2009-07-08 13:19:16 -07:00
|
|
|
/*
|
hash: Replace primary hash functions by murmurhash.
murmurhash is faster than Jenkins and slightly higher quality, so switch to
it for hashing words.
The best timings I got for hashing for data lengths of the following
numbers of 32-bit words, in seconds per 1,000,000,000 hashes, were:
words murmurhash Jenkins hash
----- ---------- ------------
1 8.4 10.4
2 10.3 10.3
3 11.2 10.7
4 12.6 18.0
5 13.9 18.3
6 15.2 18.7
In other words, murmurhash outperforms Jenkins for all input lengths other
than exactly 3 32-bit words (12 bytes). (It's understandable that Jenkins
would have a best case at 12 bytes, because Jenkins works in 12-byte
chunks.) Even in the case where Jenkins is faster, it's only by 5%. On
average within this data set, murmurhash is 15% faster, and for 4-word
input it is 30% faster.
We retain Jenkins for flow_hash_symmetric_l4() and flow_hash_fields(),
which are cases where the hash value is exposed externally.
This commit appears to improve "ovs-benchmark rate" results slightly by
a few hundred connections per second (under 1%), when used with an NVP
controller.
Signed-off-by: Ben Pfaff <blp@nicira.com>
Acked-by: Ethan Jackson <ethan@nicira.com>
2013-01-16 16:14:42 -08:00
|
|
|
* Copyright (c) 2008, 2009, 2010, 2012, 2013 Nicira, Inc.
|
2009-07-08 13:19:16 -07:00
|
|
|
*
|
2009-06-15 15:11:30 -07:00
|
|
|
* Licensed under the Apache License, Version 2.0 (the "License");
|
|
|
|
* you may not use this file except in compliance with the License.
|
|
|
|
* You may obtain a copy of the License at:
|
2009-07-08 13:19:16 -07:00
|
|
|
*
|
2009-06-15 15:11:30 -07:00
|
|
|
* http://www.apache.org/licenses/LICENSE-2.0
|
|
|
|
*
|
|
|
|
* Unless required by applicable law or agreed to in writing, software
|
|
|
|
* distributed under the License is distributed on an "AS IS" BASIS,
|
|
|
|
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
|
|
* See the License for the specific language governing permissions and
|
|
|
|
* limitations under the License.
|
2009-07-08 13:19:16 -07:00
|
|
|
*/
|
|
|
|
#ifndef HASH_H
|
|
|
|
#define HASH_H 1
|
|
|
|
|
2009-11-04 15:01:00 -08:00
|
|
|
#include <stdbool.h>
|
2009-07-08 13:19:16 -07:00
|
|
|
#include <stddef.h>
|
|
|
|
#include <stdint.h>
|
|
|
|
#include <string.h>
|
2009-10-14 17:03:55 -07:00
|
|
|
#include "util.h"
|
2009-07-08 13:19:16 -07:00
|
|
|
|
2011-08-04 16:18:59 -07:00
|
|
|
#ifdef __cplusplus
|
|
|
|
extern "C" {
|
|
|
|
#endif
|
|
|
|
|
2012-01-18 11:29:57 -08:00
|
|
|
static inline uint32_t
|
|
|
|
hash_rot(uint32_t x, int k)
|
|
|
|
{
|
|
|
|
return (x << k) | (x >> (32 - k));
|
|
|
|
}
|
2009-07-08 13:19:16 -07:00
|
|
|
|
hash: Replace primary hash functions by murmurhash.
murmurhash is faster than Jenkins and slightly higher quality, so switch to
it for hashing words.
The best timings I got for hashing for data lengths of the following
numbers of 32-bit words, in seconds per 1,000,000,000 hashes, were:
words murmurhash Jenkins hash
----- ---------- ------------
1 8.4 10.4
2 10.3 10.3
3 11.2 10.7
4 12.6 18.0
5 13.9 18.3
6 15.2 18.7
In other words, murmurhash outperforms Jenkins for all input lengths other
than exactly 3 32-bit words (12 bytes). (It's understandable that Jenkins
would have a best case at 12 bytes, because Jenkins works in 12-byte
chunks.) Even in the case where Jenkins is faster, it's only by 5%. On
average within this data set, murmurhash is 15% faster, and for 4-word
input it is 30% faster.
