Abstract

I offer you a new hash function for hash table lookup that is faster and more thorough than the one you are using now. I also give you a way to verify that it is more thorough.

The Hash

Over the past two years I've built a general hash function for hash table lookup. Most of the two dozen old hashes I've replaced have had owners who wouldn't accept a new hash unless it was a plug-in replacement for their old hash, and was demonstrably better than the old hash.

These old hashes defined my requirements:

  • The keys are unaligned variable-length byte arrays.
  • Sometimes keys are several such arrays.
  • Sometimes a set of independent hash functions were required.
  • Average key lengths ranged from 8 bytes to 200 bytes.
  • Keys might be character strings, numbers, bit-arrays, or weirder things.
  • Table sizes could be anything, including powers of 2.
  • The hash must be faster than the old one.
  • The hash must do a good job.

Without further ado, here's the fastest hash I've been able to design that meets all the requirements. The comments describe how to use it.

typedef  unsigned long  int  ub4;   /* unsigned 4-byte quantities */
typedef  unsigned       char ub1;   /* unsigned 1-byte quantities */

#define hashsize(n) ((ub4)1<<(n))
#define hashmask(n) (hashsize(n)-1)

/*
--------------------------------------------------------------------
mix -- mix 3 32-bit values reversibly.
For every delta with one or two bits set, and the deltas of all three
  high bits or all three low bits, whether the original value of a,b,c
  is almost all zero or is uniformly distributed,
* If mix() is run forward or backward, at least 32 bits in a,b,c
  have at least 1/4 probability of changing.
* If mix() is run forward, every bit of c will change between 1/3 and
  2/3 of the time.  (Well, 22/100 and 78/100 for some 2-bit deltas.)
mix() was built out of 36 single-cycle latency instructions in a 
  structure that could supported 2x parallelism, like so:
      a -= b; 
      a -= c; x = (c>>13);
      b -= c; a ^= x;
      b -= a; x = (a<<8);
      c -= a; b ^= x;
      c -= b; x = (b>>13);
      ...
  Unfortunately, superscalar Pentiums and Sparcs can't take advantage 
  of that parallelism.  They've also turned some of those single-cycle
  latency instructions into multi-cycle latency instructions.  Still,
  this is the fastest good hash I could find.  There were about 2^^68
  to choose from.  I only looked at a billion or so.
--------------------------------------------------------------------
*/
#define mix(a,b,c) \
{ \
  a -= b; a -= c; a ^= (c>>13); \
  b -= c; b -= a; b ^= (a<<8); \
  c -= a; c -= b; c ^= (b>>13); \
  a -= b; a -= c; a ^= (c>>12);  \
  b -= c; b -= a; b ^= (a<<16); \
  c -= a; c -= b; c ^= (b>>5); \
  a -= b; a -= c; a ^= (c>>3);  \
  b -= c; b -= a; b ^= (a<<10); \
  c -= a; c -= b; c ^= (b>>15); \
}

/*
--------------------------------------------------------------------
hash() -- hash a variable-length key into a 32-bit value
  k       : the key (the unaligned variable-length array of bytes)
  len     : the length of the key, counting by bytes
  initval : can be any 4-byte value
Returns a 32-bit value.  Every bit of the key affects every bit of
the return value.  Every 1-bit and 2-bit delta achieves avalanche.
About 6*len+35 instructions.

The best hash table sizes are powers of 2.  There is no need to do
mod a prime (mod is sooo slow!).  If you need less than 32 bits,
use a bitmask.  For example, if you need only 10 bits, do
  h = (h & hashmask(10));
In which case, the hash table should have hashsize(10) elements.

If you are hashing n strings (ub1 **)k, do it like this:
  for (i=0, h=0; i<n; ++i) h = hash( k[i], len[i], h);

By Bob Jenkins, 1996.  bob_jenkins@burtleburtle.net.  You may use this
code any way you wish, private, educational, or commercial.  It's free.

