Awesome
K4os.Compression.LZ4
Name | Nuget | Description |
---|---|---|
K4os.Compression.LZ4 | Block compression only | |
K4os.Compression.LZ4.Streams | Stream compression | |
K4os.Compression.LZ4.Legacy | Legacy compatibility |
LZ4
LZ4 is lossless compression algorithm, sacrificing compression ratio for compression/decompression speed. Its compression speed is ~400 MB/s per core while decompression speed reaches ~2 GB/s, not far from RAM speed limits.
This library brings LZ4 to .NET Standard compatible platforms: .NET Core, .NET Framework, Mono, Xamarin, and UWP. Well... theoretically... kind of. Currently, it targets .NET Framework 4.6.2+, .NET Standard 2.0+ and .NET 5.0+.
As it is .NET Standard 2.0+ so all this platforms should be supported although I did not test it on all of them.
LZ4 has been written by Yann Collet and original C sources can be found here
Build
./build.ps1
NOTE: technically, it could be built on Linux as well, but setup process downloads and uses some Windows tools,
like 7z.exe
and lz4.exe
. It could be adapted, but wasn't. Feel free to send PR.
Changes
Change log can be found here.
Support
Maintaining this library is outside of my daily job completely. Company I work for is not even using it, so I do this completely in my own free time.
So, if you think my work is worth something, you could support me by funding my daily caffeine dose:
(or just use PayPal)
What is 'Fast compression algorithm'?
While compression algorithms you use day-to-day to archive your data work around the speed of 10MB/s giving you quite decent compression ratios, 'fast algorithms' are designed to work 'faster than your hard drive' sacrificing compression ratio.
One of the most famous fast compression algorithms in Google's own Snappy which is advertised as 250MB/s compression, 500MB/s decompression on i7 in 64-bit mode. Fast compression algorithms help reduce network traffic / hard drive load compressing data on the fly with no noticeable latency.
I just tried to compress some sample data (Silesia Corpus) receiving:
- zlib (7zip) - 7.5M/s compression, 110MB/s decompression, 44% compression ratio
- lzma (7zip) - 1.5MB/s compression, 50MB/s decompression, 37% compression ratio
- lz4 - 280MB/s compression, 520MB/s decompression, 57% compression ratio
Note: Values above are for illustration only. they are affected by HDD read/write speed (in fact LZ4 decompression in much faster). The 'real' tests are taking HDD speed out of equation. For detailed performance tests see [Performance Testing] and [Comparison to other algorithms].
Other 'Fast compression algorithms'
There are multiple fast compression algorithms, to name a few: LZO, QuickLZ, LZF, Snappy, FastLZ. You can find comparison of them on LZ4 webpage or here
Usage
This LZ4 library can be used in two distinctive ways: to compress streams and blocks.
Use as blocks
Compression levels
enum LZ4Level
{
L00_FAST,
L03_HC, L04_HC, L05_HC, L06_HC, L07_HC, L08_HC, L09_HC,
L10_OPT, L11_OPT, L12_MAX,
}
There are multiple compression levels. LZ4 comes in 3 (4?) flavors of compression algorithms. You can notice suffixes
of those levels: FAST
, HC
, OPT
and MAX
(while MAX
is just OPT
with "ultra" settings). Please note that
compression speed drops rapidly when not using FAST
mode, while decompression speed stays the same (actually,
it is usually faster for high compression levels as there is less data to process).
Utility
static class LZ4Codec
{
static int MaximumOutputSize(int length);
}
Returns maximum size of of a block after compression. Of course, most of the time compressed data will take less space than source data, although in case of incompressible (for example: already compressed) data it may take more.
Example:
var source = new byte[1000];
var target = new byte[LZ4Codec.MaximumOutputSize(source.Length)];
//...
Compression
Block can be compressed using Encode(...)
method family. They are relatively low level functions as it is your job
to allocate all memory.
static class LZ4Codec
{
static int Encode(
byte* source, int sourceLength,
byte* target, int targetLength,
LZ4Level level = LZ4Level.L00_FAST);
static int Encode(
ReadOnlySpan<byte> source, Span<byte> target,
LZ4Level level = LZ4Level.L00_FAST);
static int Encode(
byte[] source, int sourceOffset, int sourceLength,
byte[] target, int targetOffset, int targetLength,
LZ4Level level = LZ4Level.L00_FAST);
}
All of them compress source
buffer into target
buffer and return number of bytes actually used after compression.