We retain Jenkins for flow_hash_symmetric_l4() and flow_hash_fields(),
which are cases where the hash value is exposed externally.
This commit appears to improve "ovs-benchmark rate" results slightly by
a few hundred connections per second (under 1%), when used with an NVP
controller.
Signed-off-by: Ben Pfaff <blp@nicira.com>
Acked-by: Ethan Jackson <ethan@nicira.com>
2013-01-16 16:14:42 -08:00
|
|
|
uint32_t hash_words(const uint32_t data[], size_t n_words, uint32_t basis);
|
|
|
|
uint32_t hash_bytes(const void *, size_t n_bytes, uint32_t basis);
|
|
|
|
|
|
|
|
static inline uint32_t hash_int(uint32_t x, uint32_t basis);
|
|
|
|
static inline uint32_t hash_2words(uint32_t, uint32_t);
|
2014-03-27 19:38:04 -07:00
|
|
|
static inline uint32_t hash_uint64(uint64_t);
|
2014-03-27 20:22:37 -07:00
|
|
|
static inline uint32_t hash_uint64_basis(uint64_t x, uint32_t basis);
|
hash: Replace primary hash functions by murmurhash.
murmurhash is faster than Jenkins and slightly higher quality, so switch to
it for hashing words.
The best timings I got for hashing for data lengths of the following
numbers of 32-bit words, in seconds per 1,000,000,000 hashes, were:
words murmurhash Jenkins hash
----- ---------- ------------
1 8.4 10.4
2 10.3 10.3
3 11.2 10.7
4 12.6 18.0
5 13.9 18.3
6 15.2 18.7
In other words, murmurhash outperforms Jenkins for all input lengths other
than exactly 3 32-bit words (12 bytes). (It's understandable that Jenkins
would have a best case at 12 bytes, because Jenkins works in 12-byte
chunks.) Even in the case where Jenkins is faster, it's only by 5%. On
average within this data set, murmurhash is 15% faster, and for 4-word
input it is 30% faster.
We retain Jenkins for flow_hash_symmetric_l4() and flow_hash_fields(),
which are cases where the hash value is exposed externally.
This commit appears to improve "ovs-benchmark rate" results slightly by
a few hundred connections per second (under 1%), when used with an NVP
controller.
Signed-off-by: Ben Pfaff <blp@nicira.com>
Acked-by: Ethan Jackson <ethan@nicira.com>
2013-01-16 16:14:42 -08:00
|
|
|
uint32_t hash_3words(uint32_t, uint32_t, uint32_t);
|
|
|
|
|
|
|
|
static inline uint32_t hash_boolean(bool x, uint32_t basis);
|
|
|
|
uint32_t hash_double(double, uint32_t basis);
|
|
|
|
|
|
|
|
static inline uint32_t hash_pointer(const void *, uint32_t basis);
|
|
|
|
static inline uint32_t hash_string(const char *, uint32_t basis);
|
|
|
|
|
|
|
|
/* Murmurhash by Austin Appleby,
|
|
|
|
* from http://code.google.com/p/smhasher/source/browse/trunk/MurmurHash3.cpp.
|
|
|
|
*
|
|
|
|
* The upstream license there says:
|
|
|
|
*
|
|
|
|
* // MurmurHash3 was written by Austin Appleby, and is placed in the public
|
|
|
|
* // domain. The author hereby disclaims copyright to this source code.
|
|
|
|
*
|
|
|
|
* See hash_words() for sample usage. */
|
|
|
|
|
|
|
|
static inline uint32_t mhash_add__(uint32_t hash, uint32_t data)
|
2012-01-18 11:29:57 -08:00
|
|
|
{
|
hash: Replace primary hash functions by murmurhash.
murmurhash is faster than Jenkins and slightly higher quality, so switch to
it for hashing words.
