| 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 | @@ -39,29 +39,50 @@ 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 | +#define BIG_SUM2HASH(sum) ((sum)%tablesize) |
| 35 | + |
| 36 | static void build_hash_table(struct sum_struct *s) |
| 37 | { |
| 38 | + static uint32 alloc_size; |
| 39 | int32 i; |
| 40 | |
| 41 | - if (!hash_table) { |
| 42 | - hash_table = new_array(int32, TABLESIZE); |
| 43 | + /* Dynamically calculate the hash table size so that the hash load |
| 44 | + * for big files is about 80%. This number must be odd or s2 will |
| 45 | + * not be able to span the entire set. */ |
| 46 | + tablesize = (uint32)(s->count/8) * 10 + 11; |
| 47 | + if (tablesize < TRADITIONAL_TABLESIZE) |
| 48 | + tablesize = TRADITIONAL_TABLESIZE; |
| 49 | + if (tablesize > alloc_size || tablesize < alloc_size - 16*1024) { |
| 50 | + if (hash_table) |
| 51 | + free(hash_table); |
| 52 | + hash_table = new_array(int32, tablesize); |
| 53 | if (!hash_table) |
| 54 | out_of_memory("build_hash_table"); |
| 55 | + alloc_size = tablesize; |
| 56 | } |
| 57 | |
| 58 | - memset(hash_table, 0xFF, TABLESIZE * sizeof hash_table[0]); |
| 59 | + memset(hash_table, 0xFF, tablesize * sizeof hash_table[0]); |
| 60 | |
| 61 | - for (i = 0; i < s->count; i++) { |
| 62 | - uint32 t = SUM2HASH(s->sums[i].sum1); |
| 63 | - s->sums[i].chain = hash_table[t]; |
| 64 | - hash_table[t] = i; |
| 65 | + if (tablesize == TRADITIONAL_TABLESIZE) { |
| 66 | + for (i = 0; i < s->count; i++) { |
| 67 | + uint32 t = SUM2HASH(s->sums[i].sum1); |
| 68 | + s->sums[i].chain = hash_table[t]; |
| 69 | + hash_table[t] = i; |
| 70 | + } |
| 71 | + } else { |
| 72 | + for (i = 0; i < s->count; i++) { |
| 73 | + uint32 t = BIG_SUM2HASH(s->sums[i].sum1); |
| 74 | + s->sums[i].chain = hash_table[t]; |
| 75 | + hash_table[t] = i; |
| 76 | + } |
| 77 | } |
| 78 | } |
| 79 | |
| 80 | @@ -164,11 +185,16 @@ static void hash_search(int f,struct sum |
| 81 | (double)offset, s2 & 0xFFFF, s1 & 0xFFFF); |
| 82 | } |
| 83 | |
| 84 | - i = hash_table[SUM2HASH2(s1,s2)]; |
| 85 | - if (i < 0) |
| 86 | - goto null_hash; |
| 87 | + if (tablesize == TRADITIONAL_TABLESIZE) { |
| 88 | + if ((i = hash_table[SUM2HASH2(s1,s2)]) < 0) |
| 89 | + goto null_hash; |
| 90 | + sum = (s1 & 0xffff) | (s2 << 16); |
| 91 | + } else { |
| 92 | + sum = (s1 & 0xffff) | (s2 << 16); |
| 93 | + if ((i = hash_table[BIG_SUM2HASH(sum)]) < 0) |
| 94 | + goto null_hash; |
| 95 | + } |
| 96 | |
| 97 | - sum = (s1 & 0xffff) | (s2 << 16); |
| 98 | hash_hits++; |
| 99 | do { |
| 100 | int32 l; |