数据相似性检测算法




1、引言
  "数据同步算法研究"一文研究了在网络上高效同步数据的方法,其中有个前提是文件A和B非常相似,即两者之间存在大量相同的数据。如果两个文件相似性很低,虽然这种方法依然可以正常工作,但数据同步性能却不会得到提高,甚至会有所降低。因为会产生部分元数据和网络通信消耗,这在两个文件完全不相关时尤为明显。因此,同步数据前需要计算种子文件(seed file)与目标文件之间的相似性,如果相似性大于指定阈值(通常应大于50%)则应用该数据同步算法,否则接传输文件即可。如此,可使得数据同步算法则具有较好的自适应性,在数据具有不同相似性的情形下均可进行高性能的数据同步。另外,在数据相似性检测的基础之上,可对于相似性高的数据进行数据编码处理(如Delta编码),通过一个文件给另一个文件编码的方式进行数据压缩,这是一种基于相似数据检测与编码的重复数据删除技术。
2、相似性计算  Unix diff对文档进行逐行对比来检测相似文件,它采用经典的LCS(Longest Common Subsequence,最长公共子串)算法,运用动态规划方法来计算相似性。LCS的含义是同时包含在字符串里的一个最长字符序列,LCS的长度作为这两个字符串相似性的度量。Diff算法以整行作为"字符"来计算最长公共子串,性能上比字符级的LCS算法快很多。这种方法效率很低,而且只适用文本文件的相似比较,不能直接适用于二进制文件。

  目前通常的做法是将文件相似性问题转换为集合相似性问题,如基于shingle的计算方法和基于bloom filter的计算方法,这两种方法都可适用于任何格式的数据文件。这种方式的核心思想是为每个文件提取组特征值,以特征值集合来计算相似性,从而降低计算复杂性来提高性能。shingle用特征值交集来计算相似性会导致高计算和空间开销,bloom filter技术在计算开销和匹配精度上更具势。Bloom filter所定义的集合元素是文件按照CDC(content-defined chunking)算法所切分数据块的指纹值,其相似性定义如下:
                            |fingerprints(f1) ∩ fingerprints(f2)|
  Sim(f1, f2) = ---------------------------------------------   (公式1)
                            |fingerprints(f1) ∪ fingerprints(f2)|

  另外一种方法,是将二进制文件进行切块,使用数据块指纹来表示数据块,然后将数据块映射为"字符",再应用LCS算法寻找最大公共子串并计算出相似度。其相似性定义如下:
                        2 * length(LCS(fingerprints(f1), fingerprints(f2)))
    Sim(f1, f2) = ------------------------------------------------------------------ (公式2)
                           length(fingerprints(f1)) + length(fingerprints(f2))

  上面两种相似性算法中均采用数据切分技术,数据块可以是定长或变长。为了相似性计算的精确性,实现中采用以数据块长度作为权的加权计算方法。

3、Bloom filter算法  该文件相似性计算流程如下:
      (1) 采用CDC算法将文件切分成数据块集,并为每个数据块计算MD5指纹;
      (2) 计算两个指纹集合的交集和并集,通过hashtable来实现;
      (3) 按照公式1计算文件相似性,考虑重复数据块和数据块长度来提高计算精确度。
      详细参见附录bsim源码中的file_chunk,chunk_file_process和similarity_detect函数实现。

4、LCS算法  该文件相似性计算流程如下:
    (1) 采用CDC算法将文件切分成数据块集,并为每个数据块计算MD5指纹;
    (2) 将MD5指纹串映射为"字符",则文件转换为"字符串"表示;
    (3) 应用LCS算法计算出最长公共子串,并计算其加权长度;
    (4) 按照公式2计算文件相似性,考虑重复数据块和数据块长度来提高计算精确度。
    详细参见附录bsim源码中的file_chunk,chunk_file_process,LCS和similarity_detect函数实现。

5、算法分析比较
 两种算法都对文件进行切分操作,假设文件f1切为m个块,文件f2切分成n个块。Bloom filter算法没有考虑数据块顺序,因此在相似性精确度方面要低于LCS算法,其时间和空间复杂性都是O(m + n)。相反,LCS算法考虑了数据块顺序问题,相似性度量相对精确,然而其时间和空间复杂性是O(mn),这个大大限制了应用规模。综合来看,Bloom filter算法精确度比LCS算法要低,但计算消耗要小很多,性能和适用性非常好。LCS比较适合精确的文件相似性计算,这些文件往往比较小,50MB以内比较合适。对于重复数据删除和网络数据同步来说,消重效果和性能与数据块顺序性无关,因此Bloom filter算法计算的数据相似性更适用,性能也更高。

附录:bsim.c源码(完整源码请参见deduputil源码)


[cpp]  view plain copy print ?


