812e8f2a16ff7999fc733903db80357e85b2fba8 - platform_external_zstd

commit	812e8f2a16ff7999fc733903db80357e85b2fba8	[log] [tgz]
author	mgrice <mgrice@users.noreply.github.com>	Thu Jul 11 15:31:07 2019 -0700
committer	Nick Terrell <terrelln@fb.com>	Thu Jul 11 18:31:07 2019 -0400
tree	998c0f49079367410a4e6d1867eb5b1e1493af16
parent	b01c1c679f553cadb6630618b532873e284073a5 [diff]

perf improvements for zstd decode (#1668)

* perf improvements for zstd decode

tldr: 7.5% average decode speedup on silesia corpus at compression levels 1-3 (sandy bridge)

Background: while investigating zstd perf differences between clang and gcc I noticed that even though gcc is vectorizing the loop in in wildcopy, it was not being done as well as could be done by hand.  The sites where wildcopy is invoked have an interesting distribution of lengths to be copied.  The loop trip count is rarely above 1, yet long copies are common enough to make their performance important.The code in zstd_decompress.c to invoke wildcopy handles the latter well but the gcc autovectorizer introduces a needlessly expensive startup check for vectorization.

See how GCC autovectorizes the loop here:
https://godbolt.org/z/apr0x0

Here is the code after this diff has been applied: (left hand side is the good one, right is with vectorizer on)
After: https://godbolt.org/z/OwO4F8

Note that autovectorization still does not do a good job on the optimized version, so it's turned off\
 via attribute and flag.  I found that neither attribute nor command-line flag were entirely successful in turning off vectorization, which is why there were both.

    silesia benchmark data - second triad of each file is with the original code:

