Updated documentation

diff --git a/NEWS b/NEWS
index ed414e23ac26b2f8e169c2c00c5d24c031f69201..46b170a1058fcdced6c89defcc7d9a5054900608 100644 (file)
--- a/NEWS
+++ b/NEWS
@@ -1,6 +1,7 @@
 v0.5.0
 New : dictionary builder utility
 Changed : streaming & dictionary API
+Improved : better compression of small data
 
 v0.4.7
 Improved : small compression speed improvement in HC mode

diff --git a/README.md b/README.md
index 4a339ccd2f3ead2eb607246470c5a15b1678d25d..11e41a4c8ad7c558faad1edc1d357fa463898950 100644 (file)
--- a/README.md
+++ b/README.md
@@ -7,38 +7,83 @@ It is provided as a BSD-license package, hosted on Github.
 |master      | [![Build Status](https://travis-ci.org/Cyan4973/zstd.svg?branch=master)](https://travis-ci.org/Cyan4973/zstd) |
 |dev         | [![Build Status](https://travis-ci.org/Cyan4973/zstd.svg?branch=dev)](https://travis-ci.org/Cyan4973/zstd) |
 
-For a taste of its performance, here are a few benchmark numbers from a number of compression codecs suitable for real-time. The test was completed on a Core i7-5600U @ 2.6 GHz, using m^2's [fsbench 0.14.3](http://encode.ru/threads/1371-Filesystem-benchmark?p=34029&viewfull=1#post34029) compiled with gcc 4.8.4, on the [Silesia compression corpus](http://sun.aei.polsl.pl/~sdeor/index.php?page=silesia).
+As a reference, several fast compression algorithms were tested and compared to [zlib] on a Core i7-3930K CPU @ 4.5GHz, using [lzbench], an open-source in-memory benchmark by @inikep compiled with gcc 5.2.1, on the [Silesia compression corpus].
+
+[lzbench]: https://github.com/inikep/lzbench
+[Silesia compression corpus]: http://sun.aei.polsl.pl/~sdeor/index.php?page=silesia
+
 
 |Name             | Ratio | C.speed | D.speed |
 |-----------------|-------|--------:|--------:|
 |                 |       |   MB/s  |  MB/s   |
-| **zstd 0.4**    |**2.872**|**280**| **670** |
-| [zlib] 1.2.8 -1 | 2.730 |    70   |   300   | 
-| QuickLZ 1.5.1b6 | 2.237 |   370   |   415   |
-| LZO 2.06        | 2.106 |   400   |   580   |
-| [LZ4] r131      | 2.101 |   450   |  2100   |
-| Snappy 1.1.0    | 2.091 |   330   |  1100   |
-| LZF 3.6         | 2.077 |   200   |   560   |
+|**zstd 0.4.7 -1**|**2.875**|**330**| **890** |
+| [zlib] 1.2.8 -1 | 2.730 |    95   |   360   |
+| brotli -0       | 2.708 |   220   |   430   |
+| QuickLZ 1.5     | 2.237 |   510   |   605   |
+| LZO 2.09        | 2.106 |   610   |   870   |
+| [LZ4] r131      | 2.101 |   620   |  3100   |
+| Snappy 1.1.3    | 2.091 |   480   |  1600   |
+| LZF 3.6         | 2.077 |   375   |   790   |
 
 [zlib]:http://www.zlib.net/
-[LZ4]:http://www.lz4.org/
+[LZ4]: http://www.lz4.org/
 
-Zstd can also offer stronger compression ratio at the cost of compression speed. Speed / Ratio trade-off is configurable by small increment, to fit different situations. Note however that decompression speed is preserved and remain roughly the same at all settings, a property shared by most LZ compression algorithms, such as [zlib]. The following test is run on a Core i7-3930K CPU @ 4.5GHz, using [lzbench], an open-source in-memory benchmark by inikep compiled with gcc 5.2.1, on the [Silesia compression corpus](http://sun.aei.polsl.pl/~sdeor/index.php?page=silesia).
+Zstd can also offer stronger compression ratio at the cost of compression speed. 
+Speed vs Compression trade-off is configurable by small increment. Decompression speed is preserved and remain roughly the same at all settings, a property shared by most LZ compression algorithms, such as [zlib].
 
-[lzbench]:https://github.com/inikep/lzbench
+The following test is run on a Core i7-3930K CPU @ 4.5GHz, using [lzbench], an open-source in-memory benchmark by @inikep compiled with gcc 5.2.1, on the [Silesia compression corpus].
 