See http://burtleburtle.net/bob/hash/evahash.html
Use for hash table lookup, or anything where one collision in 2^^32 is
acceptable.  Do NOT use for cryptographic purposes.
--------------------------------------------------------------------
*/

ub4 hash( k, length, initval)
register ub1 *k;        /* the key */
register ub4  length;   /* the length of the key */
register ub4  initval;  /* the previous hash, or an arbitrary value */
{
   register ub4 a,b,c,len;

   /* Set up the internal state */
   len = length;
   a = b = 0x9e3779b9;  /* the golden ratio; an arbitrary value */
   c = initval;         /* the previous hash value */

   /*---------------------------------------- handle most of the key */
   while (len >= 12)
   {
      a += (k[0] +((ub4)k[1]<<8) +((ub4)k[2]<<16) +((ub4)k[3]<<24));
      b += (k[4] +((ub4)k[5]<<8) +((ub4)k[6]<<16) +((ub4)k[7]<<24));
      c += (k[8] +((ub4)k[9]<<8) +((ub4)k[10]<<16)+((ub4)k[11]<<24));
      mix(a,b,c);
      k += 12; len -= 12;
   }

   /*------------------------------------- handle the last 11 bytes */
   c += length;
   switch(len)              /* all the case statements fall through */
   {
   case 11: c+=((ub4)k[10]<<24);
   case 10: c+=((ub4)k[9]<<16);
   case 9 : c+=((ub4)k[8]<<8);
      /* the first byte of c is reserved for the length */
   case 8 : b+=((ub4)k[7]<<24);
   case 7 : b+=((ub4)k[6]<<16);
   case 6 : b+=((ub4)k[5]<<8);
   case 5 : b+=k[4];
   case 4 : a+=((ub4)k[3]<<24);
   case 3 : a+=((ub4)k[2]<<16);
   case 2 : a+=((ub4)k[1]<<8);
   case 1 : a+=k[0];
     /* case 0: nothing left to add */
   }
   mix(a,b,c);
   /*-------------------------------------------- report the result */
   return c;
}

Most hashes can be modeled like this:

  initialize(internal state)
  for (each text block)
  {
    combine(internal state, text block);
    mix(internal state);
  }
  return postprocess(internal state);

In the new hash, mix() takes 3n of the 6n+35 instructions needed to hash n bytes. Blocks of text are combined with the internal state (a,b,c) by addition. This combining step is the rest of the hash function, consuming the remaining 3n instructions. The only postprocessing is to choose c out of (a,b,c) to be the result.

Three tricks promote speed:

  1. Mixing is done on three 4-byte registers rather than on a 1-byte quantity.
  2. Combining is done on 12-byte blocks, reducing the loop overhead.
  3. The final switch statement combines a variable-length block with the registers a,b,c without a loop.

The golden ratio really is an arbitrary value. Its purpose is to avoid mapping all zeros to all zeros.

The Hash Must Do a Good Job

The most interesting requirement was that the hash must be better than its competition. What does it mean for a hash to be good for hash table lookup?

A good hash function distributes hash values uniformly. If you don't know the keys before choosing the function, the best you can do is map an equal number of possible keys to each hash value. If keys were distributed uniformly, an excellent hash would be to choose the first few bytes of the key and use that as the hash value. Unfortunately, real keys aren't uniformly distributed. Choosing the first few bytes works quite poorly in practice.

The real requirement then is that a good hash function should distribute hash values uniformly for the keys that users actually use.

How do we test that? Let's look at some typical user data. (Since I work at Oracle, I'll use Oracle's standard example: the EMP table.)

The EMP table. Is this data uniformly distributed?
EMPNO ENAME JOB MGR HIREDATE SAL COMM DEPTNO
7369 SMITH CLERK 7902 17-DEC-80 800 20
7499 ALLEN SALESMAN 7698 20-FEB-81 1600 300 30
7521 WARD SALESMAN 7698 22-FEB-81 1250 500 30
7566 JONES MANAGER 7839 02-APR-81 2975 20
7654 MARTIN SALESMAN 7898 28-SEP-81 1250 1400 30
7698 BLAKE MANAGER 7539 01-MAY-81 2850 30
7782 CLARK MANAGER 7566 09-JUN-81 2450 10
7788 SCOTT ANALYST 7698 19-APR-87 3000 20
7839 KING PRESIDENT 17-NOV-81 5000 10
7844 TURNER SALESMAN 7698 08-SEP-81 1500 30
7876 ADAMS CLERK 7788 23-MAY-87 1100 0 20
7900 JAMES CLERK 7698 03-DEC-81 950 30
7902 FORD ANALYST 7566 03-DEC-81 3000 20
7934 MILLER CLERK 7782 23-JAN-82 1300 10

Consider each horizontal row to be a key. Some patterns appear.