    file      orig        compressedratio     encode              decode           change
    1#dickens   10192446->   4268865(2.388),       198.9MB/s           709.6MB/s
    2#dickens   10192446->   3876126(2.630),       128.7MB/s           552.5MB/s
    3#dickens   10192446->   3682956(2.767),       104.6MB/s             537MB/s
    1#dickens   10192446->   4268865(2.388),       195.4MB/s           659.5MB/s     7.60%
    2#dickens   10192446->   3876126(2.630),         127MB/s           516.3MB/s     7.01%
    3#dickens   10192446->   3682956(2.767),         105MB/s           479.5MB/s    11.99%
    1#mozilla   51220480->  20117517(2.546),       285.4MB/s           734.9MB/s
    2#mozilla   51220480->  19067018(2.686),       220.8MB/s           686.3MB/s
    3#mozilla   51220480->  18508283(2.767),       152.2MB/s           669.4MB/s
    1#mozilla   51220480->  20117517(2.546),       283.4MB/s           697.9MB/s     5.30%
    2#mozilla   51220480->  19067018(2.686),       225.9MB/s             665MB/s     3.20%
    3#mozilla   51220480->  18508283(2.767),       154.5MB/s           640.6MB/s     4.50%
    1#mr         9970564->   3840242(2.596),       262.4MB/s           899.8MB/s
    2#mr         9970564->   3600976(2.769),       181.2MB/s           717.9MB/s
    3#mr         9970564->   3563987(2.798),       116.3MB/s             620MB/s
    1#mr         9970564->   3840242(2.596),       253.2MB/s           827.3MB/s     8.76%
    2#mr         9970564->   3600976(2.769),       177.4MB/s           655.4MB/s     9.54%
    3#mr         9970564->   3563987(2.798),       111.2MB/s           564.2MB/s     9.89%
    1#nci       33553445->   2849306(11.78),       575.2MB/s ,        1335.8MB/s
    2#nci       33553445->   2890166(11.61),       509.3MB/s ,        1238.1MB/s
    3#nci       33553445->   2857408(11.74),         431MB/s ,        1210.7MB/s
    1#nci       33553445->   2849306(11.78),       565.4MB/s ,        1220.2MB/s     9.47%
    2#nci       33553445->   2890166(11.61),       508.2MB/s ,        1128.4MB/s     9.72%
    3#nci       33553445->   2857408(11.74),       429.1MB/s ,        1097.7MB/s    10.29%
    1#ooffice    6152192->   3590954(1.713),       231.4MB/s ,         662.6MB/s
    2#ooffice    6152192->   3323931(1.851),       162.8MB/s ,         592.6MB/s
    3#ooffice    6152192->   3145625(1.956),        99.9MB/s ,         549.6MB/s
    1#ooffice    6152192->   3590954(1.713),       224.7MB/s ,         624.2MB/s     6.15%
    2#ooffice    6152192->   3323931 (1.851),        155MB/s ,         564.5MB/s     4.98%
    3#ooffice    6152192->   3145625(1.956),       101.1MB/s ,         521.2MB/s     5.45%
    1#osdb      10085684->   3739042(2.697),       271.9MB/s           876.4MB/s
    2#osdb      10085684->   3493875(2.887),       208.2MB/s             857MB/s
    3#osdb      10085684->   3515831(2.869),       135.3MB/s           805.4MB/s
    1#osdb      10085684->   3739042(2.697),       257.4MB/s           793.8MB/s    10.41%
    2#osdb      10085684->   3493875(2.887),       209.7MB/s           776.1MB/s    10.42%
    3#osdb      10085684->   3515831(2.869),       130.6MB/s           727.7MB/s    10.68%
    1#reymont    6627202->   2152771(3.078),       198.9MB/s           696.2MB/s
    2#reymont    6627202->   2071140(3.200),         170MB/s           595.2MB/s
    3#reymont    6627202->   1953597(3.392),       128.5MB/s           609.7MB/s
    1#reymont    6627202->   2152771(3.078),       199.6MB/s           655.2MB/s     6.26%
    2#reymont    6627202->   2071140(3.200),       168.2MB/s           554.4MB/s     7.36%
    3#reymont    6627202->   1953597(3.392),       128.7MB/s           557.4MB/s     9.38%
    1#samba     21606400->   5510994(3.921),       338.1MB/s            1066MB/s
    2#samba     21606400->   5240208(4.123),       258.7MB/s           992.3MB/s
    3#samba     21606400->   5003358(4.318),       200.2MB/s           991.1MB/s
    1#samba     21606400->   5510994(3.921),       330.8MB/s             974MB/s     9.45%
    2#samba     21606400->   5240208(4.123),       257.9MB/s           919.4MB/s     7.93%
    3#samba     21606400->   5003358(4.318),       198.5MB/s           908.9MB/s     9.04%
    1#sao        7251944->   6256401(1.159),       194.6MB/s           602.2MB/s
    2#sao        7251944->   5808761(1.248),       128.2MB/s           532.1MB/s
    3#sao        7251944->   5556318(1.305),          73MB/s           509.4MB/s
    1#sao        7251944->   6256401(1.159),       198.7MB/s           580.7MB/s     3.70%
    2#sao        7251944->   5808761(1.248),       129.1MB/s           502.7MB/s     5.85%
    3#sao        7251944->   5556318(1.305),        74.6MB/s           493.1MB/s     3.31%
    1#webster   41458703->  13692222(3.028),       222.3MB/s             752MB/s
    2#webster   41458703->  12842646(3.228),       157.6MB/s           532.2MB/s
    3#webster   41458703->  12191964(3.400),         124MB/s           468.5MB/s
    1#webster   41458703->  13692222(3.028),       219.7MB/s             697MB/s     7.89%
    2#webster   41458703->  12842646(3.228),       153.9MB/s           495.4MB/s     7.43%
    3#webster   41458703->  12191964(3.400),       124.8MB/s           444.8MB/s     5.33%
    1#xml        5345280->    696652(7.673),         485MB/s ,        1333.9MB/s
    2#xml        5345280->    681492(7.843),       405.2MB/s ,        1237.5MB/s
    3#xml        5345280->    639057(8.364),       328.5MB/s ,        1281.3MB/s
    1#xml        5345280->    696652(7.673),       473.1MB/s ,        1232.4MB/s     8.24%
    2#xml        5345280->    681492(7.843),       398.6MB/s ,        1145.9MB/s     7.99%
    3#xml        5345280->    639057(8.364),       327.1MB/s ,          1175MB/s     9.05%
    1#x-ray      8474240->   6772557(1.251),       521.3MB/s           762.6MB/s
    2#x-ray      8474240->   6684531(1.268),       230.5MB/s           688.5MB/s
    3#x-ray      8474240->   6166679(1.374),        68.7MB/s           478.8MB/s
    1#x-ray      8474240->   6772557(1.251),       502.8MB/s           736.7MB/s     3.52%
    2#x-ray      8474240->   6684531(1.268),       224.4MB/s             662MB/s     4.00%
    3#x-ray      8474240->   6166679(1.374),        67.3MB/s           437.8MB/s     9.37%

                                                                                     7.51%

* makefile changed to only pass -fno-tree-vectorize to gcc

* <Replace this line with a title. Use 1 line only, 67 chars or less>

Don't add "no-tree-vectorize" attribute on clang (which defines __GNUC__)

* fix for warning/error with subtraction of void* pointers

* fix c90 conformance issue - ISO C90 forbids mixed declarations and code

* Fix assert for negative diff, only when there is no overlap

* fix overflow revealed in fuzzing tests

* tweak for small speed increase

5 files changed

tree: 998c0f49079367410a4e6d1867eb5b1e1493af16

README.md

Zstandard, or zstd as short version, is a fast lossless compression algorithm, targeting real-time compression scenarios at zlib-level and better compression ratios. It's backed by a very fast entropy stage, provided by Huff0 and FSE library.

The project is provided as an open-source dual BSD and GPLv2 licensed C library, and a command line utility producing and decoding .zst, .gz, .xz and .lz4 files. Should your project require another programming language, a list of known ports and bindings is provided on Zstandard homepage.

