1 /* A thread pool (C++ version). */
3 /* Copyright (c) 2004-2019 The SimGrid Team. All rights reserved. */
5 /* This program is free software; you can redistribute it and/or modify it
6 * under the terms of the license (GNU LGPL) which comes with this package. */
11 #include "src/internal_config.h" // HAVE_FUTEX_H
12 #include "src/kernel/context/Context.hpp"
14 #include <boost/optional.hpp>
15 #include <condition_variable>
20 #include <linux/futex.h>
21 #include <sys/syscall.h>
25 #include <pthread_np.h>
28 XBT_LOG_EXTERNAL_CATEGORY(xbt_parmap);
33 /** @addtogroup XBT_parmap
35 * @brief Parallel map class
38 template <typename T> class Parmap {
40 Parmap(unsigned num_workers, e_xbt_parmap_mode_t mode);
41 Parmap(const Parmap&) = delete;
42 Parmap& operator=(const Parmap&) = delete;
44 void apply(void (*fun)(T), const std::vector<T>& data);
45 boost::optional<T> next();
48 enum Flag { PARMAP_WORK, PARMAP_DESTROY };
51 * @brief Thread data transmission structure
55 ThreadData(Parmap<T>& parmap, int id) : parmap(parmap), worker_id(id) {}
61 * @brief Synchronization object (different specializations).
65 explicit Synchro(Parmap<T>& parmap) : parmap(parmap) {}
66 virtual ~Synchro() = default;
68 * @brief Wakes all workers and waits for them to finish the tasks.
70 * This function is called by the controller thread.
72 virtual void master_signal() = 0;
74 * @brief Starts the parmap: waits for all workers to be ready and returns.
76 * This function is called by the controller thread.
78 virtual void master_wait() = 0;
80 * @brief Ends the parmap: wakes the controller thread when all workers terminate.
82 * This function is called by all worker threads when they end (not including the controller).
84 virtual void worker_signal() = 0;
86 * @brief Waits for some work to process.
88 * This function is called by each worker thread (not including the controller) when it has no more work to do.
90 * @param round the expected round number
92 virtual void worker_wait(unsigned) = 0;
97 class PosixSynchro : public Synchro {
99 explicit PosixSynchro(Parmap<T>& parmap);
101 void master_signal();
103 void worker_signal();
104 void worker_wait(unsigned round);
107 std::condition_variable ready_cond;
108 std::mutex ready_mutex;
109 std::condition_variable done_cond;
110 std::mutex done_mutex;
114 class FutexSynchro : public Synchro {
116 explicit FutexSynchro(Parmap<T>& parmap) : Synchro(parmap) {}
117 void master_signal();
119 void worker_signal();
120 void worker_wait(unsigned);
123 static void futex_wait(unsigned* uaddr, unsigned val);
124 static void futex_wake(unsigned* uaddr, unsigned val);
128 class BusyWaitSynchro : public Synchro {
130 explicit BusyWaitSynchro(Parmap<T>& parmap) : Synchro(parmap) {}
131 void master_signal();
133 void worker_signal();
134 void worker_wait(unsigned);
137 static void* worker_main(void* arg);
138 Synchro* new_synchro(e_xbt_parmap_mode_t mode);
141 Flag status; /**< is the parmap active or being destroyed? */
142 unsigned work_round; /**< index of the current round */
143 std::vector<std::thread*> workers; /**< worker thread handlers */
144 unsigned num_workers; /**< total number of worker threads including the controller */
145 Synchro* synchro; /**< synchronization object */
147 unsigned thread_counter = 0; /**< number of workers that have done the work */
148 void (*fun)(const T) = nullptr; /**< function to run in parallel on each element of data */
149 const std::vector<T>* data = nullptr; /**< parameters to pass to fun in parallel */
150 std::atomic<unsigned> index; /**< index of the next element of data to pick */
154 * @brief Creates a parallel map object
155 * @param num_workers number of worker threads to create
156 * @param mode how to synchronize the worker threads
158 template <typename T> Parmap<T>::Parmap(unsigned num_workers, e_xbt_parmap_mode_t mode)
160 XBT_CDEBUG(xbt_parmap, "Create new parmap (%u workers)", num_workers);
162 /* Initialize the thread pool data structure */
163 this->status = PARMAP_WORK;
164 this->work_round = 0;
165 this->workers.resize(num_workers);
166 this->num_workers = num_workers;
167 this->synchro = new_synchro(mode);
169 /* Create the pool of worker threads (the caller of apply() will be worker[0]) */
170 this->workers[0] = nullptr;
171 XBT_ATTRIB_UNUSED unsigned int core_bind = 0;
173 for (unsigned i = 1; i < num_workers; i++) {
174 this->workers[i] = new std::thread(worker_main, new ThreadData(*this, i));
176 /* Bind the worker to a core if possible */
177 #if HAVE_PTHREAD_SETAFFINITY
178 #if HAVE_PTHREAD_NP_H /* FreeBSD ? */
180 size_t size = sizeof(cpuset_t);
183 size_t size = sizeof(cpu_set_t);
185 pthread_t pthread = this->workers[i]->native_handle();
187 CPU_SET(core_bind, &cpuset);
188 pthread_setaffinity_np(pthread, size, &cpuset);
189 if (core_bind != std::thread::hardware_concurrency() - 1)
198 * @brief Destroys a parmap
200 template <typename T> Parmap<T>::~Parmap()
202 status = PARMAP_DESTROY;
203 synchro->master_signal();
205 for (unsigned i = 1; i < num_workers; i++)
212 * @brief Applies a list of tasks in parallel.