  1. Keys often differ in only a few bits. For example, all the keys are ASCII, so the high bit of every byte is zero.
  2. Keys often consist of substrings arranged in different orders. For example, the MGR of some keys is the EMPNO of others.
  3. Length matters. The only difference between zero and no value at all may be the length of the value. Also, "aa aaa" and "aaa aa" should hash to different values.
  4. Some keys are mostly zero, with only a few bits set. (That pattern doesn't appear in this example, but it's a common pattern.)

Some patterns are easy to handle. If the length is included in the data being hashed, then lengths are not a problem. If the hash does not treat text blocks commutatively, then substrings are not a problem. Strings that are mostly zeros can be tested by listing all strings with only one bit set and checking if that set of strings produces too many collisions.

The remaining pattern is that keys often differ in only a few bits. If a hash allows small sets of input bits to cancel each other out, and the user keys differ in only those bits, then all keys will map to the same handful of hash values.

A common weakness

Usually, when a small set of input bits cancel each other out, it is because those input bits affect only a smaller set of bits in the internal state.

Consider this hash function:

  for (hash=0, i=0; i<hash; ++i)
    hash = ((hash<<5)^(hash>>27))^key[i];
  return (hash % prime);
This function maps the strings "EXXXXXB" and "AXXXXXC" to the same value. These keys differ in bit 3 of the first byte and bit 1 of the seventh byte. After the seventh bit is combined, any further postprocessing will do no good because the internal states are already the same.

Any time n input bits can only affect m output bits, and n > m, then the 2n keys that differ in those input bits can only produce 2m distinct hash values. The same is true if n input bits can only affect m bits of the internal state -- later mixing may make the 2m results look uniformly distributed, but there will still be only 2m results.

The function above has many sets of 2 bits that affect only 1 bit of the internal state. If there are n input bits, there are (n choose 2)=(n*n/2 - n/2) pairs of input bits, only a few of which match weaknesses in the function above. It is a common pattern for keys to differ in only a few bits. If those bits match one of a hash's weaknesses, which is a rare but not negligible event, the hash will do extremely bad. In most cases, though, it will do just fine. (This allows a function to slip through sanity checks, like hashing an English dictionary uniformly, while still frequently bombing on user data.) A weakness

In hashes built of repeated combine-mix steps, this is what usually causes this weakness:

  1. A small number of bits y of one input block are combined, affecting only y bits of the internal state. So far so good.
  2. The mixing step causes those y bits of the internal state to affect only z bits of the internal state.
  3. The next combining step overwrites those bits with z more input bits, cancelling out the first y input bits.
When z is smaller than the number of bits in the output, then y+z input bits have affected only z bits of the internal state, causing 2y+z possible keys to produce at most 2z hash values.

The same thing can happen in reverse:

  1. Uncombine this block, causing y block bits to unaffect y bits of the internal state.
  2. Unmix the internal state, leaving x bits unaffected by the y bits from this block.
  3. Unmixing the previous block unaffects those x bits, cancelling out this block's y bits.
If x is less than the number of bits in the output, then the 2x+y keys differing in only those x+y input bits can produce at most 2x hash values.

(If the mixing function is not a permutation of the internal state, it is not reversible. Instead, it loses information about the earlier blocks every time it is applied, so keys differing only in the first few input blocks are more likely to collide. The mixing function ought to be a permutation.)

It is easy to test whether this weakness exists: if the mixing step causes any bit of the internal state to affect fewer bits of the internal state than there are output bits, the weakness exists. This test should be run on the reverse of the mixing function as well. It can also be run with all sets of 2 internal state bits, or all sets of 3.

