-
Notifications
You must be signed in to change notification settings - Fork 136
/
hash_counter.hpp
242 lines (205 loc) · 8.13 KB
/
hash_counter.hpp
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
239
240
241
242
/* This file is part of Jellyfish.
This work is dual-licensed under 3-Clause BSD License or GPL 3.0.
You can choose between one of them if you use this work.
`SPDX-License-Identifier: BSD-3-Clause OR GPL-3.0`
*/
#ifndef __HASH_COUNTER_HPP__
#define __HASH_COUNTER_HPP__
#include <stdexcept>
#include <jellyfish/large_hash_array.hpp>
#include <jellyfish/locks_pthread.hpp>
#include <jellyfish/dumper.hpp>
/// Cooperative version of the hash_counter. In this implementation,
/// it is expected that the given number of threads will call the
/// `add` method regularly. In case the hash table is full, it gets
/// enlarged using all the threads. After the work is done, every
/// thread must call promptly the `done` method.
namespace jellyfish{ namespace cooperative {
template<typename Key, typename word = uint64_t, typename atomic_t = ::atomic::gcc, typename mem_block_t = ::allocators::mmap>
class hash_counter {
public:
typedef typename large_hash::array<Key, word, atomic_t, mem_block_t> array;
typedef typename array::key_type key_type;
typedef typename array::mapped_type mapped_type;
typedef typename array::value_type value_type;
typedef typename array::reference reference;
typedef typename array::const_reference const_reference;
typedef typename array::pointer pointer;
typedef typename array::const_pointer const_pointer;
typedef typename array::eager_iterator eager_iterator;
typedef typename array::lazy_iterator lazy_iterator;
protected:
array* ary_;
array* new_ary_;
uint16_t nb_threads_;
locks::pthread::barrier size_barrier_;
volatile uint16_t size_thid_, done_threads_;
bool do_size_doubling_;
dumper_t<array>* dumper_;
public:
hash_counter(size_t size, // Size of hash. To be rounded up to a power of 2
uint16_t key_len, // Size of key in bits
uint16_t val_len, // Size of val in bits
uint16_t nb_threads, // Number of threads accessing this hash
uint16_t reprobe_limit = 126, // Maximum reprobe
const size_t* reprobes = jellyfish::quadratic_reprobes) :
ary_(new array(size, key_len, val_len, reprobe_limit, reprobes)),
new_ary_(0),
nb_threads_(nb_threads),
size_barrier_(nb_threads),
size_thid_(0),
done_threads_(0),
do_size_doubling_(true),
dumper_(0)
{ }
~hash_counter() {
delete ary_;
}
array* ary() { return ary_; }
const array* ary() const { return ary_; }
size_t size() const { return ary_->size(); }
uint16_t key_len() const { return ary_->key_len(); }
uint16_t val_len() const { return ary_->val_len(); }
uint16_t nb_threads() const { return nb_threads_; }
uint16_t reprobe_limit() const { return ary_->max_reprobe(); }
void reset_done() { done_threads_ = 0; }
/// Whether we attempt to double the size of the hash when full.
bool do_size_doubling() const { return do_size_doubling_; }
/// Set whether we attempt to double the size of the hash when full.
void do_size_doubling(bool v) { do_size_doubling_ = v; }
/// Set dumper responsible for cleaning out the array.
void dumper(dumper_t<array> *d) { dumper_ = d; }
/// Add `v` to the entry `k`. It returns in `is_new` true if the
/// entry `k` did not exist in the hash. In `id` is returned the
/// final position of `k` in the hash array.
void add(const Key& k, uint64_t v, bool* is_new, size_t* id) {
uint64_t carry_shift = 0;
bool* is_new_ptr = is_new;
size_t* id_ptr = id;
bool is_new_void = false;
size_t id_void = false;
// while(!ary_->add(k, v, &carry_shift, is_new_ptr, id_ptr)) {
while(true) {
if(ary_->add(k, v, &carry_shift, is_new_ptr, id_ptr)) break;
handle_full_ary();
// If carry_shift == v, failed to allocate the first field for
// key, hence status of is_new and value for id are not
// determined yet. On the other hand, if carry_shift < v, we
// failed while adding extra field for large key, so the status
// of is_new and value of id are known. We do not update them in future
// calls.
if(carry_shift != v) {
is_new_ptr = &is_new_void;
id_ptr = &id_void;
v = carry_shift;
}
}
}
/// Add `v` to the entry `k`. This method is multi-thread safe. If
/// the entry for `k` does not exists, it is inserted.