If this value is negative it means that error has occurred and compression failed. In most cases mean that target
buffer is too small.
Please note, it might be tempting to use target
buffer the same size (or even one byte smaller) then source
buffer,
and use copy as a fallback. This will work just fine, yet compression into buffer that is smaller than MaximumOutputSize(source.Length)
is a little bit slower.
Example:
var source = new byte[1000];
var target = new byte[LZ4Codec.MaximumOutputSize(source.Length)];
var encodedLength = LZ4Codec.Encode(
source, 0, source.Length,
target, 0, target.Length);
Decompression
Previously compressed block can be decompressed with Decode(...)
functions.
static class LZ4Codec
{
static int Decode(
byte* source, int sourceLength,
byte* target, int targetLength);
static int Decode(
ReadOnlySpan<byte> source, Span<byte> target);
static int Decode(
byte[] source, int sourceOffset, int sourceLength,
byte[] target, int targetOffset, int targetLength);
}
You have to know upfront how much memory you need to decompress, as there is almost no way to guess it. I did not investigate theoretical maximum compression ratio, yet all-zero buffer gets compressed 245 times, therefore when decompressing output buffer would need to be 245 times bigger than input buffer. Yet, encoding itself does not store that information anywhere therefore it is your job.
var source = new byte[1000];
var target = new byte[knownOutputLength]; // or source.Length * 255 to be safe
var decoded = LZ4Codec.Decode(
source, 0, source.Length,
target, 0, target.Length);
NOTE: If I told you that decompression needs potentially 100 times more memory than original data you would think
this is insane. And it is not 100 times, it is 255 times more, so it actually is insane. Please don't do it.
This was for demonstration only. What you need is a way to store original size somehow (I'm not opinionated, do
whatever you think is right) or... you can use LZ4Pickler
(see below) or LZ4Stream
.
Pickler
Sometimes all you need is to quickly compress a small chunk of data, let's say serialized message to send it over the
network. You can use LZ4Pickler
in such case. It does encode original length within a message and handles
incompressible data (by copying).
static class LZ4Pickler
{
static byte[] Pickle(
byte[] source,
LZ4Level level = LZ4Level.L00_FAST);
static byte[] Pickle(
byte[] source, int sourceOffset, int sourceLength,
LZ4Level level = LZ4Level.L00_FAST);
static byte[] Pickle(
ReadOnlySpan<byte> source,
LZ4Level level = LZ4Level.L00_FAST);
static byte[] Pickle(
byte* source, int sourceLength,
LZ4Level level = LZ4Level.L00_FAST);
}
Example:
var source = new byte[1000];
var encoded = LZ4Pickler.Pickle(source);
var decoded = LZ4Pickler.Unpickle(encoded);
Please note that this approach is slightly slower (copy after failed compression) and has one extra memory allocation (as it resizes buffer after compression).
Streams
Stream implementation is in different package (K4os.Compression.LZ4.Streams
) as it has dependency on K4os.Hash.xxHash
.
It is fully compatible with LZ4 Frame format although not all features are supported on compression
(they are "properly" ignored on decompression).
Stream compression settings
There are some thing which can be configured when compressing data:
class LZ4EncoderSettings
{
long? ContentLength { get; set; } = null;
bool ChainBlocks { get; set; } = true;
int BlockSize { get; set; } = Mem.K64;
bool ContentChecksum { get; set; } = false;
bool BlockChecksum { get; set; } = false;
uint? Dictionary => null;
LZ4Level CompressionLevel { get; set; } = LZ4Level.L00_FAST;
int ExtraMemory { get; set; } = 0;
}
Default options are good enough so you don't change anything. Refer to original documentation for more detailed information.
Please note that ContentLength
and Dictionary
are not currently
supported and trying to use values other than defaults will throw exceptions.
Stream compression
The class responsible for compression is LZ4EncoderStream
but its usage is not obvious.
For easy access two static factory methods
has been created:
static class LZ4Stream
{
static LZ4EncoderStream Encode(
Stream stream, LZ4EncoderSettings settings = null, bool leaveOpen = false);
static LZ4EncoderStream Encode(
Stream stream, LZ4Level level, int extraMemory = 0,
bool leaveOpen = false);
}
Both of them will take a stream (a file, a network stream, a memory stream) and wrap it adding compression on top of it.
