A patch to try to make really large files get handled without bogging

down in sender-side hashtable searching.

diff --git a/segment_large_hash.diff b/segment_large_hash.diff
new file mode 100644 (file)
index 0000000..f7297bb
--- /dev/null
+++ b/segment_large_hash.diff
@@ -0,0 +1,100 @@
+This patch causes the sender to segment its use of the block-finding
+hashtable for really large files.  This avoids overloading the
+hashtable, and the ensuing slowdown that ensues from trying to find
+blocks in an overloaded hashtable.  This does mean that the sender
+won't notice some migrations of data past segment boundaries, but since
+this only affects files with ~1.6GB or more data, and the blocksize is
+already so large that we only find really large sequences of matching
+data anyway, I don't consider this that big of a loss.
+
+I also decreased the MAX_BLOCK_SIZE value to something more reasonable.
+
+To use this patch, run these commands for a successful build:
+
+    patch -p1 <patches/segment_large_hash.diff
+    ./configure                                 (optional if already run)
+    make
+
+--- old/match.c
++++ new/match.c
+@@ -46,9 +46,9 @@ static int32 *hash_table;
+ #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)
+ {
+-      int32 i;
++      int32 i, end = s->count;
+ 
+       if (!hash_table) {
+               hash_table = new_array(int32, TABLESIZE);
+@@ -58,11 +58,21 @@ static void build_hash_table(struct sum_
+ 
+       memset(hash_table, 0xFF, TABLESIZE * sizeof hash_table[0]);
+ 
+-      for (i = 0; i < s->count; i++) {
++      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;
+       }
++
++      if (verbose > 2) {
++              rprintf(FINFO, "built hash table for entries %ld - %ld\n",
++                      (long)start, (long)end - 1);
++      }
++
++      return end;
+ }
+ 
+ 
+@@ -120,8 +130,8 @@ 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;
+-      int32 k, want_i, backup;
++      OFF_T offset, end, reset = 0;
++      int32 k, want_i, backup, sum_pos = 0;
+       char sum2[SUM_LENGTH];
+       uint32 s1, s2, sum;
+       int more;
+@@ -159,6 +169,11 @@ 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);
++                      reset = sum_pos * s->blength;
++              }
++
+               if (verbose > 4) {
+                       rprintf(FINFO, "offset=%.0f sum=%04x%04x\n",
+                               (double)offset, s2 & 0xFFFF, s1 & 0xFFFF);
+@@ -336,11 +351,6 @@ 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)
+--- old/rsync.h
++++ new/rsync.h
+@@ -121,7 +121,7 @@
+ #define CHUNK_SIZE (32*1024)
+ #define MAX_MAP_SIZE (256*1024)
+ #define IO_BUFFER_SIZE (4092)
+-#define MAX_BLOCK_SIZE ((int32)1 << 29)
++#define MAX_BLOCK_SIZE ((int32)1 << 17)
+ 
+ #define IOERR_GENERAL (1<<0) /* For backward compatibility, this must == 1 */
+ #define IOERR_VANISHED        (1<<1)