The best timings I got for hashing for data lengths of the following
numbers of 32-bit words, in seconds per 1,000,000,000 hashes, were:
words murmurhash Jenkins hash
----- ---------- ------------
1 8.4 10.4
2 10.3 10.3
3 11.2 10.7
4 12.6 18.0
5 13.9 18.3
6 15.2 18.7
In other words, murmurhash outperforms Jenkins for all input lengths other
than exactly 3 32-bit words (12 bytes). (It's understandable that Jenkins
would have a best case at 12 bytes, because Jenkins works in 12-byte
chunks.) Even in the case where Jenkins is faster, it's only by 5%. On
average within this data set, murmurhash is 15% faster, and for 4-word
input it is 30% faster.
We retain Jenkins for flow_hash_symmetric_l4() and flow_hash_fields(),
which are cases where the hash value is exposed externally.
This commit appears to improve "ovs-benchmark rate" results slightly by
a few hundred connections per second (under 1%), when used with an NVP
controller.
Signed-off-by: Ben Pfaff <blp@nicira.com>
Acked-by: Ethan Jackson <ethan@nicira.com>
2013-01-16 16:14:42 -08:00
|
|
|
data *= 0xcc9e2d51;
|
|
|
|
data = hash_rot(data, 15);
|
|
|
|
data *= 0x1b873593;
|
|
|
|
return hash ^ data;
|
2012-01-18 11:29:57 -08:00
|
|
|
}
|
2009-07-08 13:19:16 -07:00
|
|
|
|
hash: Replace primary hash functions by murmurhash.
murmurhash is faster than Jenkins and slightly higher quality, so switch to
it for hashing words.
The best timings I got for hashing for data lengths of the following
numbers of 32-bit words, in seconds per 1,000,000,000 hashes, were:
words murmurhash Jenkins hash
----- ---------- ------------
1 8.4 10.4
2 10.3 10.3
3 11.2 10.7
4 12.6 18.0
5 13.9 18.3
6 15.2 18.7
In other words, murmurhash outperforms Jenkins for all input lengths other
than exactly 3 32-bit words (12 bytes). (It's understandable that Jenkins
would have a best case at 12 bytes, because Jenkins works in 12-byte
chunks.) Even in the case where Jenkins is faster, it's only by 5%. On
average within this data set, murmurhash is 15% faster, and for 4-word
input it is 30% faster.
We retain Jenkins for flow_hash_symmetric_l4() and flow_hash_fields(),
which are cases where the hash value is exposed externally.
This commit appears to improve "ovs-benchmark rate" results slightly by
a few hundred connections per second (under 1%), when used with an NVP
controller.
Signed-off-by: Ben Pfaff <blp@nicira.com>
Acked-by: Ethan Jackson <ethan@nicira.com>
2013-01-16 16:14:42 -08:00
|
|
|
static inline uint32_t mhash_add(uint32_t hash, uint32_t data)
|
2012-01-18 11:29:57 -08:00
|
|
|
{
|
hash: Replace primary hash functions by murmurhash.
murmurhash is faster than Jenkins and slightly higher quality, so switch to
it for hashing words.
The best timings I got for hashing for data lengths of the following
numbers of 32-bit words, in seconds per 1,000,000,000 hashes, were:
words murmurhash Jenkins hash
----- ---------- ------------
1 8.4 10.4
2 10.3 10.3
3 11.2 10.7
4 12.6 18.0
5 13.9 18.3
6 15.2 18.7
In other words, murmurhash outperforms Jenkins for all input lengths other
than exactly 3 32-bit words (12 bytes). (It's understandable that Jenkins
would have a best case at 12 bytes, because Jenkins works in 12-byte
chunks.) Even in the case where Jenkins is faster, it's only by 5%. On
average within this data set, murmurhash is 15% faster, and for 4-word
input it is 30% faster.
We retain Jenkins for flow_hash_symmetric_l4() and flow_hash_fields(),
which are cases where the hash value is exposed externally.
This commit appears to improve "ovs-benchmark rate" results slightly by
a few hundred connections per second (under 1%), when used with an NVP
controller.