    1. /* Copyright (C) 2010 Aigui Liu
    2.  *
    3.  * This program is free software; you can redistribute it and/or modify
    4.  * it under the terms of the GNU General Public License as published by
    5.  * the Free Software Foundation; either version 3 of the License, or
    6.  * (at your option) any later version.
    7.  *
    8.  * This program is distributed in the hope that it will be useful,
    9.  * but WITHOUT ANY WARRANTY; without even the implied warranty of
    10.  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
    11.  * GNU General Public License for more details.
    12.  *
    13.  * You should have received a copy of the GNU General Public License along
    14.  * with this program; if not, visit the http://fsf.org website.
    15.  */  
    16. #include <stdio.h>  
    17. #include <stdlib.h>  
    18. #include <string.h>  
    19. #include <sys/types.h>  
    20. #include <sys/stat.h>  
    21. #include <fcntl.h>  
    22. #include <unistd.h>  
    23. #include "hashtable.h"  
    24. #include "sync.h"  
    25. #define NEITHER       0  
    26. #define UP            1  
    27. #define LEFT          2  
    28. #define UP_AND_LEFT   3  
    29. #define MAX(x, y) (((x) > (y)) ? (x) : (y))  
    30. #define MIN(x, y) (((x) < (y)) ? (x) : (y))  
    31. #define MD5_LEN 17  
    32. enum {  
    33.     FILE1 = 0,  
    34.     FILE2  
    35. };  
    36. enum {  
    37.     LCS_NOT = 0,  
    38.     LCS_YES  
    39. };  
    40. typedef struct {  
    41.     uint32_t nr1;  
    42.     uint32_t nr2;  
    43.     uint32_t len;  
    44. } hash_entry;  
    45. typedef struct {  
    46. char **str;  
    47.     uint32_t len;  
    48. } lcs_entry;  
    49. static uint32_t sim_union = 0;  
    50. static uint32_t sim_intersect = 0;  
    51. static void usage()  
    52. {  
    53. "Usage: bsim FILE1 FILE2 CHUNK_ALGO LCS/n/n");  
    54. "Similarity detect between FILE1 and FILE2 based on block level./n");  
    55. "CHUNK_ALGO:/n");  
    56. "  FSP - fixed-size partition/n");  
    57. "  CDC - content-defined chunking/n");  
    58. "  SBC - slide block chunking/n/n");  
    59. "LCS:/n");  
    60. "  LCS_NOT - do not use LCS(longest lommon subsequence) algorithms/n");  
    61. "  LCS_YES - use LCS algorithms/n/n");  
    62. "Report bugs to <Aigui.Liu@>./n");  
    63. }  
    64. static int parse_arg(char *argname)  
    65. {  
    66. if (0 == strcmp(argname, "FSP"))  
    67. return CHUNK_FSP;  
    68. else if (0 == strcmp(argname, "CDC"))  
    69. return CHUNK_CDC;  
    70. else if (0 == strcmp(argname, "SBC"))  
    71. return CHUNK_SBC;  
    72. else if (0 == strcmp(argname, "LCS_NOT"))  
    73. return LCS_NOT;  
    74. else if (0 == strcmp(argname, "LCS_YES"))  
    75. return LCS_YES;  
    76. else  
    77. return -1;  
    78. }  
    79. static char **alloc_2d_array(int row, int col)  
    80. {  
    81. int i;  
    82. char *p, **pp;  
    83. char *)malloc(row * col * sizeof(char));  
    84. char **)malloc(row * sizeof(char *));  
    85. if (p == NULL || pp == NULL)  
    86. return NULL;  
    87. for (i = 0; i < row; i++) {  
    88.                 pp[i] = p + col * i;  
    89.         }  
    90. return pp;  
    91. }  
    92. static void free_2d_array(char **str)  
    93. {  
    94.     free(str[0]);  
    95.     free(str);  
    96. }  
    97. static void show_md5_hex(unsigned char md5_checksum[16])  
    98. {  
    99. int i;  
    100. for (i = 0; i < 16; i++) {  
    101. "%02x", md5_checksum[i]);  
    102.         }  
    103. "/n");  
    104. }  
    105. static int chunk_file_process(char *chunk_file, hashtable *htab, int which, int sim_algo, lcs_entry *le)  
    106. {  
    107. int fd, i, ret = 0;  
    108.     ssize_t rwsize;  
    109.     chunk_file_header chunk_file_hdr;  
    110.     chunk_block_entry chunk_bentry;  
    111.     hash_entry *he = NULL;  
