Kanzi

Kanzi is a modern, modular, expandable and efficient lossless data compressor implemented in Go.

• modern: state-of-the-art algorithms are implemented and multi-core CPUs can take advantage of the built-in multi-tasking.
• modular: entropy codec and a combination of transforms can be provided at runtime to best match the kind of data to compress.
• expandable: clean design with heavy use of interfaces as contracts makes integrating and expanding the code easy. No dependencies.
• efficient: the code is optimized for efficiency (trade-off between compression ratio and speed).

Unlike the most common lossless data compressors, Kanzi uses a variety of different compression algorithms and supports a wider range of compression ratios as a result. Most usual compressors do not take advantage of the many cores and threads available on modern CPUs (what a waste!). Kanzi is concurrent by design and uses threads to compress several blocks in parallel. It is not compatible with standard compression formats.

Kanzi is a lossless data compressor, not an archiver. It uses checksums (optional but recommended) to validate data integrity but does not have a mechanism for data recovery. It also lacks data deduplication across files. However, Kanzi generates a bitstream that is seekable (one or several consecutive blocks can be decompressed without the need for the whole bitstream to be decompressed).

For more details, see Wiki and Q&A

See how to reuse the code here: https://github.com/flanglet/kanzi-go/wiki/Using-and-extending-the-code

There is a C++ implementation available here: https://github.com/flanglet/kanzi-cpp

There is Java implementation available here: https://github.com/flanglet/kanzi

Why Kanzi

There are many excellent, open-source lossless data compressors available already.

If gzip is starting to show its age, zstd and brotli are open-source, standardized and used daily by millions of people. Zstd is incredibly fast and probably the best choice in many cases. There are a few scenarios where Kanzi can be a better choice:

• gzip, lzma, brotli, zstd are all LZ based. It means that they can reach certain compression ratios only. Kanzi also makes use of BWT and CM which can compress beyond what LZ can do.

• These LZ based compressors are well suited for software distribution (one compression / many decompressions) due to their fast decompression (but low compression speed at high compression ratios). There are other scenarios where compression speed is critical: when data is generated before being compressed and consumed (one compression / one decompression) or during backups (many compressions / one decompression).

• Kanzi has built-in customized data transforms (multimedia, utf, text, dna, ...) that can be chosen and combined at compression time to better compress specific kinds of data.

• Kanzi can take advantage of the multiple cores of a modern CPU to improve performance

• Implementing a new transform or entropy codec (to either test an idea or improve compression ratio on specific kinds of data) is simple.

Benchmarks

AWS c5a8xlarge: AMD EPYC 7R32 (32 vCPUs), 64 GB RAM

go 1.21.10

Ubuntu 24.04 LTS

Kanzi version 2.3.0

On this machine, Kanzi uses up to 16 threads (half of CPUs by default).

bzip3 and zpaq use 16 threads. zstd uses 16 threads for compression and 1 for decompression, other compressors are single threaded.

The default block size at level 9 is 32MB, severely limiting the number of threads in use, especially with enwik8, but all tests are performed with default values.

silesia.tar

Download at http://sun.aei.polsl.pl/~sdeor/corpus/silesia.zip

Compressor	Encoding (sec)	Decoding (sec)	Size
Original			211,957,760
s2 -cpu 16	0.494	0.868	86,650,932
Kanzi -l 1	0.504	0.366	80,277,212
Lz4 1.9.5 -4	0.321	0.330	79,912,419
s2 -cpu 16 -slower	1.517	0.868	79,555,929
Zstd 1.5.6 -2 -T16	0.151	0.271	69,556,157
Kanzi -l 2	0.533	0.263	68,195,845
Brotli 1.1.0 -2	1.749	0.761	68,041,629
Gzip 1.12 -9	20.09	1.403	67,652,449
Kanzi -l 3	1.057	0.359	65,613,695
Zstd 1.5.6 -5 -T16	0.356	0.289	63,131,656
Kanzi -l 4	1.125	0.519	61,249,959
Zstd 1.5.5 -9 -T16	0.690	0.278	59,429,335
Brotli 1.1.0 -6	8.388	0.677	58,571,909
Zstd 1.5.6 -13 -T16	3.244	0.272	58,041,112
Brotli 1.1.0 -9	70.07	0.677	56,376,419
Bzip2 1.0.8 -9	16.94	6.734	54,572,500
Kanzi -l 5	2.161	1.043	54,039,773
Zstd 1.5.6 -19 -T16	20.87	0.303	52,889,925
Kanzi -l 6	2.779	2.056	49,567,817
Lzma 5.4.5 -9	95.97	3.172	48,745,354
Kanzi -l 7	3.738	2.888	47,520,629
bzip3 1.3.2.r4-gb2d61e8 -j 16	2.682	3.221	47,237,088
Kanzi -l 8	19.47	20.01	43,167,429
Kanzi -l 9	45.17	45.55	41,497,835
zpaq 7.15 -m5 -t16	213.8	213.8	40,050,429

enwik8

Download at https://mattmahoney.net/dc/enwik8.zip

Compressor	Encoding (sec)	Decoding (sec)	Size
Original			100,000,000
Kanzi -l 1	0.425	0.149	43,746,017
Kanzi -l 2	0.432	0.179	37,816,913
Kanzi -l 3	0.683	0.245	33,865,383
Kanzi -l 4	0.621	0.365	29,597,577
Kanzi -l 5	0.808	0.437	26,528,023
Kanzi -l 6	1.212	0.916	24,076,674
Kanzi -l 7	2.321	2.755	22,817,373
Kanzi -l 8	12.52	12.27	21,181,983
Kanzi -l 9	32.24	32.27	20,035,138

Round-trip scores for LZ

Below is a table showing silesia.tar compressed using different LZ compressors (no entropy) in single-threaded mode.

The efficiency score is computed as such: score(lambda) = compTime + 2 x decompTime + 10^-lambda x compSize

A lower score is better. Best scores are in bold.

Tested on Ubuntu 22.04.4 LTS, i7-7700K CPU @ 4.20GHz, 32 GB RAM

Compressor	Encoding (sec)	Decoding (sec)	Size	Score(5)	Score(6)	Score(7)
Lz4	0.91	0.53	98375894	985.73	100.35	11.81
s2 -cpu 1	0.81	0.40	86646819	868.08	88.25	10.27
Kanzi 2.3 -t lz -j 1	1.30	0.33	83355862	835.52	85.31	10.29
Kanzi 2.3 -t lzx -j 1	1.58	0.30	81485228	817.04	83.67	10.33

References: S2 LZ4 (go)

Build

Using formal releases is recommended (see https://github.com/flanglet/kanzi-go/releases).

go install github.com/flanglet/kanzi-go/v2/app@v2.3.0

Otherwise, to build manually from the latest tag, follow the instructions below:

git clone https://github.com/flanglet/kanzi-go.git

cd kanzi-go/v2/app

go build Kanzi.go BlockCompressor.go BlockDecompressor.go InfoPrinter.go

Credits

Matt Mahoney, Yann Collet, Jan Ondrus, Yuta Mori, Ilya Muravyov, Neal Burns, Fabian Giesen, Jarek Duda, Ilya Grebnov

Disclaimer

Use at your own risk. Always keep a copy of your original files.