Signed-off-by: Ben Pfaff <blp@nicira.com>
Acked-by: Ethan Jackson <ethan@nicira.com>
2013-01-16 16:14:42 -08:00
|
|
|
hash = mhash_add__(hash, data);
|
|
|
|
hash = hash_rot(hash, 13);
|
|
|
|
return hash * 5 + 0xe6546b64;
|
2012-01-18 11:29:57 -08:00
|
|
|
}
|
2009-07-08 13:19:16 -07:00
|
|
|
|
2014-07-04 07:57:18 -07:00
|
|
|
static inline uint32_t mhash_finish(uint32_t hash, uint32_t n_bytes)
|
hash: Replace primary hash functions by murmurhash.
murmurhash is faster than Jenkins and slightly higher quality, so switch to
it for hashing words.
The best timings I got for hashing for data lengths of the following
numbers of 32-bit words, in seconds per 1,000,000,000 hashes, were:
words murmurhash Jenkins hash
----- ---------- ------------
1 8.4 10.4
2 10.3 10.3
3 11.2 10.7
4 12.6 18.0
5 13.9 18.3
6 15.2 18.7
In other words, murmurhash outperforms Jenkins for all input lengths other
than exactly 3 32-bit words (12 bytes). (It's understandable that Jenkins
would have a best case at 12 bytes, because Jenkins works in 12-byte
chunks.) Even in the case where Jenkins is faster, it's only by 5%. On
average within this data set, murmurhash is 15% faster, and for 4-word
input it is 30% faster.
We retain Jenkins for flow_hash_symmetric_l4() and flow_hash_fields(),
which are cases where the hash value is exposed externally.
This commit appears to improve "ovs-benchmark rate" results slightly by
a few hundred connections per second (under 1%), when used with an NVP
controller.
Signed-off-by: Ben Pfaff <blp@nicira.com>
Acked-by: Ethan Jackson <ethan@nicira.com>
2013-01-16 16:14:42 -08:00
|
|
|
{
|
|
|
|
hash ^= n_bytes;
|
|
|
|
hash ^= hash >> 16;
|
|
|
|
hash *= 0x85ebca6b;
|
|
|
|
hash ^= hash >> 13;
|
|
|
|
hash *= 0xc2b2ae35;
|
|
|
|
hash ^= hash >> 16;
|
|
|
|
return hash;
|
|
|
|
}
|
2009-07-08 13:19:16 -07:00
|
|
|
|
2014-07-04 07:57:18 -07:00
|
|
|
static inline uint32_t hash_add(uint32_t hash, uint32_t data)
|
|
|
|
{
|
|
|
|
return mhash_add(hash, data);
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline uint32_t hash_finish(uint32_t hash, uint32_t final)
|
|
|
|
{
|
|
|
|
return mhash_finish(hash, final);
|
|
|
|
}
|
|
|
|
|
2009-07-08 13:19:16 -07:00
|
|
|
static inline uint32_t hash_string(const char *s, uint32_t basis)
|
|
|
|
{
|
|
|
|
return hash_bytes(s, strlen(s), basis);
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline uint32_t hash_int(uint32_t x, uint32_t basis)
|
|
|
|
{
|
hash: Replace primary hash functions by murmurhash.
murmurhash is faster than Jenkins and slightly higher quality, so switch to
it for hashing words.
The best timings I got for hashing for data lengths of the following
numbers of 32-bit words, in seconds per 1,000,000,000 hashes, were:
words murmurhash Jenkins hash
----- ---------- ------------
1 8.4 10.4
2 10.3 10.3
3 11.2 10.7
4 12.6 18.0
5 13.9 18.3
6 15.2 18.7
In other words, murmurhash outperforms Jenkins for all input lengths other
than exactly 3 32-bit words (12 bytes). (It's understandable that Jenkins
would have a best case at 12 bytes, because Jenkins works in 12-byte
chunks.) Even in the case where Jenkins is faster, it's only by 5%. On
average within this data set, murmurhash is 15% faster, and for 4-word
input it is 30% faster.
We retain Jenkins for flow_hash_symmetric_l4() and flow_hash_fields(),
which are cases where the hash value is exposed externally.
This commit appears to improve "ovs-benchmark rate" results slightly by
a few hundred connections per second (under 1%), when used with an NVP
controller.