Development branch status:

Benchmarks

For reference, several fast compression algorithms were tested and compared on a server running Arch Linux (Linux version 5.0.5-arch1-1), with a Core i9-9900K CPU @ 5.0GHz, using lzbench, an open-source in-memory benchmark by @inikep compiled with gcc 8.2.1, on the Silesia compression corpus.

Compressor name	Ratio	Compression	Decompress.
zstd 1.4.0 -1	2.884	530 MB/s	1360 MB/s
zlib 1.2.11 -1	2.743	110 MB/s	440 MB/s
brotli 1.0.7 -0	2.701	430 MB/s	470 MB/s
quicklz 1.5.0 -1	2.238	600 MB/s	800 MB/s
lzo1x 2.09 -1	2.106	680 MB/s	950 MB/s
lz4 1.8.3	2.101	800 MB/s	4220 MB/s
snappy 1.1.4	2.073	580 MB/s	2020 MB/s
lzf 3.6 -1	2.077	440 MB/s	930 MB/s

Zstd can also offer stronger compression ratios at the cost of compression speed. Speed vs Compression trade-off is configurable by small increments. Decompression speed is preserved and remains roughly the same at all settings, a property shared by most LZ compression algorithms, such as zlib or lzma.

The following tests were run on a server running Linux Debian (Linux version 4.14.0-3-amd64) with a Core i7-6700K CPU @ 4.0GHz, using lzbench, an open-source in-memory benchmark by @inikep compiled with gcc 7.3.0, on the Silesia compression corpus.

Compression Speed vs Ratio	Decompression Speed

A few other algorithms can produce higher compression ratios at slower speeds, falling outside of the graph. For a larger picture including slow modes, click on this link.

The case for Small Data compression

Previous charts provide results applicable to typical file and stream scenarios (several MB). Small data comes with different perspectives.

The smaller the amount of data to compress, the more difficult it is to compress. This problem is common to all compression algorithms, and reason is, compression algorithms learn from past data how to compress future data. But at the beginning of a new data set, there is no "past" to build upon.

To solve this situation, Zstd offers a training mode, which can be used to tune the algorithm for a selected type of data. Training Zstandard is achieved by providing it with a few samples (one file per sample). The result of this training is stored in a file called "dictionary", which must be loaded before compression and decompression. Using this dictionary, the compression ratio achievable on small data improves dramatically.

The following example uses the github-users sample set, created from github public API. It consists of roughly 10K records weighing about 1KB each.

Compression Ratio	Compression Speed	Decompression Speed

These compression gains are achieved while simultaneously providing faster compression and decompression speeds.

Training works if there is some correlation in a family of small data samples. The more data-specific a dictionary is, the more efficient it is (there is no universal dictionary). Hence, deploying one dictionary per type of data will provide the greatest benefits. Dictionary gains are mostly effective in the first few KB. Then, the compression algorithm will gradually use previously decoded content to better compress the rest of the file.

Dictionary compression How To:

Create the dictionary
zstd --train FullPathToTrainingSet/* -o dictionaryName
Compress with dictionary
zstd -D dictionaryName FILE
Decompress with dictionary
zstd -D dictionaryName --decompress FILE.zst

Build instructions

Makefile

If your system is compatible with standard make (or gmake), invoking make in root directory will generate zstd cli in root directory.

Other available options include:

make install : create and install zstd cli, library and man pages
make check : create and run zstd, tests its behavior on local platform

cmake

A cmake project generator is provided within build/cmake. It can generate Makefiles or other build scripts to create zstd binary, and libzstd dynamic and static libraries.

By default, CMAKE_BUILD_TYPE is set to Release.

Meson

A Meson project is provided within build/meson. Follow build instructions in that directory.

You can also take a look at .travis.yml file for an example about how Meson is used to build this project.

Note that default build type is release.

Visual Studio (Windows)

Going into build directory, you will find additional possibilities:

Projects for Visual Studio 2005, 2008 and 2010.
- VS2010 project is compatible with VS2012, VS2013, VS2015 and VS2017.
Automated build scripts for Visual compiler by @KrzysFR, in build/VS_scripts, which will build zstd cli and libzstd library without any need to open Visual Studio solution.

Buck

You can build the zstd binary via buck by executing: buck build programs:zstd from the root of the repo. The output binary will be in buck-out/gen/programs/.

Status

Zstandard is currently deployed within Facebook. It is used continuously to compress large amounts of data in multiple formats and use cases. Zstandard is considered safe for production environments.

License

Zstandard is dual-licensed under BSD and GPLv2.

Contributing

The "dev" branch is the one where all contributions are merged before reaching "master". If you plan to propose a patch, please commit into the "dev" branch, or its own feature branch. Direct commit to "master" are not permitted. For more information, please read CONTRIBUTING.