Another way this weakness can happen is if any bit in the final input block does not affect every bit of the output. (The user might choose to use only the unaffected output bit, then that's 1 input bit that affects 0 output bits.)

A Survey of Hash Functions

We now have a new hash function and some theory for evaluating hash functions. Let's see how various hash functions stack up.

Additive Hash

ub4 additive(char *key, ub4 len, ub4 prime)
{
  ub4 hash, i;
  for (hash=len, i=0; i<len; ++i) 
    hash += key[i];
  return (hash % prime);
}
This takes 5n+3 instructions. There is no mixing step. The combining step handles one byte at a time. Input bytes commute. The table length must be prime, and can't be much bigger than one byte because the value of variable hash is never much bigger than one byte.

Rotating Hash

ub4 rotating(char *key, ub4 len, ub4 prime)
{
  ub4 hash, i;
  for (hash=len, i=0; i<len; ++i)
    hash = (hash<<4)^(hash>>28)^key[i];
  return (hash % prime);
}
This takes 8n+3 instructions. This is the same as the additive hash, except it has a mixing step (a circular shift by 4) and the combining step is exclusive-or instead of addition. The table size is a prime, but the prime can be any size. On machines with a rotate (such as the Intel x86 line) this is 6n+2 instructions. I have seen the (hash % prime) replaced with
  hash = (hash ^ (hash>>10) ^ (hash>>20)) & mask;
eliminating the % and allowing the table size to be a power of 2, making this 6n+6 instructions. % can be very slow, for example it is 230 times slower than addition on a Sparc 20.

One-at-a-Time Hash

ub4 one_at_a_time(char *key, ub4 len)
{
  ub4   hash, i;
  for (hash=0, i=0; i<len; ++i)
  {
    hash += key[i];
    hash += (hash << 10);
    hash ^= (hash >> 6);
  }
  hash += (hash << 3);
  hash ^= (hash >> 11);
  hash += (hash << 15);
  return (hash & mask);
} 
This is similar to the rotating hash, but it actually mixes the internal state. It takes 9n+9 instructions and produces a full 4-byte result. Preliminary analysis suggests there are no funnels.

This hash was not in the original Dr. Dobb's article. I implemented it to fill a set of requirements posed by Colin Plumb. Colin ended up using an even simpler (and weaker) hash that was sufficient for his purpose.

Bernstein's hash

I need to fill this in. Basically, hash = hash*33 + key[i]. If your keys are lowercase English words, this will fit 6 characters into a 32-bit hash with no collisions (you'd have to compare all 32 bits). If your keys are mixed case English words, hash*65+key[i] fits 5 characters into a 32-bit hash with no collisions. That means this type of hash can produce (for the right type of keys) fewer collisions than a hash that gives a more truly random distribution. If you don't have the right type of keys, or if you don't use all the bits, these hashes have lots of easily detectable flaws. Use with care.

FNV Hash

I need to fill this in too. Search the web for FNV hash. It's faster than my hash on Intel (because Intel has fast multiplication), and preliminary tests suggested it has decent distributions. I have three complaints against it. First, increasing the result size by one bit gives you a completely different hash. If you use a hash table that grows by increasing the result size by one bit, one old bucket maps across the entire new table, not to just two new buckets. If your algorithm has a sliding pointer for which buckets have been split, that just won't work with FNV. Second, it's linear. That means that widely separated things can cancel each other out at least as easily as nearby things. Third, since multiplication only affects higher bits, the lowest 7 bits of the state are never affected by the 8th bit of all the input bytes.

On the plus side, very nearby things never cancel each other out at all. This makes FNV a good choice for hashing very short keys (like single English words). FNV is more robust than Bernstein's hash. It can handle any byte values, not just ASCII characters.

ub4 fnv()
{
  /* I need to fill this in */
}

Paul Hsieh's hash

I haven't had time to analyze this one, but it looks rather promising. It's kind of a cross between that big hash at the start of this article and my one-at-a-time hash. Paul's timed it and it was faster than any of my hashes. It has a 4-byte internal state that it does light nonlinear mixing after every combine. That's good. It combines 2-byte blocks with its 4-byte state, which is something I'd never tried. (FNV and CRC do something similar. Their input blocks are all smaller than their state, and they mix their state after each input block, which makes it impossible for consecutive input blocks to cancel.)