213 * @param fun the function to call in parallel
214 * @param data each element of this vector will be passed as an argument to fun
216 template <typename T> void Parmap<T>::apply(void (*fun)(T), const std::vector<T>& data)
218 /* Assign resources to worker threads (we are maestro here)*/
222 this->synchro->master_signal(); // maestro runs futex_wake to wake all the minions (the working threads)
223 this->work(); // maestro works with its minions
224 this->synchro->master_wait(); // When there is no more work to do, then maestro waits for the last minion to stop
225 XBT_CDEBUG(xbt_parmap, "Job done"); // ... and proceeds
229 * @brief Returns a next task to process.
231 * Worker threads call this function to get more work.
233 * @return the next task to process, or throws a std::out_of_range exception if there is no more work
235 template <typename T> boost::optional<T> Parmap<T>::next()
237 unsigned index = this->index.fetch_add(1, std::memory_order_relaxed);
238 if (index < this->data->size())
239 return (*this->data)[index];
245 * @brief Main work loop: applies fun to elements in turn.
247 template <typename T> void Parmap<T>::work()
249 unsigned length = this->data->size();
250 unsigned index = this->index.fetch_add(1, std::memory_order_relaxed);
251 while (index < length) {
252 this->fun((*this->data)[index]);
253 index = this->index.fetch_add(1, std::memory_order_relaxed);
258 * Get a synchronization object for given mode.
259 * @param mode the synchronization mode
261 template <typename T> typename Parmap<T>::Synchro* Parmap<T>::new_synchro(e_xbt_parmap_mode_t mode)
263 if (mode == XBT_PARMAP_DEFAULT) {
265 mode = XBT_PARMAP_FUTEX;
267 mode = XBT_PARMAP_POSIX;
272 case XBT_PARMAP_POSIX:
273 res = new PosixSynchro(*this);
275 case XBT_PARMAP_FUTEX:
277 res = new FutexSynchro(*this);
279 xbt_die("Futex is not available on this OS.");
282 case XBT_PARMAP_BUSY_WAIT:
283 res = new BusyWaitSynchro(*this);
291 /** @brief Main function of a worker thread */
292 template <typename T> void* Parmap<T>::worker_main(void* arg)
294 ThreadData* data = static_cast<ThreadData*>(arg);
295 Parmap<T>& parmap = data->parmap;
297 smx_context_t context = SIMIX_context_new(std::function<void()>(), nullptr, nullptr);
298 kernel::context::Context::set_current(context);
300 XBT_CDEBUG(xbt_parmap, "New worker thread created");
302 /* Worker's main loop */
304 round++; // New scheduling round
305 parmap.synchro->worker_wait(round);
306 if (parmap.status == PARMAP_DESTROY)
309 XBT_CDEBUG(xbt_parmap, "Worker %d got a job", data->worker_id);
311 parmap.synchro->worker_signal();
312 XBT_CDEBUG(xbt_parmap, "Worker %d has finished", data->worker_id);
314 /* We are destroying the parmap */
320 template <typename T> Parmap<T>::PosixSynchro::PosixSynchro(Parmap<T>& parmap) : Synchro(parmap)
324 template <typename T> Parmap<T>::PosixSynchro::~PosixSynchro()
328 template <typename T> void Parmap<T>::PosixSynchro::master_signal()
330 std::unique_lock<std::mutex> lk(ready_mutex);
331 this->parmap.thread_counter = 1;
332 this->parmap.work_round++;
333 /* wake all workers */
334 ready_cond.notify_all();
337 template <typename T> void Parmap<T>::PosixSynchro::master_wait()
339 std::unique_lock<std::mutex> lk(done_mutex);
340 while (this->parmap.thread_counter < this->parmap.num_workers) {
