, tables sized to the free memory | slower still; needs the same memory to decompress |
Levels 7 to 10 are symmetric, like zpaq and cmix: opening the archive takes about as long
as making it. Use them for archives you keep, not for files you open every day. Data that
is as dense as noise (archives, encrypted data) is stored as it is at levels 1 to 9;
-F codes it anyway.
以上為 README 的開頭部分。其餘內容請閱讀原文。 檢視完整 README ↗