    112. /* parse chunk file */  
    113.     fd = open(chunk_file, O_RDONLY);  
    114. if (-1 == fd) {  
    115. return -1;  
    116.     }  
    117.     rwsize = read(fd, &chunk_file_hdr, CHUNK_FILE_HEADER_SZ);  
    118. if (rwsize != CHUNK_FILE_HEADER_SZ) {  
    119.         ret = -1;  
    120. goto _CHUNK_FILE_PROCESS_EXIT;  
    121.     }  
    122. if (sim_algo == LCS_YES) {  
    123.         le->str = alloc_2d_array(chunk_file_hdr.block_nr, MD5_LEN);  
    124. if (le->str == NULL) {  
    125.             ret = -1;  
    126. goto _CHUNK_FILE_PROCESS_EXIT;  
    127.         }  
    128.         le->len = chunk_file_hdr.block_nr;  
    129.     }  
    130. for(i = 0; i < chunk_file_hdr.block_nr; i++) {  
    131.         rwsize = read(fd, &chunk_bentry, CHUNK_BLOCK_ENTRY_SZ);  
    132. if (rwsize != CHUNK_BLOCK_ENTRY_SZ) {  
    133.             ret = -1;  
    134. goto _CHUNK_FILE_PROCESS_EXIT;  
    135.         }  
    136. void *)chunk_bentry.md5, htab);  
    137. if (he == NULL) {  
    138. sizeof(hash_entry));  
    139.             he->nr1 = he->nr2 = 0;  
    140.             he->len = chunk_bentry.len;  
    141.         }  
    142.         (which == FILE1) ? he->nr1++ : he->nr2++;  
    143. /* insert or update hash entry */  
    144. void *)strdup(chunk_bentry.md5), (void *)he, htab);  
    145. if (sim_algo == LCS_YES) {  
    146.             memcpy(le->str[i], chunk_bentry.md5, MD5_LEN);  
    147.         }  
    148.     }  
    149. _CHUNK_FILE_PROCESS_EXIT:  
    150.     close(fd);  
    151. return ret;  
    152. }  
    153. uint32_t LCS(char** a, int n, char** b, int m, hashtable *htab)   
    154. {  
    155. int** S;  
    156. int** R;  
    157. int ii;  
    158. int jj;  
    159. int pos;  
    160.         uint32_t len = 0;  
    161.     hash_entry *he = NULL;  
    162. /* Memory allocation */  
    163. int **)malloc( (n+1) * sizeof(int *) );  
    164. int **)malloc( (n+1) * sizeof(int *) );  
    165. if (S == NULL || R == NULL) {  
    166. "malloc for S and R in LCS");  
    167.         exit(0);  
    168.     }  
    169. for(ii = 0; ii <= n; ++ii) {  
    170. int*) malloc( (m+1) * sizeof(int) );  
    171. int*) malloc( (m+1) * sizeof(int) );  
    172. if (S[ii] == NULL || R[ii] == NULL) {  
    173. "malloc for S[ii] and R[ii] in LCS");  
    174.             exit(0);  
    175.         }  
    176.         }  
    177. /* It is important to use <=, not <.  The next two for-loops are initialization */  
    178. for(ii = 0; ii <= n; ++ii) {  
    179.                 S[ii][0] = 0;  
    180.                 R[ii][0] = UP;  
    181.         }  
    182. for(jj = 0; jj <= m; ++jj) {  
    183.                 S[0][jj] = 0;  
    184.                 R[0][jj] = LEFT;  
    185.         }  
    186. /* This is the main dynamic programming loop that computes the score and */  
    187. /* backtracking arrays. */  
    188. for(ii = 1; ii <= n; ++ii) {  
    189. for(jj = 1; jj <= m; ++jj) {  
    190. if (strcmp(a[ii-1], b[jj-1]) == 0) {  
    191.                                 S[ii][jj] = S[ii-1][jj-1] + 1;  
    192.                                 R[ii][jj] = UP_AND_LEFT;  
    193.                         }  
    194. else {  
    195.                                 S[ii][jj] = S[ii-1][jj-1] + 0;  
    196.                                 R[ii][jj] = NEITHER;  
    197.                         }  
    198. if( S[ii-1][jj] >= S[ii][jj] ) {  
    199.                                 S[ii][jj] = S[ii-1][jj];  
    200.                                 R[ii][jj] = UP;  
    201.                         }  
    202. if( S[ii][jj-1] >= S[ii][jj] ) {  
    203.                                 S[ii][jj] = S[ii][jj-1];  
    204.                                 R[ii][jj] = LEFT;  
    205.                         }  
    206.                 }  
    207.         }  
