The v1.3.4 release of Zstandard is focused on performance, and offers nice speed boost in most scenarios.

Asynchronous compression by default for zstd CLI

zstd cli will now performs compression in parallel with I/O operations by default. This requires multi-threading capability (which is also enabled by default).
It doesn't sound like much, but effectively improves throughput by 20-30%, depending on compression level and underlying I/O performance.

For example, on a Mac OS-X laptop with an Intel Core i7-5557U CPU @ 3.10GHz, running time zstd enwik9 at default compression level (2) on a SSD gives the following :

This is a nice boost to all scripts using zstd cli, typically in network or storage tasks. The effect is even more pronounced at faster compression setting, since the CLI overlaps a proportionally higher share of compression with I/O.

Previous default behavior (blocking single thread) is still available, accessible through --single-thread long command. It's also the only mode available when no multi-threading capability is detected.

General speed improvements

Some core routines have been refined to provide more speed on newer cpus, making better use of their out-of-order execution units. This is more sensible on the decompression side, and even more so with gcc compiler.

Example on the same platform, running in-memory benchmark zstd -b1 silesia.tar :

Faster compression levels

So far, compression level 1 has been the fastest one available. Starting with v1.3.4, there will be additional choices. Faster compression levels can be invoked using negative values.
On the command line, the equivalent one can be triggered using --fast[=#] command.

Negative compression levels sample data more sparsely, and disable Huffman compression of literals, translating into faster decoding speed.

It's possible to create one's own custom fast compression level
by using strategy ZSTD_fast, increasing ZSTD_p_targetLength to desired value,
and turning on or off literals compression, using ZSTD_p_compressLiterals.

Performance is generally on par or better than other high speed algorithms. On below benchmark (compressing silesia.tar on an Intel Core i7-6700K CPU @ 4.00GHz) , it ends up being faster and stronger on all metrics compared with quicklz and snappy at --fast=2. It also compares favorably to lzo with --fast=3. lz4 still offers a better speed / compression combo, with zstd --fast=4 approaching close.

Applications which were considering Zstandard but were worried of being CPU-bounded are now able to shift the load from CPU to bandwidth on a larger scale, and may even vary temporarily their choice depending on local conditions (to deal with some sudden workload surge for example).

Long Range Mode with Multi-threading

zstd-1.3.2 introduced the long range mode, capable to deduplicate long distance redundancies in a large data stream, a situation typical in backup scenarios for example. But its usage in association with multi-threading was discouraged, due to inefficient use of memory.
zstd-1.3.4 solves this issue, by making long range match finder run in serial mode, like a pre-processor, before passing its result to backend compressors (regular zstd). Memory usage is now bounded to the maximum of the long range window size, and the memory that zstdmt would require without long range matching. As the long range mode runs at about 200 MB/s, depending on the number of cores available, it's possible to tune compression level to match the LRM speed, which becomes the upper limit.

zstd -T0  -5  --long    file # autodetect threads, level 5, 128 MB window
zstd -T16 -10 --long=31 file # 16 threads, level 10, 2 GB window

As illustration, benchmarks of the two files "Linux 4.7 - 4.12" and "Linux git" from the 1.3.2 release are shown below. All compressors are run with 16 threads, except "zstd single 2 GB". zstd compressors are run with either a 128 MB or 2 GB window size, and lrzip compressor is run with lzo, gzip, and xz backends. The benchmarks were run on a 16 core Sandy Bridge @ 2.2 GHz.

The association of Long Range Mode with multi-threading is pretty compelling for large stream scenarios.

Miscellaneous

This release also brings its usual list of small improvements and bug fixes, as detailed below :

• perf: faster speed (especially decoding speed) on recent cpus (haswell+)
• perf: much better performance associating --long with multi-threading, by @terrelln
• perf: better compression at levels 13-15
• cli : asynchronous compression by default, for faster experience (use --single-thread for former behavior)
• cli : smoother status report in multi-threading mode
• cli : added command --fast=#, for faster compression modes
• cli : fix crash when not overwriting existing files, by Pádraig Brady (@pixelb)
• api : nbThreads becomes nbWorkers : 1 triggers asynchronous mode
• api : compression levels can be negative, for even more speed
• api : ZSTD_getFrameProgression() : get precise progress status of ZSTDMT anytime
• api : ZSTDMT can accept new compression parameters during compression
• api : implemented all advanced dictionary decompression prototypes
• build: improved meson recipe, by Shawn Landden (@shawnl)
• build: VS2017 scripts, by @HaydnTrigg
• misc: all /contrib projects fixed
• misc: added /contrib/docker script by @gyscos

This is bugfix release, mostly focused on cleaning several detrimental corner cases scenarios.
It is nonetheless a recommended upgrade.

