LZAV - Fast Data Compression Algorithm (in C/C++)

Introduction

LZAV is a fast general-purpose in-memory data compression algorithm based on now-classic LZ77 lossless data compression method. LZAV holds a good position on the Pareto landscape of factors, among many similar in-memory (non-streaming) compression algorithms.

LZAV algorithm's code is portable, cross-platform, scalar, header-only, inlineable C (C++ compatible). It supports big- and little-endian platforms, and any memory alignment models. The algorithm is efficient on both 32- and 64-bit platforms. Incompressible data almost does not expand.

LZAV does not sacrifice internal out-of-bounds (OOB) checks for decompression speed. This means that LZAV can be used in strict conditions where OOB memory writes (and especially reads) that lead to a trap, are unacceptable (e.g., real-time, system, server software). LZAV can be used safely (causing no crashing nor UB) even when decompressing malformed or damaged compressed data. Which means that LZAV does not require calculation of a checksum (or hash) of the compressed data. Only a checksum of the uncompressed data may be required, depending on application's guarantees.

The internal functions available in the lzav.h file allow you to easily implement, and experiment with, your own compression algorithms. LZAV stream format and decompressor have a potential of high decompression speeds and compression ratios, which depends on the way data is compressed.

Usage Information

To compress data:

#include "lzav.h"

int max_len = lzav_compress_bound( src_len );
void* comp_buf = malloc( max_len );
int comp_len = lzav_compress_default( src_buf, comp_buf, src_len, max_len );

if( comp_len == 0 && src_len != 0 )
{
    // Error handling.
}

To decompress data:

#include "lzav.h"

void* decomp_buf = malloc( src_len );
int l = lzav_decompress( comp_buf, decomp_buf, comp_len, src_len );

if( l < 0 )
{
    // Error handling.
}

To compress data with a higher ratio, for non-time-critical uses (e.g., compression of application's static assets):

#include "lzav.h"

int max_len = lzav_compress_bound_hi( src_len ); // Note another bound function!
void* comp_buf = malloc( max_len );
int comp_len = lzav_compress_hi( src_buf, comp_buf, src_len, max_len );

if( comp_len == 0 && src_len != 0 )
{
    // Error handling.
}

LZAV algorithm and its source code (which is ISO C99) were quality-tested with: Clang, GCC, MSVC, Intel C++ compilers; on x86, x86-64 (Intel, AMD), AArch64 (Apple Silicon) architectures; Windows 10, AlmaLinux 9.3, macOS 14.6.1.

Comparisons

The tables below present performance ballpark numbers of LZAV algorithm (based on Silesia dataset).

While LZ4 there seems to be compressing faster, LZAV comparably provides 14.2% memory storage cost savings. This is a significant benefit in database and file system use cases since compression is only about 35% slower while CPUs rarely run at their maximum capacity anyway, and disk I/O times are reduced due to a better compression. In general, LZAV holds a very strong position in this class of data compression algorithms, if one considers all factors: compression and decompression speeds, compression ratio, and not less important - code maintainability: LZAV is maximally portable and has a rather small independent codebase.

Performance of LZAV is not limited to the presented ballpark numbers. Depending on the data being compressed, LZAV can achieve 800 MB/s compression and 5000 MB/s decompression speeds. Incompressible data decompresses at 10000 MB/s rate, which is not far from the "memcpy". There are cases like the enwik9 dataset where LZAV provides 21.2% higher memory storage savings compared to LZ4. However, on small data (below 50 KB), compression ratio difference between LZAV and LZ4 diminishes, and LZ4 may have some advantage.

LZAV algorithm's geomean performance on a variety of datasets is 530 +/- 150 MB/s compression and 3800 +/- 1300 MB/s decompression speeds, on 4+ GHz 64-bit processors released since 2019. Note that the algorithm exhibits adaptive qualities, and its actual performance depends on the data being compressed. LZAV may show an exceptional performance on your specific data, including, but not limited to: sparse databases, log files, HTML/XML files.

