| 1 | This patch makes the processing of large really large files more efficient |
| 2 | by making sure that the sender's hash table is large enough to hold all the |
| 3 | checksum entries without being overloaded. |
| 4 | |
| 5 | Updated to use the current hashtable method when possible, and the new |
| 6 | hashtable method (which requires a modulus calculation for up to every byte |
| 7 | of the source file) only on large files that need a larger hashtable size. |
| 8 | This avoids slowing down files that don't need the extra-large hashtable. |
| 9 | |
| 10 | This was updated for the latest codebase from a patch written by Shachar |
| 11 | Shemesh. |
| 12 | |
| 13 | To use this patch, run these commands for a successful build: |
| 14 | |
| 15 | patch -p1 <patches/dynamic_hash.diff |
| 16 | ./configure (optional if already run) |
| 17 | make |
| 18 | |
| 19 | --- old/match.c |
| 20 | +++ new/match.c |
| 21 | @@ -39,40 +39,51 @@ static int total_matches; |
| 22 | |
| 23 | extern struct stats stats; |
| 24 | |
| 25 | -#define TABLESIZE (1<<16) |
| 26 | +#define TRADITIONAL_TABLESIZE (1<<16) |
| 27 | |
| 28 | +static uint32 tablesize; |
| 29 | static int32 *hash_table; |
| 30 | |
| 31 | #define SUM2HASH2(s1,s2) (((s1) + (s2)) & 0xFFFF) |
| 32 | #define SUM2HASH(sum) SUM2HASH2((sum)&0xFFFF,(sum)>>16) |
| 33 | |
| 34 | -static int32 build_hash_table(struct sum_struct *s, int32 start) |
| 35 | +#define BIG_SUM2HASH(sum) ((sum)%tablesize) |
| 36 | + |
| 37 | +static void build_hash_table(struct sum_struct *s) |
| 38 | { |
| 39 | - int32 i, end = s->count; |
| 40 | + static uint32 alloc_size; |
| 41 | + int32 i; |
| 42 | |
| 43 | - if (!hash_table) { |
| 44 | - hash_table = new_array(int32, TABLESIZE); |
| 45 | + /* Dynamically calculate the hash table size so that the hash load |
| 46 | + * for big files is about 80%. A number greater than the traditional |
| 47 | + * size must be odd or s2 will not be able to span the entire set. */ |
| 48 | + tablesize = (uint32)(s->count/8) * 10 + 11; |
| 49 | + if (tablesize < TRADITIONAL_TABLESIZE) |
| 50 | + tablesize = TRADITIONAL_TABLESIZE; |
| 51 | + if (tablesize > alloc_size || tablesize < alloc_size - 16*1024) { |
| 52 | + if (hash_table) |
| 53 | + free(hash_table); |
| 54 | + hash_table = new_array(int32, tablesize); |
| 55 | if (!hash_table) |
| 56 | out_of_memory("build_hash_table"); |
| 57 | + alloc_size = tablesize; |
| 58 | } |
| 59 | |
| 60 | - memset(hash_table, 0xFF, TABLESIZE * sizeof hash_table[0]); |
| 61 | - |
| 62 | - if (end - start > TABLESIZE*8/10) |
| 63 | - end = start + TABLESIZE*8/10; |
| 64 | - |
| 65 | - for (i = start; i < end; i++) { |
| 66 | - uint32 t = SUM2HASH(s->sums[i].sum1); |
| 67 | - s->sums[i].chain = hash_table[t]; |
| 68 | - hash_table[t] = i; |
| 69 | - } |
| 70 | + memset(hash_table, 0xFF, tablesize * sizeof hash_table[0]); |
| 71 | |
| 72 | - if (verbose > 2) { |
| 73 | - rprintf(FINFO, "built hash table for entries %ld - %ld\n", |
| 74 | - (long)start, (long)end - 1); |
| 75 | + if (tablesize == TRADITIONAL_TABLESIZE) { |
| 76 | + for (i = 0; i < s->count; i++) { |
| 77 | + uint32 t = SUM2HASH(s->sums[i].sum1); |
| 78 | + s->sums[i].chain = hash_table[t]; |
| 79 | + hash_table[t] = i; |
| 80 | + } |
| 81 | + } else { |
| 82 | + for (i = 0; i < s->count; i++) { |
| 83 | + uint32 t = BIG_SUM2HASH(s->sums[i].sum1); |
| 84 | + s->sums[i].chain = hash_table[t]; |
| 85 | + hash_table[t] = i; |
| 86 | + } |
| 87 | } |
| 88 | - |
| 89 | - return end; |
| 90 | } |
| 91 | |
| 92 | |
| 93 | @@ -130,8 +141,8 @@ static void matched(int f, struct sum_st |
| 94 | static void hash_search(int f,struct sum_struct *s, |
| 95 | struct map_struct *buf, OFF_T len) |
| 96 | { |
| 97 | - OFF_T offset, end, reset = 0; |
| 98 | - int32 k, want_i, backup, sum_pos = 0; |
| 99 | + OFF_T offset, end; |
| 100 | + int32 k, want_i, backup; |
| 101 | char sum2[SUM_LENGTH]; |
| 102 | uint32 s1, s2, sum; |
| 103 | int more; |
| 104 | @@ -169,24 +180,21 @@ static void hash_search(int f,struct sum |
| 105 | int done_csum2 = 0; |
| 106 | int32 i; |
| 107 | |
| 108 | - if (offset >= reset) { |
| 109 | - sum_pos = build_hash_table(s, sum_pos); |
| 110 | - if (sum_pos == s->count) |
| 111 | - reset = len; |
| 112 | - else |
| 113 | - reset = sum_pos * s->blength; |
| 114 | - } |
| 115 | - |
| 116 | if (verbose > 4) { |
| 117 | rprintf(FINFO, "offset=%.0f sum=%04x%04x\n", |
| 118 | (double)offset, s2 & 0xFFFF, s1 & 0xFFFF); |
| 119 | } |
| 120 | |
| 121 | - i = hash_table[SUM2HASH2(s1,s2)]; |
| 122 | - if (i < 0) |
| 123 | - goto null_hash; |
| 124 | + if (tablesize == TRADITIONAL_TABLESIZE) { |
| 125 | + if ((i = hash_table[SUM2HASH2(s1,s2)]) < 0) |
| 126 | + goto null_hash; |
| 127 | + sum = (s1 & 0xffff) | (s2 << 16); |
| 128 | + } else { |
| 129 | + sum = (s1 & 0xffff) | (s2 << 16); |
| 130 | + if ((i = hash_table[BIG_SUM2HASH(sum)]) < 0) |
| 131 | + goto null_hash; |
| 132 | + } |
| 133 | |
| 134 | - sum = (s1 & 0xffff) | (s2 << 16); |
| 135 | hash_hits++; |
| 136 | do { |
| 137 | int32 l; |
| 138 | @@ -354,6 +362,11 @@ void match_sums(int f, struct sum_struct |
| 139 | } |
| 140 | |
| 141 | if (len > 0 && s->count > 0) { |
| 142 | + build_hash_table(s); |
| 143 | + |
| 144 | + if (verbose > 2) |
| 145 | + rprintf(FINFO,"built hash table\n"); |
| 146 | + |
| 147 | hash_search(f, s, buf, len); |
| 148 | |
| 149 | if (verbose > 2) |