Changes Summary

• perf: improved zstd_opt strategy (levels 16-19)
• fix : bug #944 : multithreading with shared ditionary and large data, reported by @gsliepen
• cli : change : -o can be combined with multiple inputs, by @terrelln
• cli : fix : content size written in header by default
• cli : fix : improved LZ4 format support, by @felixhandte
• cli : new : hidden command -b -S, to benchmark multiple files and generate one result per file
• api : change : when setting pledgedSrcSize, use ZSTD_CONTENTSIZE_UNKNOWN macro value to mean "unknown"
• api : fix : support large skippable frames, by @terrelln
• api : fix : re-using context could result in suboptimal block size in some corner case scenarios
• api : fix : streaming interface was adding a useless 3-bytes null block to small frames
• build: fix : compilation under rhel6 and centos6, reported by @pixelb
• build: added check target
• build: improved meson support, by @shawnl

Zstandard Long Range Match Finder

Zstandard has a new long range match finder written by Facebook's intern Stella Lau (@stellamplau), which specializes on finding long matches in the distant past. It integrates seamlessly with the regular compressor, and the output can be decompressed just like any other Zstandard compressed data.

The long range match finder adds minimal overhead to the compressor, works with any compression level, and maintains Zstandard's blazingly fast decompression speed. However, since the window size is larger, it requires more memory for compression and decompression.

To go along with the long range match finder, we've increased the maximum window size to 2 GB. The decompressor only accepts window sizes up to 128 MB by default, but zstd -d --memory=2GB will decompress window sizes up to 2 GB.

Example usage

# 128 MB window size
zstd -1 --long file
zstd -d file.zst

# 2 GB window size (window log = 31)
zstd -6 --long=31 file
zstd -d --long=31 file.zst
# OR
zstd -d --memory=2GB file.zst

ZSTD_CCtx *cctx = ZSTD_createCCtx();
ZSTD_CCtx_setParameter(cctx, ZSTD_p_compressionLevel, 19);
ZSTD_CCtx_setParameter(cctx, ZSTD_p_enableLongDistanceMatching, 1); // Sets windowLog=27
ZSTD_CCtx_setParameter(cctx, ZSTD_p_windowLog, 30); // Optionally increase the window log
ZSTD_compress_generic(cctx, &out, &in, ZSTD_e_end);

ZSTD_DCtx *dctx = ZSTD_createDCtx();
ZSTD_DCtx_setMaxWindowSize(dctx, 1 << 30);
ZSTD_decompress_generic(dctx, &out, &in);

Benchmarks

We compared the zstd long range matcher to zstd and lrzip. The benchmarks were run on an AMD Ryzen 1800X (8 cores with 16 threads at 3.6 GHz).

Compressors

• zstd — The regular Zstandard compressor.
• zstd 128 MB — The Zstandard compressor with a 128 MB window size.
• zstd 2 GB — The Zstandard compressor with a 2 GB window size.
• lrzip xz — The lrzip compressor with default options, which uses the xz backend at level 7 with 16 threads.
• lrzip xz single — The lrzip compressor with a single-threaded xz backend at level 7.
• lrzip zstd — The lrzip compressor without a backend, then its output is compressed by zstd (not multithreaded).

Files

• Linux 4.7 - 4.12 — This file consists of the uncompressed tarballs of the six Linux kernel release from 4.7 to 4.12 concatenated together in order. This file is extremely compressible if the compressor can match against the previous versions well.
• Linux git — This file is a tarball of the linux repo, created by git clone https://github.com/torvalds/linux && tar -cf linux-git.tar linux/. This file gets a small benefit from long range matching. This file shows how the long range matcher performs when there isn't too many matches to find.