It is also worth noting that compression methods like LZAV and LZ4 usually have an advantage over dictionary- and entropy-based coding in that hash-table-based compression has a small operation and memory overhead while the classic LZ77 decompression has no overhead at all - this is especially relevant for smaller data.

For a more comprehensive in-memory compression algorithms benchmark you may visit lzbench.

Apple clang 15.0.0 arm64, macOS 14.6.1, Apple M1, 3.5 GHz

Silesia compression corpus

Compressor	Compression	Decompression	Ratio %
LZAV 4.5	580 MB/s	3800 MB/s	40.81
LZ4 1.9.4	700 MB/s	4570 MB/s	47.60
Snappy 1.1.10	495 MB/s	3230 MB/s	48.22
LZF 3.6	395 MB/s	800 MB/s	48.15
LZAV 4.5 HI	122 MB/s	3670 MB/s	35.67
LZ4HC 1.9.4 -9	40 MB/s	4360 MB/s	36.75

LLVM clang 16.0.6 x86-64, AlmaLinux 9.3, Xeon E-2386G (RocketLake), 5.1 GHz

Silesia compression corpus

Compressor	Compression	Decompression	Ratio %
LZAV 4.5	565 MB/s	3400 MB/s	40.81
LZ4 1.9.4	845 MB/s	4960 MB/s	47.60
Snappy 1.1.10	690 MB/s	3360 MB/s	48.22
LZF 3.6	455 MB/s	1020 MB/s	48.15
LZAV 4.5 HI	105 MB/s	3270 MB/s	35.67
LZ4HC 1.9.4 -9	43 MB/s	4890 MB/s	36.75

LLVM clang-cl 16.0.4 x86-64, Windows 10, Ryzen 3700X (Zen2), 4.2 GHz

Silesia compression corpus

Compressor	Compression	Decompression	Ratio %
LZAV 4.5	505 MB/s	3100 MB/s	40.81
LZ4 1.9.4	680 MB/s	4300 MB/s	47.60
Snappy 1.1.10	425 MB/s	2430 MB/s	48.22
LZF 3.6	320 MB/s	700 MB/s	48.15
LZAV 4.5 HI	106 MB/s	3020 MB/s	35.67
LZ4HC 1.9.4 -9	36 MB/s	4100 MB/s	36.75

P.S. Popular Zstd's benchmark was not included here, because it is not a pure LZ77, much harder to integrate, and has a much larger code size - a different league, close to zlib. Here are author's Zstd measurements with TurboBench, on Ryzen 3700X, on Silesia dataset:

Compressor	Compression	Decompression	Ratio %
zstd 1.5.5 -1	460 MB/s	1870 MB/s	41.0
zstd 1.5.5 1	436 MB/s	1400 MB/s	34.6

Notes

1. LZAV API is not equivalent to LZ4 nor Snappy API. For example, the "dstl" parameter in the decompressor should specify the original uncompressed length, which should have been previously stored in some way, independent of LZAV.

2. Run-time memory sanitizers like Valgrind and Dr.Memory may generate the "uninitialized read" warning in decompressor's block type 1 handler. This is an expected behavior, and not a bug - this happens due to SIMD optimizations that read bytes from the output buffer (within its valid range) which were not yet initialized.

3. Compared to Clang, other compilers systematically produce 5% slower LZAV code. Compiler architecture tuning (other than generic x86-64) may produce varying, including counter-productive, results.

4. From a technical point of view, peak decompression speeds of LZAV have an implicit limitation arising from its more complex stream format, compared to LZ4: LZAV decompression requires more code branching. Another limiting factor is a rather big 8 MiB LZ77 window which is not CPU cache-friendly. On the other hand, without these features it would not be possible to achieve competitive compression ratios while having fast compression speeds.

5. LZAV supports compression of continuous data blocks of up to 2 GB. Larger data should be compressed in chunks of at least 32 MB. Using smaller chunks may reduce the achieved compression ratio.

Thanks

• Paul Dreik, for finding memcpy UB in the decompressor.