    208. /* The length of the longest substring is S[n][m] */  
    209.         ii = n;  
    210.         jj = m;  
    211.         pos = S[ii][jj];  
    212. /* Trace the backtracking matrix. */  
    213. while( ii > 0 || jj > 0 ) {  
    214. if( R[ii][jj] == UP_AND_LEFT ) {  
    215.                         ii--;  
    216.                         jj--;  
    217. //lcs[pos--] = a[ii];  
    218. void *)a[ii], htab);  
    219.             len += ((he == NULL) ? 0: he->len);  
    220.                 }  
    221. else if( R[ii][jj] == UP ) {  
    222.                         ii--;  
    223.                 }  
    224. else if( R[ii][jj] == LEFT ) {  
    225.                         jj--;  
    226.                 }  
    227.         }  
    228. for(ii = 0; ii <= n; ++ii ) {  
    229.                 free(S[ii]);  
    230.                 free(R[ii]);  
    231.         }  
    232.         free(S);  
    233.         free(R);  
    234. return len;  
    235. }  
    236. int hash_callback(void *key, void *data)  
    237. {  
    238.     hash_entry *he = (hash_entry *)data;  
    239.     sim_union += (he->len * (he->nr1 + he->nr2));  
    240.     sim_intersect += (he->len * MIN(he->nr1, he->nr2));  
    241. }  
    242. static float similarity_detect(hashtable *htab, char **str1, int n, char **str2, int m, int sim_algo)  
    243. {  
    244.     uint32_t lcs_len = 0;  
    245.     hash_for_each_do(htab, hash_callback);  
    246. if (sim_algo == LCS_YES) {  
    247.         lcs_len = LCS(str1, n, str2, m, htab);  
    248. return lcs_len * 2.0 / sim_union;  
    249. else { /* LCS_NOT */  
    250. return sim_intersect * 2.0 / sim_union;  
    251.     }  
    252. }  
    253. int main(int argc, char *argv[])  
    254. {  
    255. int chunk_algo = CHUNK_CDC;  
    256. int sim_algo = LCS_NOT;  
    257. char *file1 = NULL;  
    258. char *file2 = NULL;  
    259.     lcs_entry le1, le2;  
    260. char tmpname[NAME_MAX_SZ] = {0};  
    261. char template[] = "deduputil_bsim_XXXXXX";  
    262.     hashtable *htab = NULL;  
    263. int ret = 0;  
    264. if (argc < 5) {  
    265.         usage();  
    266. return -1;  
    267.     }  
    268. /* parse chunk algorithms */  
    269.     file1 = argv[1];  
    270.     file2 = argv[2];  
    271.     chunk_algo = parse_arg(argv[3]);  
    272.     sim_algo = parse_arg(argv[4]);  
    273. if (chunk_algo == -1 || sim_algo == -1) {  
    274.         usage();  
    275. return -1;  
    276.     }  
    277.     htab = create_hashtable(HASHTABLE_BUCKET_SZ);  
    278. if (htab == NULL) {  
    279. "create hashtabke failed/n");  
    280. return -1;  
    281.     }  
    282. /* chunk file1 and file2 into blocks */  
    283. "/tmp/%s_%d", mktemp(template), getpid());  
    284.     ret = file_chunk(file1, tmpname, chunk_algo);  
    285. if (0 != ret) {  
    286. "chunk %s failed/n", file1);  
    287. goto _BENCODE_EXIT;  
    288.     }  
    289.     le1.str = NULL;  
    290.     ret = chunk_file_process(tmpname, htab, FILE1, sim_algo, &le1);  
    291. if (ret != 0) {  
    292. "pasre %s failed/n", file1);  
    293. goto _BENCODE_EXIT;  
    294.     }  
    295.     ret = file_chunk(file2, tmpname, chunk_algo);  
    296. if (0 != ret){  
    297. "chunk %s failed/n", file2);  
    298. goto _BENCODE_EXIT;  
    299.     }  
    300.     le2.str = NULL;  
    301.     ret = chunk_file_process(tmpname, htab, FILE2, sim_algo, &le2);  
    302. if (ret != 0) {  
    303. "pasre %s failed/n", file2);  
    304. goto _BENCODE_EXIT;  
    305.     }  
    306. "similarity = %.4f/n", similarity_detect(htab, le1.str, le1.len, le2.str, le2.len, sim_algo));  
    307. _BENCODE_EXIT:  
    308.     unlink(tmpname);  
    309.     hash_free(htab);  
    310. if (le1.str) free_2d_array(le1.str);  
    311. if (le2.str) free_2d_array(le2.str);  
    312. return ret;  
    313. }