Example:
using (var source = File.OpenRead(filename))
using (var target = LZ4Stream.Encode(File.Create(filename + ".lz4")))
{
source.CopyTo(target);
}
Stream decompression settings
Decompression settings are pretty simple and class has been added for symmetry with LZ4EncoderSettings
.
class LZ4DecoderSettings
{
int ExtraMemory { get; set; } = 0;
}
Adding extra memory to decompression process may increase decompression speed. Not significantly, though so there is no reason to stress about it too much.
Stream decompression
Same as before, there are two static factory methods to wrap existing stream and provide decompression.
static class LZ4Stream
{
static LZ4DecoderStream Decode(
Stream stream, LZ4DecoderSettings settings = null, bool leaveOpen = false);
static LZ4DecoderStream Decode(
Stream stream, int extraMemory, bool leaveOpen = false);
}
Example:
using (var source = LZ4Stream.Decode(File.OpenRead(filename + ".lz4")))
using (var target = File.Create(filename))
{
source.CopyTo(target);
}
Please note that stream decompression is (at least I hope it is) fully compatible with original specification.
Well, it does not handle predefined dictionaries but lz4.exe
does not either. All the other features which are
not implemented yet (ContentLength
, ContentChecksum
, BlockChecksum
) are just gracefully ignored but does
not cause decompression to fail.
Other stream-like data structures
As per version 1.3-beta new stream abstractions has been added (note, it has both sync and async methods, but here I'm listing sync ones only):
interface ILZ4FrameReader: IDisposable
{
bool OpenFrame();
long? GetFrameLength();
int ReadOneByte();
int ReadManyBytes(Span<byte> buffer, bool interactive = false);
long GetBytesRead();
void CloseFrame();
}
interface ILZ4FrameWriter: IDisposable
{
bool OpenFrame();
void WriteOneByte(byte value);
void WriteManyBytes(ReadOnlySpan<byte> buffer);
long GetBytesWritten();
void CloseFrame();
}
This allows to adapt any stream-like data structure to LZ4 compression/decompression, which currently is:
Span
and ReadOnlySpan
(limited support), Memory
and ReadOnlyMemory
, ReadOnlySequence
, BufferWriter
,
Stream
, PipeReader
, and PipeWriter
.
This mechanism is extendable so implementing stream-like approach for other data structures will be possible (although not trivial).
Factory methods for creating ILZ4FrameReader
and ILZ4FrameWriter
are available on LZFrame
class:
static class LZ4Frame
{
// Decode
static void Decode<TBufferWriter>(
ReadOnlySpan<byte> source, TBufferWriter target, int extraMemory = 0);
static ByteMemoryLZ4FrameReader Decode(
ReadOnlyMemory<byte> memory, int extraMemory = 0);
static ByteSequenceLZ4FrameReader Decode(
ReadOnlySequence<byte> sequence, int extraMemory = 0);
static StreamLZ4FrameReader Decode(
Stream stream, int extraMemory = 0, bool leaveOpen = false);
static PipeLZ4FrameReader Decode(
PipeReader reader, int extraMemory = 0, bool leaveOpen = false);
// Encode
static int Encode(
ReadOnlySequence<byte> source, Span<byte> target, LZ4EncoderSettings? settings = default);
static int Encode(
Span<byte> source, Span<byte> target, LZ4EncoderSettings? settings = default);
static int Encode(
Action<ILZ4FrameWriter> source, Span<byte> target, LZ4EncoderSettings? settings = default);
static ByteSpanLZ4FrameWriter Encode(
byte* target, int length, LZ4EncoderSettings? settings = default);
static ByteMemoryLZ4FrameWriter Encode(
Memory<byte> target, LZ4EncoderSettings? settings = default);
static ByteBufferLZ4FrameWriter<TBufferWriter> Encode<TBufferWriter>(
TBufferWriter target, LZ4EncoderSettings? settings = default);
static ByteBufferLZ4FrameWriter Encode(
IBufferWriter<byte> target, LZ4EncoderSettings? settings = default);
static StreamLZ4FrameWriter Encode(
Stream target, LZ4EncoderSettings? settings = default, bool leaveOpen = false);
static PipeLZ4FrameWriter Encode(
PipeWriter target, LZ4EncoderSettings? settings = default, bool leaveOpen = false);
}
Performance for small frames
Lot of LZ4 usage are small frames, like network packets, or small files. As this is still work not finished, and there is more memory allocation than I would like, performance is getting much better.
Please note, LZPickler
is the fastest option, it is just not portable, as it has been developed by me, for my own needs.
