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
|
// Copyright Vespa.ai. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
#include "sequencedtaskexecutor.h"
#include "adaptive_sequenced_executor.h"
#include "singleexecutor.h"
#include <vespa/vespalib/util/atomic.h>
#include <vespa/vespalib/util/threadstackexecutor.h>
#include <vespa/vespalib/util/blockingthreadstackexecutor.h>
#include <vespa/vespalib/util/size_literals.h>
#include <vespa/vespalib/stllike/hashtable.h>
#include <cassert>
namespace vespalib {
namespace {
constexpr uint8_t MAGIC = 255;
constexpr uint32_t NUM_PERFECT_PER_EXECUTOR = 8;
constexpr uint16_t INVALID_KEY = 0x8000;
bool
isLazy(const std::vector<std::unique_ptr<vespalib::SyncableThreadExecutor>> & executors) {
for (const auto &executor : executors) {
if (dynamic_cast<const vespalib::SingleExecutor *>(executor.get()) == nullptr) {
return false;
}
}
return true;
}
ssize_t
find(uint16_t key, const uint16_t values[], size_t numValues) {
for (size_t i(0); i < numValues; i++) {
auto value = vespalib::atomic::load_ref_relaxed(values[i]);
if (key == value) {
return i;
}
if (INVALID_KEY == value) {
return -1;
}
}
return -1;
}
}
std::unique_ptr<ISequencedTaskExecutor>
SequencedTaskExecutor::create(Runnable::init_fun_t func, uint32_t threads) {
return create(func, threads, 1000);
}
std::unique_ptr<ISequencedTaskExecutor>
SequencedTaskExecutor::create(Runnable::init_fun_t func, uint32_t threads, uint32_t taskLimit) {
return create(func, threads, taskLimit, true, OptimizeFor::LATENCY);
}
std::unique_ptr<ISequencedTaskExecutor>
SequencedTaskExecutor::create(Runnable::init_fun_t func, uint32_t threads, uint32_t taskLimit, bool is_task_limit_hard, OptimizeFor optimize) {
return create(func, threads, taskLimit, is_task_limit_hard, optimize, 0);
}
std::unique_ptr<ISequencedTaskExecutor>
SequencedTaskExecutor::create(Runnable::init_fun_t func, uint32_t threads, uint32_t taskLimit,
bool is_task_limit_hard, OptimizeFor optimize, uint32_t kindOfWatermark)
{
if (optimize == OptimizeFor::ADAPTIVE) {
size_t num_strands = std::min(taskLimit, threads*32);
return std::make_unique<AdaptiveSequencedExecutor>(num_strands, threads, kindOfWatermark, taskLimit, is_task_limit_hard);
} else {
auto executors = std::vector<std::unique_ptr<SyncableThreadExecutor>>();
executors.reserve(threads);
for (uint32_t id = 0; id < threads; ++id) {
if (optimize == OptimizeFor::THROUGHPUT) {
uint32_t watermark = (kindOfWatermark == 0) ? taskLimit / 10 : kindOfWatermark;
executors.push_back(std::make_unique<SingleExecutor>(func, taskLimit, is_task_limit_hard, watermark, 100ms));
} else {
if (is_task_limit_hard) {
executors.push_back(std::make_unique<BlockingThreadStackExecutor>(1, taskLimit, func));
} else {
executors.push_back(std::make_unique<ThreadStackExecutor>(1, func));
}
}
}
return std::unique_ptr<ISequencedTaskExecutor>(new SequencedTaskExecutor(std::move(executors)));
}
}
SequencedTaskExecutor::~SequencedTaskExecutor()
{
sync_all();
}
SequencedTaskExecutor::SequencedTaskExecutor(std::vector<std::unique_ptr<vespalib::SyncableThreadExecutor>> executors)
: ISequencedTaskExecutor(executors.size()),
_executors(std::move(executors)),
_lazyExecutors(isLazy(_executors)),