#undef getshort
#if (defined(__GNUC__) && defined(__i386__)) || defined(__WATCOMC__)
#define getshort(d) (*((const unsigned short *) (d)))
#endif
#if defined(_MSC_VER) || defined(__BORLANDC__)
#define getshort(d) (*((const unsigned short *) (d)))
#endif

#if !defined(getshort)
#define getshort(d) ((((const unsigned char *) (d))[1] << 8UL)\
                     +((const unsigned char *) (d))[0])
#endif

unsigned int SuperFastHash (const char * data, int len) {
unsigned long hash;
int i, rem;

    if (len <= 0 || data == NULL) return 0;

    hash = (unsigned long) len; /* Avoid 0 -> 0 trap */
    rem = len & 3;
    len >>= 2;

    /* Main loop */
    for (i=0; i < len; i++) {
        hash += getshort (data);
        data += 2*sizeof (char);
        hash ^= hash << 16;
        hash ^= getshort (data) << 11;
        data += 2*sizeof (char);
        hash += hash >> 11;
    }

    /* Handle end cases */
    switch (rem) {
        case 3: hash += getshort (data);
                hash ^= hash << 16;
                hash ^= data[2 * sizeof (char)] << 18;
                hash += hash >> 11;
                break;
        case 2: hash += getshort (data);
                hash ^= hash << 11;
                hash += hash >> 17;
                break;
        case 1: hash += *data;
                hash ^= hash << 10;
                hash += hash >> 1;
    }

    /* Force "avalanching" of final 127 bits */
    hash ^= hash << 3;
    hash += hash >> 5;
    hash ^= hash << 2;
    hash += hash >> 15;
    hash ^= hash << 10;

    return hash;
}

Pearson's Hash

char pearson(char *key, ub4 len, char tab[256])
{
  char hash;
  ub4  i;
  for (hash=len, i=0; i<len; ++i) 
    hash=tab[hash^key[i]];
  return (hash);
}
This preinitializes tab[] to an arbitrary permutation of 0 .. 255. It takes 6n+2 instructions, but produces only a 1-byte result. Larger results can be made by running it several times with different initial hash values.

CRC Hashing

ub4 crc(char *key, ub4 len, ub4 mask, ub4 tab[256])
{
  ub4 hash, i;
  for (hash=len, i=0; i<len; ++i)
    hash = (hash<<8)^tab[(hash>>24)^key[i]];
  return (hash & mask);
}
This takes 9n+3 instructions. tab is initialized to simulate a maximal-length Linear Feedback Shift Register (LFSR). I used a 32-bit state with a polynomial of 0x04c11db7 for the tests.

Generalized CRC Hashing

This takes 9n+3 instructions. It is the same as CRC hashing except it fills tab[] with arbitrary values. The values in the low bytes must form a permutation of 0..255.

Universal Hashing

ub4 universal(char *key, ub4 len, ub4 mask, ub4 tab[MAXBITS])
{
  ub4 hash, i;
  for (hash=len, i=0; i<(length<<3); i+=8)
  {
    register char k = key[i>>3];
    if (k&0x01) hash ^= tab[i+0];
    if (k&0x02) hash ^= tab[i+1];
    if (k&0x04) hash ^= tab[i+2];
    if (k&0x08) hash ^= tab[i+3];
    if (k&0x10) hash ^= tab[i+4];
    if (k&0x20) hash ^= tab[i+5];
    if (k&0x40) hash ^= tab[i+6];
    if (k&0x80) hash ^= tab[i+7];
  }
  return (hash & mask);
}
This takes 52n+3 instructions. The size of tab[] is the maximum number of input bits. Values in tab[] are chosen at random.

