Fix performance problems in multi-batch hash joins by ensuring that we select

a well-randomized batch number even when given a poorly-randomized hash value.
This is a bit inefficient but seems the only practical solution given the
constraint that we can't change the hash functions in released branches.
Per report from Joseph Shraibman.

Applied to 8.1 and 8.2 only --- HEAD is getting a cleaner fix, and 8.0 and
before use different coding that seems less vulnerable.

diff --git a/src/backend/access/hash/hashfunc.c b/src/backend/access/hash/hashfunc.c
index d8e5732695e449445442fc483600c2a6a27c6f4d..89507fb8162c16c370253b67794a84b84ad38f5c 100644 (file)
--- a/src/backend/access/hash/hashfunc.c
+++ b/src/backend/access/hash/hashfunc.c
@@ -267,6 +267,31 @@ hash_any(register const unsigned char *k, register int keylen)
                        /* case 0: nothing left to add */
        }
        mix(a, b, c);
+
+       /* report the result */
+       return UInt32GetDatum(c);
+}
+
+/*
+ * hash_uint32() -- hash a 32-bit value
+ *
+ * This has the same result (at least on little-endian machines) as
+ *             hash_any(&k, sizeof(uint32))
+ * but is faster and doesn't force the caller to store k into memory.
+ */
+Datum
+hash_uint32(uint32 k)
+{
+       register uint32 a,
+                               b,
+                               c;
+
+       a = 0x9e3779b9 + k;
+       b = 0x9e3779b9;
+       c = 3923095 + (uint32) sizeof(uint32);
+
+       mix(a, b, c);
+
        /* report the result */
        return UInt32GetDatum(c);
 }

diff --git a/src/backend/executor/nodeHash.c b/src/backend/executor/nodeHash.c
index d0fe9704d1a9d455e5d8650e640bc343e68e0b1e..40193ad085c26327af2888f46e5b1060b99c85e5 100644 (file)
--- a/src/backend/executor/nodeHash.c
+++ b/src/backend/executor/nodeHash.c
@@ -20,6 +20,7 @@
  */
 #include "postgres.h"
 
+#include "access/hash.h"
 #include "executor/execdebug.h"
 #include "executor/hashjoin.h"
 #include "executor/instrument.h"
@@ -713,9 +714,11 @@ ExecHashGetHashValue(HashJoinTable hashtable,
  * chains), and must only cause the batch number to remain the same or
  * increase.  Our algorithm is
  *             bucketno = hashvalue MOD nbuckets
- *             batchno = (hashvalue DIV nbuckets) MOD nbatch
- * where nbuckets should preferably be prime so that all bits of the
- * hash value can affect both bucketno and batchno.
+ *             batchno = hash_uint32(hashvalue) MOD nbatch
+ * which gives reasonably independent bucket and batch numbers in the face
+ * of some rather poorly-implemented hash functions in hashfunc.c.  (This
+ * will change in PG 8.3.)
+ *
  * nbuckets doesn't change over the course of the join.
  *
  * nbatch is always a power of 2; we increase it only by doubling it.  This
@@ -734,7 +737,7 @@ ExecHashGetBucketAndBatch(HashJoinTable hashtable,
        {
                *bucketno = hashvalue % nbuckets;
                /* since nbatch is a power of 2, can do MOD by masking */
-               *batchno = (hashvalue / nbuckets) & (nbatch - 1);
+               *batchno = hash_uint32(hashvalue) & (nbatch - 1);
        }
        else
        {

diff --git a/src/include/access/hash.h b/src/include/access/hash.h
index 6f7863bd676a84311a3d9da7e0af86a8b2ecc2f7..1b9781336355a8edab197e3258454100c51f75df 100644 (file)
--- a/src/include/access/hash.h
+++ b/src/include/access/hash.h
@@ -263,6 +263,7 @@ extern Datum hashname(PG_FUNCTION_ARGS);
 extern Datum hashtext(PG_FUNCTION_ARGS);
 extern Datum hashvarlena(PG_FUNCTION_ARGS);
 extern Datum hash_any(register const unsigned char *k, register int keylen);
+extern Datum hash_uint32(uint32 k);
 
 /* private routines */