_component2IdPerfect(std::make_unique<PerfectKeyT[]>(getNumExecutors()*NUM_PERFECT_PER_EXECUTOR)),
_component2IdImperfect(vespalib::hashtable_base::getModuloStl(getNumExecutors()*NUM_PERFECT_PER_EXECUTOR), MAGIC),
_mutex(),
_nextId(0)
{
assert(getNumExecutors() < 256);
for (size_t i(0); i < getNumExecutors() * NUM_PERFECT_PER_EXECUTOR; i++) {
_component2IdPerfect[i] = INVALID_KEY;
}
}
void
SequencedTaskExecutor::setTaskLimit(uint32_t taskLimit)
{
for (const auto &executor : _executors) {
executor->setTaskLimit(taskLimit);
}
}
void
SequencedTaskExecutor::executeTask(ExecutorId id, vespalib::Executor::Task::UP task)
{
assert(id.getId() < _executors.size());
auto rejectedTask = _executors[id.getId()]->execute(std::move(task));
assert(!rejectedTask);
}
void
SequencedTaskExecutor::sync_all() {
wakeup();
for (auto &executor : _executors) {
executor->sync();
}
}
void
SequencedTaskExecutor::wakeup() {
if (_lazyExecutors) {
for (auto &executor : _executors) {
//Enforce parallel wakeup of napping executors.
executor->wakeup();
}
}
}
ExecutorStats
SequencedTaskExecutor::getStats()
{
ExecutorStats accumulatedStats;
for (auto &executor : _executors) {
accumulatedStats.aggregate(executor->getStats());
}
return accumulatedStats;
}
std::vector<ExecutorStats>
SequencedTaskExecutor::get_raw_stats()
{
std::vector<ExecutorStats> result;
for (auto& executor : _executors) {
result.push_back(executor->getStats());
}
return result;
}
ISequencedTaskExecutor::ExecutorId
SequencedTaskExecutor::getExecutorId(uint64_t componentId) const {
auto id = getExecutorIdPerfect(componentId);
return id ? id.value() : getExecutorIdImPerfect(componentId);
}
std::optional<ISequencedTaskExecutor::ExecutorId>
SequencedTaskExecutor::getExecutorIdPerfect(uint64_t componentId) const {
PerfectKeyT key = componentId & 0x7fff;
ssize_t pos = find(key, _component2IdPerfect.get(), getNumExecutors() * NUM_PERFECT_PER_EXECUTOR);
if (pos < 0) {
std::unique_lock guard(_mutex);
pos = find(key, _component2IdPerfect.get(), getNumExecutors() * NUM_PERFECT_PER_EXECUTOR);
if (pos < 0) {
pos = find(INVALID_KEY, _component2IdPerfect.get(), getNumExecutors() * NUM_PERFECT_PER_EXECUTOR);
if (pos >= 0) {
vespalib::atomic::store_ref_relaxed(_component2IdPerfect[pos], key);
} else {
// There was a race for the last spots
return std::optional<ISequencedTaskExecutor::ExecutorId>();
}
}
}
return std::optional<ISequencedTaskExecutor::ExecutorId>(ExecutorId(pos % getNumExecutors()));
}
ISequencedTaskExecutor::ExecutorId
SequencedTaskExecutor::getExecutorIdImPerfect(uint64_t componentId) const {
uint32_t shrunkId = componentId % _component2IdImperfect.size();
uint8_t executorId = vespalib::atomic::load_ref_relaxed(_component2IdImperfect[shrunkId]);
if (executorId == MAGIC) {
std::lock_guard guard(_mutex);
if (vespalib::atomic::load_ref_relaxed(_component2IdImperfect[shrunkId]) == MAGIC) {
vespalib::atomic::store_ref_relaxed(_component2IdImperfect[shrunkId], _nextId % getNumExecutors());
_nextId++;
}
executorId = _component2IdImperfect[shrunkId];
}
return ExecutorId(executorId);
}
const vespalib::ThreadExecutor*
SequencedTaskExecutor::first_executor() const
{
if (_executors.empty()) {
return nullptr;
}
return _executors.front().get();
}
} // namespace search
|