 Compression Speed vs Ratio | Decompression Speed
 ---------------------------|--------------------
 ![Compression Speed vs Ratio](images/CSpeed.png "Compression Speed vs Ratio") | ![Decompression Speed](images/DSpeed.png "Decompression Speed")
 
 
-Zstd entropy stage is provided by [Huff0 and FSE, from Finite State Entropy library](https://github.com/Cyan4973/FiniteStateEntropy).
+### The case for Small Data compression
 
-Its memory requirement can be configured to fit into low-memory hardware configurations, or servers handling multiple connections/contexts in parallel.
+The above chart is applicable to large files or large streams scenarios (200 MB in this case).
+Small data (< 64 KB) come with different perspectives.
+The smaller the amount of data to compress, the more difficult it is to achieve any significant compression.
+On reaching the 1 KB region, it becomes almost impossible to compress anything.
+This problem is common to all compression algorithms, and throwing CPU power at it achieves no significant gains.
+
+The reason is, compression algorithms learn from past data how to compress future data.
+But at the beginning of a new file, there is no "past" to build upon.
+
+[Starting with 0.5](https://github.com/Cyan4973/zstd/releases), Zstd now offers [a _Dictionary Builder_ tool](https://github.com/Cyan4973/zstd/tree/master/dictBuilder).
+It can be used to train the algorithm to fit a selected type of data, by providing it with some samples.
+The result is a file (or a byte buffer) called "dictionary", which can be loaded before compression and decompression.
+By using this dictionary, the compression ratio achievable on small data improves dramatically :
+
+| Collection Name    | Direct compression | Dictionary Compression | Gains  | Average unit | Range       |
+| ---------------    | ------------------ | ---------------------- | -----  | ------------:| -----       |
+| Small JSON records | x1.331 - x1.366   | x5.860 - x6.830        | ~ x4.7 | 300          | 200 - 400   |
+| Mercurial events   | x2.322 - x2.538    | x3.377 - x4.462        | ~ x1.5 | 1.5 KB       | 20 - 200 KB |     
+| Large JSON docs    | x3.813 - x4.043    | x8.935 - x13.366       | ~ x2.8 | 6 KB         | 800 - 20 KB |     
+
+It has to be noted that these compression gains are achieved without any speed loss, and even some faster decompression processing.
+
+Dictionary work if there is some correlation in a family of small data (there is no _universal dictionary_).
+Hence, deploying one dictionary per type of data will provide the greater benefits.
+
+Large documents will benefit proportionally less, since dictionary gains are mostly effective in the first few KB.
+Then there is enough history to build upon, and the compression algorithm can rely on it to compress the rest of the file.
 
-Zstd has not yet reached "stable format" status. It doesn't guarantee yet that its current compressed format will remain stable and supported in future versions. During this period, it can still change to adapt new optimizations still being investigated. "Stable Format" is projected sometimes early 2016. 
 
-That being said, the library is now fairly robust, able to withstand hazards situations, including invalid inputs. The library reliability has been tested using [Fuzz Testing](https://en.wikipedia.org/wiki/Fuzz_testing), with both [internal tools](programs/fuzzer.c) and [external ones](http://lcamtuf.coredump.cx/afl). Therefore, it seems now safe to test Zstandard even within production environments.
+### Status
+
+Zstd has not yet reached "stable format" status. It doesn't guarantee yet that its current compression format will remain stable in future versions. During this period, it can still change to adapt new optimizations still being investigated. "Stable Format" is projected H1 2016, and will be tagged `v1.0`.
+
+That being said, the library is now fairly robust, able to withstand hazards situations, including invalid inputs. It also features legacy support, so that documents compressed with current and previous version of zstd can still be decoded in the future. 
+Library reliability has been tested using [Fuzz Testing](https://en.wikipedia.org/wiki/Fuzz_testing), with both [internal tools](programs/fuzzer.c) and [external ones](http://lcamtuf.coredump.cx/afl). Therefore, Zstandard is not considered safe for testings, even within production environments.
 
 ### Branch Policy
+
 The "dev" branch is the one where all contributions will be 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.
+
+
+### Trivia
+
+Zstd entropy stage is provided by [Huff0 and FSE, from Finite State Entropy library](https://github.com/Cyan4973/FiniteStateEntropy).
+
+Its memory requirement can be configured to fit into low-memory hardware configurations, or servers handling multiple connections/contexts in parallel.
+

diff --git a/programs/zstdcli.c b/programs/zstdcli.c
index 564686c622ad68b19ce6eb1143bd683e3462da2e..1fe1b8b3204ecf28c7ef67a0ce43f7eb48df3451 100644 (file)
--- a/programs/zstdcli.c
+++ b/programs/zstdcli.c
@@ -124,7 +124,8 @@ static int usage(const char* programName)
     DISPLAY( "          with no FILE, or when FILE is - , read standard input\n");
     DISPLAY( "Arguments :\n");
     DISPLAY( " -#     : # compression level (1-19, default:1) \n");
-    DISPLAY( " -d     : decompression (default for %s extension)\n", ZSTD_EXTENSION);
+    DISPLAY( " -d     : decompression \n");
+    DISPLAY( " -D file: use `file` as Dictionary \n");
     //DISPLAY( " -z     : force compression\n");
     DISPLAY( " -f     : overwrite output without prompting \n");
     DISPLAY( " -h/-H  : display help/long help and exit\n");
@@ -142,7 +143,6 @@ static int usage_advanced(const char* programName)
     DISPLAY( " -q     : suppress warnings; specify twice to suppress errors too\n");
     DISPLAY( " -m     : multiple input filenames mode \n");
     DISPLAY( " -c     : force write to standard output, even if it is the console\n");
-    DISPLAY( " -D file: use file content as Dictionary \n");
 #ifndef ZSTD_NOBENCH
     DISPLAY( "Benchmark arguments :\n");
     DISPLAY( " -b#    : benchmark file(s), using # compression level (default : 1) \n");