///
/// @param k Key to add to
/// @param v Value to add
inline void add(const Key& k, uint64_t v) {
bool is_new;
size_t id;
add(k, v, &is_new, &id);
}
/// Insert the key `k` in the hash. The value is not changed or set
/// to 0 if not already in the hash.
///
/// @param k Key to insert
inline void set(const Key& k) {
bool is_new;
size_t id;
set(k, &is_new, &id);
}
/// Insert the key `k` in the hash. The value is not changed or set
/// to 0 if not already in the hash. Set `is_new` to true if `k` did
/// not already exist in the hash. In `id` is returned the final
/// position of `k` in the hash.
void set(const Key& k, bool* is_new, size_t* id) {
while(!ary_->set(k, is_new, id))
handle_full_ary();
}
/// Update the value of key `k` by adding `v`, if `k` is already
/// present in the hash, otherwise this nothing happens. Returns
/// true if `k` is already in the hash, false otherwise.
bool update_add(const Key& k, uint64_t v) {
Key tmp_key;
return update_add(k, v, tmp_key);
}
bool update_add(const Key& k, uint64_t v, Key& tmp_key) {
uint64_t carry_shift = 0;
while(true) {
if(ary_->update_add(k, v, &carry_shift, tmp_key))
return true;
if(carry_shift == v)
return false;
handle_full_ary();
v = carry_shift;
}
}
/// Signify that thread is done and wait for all threads to be done.
void done() {
atomic_t::fetch_add(&done_threads_, (uint16_t)1);
while(!handle_full_ary()) ;
}
protected:
// Double the size of the hash and return false. Unless all the
// thread have reported they are done, in which case do nothing and
// return true.
bool handle_full_ary() {
bool serial_thread = size_barrier_.wait();
if(done_threads_ >= nb_threads_) // All done?
return true;
bool success = false;
if(do_size_doubling_)
success = success || double_size(serial_thread);
if(!success && dumper_) {
if(serial_thread)
dumper_->dump(ary_);
success = true;
size_barrier_.wait();
}
if(!success)
throw std::runtime_error("Hash full");
return false;
}
bool double_size(bool serial_thread) {
if(serial_thread) {// Allocate new array for size doubling
try {
if(ary_->key_len() >= sizeof(size_t) * 8 || ary_->size() < ((size_t)1 << ary_->key_len())) {
// Increase number of keys
new_ary_ = new array(ary_->size() * 2, ary_->key_len(), ary_->val_len(),
ary_->max_reprobe(), ary_->reprobes());
} else {
new_ary_ = new array(ary_->size(), ary_->key_len(), ary_->val_len() + 1,
ary_->max_reprobe(), ary_->reprobes());
}
} catch(typename array::ErrorAllocation &e) {
new_ary_ = 0;
}
}
size_thid_ = 0;
size_barrier_.wait();
array* my_ary = *(array* volatile*)&new_ary_;
if(!my_ary) // Allocation failed
return false;
// Copy data from old to new
uint16_t id = atomic_t::fetch_add(&size_thid_, (uint16_t)1);
eager_iterator it = ary_->eager_slice(id, nb_threads_);
while(it.next())
my_ary->add(it.key(), it.val());
size_barrier_.wait();
if(serial_thread) { // Set new ary to be current and free old
delete ary_;
ary_ = new_ary_;
}
// Done. Last sync point
size_barrier_.wait();
return true;
}
};
} } // namespace jellyfish { namespace cooperative {
#endif /* __HASH_COUNTER_HPP__ */