Results

Both zstd and zstd 128 MB don't have large enough of a window size to compress Linux 4.7 - 4.12 well. zstd 2 GB compresses the fastest, and slightly better than lrzip-zstd. lrzip-xz compresses the best, and at a reasonable speed with multithreading enabled. The place where zstd shines is decompression ease and speed. Since it is just regular Zstandard compressed data, it is decompressed by the highly optimized decompressor.

The Linux git file shows that the long range matcher maintains good compression and decompression speed, even when there are far less long range matches. The decompression speed takes a small hit because it has to look further back to reconstruct the matches.

Compression Ratio vs Speed	Decompression Speed

Implementation details

The long distance match finder was inspired by great work from Con Kolivas' lrzip, which in turn was inspired by Andrew Tridgell's rzip. Also, let's mention Bulat Ziganshin's srep, which we have not been able to test unfortunately (site down), but the discussions on encode.ru proved great sources of inspiration.

Therefore, many similar mechanisms are adopted, such as using a rolling hash, and filling a hash table divided into buckets of entries.

That being said, we also made different choices, with the goal to favor speed, as can be observed in benchmark. The rolling hash formula is selected for computing efficiency. There is a restrictive insertion policy, which only inserts candidates that respect a mask condition. The insertion policy allows us to skip the hash table in the common case that a match isn't present. Confirmation bits are saved, to only check for matches when there is a strong presumption of success. These and a few more details add up to make zstd's long range matcher a speed-oriented implementation.

The biggest difference though is that the long range matcher is blended into the regular compressor, producing a single valid zstd frame, undistinguishable from normal operation (except obviously for the larger window size). This makes decompression a single pass process, preserving its speed property.

More details are available directly in source code, at lib/compress/zstd_ldm.c.

Future work

This is a first implementation, and it still has a few limitations, that we plan to lift in the future.

The long range matcher doesn't interact well with multithreading. Due to the way zstd multithreading is currently implemented, memory usage will scale with the window size times the number of threads, which is a problem for large window sizes. We plan on supporting multithreaded long range matching with reasonable memory usage in a future version.

Secondly, Zstandard is currently limited to a 2 GB window size because of indexer's design. While this is a significant update compared to previous 128 MB limit, we believe this limitation can be lifted altogether, with some structural changes in the indexer. However, it also means that window size would become really big, with knock-off consequences on memory usage. So, to reduce this load, we will have to consider memory map as a complementary way to reference past content in the uncompressed file.

Detailed list of changes

• new : long range mode, using --long command, by Stella Lau (@stellamplau)
• new : ability to generate and decode magicless frames (#591)
• changed : maximum nb of threads reduced to 200, to avoid address space exhaustion in 32-bits mode
• fix : multi-threading compression works with custom allocators, by @terrelln
• fix : a rare compression bug when compression generates very large distances and bunch of other conditions (only possible at --ultra -22)
• fix : 32-bits build can now decode large offsets (levels 21+)
• cli : added LZ4 frame support by default, by Felix Handte (@felixhandte)
• cli : improved --list output
• cli : new : can split input file for dictionary training, using command -B#
• cli : new : clean operation artefact on Ctrl-C interruption (#854)
• cli : fix : do not change /dev/null permissions when using command -t with root access, reported by @mike155 (#851)
• cli : fix : write file size in header in multiple-files mode
• api : added macro ZSTD_COMPRESSBOUND() for static allocation
• api : experimental : new advanced decompression API
• api : fix : sizeof_CCtx() used to over-estimate
• build: fix : compilation works with -mbmi (#868)
• build: fix : no-multithread variant compiles without pool.c dependency, reported by Mitchell Blank Jr (@mitchblank) (#819)
• build: better compatibility with reproducible builds, by Bernhard M. Wiedemann (@bmwiedemann) (#818)
• example : added streaming_memory_usage
• license : changed /examples license to BSD + GPLv2
• license : fix a few header files to reflect new license (#825)

Warning

bug #944 : v1.3.2 is known to produce corrupted data in the following scenario, requiring all these conditions simultaneously :

• compression using multi-threading
• with a dictionary
• on "large enough" files (several MB, exact threshold depends on compression level)

Note that dictionary is meant to help compression of small files (a few KB), while multi-threading is only useful for large files, so it's pretty rare to need both at the same time. Nonetheless, if your application happens to trigger this situation, it's recommended to skip v1.3.2 for a newer version. At the time of this warning, the dev branch is known to work properly for the same scenario.

