This patch makes the processing of large really large files more efficient
by making sure that the sender's hash table is large enough to hold all the
-checksum entries without being overloaded. Unfortunately, the code adds a
-modulus calculation for (up to) every byte of the source file, which slows
-down the code for normal file sizes (e.g. 4 CPU seconds slower on a Pentium
-III when copying a 65 MB file without very much matching data).
+checksum entries without being overloaded.
-This was udapted for the latest codebase from a patch written by Shachar
+Updated to use the current hashtable method when possible, and the new
+hashtable method (which requires a modulus calculation for up to every byte
+of the source file) only on large files that need a larger hashtable size.
+This avoids slowing down files that don't need the extra-large hashtable.
+
+This was updated for the latest codebase from a patch written by Shachar
Shemesh.
+To use this patch, run these commands for a successful build:
+
+ patch -p1 <patches/dynamic_hash.diff
+ ./configure (optional if already run)
+ make
+
--- old/match.c
+++ new/match.c
-@@ -37,24 +37,31 @@ static int total_matches;
+@@ -39,40 +39,51 @@ static int total_matches;
extern struct stats stats;
-#define TABLESIZE (1<<16)
--
++#define TRADITIONAL_TABLESIZE (1<<16)
+
+static uint32 tablesize;
static int32 *hash_table;
--#define SUM2HASH2(s1,s2) (((s1) + (s2)) & 0xFFFF)
--#define SUM2HASH(sum) SUM2HASH2((sum)&0xFFFF,(sum)>>16)
-+#define SUM2HASH(sum) ((sum)%tablesize)
+ #define SUM2HASH2(s1,s2) (((s1) + (s2)) & 0xFFFF)
+ #define SUM2HASH(sum) SUM2HASH2((sum)&0xFFFF,(sum)>>16)
- static void build_hash_table(struct sum_struct *s)
+-static int32 build_hash_table(struct sum_struct *s, int32 start)
++#define BIG_SUM2HASH(sum) ((sum)%tablesize)
++
++static void build_hash_table(struct sum_struct *s)
{
- int32 i;
-+ uint32 prior_size = tablesize;
+- int32 i, end = s->count;
++ static uint32 alloc_size;
++ int32 i;
- if (!hash_table) {
- hash_table = new_array(int32, TABLESIZE);
+ * for big files is about 80%. This number must be odd or s2 will
+ * not be able to span the entire set. */
+ tablesize = (uint32)(s->count/8) * 10 + 11;
-+ if (tablesize < 65537)
-+ tablesize = 65537; /* a prime number */
-+ if (tablesize != prior_size) {
++ if (tablesize < TRADITIONAL_TABLESIZE)
++ tablesize = TRADITIONAL_TABLESIZE;
++ if (tablesize > alloc_size || tablesize < alloc_size - 16*1024) {
+ if (hash_table)
+ free(hash_table);
+ hash_table = new_array(int32, tablesize);
if (!hash_table)
out_of_memory("build_hash_table");
++ alloc_size = tablesize;
}
- memset(hash_table, 0xFF, TABLESIZE * sizeof hash_table[0]);
+-
+- if (end - start > TABLESIZE*8/10)
+- end = start + TABLESIZE*8/10;
+-
+- for (i = start; i < end; i++) {
+- uint32 t = SUM2HASH(s->sums[i].sum1);
+- s->sums[i].chain = hash_table[t];
+- hash_table[t] = i;
+- }
+ memset(hash_table, 0xFF, tablesize * sizeof hash_table[0]);
- for (i = 0; i < s->count; i++) {
- uint32 t = SUM2HASH(s->sums[i].sum1);
-@@ -162,11 +169,11 @@ static void hash_search(int f,struct sum
+- if (verbose > 2) {
+- rprintf(FINFO, "built hash table for entries %ld - %ld\n",
+- (long)start, (long)end - 1);
++ if (tablesize == TRADITIONAL_TABLESIZE) {
++ for (i = 0; i < s->count; i++) {
++ uint32 t = SUM2HASH(s->sums[i].sum1);
++ s->sums[i].chain = hash_table[t];
++ hash_table[t] = i;
++ }
++ } else {
++ for (i = 0; i < s->count; i++) {
++ uint32 t = BIG_SUM2HASH(s->sums[i].sum1);
++ s->sums[i].chain = hash_table[t];
++ hash_table[t] = i;
++ }
+ }
+-
+- return end;
+ }
+
+
+@@ -130,8 +141,8 @@ static void matched(int f, struct sum_st
+ static void hash_search(int f,struct sum_struct *s,
+ struct map_struct *buf, OFF_T len)
+ {
+- OFF_T offset, end, reset = 0;
+- int32 k, want_i, backup, sum_pos = 0;
++ OFF_T offset, end;
++ int32 k, want_i, backup;
+ char sum2[SUM_LENGTH];
+ uint32 s1, s2, sum;
+ int more;
+@@ -169,24 +180,21 @@ static void hash_search(int f,struct sum
+ int done_csum2 = 0;
+ int32 i;
+
+- if (offset >= reset) {
+- sum_pos = build_hash_table(s, sum_pos);
+- if (sum_pos == s->count)
+- reset = len;
+- else
+- reset = sum_pos * s->blength;
+- }
+-
+ if (verbose > 4) {
+ rprintf(FINFO, "offset=%.0f sum=%04x%04x\n",
(double)offset, s2 & 0xFFFF, s1 & 0xFFFF);
}
- i = hash_table[SUM2HASH2(s1,s2)];
-+ sum = (s1 & 0xffff) | (s2 << 16);
-+ i = hash_table[SUM2HASH(sum)];
- if (i < 0)
- goto null_hash;
+- if (i < 0)
+- goto null_hash;
++ if (tablesize == TRADITIONAL_TABLESIZE) {
++ if ((i = hash_table[SUM2HASH2(s1,s2)]) < 0)
++ goto null_hash;
++ sum = (s1 & 0xffff) | (s2 << 16);
++ } else {
++ sum = (s1 & 0xffff) | (s2 << 16);
++ if ((i = hash_table[BIG_SUM2HASH(sum)]) < 0)
++ goto null_hash;
++ }
- sum = (s1 & 0xffff) | (s2 << 16);
hash_hits++;
do {
int32 l;
+@@ -354,6 +362,11 @@ void match_sums(int f, struct sum_struct
+ }
+
+ if (len > 0 && s->count > 0) {
++ build_hash_table(s);
++
++ if (verbose > 2)
++ rprintf(FINFO,"built hash table\n");
++
+ hash_search(f, s, buf, len);
+
+ if (verbose > 2)
Matt McCutchen's Web Site / rsync / rsync-patches.git / blobdiff