If you need to use LZ4Frame
format (the official streaming format) you will by pleased to know that that
with new abstraction it can be much faster. So far, people needed to use Stream
even if data was in memory already:
using var source = new MemoryStream(encoded);
using var decoder = LZ4Stream.Decode(source);
using var target = new MemoryStream();
decoder.CopyTo(target);
var decoded = target.ToArray();
Now it is simpler, and faster:
var decoded = LZ4Frame.Decode(encoded.AsSpan(), new ArrayBufferWriter<byte>()).WrittenMemory.ToArray();
ArrayBufferWriter<T>
is a go-to implementation of IBufferWriter<T>
, but you may want specialized implementation, if performance is critical.
It is quite fast, but it seems is relatively relaxed about allocating memory.
If you want to reduce GC usage, implement your own IBufferWriter<T>
and test it.
Below decoding small blocks:
using var source = new MemoryStream(_encoded);
using var decoder = LZ4Stream.Decode(source);
_ = decoder.Read(_decoded, 0, _decoded.Length);
vs
using var decoder = LZ4Frame.Decode(_encoded);
_ = decoder.ReadManyBytes(_decoded.AsSpan());
shows that frame reader is much faster, and is allocating less memory, also please note no Gen1 or Get2 allocations.
BenchmarkDotNet=v0.13.2, OS=Windows 10 (10.0.19044.1889/21H2/November2021Update)
AMD Ryzen 5 3600, 1 CPU, 12 logical and 6 physical cores
.NET SDK=6.0.300
[Host] : .NET 5.0.17 (5.0.1722.21314), X64 RyuJIT AVX2
DefaultJob : .NET 5.0.17 (5.0.1722.21314), X64 RyuJIT AVX2
Method | Size | Mean | Ratio | Gen0 | Gen1 | Gen2 | Alloc Ratio |
---|---|---|---|---|---|---|---|
UseStream | 128 | 1,173.6 ns | 1.00 | 1.2703 | 1.2074 | 1.2074 | 1.00 |
UseFrameReader | 128 | 466.2 ns | 0.40 | 0.0525 | - | - | 0.85 |
UseStream | 1024 | 1,593.2 ns | 1.00 | 1.6575 | 1.5945 | 1.5945 | 1.00 |
UseFrameReader | 1024 | 756.7 ns | 0.47 | 0.0525 | - | - | 0.85 |
UseStream | 8192 | 5,081.0 ns | 1.00 | 5.1956 | 5.1270 | 5.1270 | 1.00 |
UseFrameReader | 8192 | 3,836.3 ns | 0.76 | 0.0458 | - | - | 0.82 |
Legacy (lz4net) compatibility
There is a separate package for those who used lz4net before and still need to access files generated with it:
static class LZ4Legacy
{
static LZ4Stream Encode(
Stream innerStream,
bool highCompression = false,
int blockSize = 1024 * 1024,
bool leaveOpen = false);
static LZ4Stream Decode(
Stream innerStream,
bool leaveOpen = false);
}
This provide access to streams written by lz4net
. Please note, that interface is not compatible, but the file format is.
Example:
using (var source = LZ4Legacy.Decode(File.OpenRead(filename + ".old")))
using (var target = LZ4Stream.Encode(File.Create(filename + ".new")))
{
source.CopyTo(target);
}
Code above will read old (lz4net) format and write to new format (lz4_Frame_format).
Memory pooling
I've added memory block pooling to most of the classes. It is enabled by default, but comes with potential danger: this pooled memory gets pinned and may be problematic in some scenarios, for example with long lived streams.
LZ4 is used most of the time for small packages, "in-and-out 20 minutes adventure", so pinning pooling memory is not a problem.
As usual: it depends.
If you want to change the maximum pooled array use PinnedMemory.MaxPooledSize
. You can set it to 0 to disable pooling.
ARMv7, IL2CPP, Unity
Apparently ARMv7 does not handle unaligned access:
[...] It looks like the code here will do an unaligned memory access, which is not allowed on armv7, hence the crash. Mono works in this case because it generates code that is less efficient than IL2CPP, and does only aligned memory access. With IL2CPP, we have chosen to convert the C# code as-is, so that the generated code will do unaligned access if the C# code does. [...]
I've adapted 32-bit algorithm to use aligned access only (64-bit version still tries to maximise speed by using unaligned access). Think about 32-bit as "compatibility mode". It case of alignment related problems force 32-bit mode as soon as possible with:
LZ4Codec.Enforce32 = true;
Issues
Please use this template when raising one. Try to be as helpful as possible to help me reproduce it.