@Cyan4973 Cyan4973 released this Mar 26, 2018 · 1261 commits to master since this release

Assets 6

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 :

Version real time
1.3.3 9.2s
1.3.4 --single-thread 8.8s
1.3.4 (asynchronous) 7.5s

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 :

Version C.Speed D.Speed
1.3.3 llvm9 290 MB/s 660 MB/s
1.3.4 llvm9 304 MB/s 700 MB/s (+6%)
1.3.3 gcc7 280 MB/s 710 MB/s
1.3.4 gcc7 300 MB/s 890 MB/s (+25%)

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.

name ratio compression decompression
zstd 1.3.4 --fast=5 1.996 770 MB/s 2060 MB/s
lz4 1.8.1 2.101 750 MB/s 3700 MB/s
zstd 1.3.4 --fast=4 2.068 720 MB/s 2000 MB/s
zstd 1.3.4 --fast=3 2.153 675 MB/s 1930 MB/s
lzo1x 2.09 -1 2.108 640 MB/s 810 MB/s
zstd 1.3.4 --fast=2 2.265 610 MB/s 1830 MB/s
quicklz 1.5.0 -1 2.238 540 MB/s 720 MB/s
snappy 1.1.4 2.091 530 MB/s 1820 MB/s
zstd 1.3.4 --fast=1 2.431 530 MB/s 1770 MB/s
zstd 1.3.4 -1 2.877 470 MB/s 1380 MB/s
brotli 1.0.2 -0 2.701 410 MB/s 430 MB/s
lzf 3.6 -1 2.077 400 MB/s 860 MB/s
zlib 1.2.11 -1 2.743 110 MB/s 400 MB/s

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.

Linux 4.7 - 12 compression ratio vs speed
Linux git compression ratio vs speed

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


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