| 1 |
## Edit Distance Match Finder |
| 2 |
|
| 3 |
``` |
| 4 |
/* This match finder leverages techniques used in file comparison algorithms |
| 5 |
* to find matches between a dictionary and a source file. |
| 6 |
* |
| 7 |
* The original motivation for studying this approach was to try and optimize |
| 8 |
* Zstandard for the use case of patching: the most common scenario being |
| 9 |
* updating an existing software package with the next version. When patching, |
| 10 |
* the difference between the old version of the package and the new version |
| 11 |
* is generally tiny (most of the new file will be identical to |
| 12 |
* the old one). In more technical terms, the edit distance (the minimal number |
| 13 |
* of changes required to take one sequence of bytes to another) between the |
| 14 |
* files would be small relative to the size of the file. |
| 15 |
* |
| 16 |
* Various 'diffing' algorithms utilize this notion of edit distance and |
| 17 |
* the corresponding concept of a minimal edit script between two |
| 18 |
* sequences to identify the regions within two files where they differ. |
| 19 |
* The core algorithm used in this match finder is described in: |
| 20 |
* |
| 21 |
* "An O(ND) Difference Algorithm and its Variations", Eugene W. Myers, |
| 22 |
* Algorithmica Vol. 1, 1986, pp. 251-266, |
| 23 |
* <https://doi.org/10.1007/BF01840446>. |
| 24 |
* |
| 25 |
* Additional algorithmic heuristics for speed improvement have also been included. |
| 26 |
* These we inspired from implementations of various regular and binary diffing |
| 27 |
* algorithms such as GNU diff, bsdiff, and Xdelta. |
| 28 |
* |
| 29 |
* Note: after some experimentation, this approach proved to not provide enough |
| 30 |
* utility to justify the additional CPU used in finding matches. The one area |
| 31 |
* where this approach consistently outperforms Zstandard even on level 19 is |
| 32 |
* when compressing small files (<10 KB) using an equally small dictionary that |
| 33 |
* is very similar to the source file. For the use case that this was intended, |
| 34 |
* (large similar files) this approach by itself took 5-10X longer than zstd-19 and |
| 35 |
* generally resulted in 2-3X larger files. The core advantage that zstd-19 has |
| 36 |
* over this approach for match finding is the overlapping matches. This approach |
| 37 |
* cannot find any. |
| 38 |
* |
| 39 |
* I'm leaving this in the contrib section in case this ever becomes interesting |
| 40 |
* to explore again. |
| 41 |
* */ |
| 42 |
``` |