341 /* wait for all workers to be ready */
346 template <typename T> void Parmap<T>::PosixSynchro::worker_signal()
348 std::unique_lock<std::mutex> lk(done_mutex);
349 this->parmap.thread_counter++;
350 if (this->parmap.thread_counter == this->parmap.num_workers) {
351 /* all workers have finished, wake the controller */
352 done_cond.notify_one();
356 template <typename T> void Parmap<T>::PosixSynchro::worker_wait(unsigned round)
358 std::unique_lock<std::mutex> lk(ready_mutex);
359 /* wait for more work */
360 while (this->parmap.work_round != round) {
366 template <typename T> inline void Parmap<T>::FutexSynchro::futex_wait(unsigned* uaddr, unsigned val)
368 XBT_CVERB(xbt_parmap, "Waiting on futex %p", uaddr);
369 syscall(SYS_futex, uaddr, FUTEX_WAIT_PRIVATE, val, nullptr, nullptr, 0);
372 template <typename T> inline void Parmap<T>::FutexSynchro::futex_wake(unsigned* uaddr, unsigned val)
374 XBT_CVERB(xbt_parmap, "Waking futex %p", uaddr);
375 syscall(SYS_futex, uaddr, FUTEX_WAKE_PRIVATE, val, nullptr, nullptr, 0);
378 template <typename T> void Parmap<T>::FutexSynchro::master_signal()
380 __atomic_store_n(&this->parmap.thread_counter, 1, __ATOMIC_SEQ_CST);
381 __atomic_add_fetch(&this->parmap.work_round, 1, __ATOMIC_SEQ_CST);
382 /* wake all workers */
383 futex_wake(&this->parmap.work_round, std::numeric_limits<int>::max());
386 template <typename T> void Parmap<T>::FutexSynchro::master_wait()
388 unsigned count = __atomic_load_n(&this->parmap.thread_counter, __ATOMIC_SEQ_CST);
389 while (count < this->parmap.num_workers) {
390 /* wait for all workers to be ready */
391 futex_wait(&this->parmap.thread_counter, count);
392 count = __atomic_load_n(&this->parmap.thread_counter, __ATOMIC_SEQ_CST);
396 template <typename T> void Parmap<T>::FutexSynchro::worker_signal()
398 unsigned count = __atomic_add_fetch(&this->parmap.thread_counter, 1, __ATOMIC_SEQ_CST);
399 if (count == this->parmap.num_workers) {
400 /* all workers have finished, wake the controller */
401 futex_wake(&this->parmap.thread_counter, std::numeric_limits<int>::max());
405 template <typename T> void Parmap<T>::FutexSynchro::worker_wait(unsigned round)
407 unsigned work_round = __atomic_load_n(&this->parmap.work_round, __ATOMIC_SEQ_CST);
408 /* wait for more work */
409 while (work_round != round) {
410 futex_wait(&this->parmap.work_round, work_round);
411 work_round = __atomic_load_n(&this->parmap.work_round, __ATOMIC_SEQ_CST);
416 template <typename T> void Parmap<T>::BusyWaitSynchro::master_signal()
418 __atomic_store_n(&this->parmap.thread_counter, 1, __ATOMIC_SEQ_CST);
419 __atomic_add_fetch(&this->parmap.work_round, 1, __ATOMIC_SEQ_CST);
422 template <typename T> void Parmap<T>::BusyWaitSynchro::master_wait()
424 while (__atomic_load_n(&this->parmap.thread_counter, __ATOMIC_SEQ_CST) < this->parmap.num_workers) {
425 std::this_thread::yield();
429 template <typename T> void Parmap<T>::BusyWaitSynchro::worker_signal()
431 __atomic_add_fetch(&this->parmap.thread_counter, 1, __ATOMIC_SEQ_CST);
434 template <typename T> void Parmap<T>::BusyWaitSynchro::worker_wait(unsigned round)
436 /* wait for more work */
437 while (__atomic_load_n(&this->parmap.work_round, __ATOMIC_SEQ_CST) != round) {
438 std::this_thread::yield();