This is the mail archive of the gdb-patches@sources.redhat.com mailing list for the GDB project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [patch/rfc] Tweak bcache performance


Comments? Baring them, I'll commit in a few days.


Wrong patch attached? It doesn't answer my question, but the right patch might :)

Ugger.


I would like to know why bcache has significantly less memory overhead. I haven't looked at this code in a while; I'll find some time to dig in
and remind myself.

Andrew


2003-11-03  Andrew Cagney  <cagney@redhat.com>

	* bcache.h: Update copyright.  Add comments on bcache VS hashtab.
	* bcache.c (struct bstring): Make "length" an unsigned short, add
	"half_hash".
	(struct bcache): Add "half_hash_error_count".
	(bcache): Compute and save the "half_hash".  Compare the
	"half_hash" before comparing the length.  Update
	half_hash_error_count.

Index: bcache.c
===================================================================
RCS file: /cvs/src/src/gdb/bcache.c,v
retrieving revision 1.11
diff -u -r1.11 bcache.c
--- bcache.c	3 Nov 2003 17:03:11 -0000	1.11
+++ bcache.c	3 Nov 2003 19:51:14 -0000
@@ -38,8 +38,15 @@
 
 struct bstring
 {
+  /* Hash chain.  */
   struct bstring *next;
-  size_t length;
+  /* Assume the data length is no more than 64k.  */
+  unsigned short length;
+  /* The half hash hack.  This contains the upper 16 bits of the hash
+     value and is used as a pre-check when comparing two strings and
+     avoids the need to do length or memcmp calls.  It proves to be
+     roughly 100% effective.  */
+  unsigned short half_hash;
 
   union
   {
@@ -79,6 +86,10 @@
      expand_hash_count.  */
   unsigned long expand_count;
   unsigned long expand_hash_count;
+  /* Number of times that the half-hash compare hit (compare the upper
+     16 bits of hash values) hit, but the corresponding combined
+     length/data compare missed.  */
+  unsigned long half_hash_miss_count;
 };
 
 /* The old hash function was stolen from SDBM. This is what DB 3.0 uses now,
@@ -187,6 +198,8 @@
 void *
 bcache (const void *addr, int length, struct bcache *bcache)
 {
+  unsigned long full_hash;
+  unsigned short half_hash;
   int hash_index;
   struct bstring *s;
 
@@ -197,13 +210,24 @@
   bcache->total_count++;
   bcache->total_size += length;
 
-  hash_index = hash (addr, length) % bcache->num_buckets;
-
-  /* Search the hash bucket for a string identical to the caller's.  */
+  full_hash = hash (addr, length);
+  half_hash = (full_hash >> 16);
+  hash_index = full_hash % bcache->num_buckets;
+
+  /* Search the hash bucket for a string identical to the caller's.
+     As a short-circuit first compare the upper part of each hash
+     values.  */
   for (s = bcache->bucket[hash_index]; s; s = s->next)
-    if (s->length == length
-	&& ! memcmp (&s->d.data, addr, length))
-      return &s->d.data;
+    {
+      if (s->half_hash == half_hash)
+	{
+	  if (s->length == length
+	      && ! memcmp (&s->d.data, addr, length))
+	    return &s->d.data;
+	  else
+	    bcache->half_hash_miss_count++;
+	}
+    }
 
   /* The user's string isn't in the list.  Insert it after *ps.  */
   {
@@ -212,6 +236,7 @@
     memcpy (&new->d.data, addr, length);
     new->length = length;
     new->next = bcache->bucket[hash_index];
+    new->half_hash = half_hash;
     bcache->bucket[hash_index] = new;
 
     bcache->unique_count++;
@@ -378,6 +403,8 @@
 		   c->expand_count);
   printf_filtered ("    Hash table hashes:         %lu\n",
 		   c->total_count + c->expand_hash_count);
+  printf_filtered ("    Half hash misses:          %lu\n",
+		   c->half_hash_miss_count);
   printf_filtered ("    Hash table population:     ");
   print_percentage (occupied_buckets, c->num_buckets);
   printf_filtered ("    Median hash chain length:  %3d\n",
Index: bcache.h
===================================================================
RCS file: /cvs/src/src/gdb/bcache.h,v
retrieving revision 1.6
diff -u -r1.6 bcache.h
--- bcache.h	12 Jul 2002 15:23:10 -0000	1.6
+++ bcache.h	3 Nov 2003 19:51:14 -0000
@@ -2,7 +2,7 @@
    Written by Fred Fish <fnf@cygnus.com>
    Rewritten by Jim Blandy <jimb@cygnus.com>
 
-   Copyright 1999, 2000, 2002 Free Software Foundation, Inc.
+   Copyright 1999, 2000, 2002, 2003 Free Software Foundation, Inc.
 
