1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
|
// Copyright (c) 2013, Facebook, Inc. All rights reserved.
// This source code is licensed under both the GPLv2 (found in the
// COPYING file in the root directory) and Apache 2.0 License
// (found in the LICENSE.Apache file in the root directory).
//
#pragma once
#ifndef ROCKSDB_LITE
#include <assert.h>
#include <list>
#include <vector>
#ifdef OS_LINUX
#include <sys/mman.h>
#endif
#include "rocksdb/env.h"
#include "util/mutexlock.h"
namespace ROCKSDB_NAMESPACE {
// HashTable<T, Hash, Equal>
//
// Traditional implementation of hash table with synchronization built on top
// don't perform very well in multi-core scenarios. This is an implementation
// designed for multi-core scenarios with high lock contention.
//
// |<-------- alpha ------------->|
// Buckets Collision list
// ---- +----+ +---+---+--- ...... ---+---+---+
// / | |--->| | | | | |
// / +----+ +---+---+--- ...... ---+---+---+
// / | |
// Locks/ +----+
// +--+/ . .
// | | . .
// +--+ . .
// | | . .
// +--+ . .
// | | . .
// +--+ . .
// \ +----+
// \ | |
// \ +----+
// \ | |
// \---- +----+
//
// The lock contention is spread over an array of locks. This helps improve
// concurrent access. The spine is designed for a certain capacity and load
// factor. When the capacity planning is done correctly we can expect
// O(load_factor = 1) insert, access and remove time.
//
// Micro benchmark on debug build gives about .5 Million/sec rate of insert,
// erase and lookup in parallel (total of about 1.5 Million ops/sec). If the
// blocks were of 4K, the hash table can support a virtual throughput of
// 6 GB/s.
//
// T Object type (contains both key and value)
// Hash Function that returns an hash from type T
// Equal Returns if two objects are equal
// (We need explicit equal for pointer type)
//
template <class T, class Hash, class Equal>
class HashTable {
public:
explicit HashTable(const size_t capacity = 1024 * 1024,
const float load_factor = 2.0, const uint32_t nlocks = 256)
: nbuckets_(
static_cast<uint32_t>(load_factor ? capacity / load_factor : 0)),
nlocks_(nlocks) {
// pre-conditions
assert(capacity);
assert(load_factor);
assert(nbuckets_);
assert(nlocks_);
buckets_.reset(new Bucket[nbuckets_]);
#ifdef OS_LINUX
mlock(buckets_.get(), nbuckets_ * sizeof(Bucket));
#endif
// initialize locks
locks_.reset(new port::RWMutex[nlocks_]);
#ifdef OS_LINUX
mlock(locks_.get(), nlocks_ * sizeof(port::RWMutex));
#endif
// post-conditions
assert(buckets_);
assert(locks_);
}
virtual ~HashTable() { AssertEmptyBuckets(); }
//
// Insert given record to hash table
//
bool Insert(const T& t) {
const uint64_t h = Hash()(t);
const uint32_t bucket_idx = h % nbuckets_;
const uint32_t lock_idx = bucket_idx % nlocks_;
WriteLock _(&locks_[lock_idx]);
auto& bucket = buckets_[bucket_idx];
return Insert(&bucket, t);
}
// Lookup hash table
//
// Please note that read lock should be held by the caller. This is because
// the caller owns the data, and should hold the read lock as long as he
// operates on the data.
bool Find(const T& t, T* ret, port::RWMutex** ret_lock) {
const uint64_t h = Hash()(t);
const uint32_t bucket_idx = h % nbuckets_;
const uint32_t lock_idx = bucket_idx % nlocks_;
port::RWMutex& lock = locks_[lock_idx];
lock.ReadLock();
auto& bucket = buckets_[bucket_idx];
if (Find(&bucket, t, ret)) {
*ret_lock = &lock;
return true;
}
lock.ReadUnlock();
return false;
}
//
// Erase a given key from the hash table
//
bool Erase(const T& t, T* ret) {
const uint64_t h = Hash()(t);
const uint32_t bucket_idx = h % nbuckets_;
const uint32_t lock_idx = bucket_idx % nlocks_;
WriteLock _(&locks_[lock_idx]);
auto& bucket = buckets_[bucket_idx];
return Erase(&bucket, t, ret);
}
// Fetch the mutex associated with a key
// This call is used to hold the lock for a given data for extended period of
// time.
port::RWMutex* GetMutex(const T& t) {
const uint64_t h = Hash()(t);
const uint32_t bucket_idx = h % nbuckets_;
const uint32_t lock_idx = bucket_idx % nlocks_;
return &locks_[lock_idx];
}
void Clear(void (*fn)(T)) {
for (uint32_t i = 0; i < nbuckets_; ++i) {
const uint32_t lock_idx = i % nlocks_;
WriteLock _(&locks_[lock_idx]);
for (auto& t : buckets_[i].list_) {
(*fn)(t);
}
buckets_[i].list_.clear();
}
}
protected:
// Models bucket of keys that hash to the same bucket number
struct Bucket {
std::list<T> list_;
};
// Substitute for std::find with custom comparator operator
typename std::list<T>::iterator Find(std::list<T>* list, const T& t) {
for (auto it = list->begin(); it != list->end(); ++it) {
if (Equal()(*it, t)) {
return it;
}
}
return list->end();
}
bool Insert(Bucket* bucket, const T& t) {
// Check if the key already exists
auto it = Find(&bucket->list_, t);
if (it != bucket->list_.end()) {
return false;
}
// insert to bucket
bucket->list_.push_back(t);
return true;
}
bool Find(Bucket* bucket, const T& t, T* ret) {
auto it = Find(&bucket->list_, t);
if (it != bucket->list_.end()) {
if (ret) {
*ret = *it;
}
return true;
}
return false;
}
bool Erase(Bucket* bucket, const T& t, T* ret) {
auto it = Find(&bucket->list_, t);
if (it != bucket->list_.end()) {
if (ret) {
*ret = *it;
}
bucket->list_.erase(it);
return true;
}
return false;
}
// assert that all buckets are empty
void AssertEmptyBuckets() {
#ifndef NDEBUG
for (size_t i = 0; i < nbuckets_; ++i) {
WriteLock _(&locks_[i % nlocks_]);
assert(buckets_[i].list_.empty());
}
#endif
}
const uint32_t nbuckets_; // No. of buckets in the spine
std::unique_ptr<Bucket[]> buckets_; // Spine of the hash buckets
const uint32_t nlocks_; // No. of locks
std::unique_ptr<port::RWMutex[]> locks_; // Granular locks
};
} // namespace ROCKSDB_NAMESPACE
#endif
|