Zstandard Fuzz Corpora

• New license : BSD + GPLv2
• perf: substantially decreased memory usage in Multi-threading mode, thanks to reports by Tino Reichardt (@mcmilk)
• perf: Multi-threading supports up to 256 threads. Cap at 256 when more are requested (#760)
• cli : improved and fixed --list command, by @ib (#772)
• cli : command -vV lists supported formats, by @ib (#771)
• build : fixed binary variants, reported by @svenha (#788)
• build : fix Visual compilation for non x86/x64 targets, reported by @GregSlazinski (#718)
• API exp : breaking change : ZSTD_getframeHeader() provides more information
• API exp : breaking change : pinned down values of error codes
• doc : fixed huffman example, by Ulrich Kunitz (@ulikunitz)
• new : contrib/adaptive-compression, I/O driven compression level, by Paul Cruz (@paulcruz74)
• new : contrib/long_distance_matching, statistics tool by Stella Lau (@stellamplau)
• updated : contrib/linux-kernel, by Nick Terrell (@terrelln)

cli : new : --list command, by @paulcruz74
cli : changed : xz/lzma support enabled by default
cli : changed : -t * continue processing list after a decompression error
API : added : ZSTD_versionString()
API : promoted to stable status : ZSTD_getFrameContentSize(), by @iburinoc
API exp : new advanced API : ZSTD_compress_generic(), ZSTD_CCtx_setParameter()
API exp : new : API for static or external allocation : ZSTD_initStatic?Ctx()
API exp : added : ZSTD_decompressBegin_usingDDict(), requested by @Crazee (#700)
API exp : clarified memory estimation / measurement functions.
API exp : changed : strongest strategy renamed ZSTD_btultra, fastest strategy ZSTD_fast set to 1
Improved : reduced stack memory usage, by @terrelln and @stellamplau
tools : decodecorpus can generate random dictionary-compressed samples, by @paulcruz74
new : contrib/seekable_format, demo and API, by @iburinoc
changed : contrib/linux-kernel, updated version and license, by @terrelln

Major features :

• Multithreading is enabled by default in the cli. Use -T# to select nb of thread. To disable multithreading, build target zstd-nomt or compile with HAVE_THREAD=0.
• New dictionary builder named "cover" with improved quality (produces better compression ratio), by @terrelln. Legacy dictionary builder remains available, using --train-legacy command.

Other changes :
cli : new : command -T0 means "detect and use nb of cores", by @iburinoc
cli : new : zstdmt symlink hardwired to zstd -T0
cli : new : command --threads=# (#671)
cli : new : commands --train-cover and --train-legacy, to select dictionary algorithm and parameters
cli : experimental targets zstd4 and xzstd4, supporting lz4 format, by @iburinoc
cli : fix : does not output compressed data on console
cli : fix : ignore symbolic links unless --force specified,
API : breaking change : ZSTD_createCDict_advanced() uses compressionParameters as argument
API : added : prototypes ZSTD_*_usingCDict_advanced(), for direct control over frameParameters.
API : improved: ZSTDMT_compressCCtx() reduced memory usage
API : fix : ZSTDMT_compressCCtx() now provides srcSize in header (#634)
API : fix : src size stored in frame header is controlled at end of frame
API : fix : enforced consistent rules for pledgedSrcSize==0 (#641)
API : fix : error code GENERIC replaced by dstSizeTooSmall when appropriate
build: improved cmake script, by @Majlen
build: enabled Multi-threading support for *BSD, by @bapt
tools: updated paramgrill. Command -O# provides best parameters for sample and speed target.
new : contrib/linux-kernel version, by @terrelln