    This file is part of GDB.
 
@@ -48,20 +48,95 @@
    You shouldn't modify the strings you get from a bcache, because:
 
    - You don't necessarily know who you're sharing space with.  If I
-     stick eight bytes of text in a bcache, and then stick an
-     eight-byte structure in the same bcache, there's no guarantee
-     those two objects don't actually comprise the same sequence of
-     bytes.  If they happen to, the bcache will use a single byte
-     string for both of them.  Then, modifying the structure will
-     change the string.  In bizarre ways.
+   stick eight bytes of text in a bcache, and then stick an eight-byte
+   structure in the same bcache, there's no guarantee those two
+   objects don't actually comprise the same sequence of bytes.  If
+   they happen to, the bcache will use a single byte string for both
+   of them.  Then, modifying the structure will change the string.  In
+   bizarre ways.
 
    - Even if you know for some other reason that all that's okay,
-     there's another problem.  A bcache stores all its strings in a
-     hash table.  If you modify a string's contents, you will probably
-     change its hash value.  This means that the modified string is
-     now in the wrong place in the hash table, and future bcache
-     probes will never find it.  So by mutating a string, you give up
-     any chance of sharing its space with future duplicates.  */
+   there's another problem.  A bcache stores all its strings in a hash
+   table.  If you modify a string's contents, you will probably change
+   its hash value.  This means that the modified string is now in the
+   wrong place in the hash table, and future bcache probes will never
+   find it.  So by mutating a string, you give up any chance of
+   sharing its space with future duplicates.
+
+
+   Size of bcache VS hashtab:
+
+   For bcache, the most critical cost is size (or more exactly the
+   overhead added by the bcache).  It turns out that the bcache is
+   remarkably efficient.
+
+   Assuming a 32-bit system (the hash table slots are 4 bytes),
+   ignoring alignment, and limit strings to 255 bytes (1 byte length)
+   we get ...
+
+   bcache: This uses a separate linked list to track the hash chain.
+   The numbers show roughly 100% occupancy of the hash table and an
+   average chain length of 4.  Spreading the slot cost over the 4
+   chain elements:
+
+   4 (slot) / 4 (chain length) + 1 (length) + 4 (chain) = 6 bytes
+
+   hashtab: This uses a more traditional re-hash algorithm where the
+   chain is maintained within the hash table.  The table occupancy is
+   kept below 75% but we'll assume its perfect:
+
+   4 (slot) x 4/3 (occupancy) +  1 (length) = 6 1/3 bytes
+
+   So a perfect hashtab has just slightly larger than an average
+   bcache.
+
+   It turns out that an average hashtab is far worse.  Two things
+   hurt:
+
+   - Hashtab's occupancy is more like 50% (it ranges between 38% and
+   75%) giving a per slot cost of 4x2 vs 4x4/3.
+
+   - the string structure needs to be aligned to 8 bytes which for
+   hashtab wastes 7 bytes, while for bcache wastes only 3.
+
+   This gives:
+
+   hashtab: 4 x 2 + 1 + 7 = 16 bytes
+
+   bcache 4 / 4 + 1 + 4 + 3 = 9 bytes
+
+   The numbers of GDB debugging GDB support this.  ~40% vs ~70% overhead.
+
+
+   Speed of bcache VS hashtab (the half hash hack):
+
+   While hashtab has a typical chain length of 1, bcache has a chain
+   length of round 4.  This means that the bcache will require
+   something like double the number of compares after that initial
+   hash.  In both cases the comparison takes the form:
+
+   a.length == b.length && memcmp (a.data, b.data, a.length) == 0
+
+   That is lengths are checked before doing the memcmp.
+
+   For GDB debugging GDB, it turned out that all lengths were 24 bytes
+   (no C++ so only psymbols were cached) and hence, all compares
+   required a call to memcmp.  As a hack, two bytes of padding
+   (mentioned above) are used to store the upper 16 bits of the
+   string's hash value and then that is used in the comparison vis:
+
+   a.half_hash == b.half_hash && a.length == b.length && memcmp
+   (a.data, b.data, a.length)
+
+   The numbers from GDB debugging GDB show this to be a remarkable
+   100% effective (only necessary length and memcmp tests being
+   performed).
+
+   Mind you, looking at the wall clock, the same GDB debugging GDB
+   showed only marginal speed up (0.780 vs 0.773s).  Seems GDB is too
+   busy doing something else :-(
+  
+*/
 
 
 struct bcache;

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]