Zobrist Hashing

ub4 zobrist( char *key, ub4 len, ub4 mask, ub4 tab[MAXBYTES][256])
{
  ub4 hash, i;
  for (hash=len, i=0; i<len; ++i)
    hash ^= tab[i][key[i]];
  return (hash & mask);
}
This takes 10n+3 instructions. The size of tab[][256] is the maximum number of input bytes. Values of tab[][256] are chosen at random. Universal hashing can be implemented by a Zobrist hash with carefully chosen table values.

My Hash

This takes 6n+35 instructions.

MD4

This takes 9.5n+230 instructions. MD4 is a hash designed for cryptography by Ron Rivest. It takes 420 instructions to hash a block of 64 aligned bytes. I combined that with my hash's method of putting unaligned bytes into registers, adding 3n instructions. MD4 is overkill for hash table lookup.

The table below compares all these hash functions.

NAME
is the name of the hash.
SIZE-1000
is the smallest reasonable hash table size greater than 1000.
SPEED
is the speed of the hash, measured in instructions required to produce a hash value for a table with SIZE-1000 buckets. It is assumed the machine has a rotate instruction.
INLINE
This is the speed assuming the hash is inlined in a loop that has to walk through all the characters anyways, such as a tokenizer. Such a loop doesn't always exist, and even when it does inlining isn't always possible. Some hashes (my new hash and MD4) work on blocks larger than a character. Inlining a hash removes 3n+1 instructions of loop overhead. It also removes the n instructions needed to get the characters out of the key array and into a register. It also means the length isn't known. Inlining offers other advantages. It allows the string to be converted to uppercase, and/or to unicode, before the hash is performed without the expense of an extra loop or a temporary buffer.
FUNNEL-15
is the largest set of input bits affecting the smallest set of internal state bits when mapping 15-byte keys into a 1-byte result.
FUNNEL-100
is the largest set of input bits affecting the smallest set of internal state bits when mapping 100-byte keys into a 32-bit result.
COLLIDE-32
is the number of collisions found when a dictionary of 38,470 English words was hashed into a 32-bit result. (The expected number of collisions is 0.2 .)
COLLIDE-1000
is a chi2 measure of how well the hash did at mapping the 38470-word dictionary into the SIZE-1000 table. (A chi2 measure greater than +3 is significantly worse than a random mapping; less than -3 is significantly better than a random mapping; in between is just random fluctuations.)
Comparison of several hash functions
NAME SIZE-1000 SPEED INLINE FUNNEL-15 FUNNEL-100 COLLIDE-32 COLLIDE-1000
Additive 1009 5n+3 n+2 15 into 2 100 into 2 37006 +806.02
Rotating 1009 6n+3 2n+2 4 into 1 25 into 1 24 +1.24
One-at-a-Time 1024 9n+9 5n+8 none none 0 -0.09
Pearson 1024 12n+5 4n+3 none none 0 +1.65
CRC 1024 9n+3 5n+2 2 into 1 11 into 10 1 +0.16
Generalized 1024 9n+3 5n+2 none none 0 +0.79
Universal 1024 52n+3 48n+2 4 into 3 50 into 28 0 +0.20
Zobrist 1024 10n+3 6n+2 none none 1 -0.03
My Hash 1024 6n+35 N/A none none 0 +0.33
MD4 1024 9.5n+230 N/A none none 1 +0.73

From the measurements we can conclude that the Additive and Rotating hash were noticably bad for 32-bit results, and only the Additive hash was noticably bad for 10-bit results. If inlining is possible, the Rotating hash was the fastest acceptable hash, followed by Pearson or the Generalized CRC (if table lookup is OK) or One-at-a-Time (if table lookup is not OK). If inlining is not possible, it's a draw between My Hash and the Rotating hash. Note that, for many applications, the Rotating hash is noticably bad and should not be used. Table lengths should always be a power of 2 because that's faster than prime lengths and all acceptable hashes allow it.

Conclusion

A common weakness in hash function is for a small set of input bits to cancel each other out. There is an efficient test to detect most such weaknesses, and many functions pass this test. I gave code for the fastest such function I could find. Hash functions without this weakness work equally well on all classes of keys.

Testimonials:



Posted by BAGE