cli : new : can compress in *.gz format, using --format=gzip command, by @inikep
cli : new : advanced benchmark command --priority=rt
cli : fix : write on sparse-enabled file systems in 32-bits mode, by @ds77
cli : fix : --rm remains silent when input is stdin
cli : experimental xzstd target, with support for xz/lzma decoding, by @inikep
speed : improved decompression speed in streaming mode for single pass scenarios (+5%)
memory : DDict (decompression dictionary) memory usage down from 150 KB to 20 KB
arch : 32-bits variant able to generate and decode very long matches (>32 MB), by @iburinoc
API : new : ZSTD_findFrameCompressedSize(), ZSTD_getFrameContentSize(), ZSTD_findDecompressedSize()
API : changed : dropped support of legacy versions <= v0.3 (can be selected by modifying ZSTD_LEGACY_SUPPORT value)
build: new: meson build system in contrib/meson, by @dimkr
build: improved cmake script, by @Majlen
build: added -Wformat-security flag, as recommended by @pixelb
doc : new : doc/educational_decoder, by @iburinoc

Warning : the experimental target zstdmt contained in this release has an issue when using multiple threads on large enough files, which makes it generate buggy header. While fixing the header after the fact is possible, it's much better to avoid the issue. This can be done by using zstdmt in pipe mode :
cat file | zstdmt -T2 -o file.zst
This issue is fixed in current dev branch, so alternatively, create zstdmt from dev branch.

Note : pre-compiled Windows binaries attached below contain the fix for zstdmt

cli : zstd can decompress .gz files (can be disabled with make zstd-nogz or make HAVE_ZLIB=0)
cli : new : experimental target make zstdmt, with multi-threading support
cli : new : improved dictionary builder "cover" (experimental), by @terrelln, based on previous work by @ot
cli : new : advanced commands for detailed parameters, by @inikep
cli : fix zstdless on Mac OS-X, by @apjanke
cli : fix #232 "compress non-files"
API : new : lib/compress/ZSTDMT_compress.h multithreading API (experimental)
API : new : ZSTD_create?Dict_byReference(), requested by Bartosz Taudul
API : new : ZDICT_finalizeDictionary()
API : fix : ZSTD_initCStream_usingCDict() properly writes dictID into frame header, by @indygreg (#511)
API : fix : all symbols properly exposed in libzstd, by @terrelln
build : support for Solaris target, by @inikep
doc : clarified specification, by @iburinoc

Sample set for reference dictionary compression benchmark

# Download and expand sample set 
wget https://github.com/facebook/zstd/releases/download/v1.1.3/github_users_sample_set.tar.zst
zstd -d github_users_sample_set.tar.zst
tar xf github_users_sample_set.tar

# benchmark sample set with and without dictionary compression
zstd -b1 -r github
zstd --train -r github
zstd -b1 -r github -D dictionary

# rebuild sample set archive
tar cf github_users_sample_set.tar github
zstd -f --ultra -22 github_users_sample_set.tar

new : programs/gzstd , combined *.gz and *.zst decoder, by @inikep
new : zstdless, less on compressed *.zst files
new : zstdgrep, grep on compressed *.zst files
fixed : zstdcat

cli : new : preserve file attributes
cli : fixed : status displays total amount decoded, even for file consisting of multiple frames (like pzstd)
lib : improved : faster decompression speed at ultra compression settings and 32-bits mode
lib : changed : only public ZSTD_ symbols are now exposed in dynamic library
lib : changed : reduced usage of stack memory
lib : fixed : several corner case bugs, by @terrelln
API : streaming : decompression : changed : automatic implicit reset when chain-decoding new frames without init
API : experimental : added : dictID retrieval functions, and ZSTD_initCStream_srcSize()
API : zbuff : changed : prototypes now generate deprecation warnings
zlib_wrapper : added support for gz* functions, by @inikep
install : better compatibility with FreeBSD, by @DimitryAndric
source tree : changed : zbuff source files moved to lib/deprecated