Signed-off-by: Ben Pfaff <blp@nicira.com>
Acked-by: Ethan Jackson <ethan@nicira.com>
2013-01-16 16:14:42 -08:00
|
|
|
return hash_2words(x, basis);
|
2009-07-08 13:19:16 -07:00
|
|
|
}
|
|
|
|
|
2009-11-04 15:01:00 -08:00
|
|
|
/* An attempt at a useful 1-bit hash function. Has not been analyzed for
|
|
|
|
* quality. */
|
|
|
|
static inline uint32_t hash_boolean(bool x, uint32_t basis)
|
|
|
|
{
|
2011-08-04 16:18:59 -07:00
|
|
|
const uint32_t P0 = 0xc2b73583; /* This is hash_int(1, 0). */
|
|
|
|
const uint32_t P1 = 0xe90f1258; /* This is hash_int(2, 0). */
|
2012-01-18 11:29:57 -08:00
|
|
|
return (x ? P0 : P1) ^ hash_rot(basis, 1);
|
2009-11-04 15:01:00 -08:00
|
|
|
}
|
|
|
|
|
2009-11-04 15:00:28 -08:00
|
|
|
static inline uint32_t hash_pointer(const void *p, uint32_t basis)
|
|
|
|
{
|
|
|
|
/* Often pointers are hashed simply by casting to integer type, but that
|
|
|
|
* has pitfalls since the lower bits of a pointer are often all 0 for
|
|
|
|
* alignment reasons. It's hard to guess where the entropy really is, so
|
|
|
|
* we give up here and just use a high-quality hash function.
|
|
|
|
*
|
|
|
|
* The double cast suppresses a warning on 64-bit systems about casting to
|
|
|
|
* an integer to different size. That's OK in this case, since most of the
|
|
|
|
* entropy in the pointer is almost certainly in the lower 32 bits. */
|
|
|
|
return hash_int((uint32_t) (uintptr_t) p, basis);
|
|
|
|
}
|
|
|
|
|
hash: Replace primary hash functions by murmurhash.
murmurhash is faster than Jenkins and slightly higher quality, so switch to
it for hashing words.
The best timings I got for hashing for data lengths of the following
numbers of 32-bit words, in seconds per 1,000,000,000 hashes, were:
words murmurhash Jenkins hash
----- ---------- ------------
1 8.4 10.4
2 10.3 10.3
3 11.2 10.7
4 12.6 18.0
5 13.9 18.3
6 15.2 18.7
In other words, murmurhash outperforms Jenkins for all input lengths other
than exactly 3 32-bit words (12 bytes). (It's understandable that Jenkins
would have a best case at 12 bytes, because Jenkins works in 12-byte
chunks.) Even in the case where Jenkins is faster, it's only by 5%. On
average within this data set, murmurhash is 15% faster, and for 4-word
input it is 30% faster.
We retain Jenkins for flow_hash_symmetric_l4() and flow_hash_fields(),
which are cases where the hash value is exposed externally.
This commit appears to improve "ovs-benchmark rate" results slightly by
a few hundred connections per second (under 1%), when used with an NVP
controller.
Signed-off-by: Ben Pfaff <blp@nicira.com>
Acked-by: Ethan Jackson <ethan@nicira.com>
2013-01-16 16:14:42 -08:00
|
|
|
static inline uint32_t hash_2words(uint32_t x, uint32_t y)
|
2012-08-21 14:26:23 -07:00
|
|
|
{
|
2014-07-04 07:57:18 -07:00
|
|
|
return hash_finish(hash_add(hash_add(x, 0), y), 8);
|
2012-08-21 14:26:23 -07:00
|
|
|
}
|
|
|
|
|
2014-03-27 19:38:04 -07:00
|
|
|
static inline uint32_t hash_uint64(const uint64_t x)
|
|
|
|
{
|
|
|
|
return hash_2words((uint32_t)(x >> 32), (uint32_t)x);
|
|
|
|
}
|
|
|
|
|
2014-03-27 20:22:37 -07:00
|
|
|
static inline uint32_t hash_uint64_basis(const uint64_t x,
|
|
|
|
const uint32_t basis)
|
|
|
|
{
|
|
|
|
return hash_3words((uint32_t)(x >> 32), (uint32_t)x, basis);
|
|
|
|
}
|
2011-08-04 16:18:59 -07:00
|
|
|
#ifdef __cplusplus
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
2009-07-08 13:19:16 -07:00
|
|
|
#endif /* hash.h */
|