New : cli commands -M#, --memory=, --memlimit=, --memlimit-decompress= to limit allowed memory consumption during decompression
New : doc/zstd_manual.html, by @inikep
Improved : slightly better compression ratio at --ultra levels (>= 20)
Improved : better memory usage when using streaming compression API, thanks to @Rogier-5 report
Added : API : ZSTD_initCStream_usingCDict(), ZSTD_initDStream_usingDDict() (experimental section)
Added : examples/multiple_streaming_compression.c
Changed : zstd_errors.h is now installed within /include (and replaces errors_public.h)
Updated man page
Fixed : several sanitizer warnings, by @terrelln
Fixed : zstd-small, zstd-compress and zstd-decompress compilation targets

New : pzstd , parallel version of zstd, by @terrelln

added : NetBSD install target (#338)
Improved : speed for batches of small files
Improved : speed of zlib wrapper, by @inikep
Changed : libzstd on Windows supports legacy formats, by @KrzysFR
Fixed : CLI -d output to stdout by default when input is stdin (#322)
Fixed : CLI correctly detects console on Mac OS-X
Fixed : CLI supports recursive mode -r on Mac OS-X
Fixed : Legacy decoders use unified error codes, reported by benrg (#341), fixed by @inikep
Fixed : compatibility with OpenBSD, reported by@juanfra684 (#319)
Fixed : compatibility with Hurd, by @inikep (#365)
Fixed : zstd-pgo, reported by @octoploid (#329)

Change Licensing, all project is now BSD, copyright Facebook
Added Patent Grant
Small decompression speed improvement
API : Streaming API supports legacy format
API : New : ZDICT_getDictID(), ZSTD_sizeof_{CCtx, DCtx, CStream, DStream}(), ZSTD_setDStreamParamter()
CLI supports legacy formats v0.4+
Fixed : compression fails on certain huge files, reported by Jesse McGrew
Enhanced documentation, by @inikep

New streaming API
Changed : --ultra now enables levels beyond 19
Changed : -i# now selects benchmark time in second
Fixed : ZSTD_compress* can now compress > 4 GB in a single pass, reported by Nick Terrell
Fixed : speed regression on specific patterns (#272)
Fixed : support for Z_SYNC_FLUSH, by @dak-evanti-ru (#291)
Fixed : ICC compilation, by @inikep

Final compression format
Improved : better speed on clang and gcc -O2, thanks to @ebiggers
New : Build on FreeBSD and DragonFly, thanks to @jrmarino
Changed : modified API : ZSTD_compressEnd()
Fixed : legacy mode with ZSTD_HEAPMODE=0, by @gymdis
Fixed : premature end of frame when zero-sized raw block, reported by @ebiggers
Fixed : large dictionaries (> 384 KB), reported by Ilona Papava
Fixed : checksum correctly checked in single-pass mode
Fixed : combined --test amd --rm, reported by @amnilsson
Modified : minor compression level adaptations
Updated : compression format specification to v0.2.0
changed : zstd.h moved to /lib directory

Same as v0.6.1
with added ability to decode v0.7x and v0.8x streams (forward compatibility)

Same as v0.7.4
with added ability to decode v0.8x streams (forward compatibility)

Modified : default compression level is now 3 for CLI
Added : homebrew for Mac, by @cadedaniel
Fixed : segfault when using small dictionaries, reported by Felix Handte
Added : more examples
Updated : specification, to v0.1.1

New : compression format specification zstd_compression_format.md
New : -- separator, stating that all following arguments are file names. Suggested by @chipturner
New : ZSTD_getDecompressedSize()
New : OpenBSD target, by @juanfra684
New : examples directory
fixed : dictBuilder using HC levels, reported by Bartosz Taudul
fixed : legacy support from ZSTD_decompress_usingDDict(), reported by Felix Handte
fixed : multi-blocks decoding with intermediate uncompressed blocks, reported by @GregSlazinski
modified : removed "mem.h" and "error_public.h" dependencies from "zstd.h" (experimental section)
modified : legacy functions no longer need magic number

fixed : ZSTD_decompressBlock() using multiple consecutive blocks. Reported by @GregSlazinski
fixed : potential segfault on very large files (many gigabytes). Reported by @chipturner
fixed : CLI displays system error message when destination file cannot be created (#231). Reported by @chipturner
fixed : leak in some fail scenario in dictionary builder, reported by @nemequ