1 /* A thread pool (C++ version). */
3 /* Copyright (c) 2004-2018 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>
17 #include <linux/futex.h>
18 #include <sys/syscall.h>
21 XBT_LOG_EXTERNAL_CATEGORY(xbt_parmap);
26 /** \addtogroup XBT_parmap
28 * \brief Parallel map class
31 template <typename T> class Parmap {
33 Parmap(unsigned num_workers, e_xbt_parmap_mode_t mode);
34 Parmap(const Parmap&) = delete;
35 Parmap& operator=(const Parmap&) = delete;
37 void apply(void (*fun)(T), const std::vector<T>& data);
38 boost::optional<T> next();
41 enum Flag { PARMAP_WORK, PARMAP_DESTROY };
44 * \brief Thread data transmission structure
48 ThreadData(Parmap<T>& parmap, int id) : parmap(parmap), worker_id(id) {}
54 * \brief Synchronization object (different specializations).
58 explicit Synchro(Parmap<T>& parmap) : parmap(parmap) {}
59 virtual ~Synchro() = default;
61 * \brief Wakes all workers and waits for them to finish the tasks.
63 * This function is called by the controller thread.
65 virtual void master_signal() = 0;
67 * \brief Starts the parmap: waits for all workers to be ready and returns.
69 * This function is called by the controller thread.
71 virtual void master_wait() = 0;
73 * \brief Ends the parmap: wakes the controller thread when all workers terminate.
75 * This function is called by all worker threads when they end (not including the controller).
77 virtual void worker_signal() = 0;
79 * \brief Waits for some work to process.
81 * This function is called by each worker thread (not including the controller) when it has no more work to do.
83 * \param round the expected round number
85 virtual void worker_wait(unsigned) = 0;
90 class PosixSynchro : public Synchro {
92 explicit PosixSynchro(Parmap<T>& parmap);
97 void worker_wait(unsigned round);
100 xbt_os_cond_t ready_cond;
101 xbt_os_mutex_t ready_mutex;
102 xbt_os_cond_t done_cond;
103 xbt_os_mutex_t done_mutex;
107 class FutexSynchro : public Synchro {
109 explicit FutexSynchro(Parmap<T>& parmap) : Synchro(parmap) {}
110 void master_signal();
112 void worker_signal();
113 void worker_wait(unsigned);
116 static void futex_wait(unsigned* uaddr, unsigned val);
117 static void futex_wake(unsigned* uaddr, unsigned val);
121 class BusyWaitSynchro : public Synchro {
123 explicit BusyWaitSynchro(Parmap<T>& parmap) : Synchro(parmap) {}
124 void master_signal();
126 void worker_signal();
127 void worker_wait(unsigned);
130 static void* worker_main(void* arg);
131 Synchro* new_synchro(e_xbt_parmap_mode_t mode);
134 Flag status; /**< is the parmap active or being destroyed? */
135 unsigned work_round; /**< index of the current round */
136 xbt_os_thread_t* workers; /**< worker thread handlers */
137 unsigned num_workers; /**< total number of worker threads including the controller */
138 Synchro* synchro; /**< synchronization object */
140 unsigned thread_counter = 0; /**< number of workers that have done the work */
141 void (*fun)(const T) = nullptr; /**< function to run in parallel on each element of data */
142 const std::vector<T>* data = nullptr; /**< parameters to pass to fun in parallel */
143 std::atomic<unsigned> index; /**< index of the next element of data to pick */
147 * \brief Creates a parallel map object
148 * \param num_workers number of worker threads to create
149 * \param mode how to synchronize the worker threads
151 template <typename T> Parmap<T>::Parmap(unsigned num_workers, e_xbt_parmap_mode_t mode)
153 XBT_CDEBUG(xbt_parmap, "Create new parmap (%u workers)", num_workers);
155 /* Initialize the thread pool data structure */
156 this->status = PARMAP_WORK;
157 this->work_round = 0;
158 this->workers = new xbt_os_thread_t[num_workers];
159 this->num_workers = num_workers;
160 this->synchro = new_synchro(mode);
162 /* Create the pool of worker threads */
163 this->workers[0] = nullptr;
164 #if HAVE_PTHREAD_SETAFFINITY
167 for (unsigned i = 1; i < num_workers; i++) {
168 ThreadData* data = new ThreadData(*this, i);
169 this->workers[i] = xbt_os_thread_create(nullptr, worker_main, data, nullptr);
170 #if HAVE_PTHREAD_SETAFFINITY
171 xbt_os_thread_bind(this->workers[i], core_bind);
172 if (core_bind != xbt_os_get_numcores() - 1)
181 * \brief Destroys a parmap
183 template <typename T> Parmap<T>::~Parmap()
185 status = PARMAP_DESTROY;
186 synchro->master_signal();
188 for (unsigned i = 1; i < num_workers; i++)
189 xbt_os_thread_join(workers[i], nullptr);
196 * \brief Applies a list of tasks in parallel.
197 * \param fun the function to call in parallel
198 * \param data each element of this vector will be passed as an argument to fun
200 template <typename T> void Parmap<T>::apply(void (*fun)(T), const std::vector<T>& data)
202 /* Assign resources to worker threads (we are maestro here)*/
206 this->synchro->master_signal(); // maestro runs futex_wake to wake all the minions (the working threads)
207 this->work(); // maestro works with its minions
208 this->synchro->master_wait(); // When there is no more work to do, then maestro waits for the last minion to stop
209 XBT_CDEBUG(xbt_parmap, "Job done"); // ... and proceeds
213 * \brief Returns a next task to process.
215 * Worker threads call this function to get more work.
217 * \return the next task to process, or throws a std::out_of_range exception if there is no more work
219 template <typename T> boost::optional<T> Parmap<T>::next()
221 unsigned index = this->index.fetch_add(1, std::memory_order_relaxed);
222 if (index < this->data->size())
223 return (*this->data)[index];
229 * \brief Main work loop: applies fun to elements in turn.
231 template <typename T> void Parmap<T>::work()
233 unsigned length = this->data->size();
234 unsigned index = this->index.fetch_add(1, std::memory_order_relaxed);
235 while (index < length) {
236 this->fun((*this->data)[index]);
237 index = this->index.fetch_add(1, std::memory_order_relaxed);
242 * Get a synchronization object for given mode.
243 * \param mode the synchronization mode
245 template <typename T> typename Parmap<T>::Synchro* Parmap<T>::new_synchro(e_xbt_parmap_mode_t mode)
247 if (mode == XBT_PARMAP_DEFAULT) {
249 mode = XBT_PARMAP_FUTEX;
251 mode = XBT_PARMAP_POSIX;
256 case XBT_PARMAP_POSIX:
257 res = new PosixSynchro(*this);
259 case XBT_PARMAP_FUTEX:
261 res = new FutexSynchro(*this);
263 xbt_die("Futex is not available on this OS.");
266 case XBT_PARMAP_BUSY_WAIT:
267 res = new BusyWaitSynchro(*this);
276 * \brief Main function of a worker thread.
278 template <typename T> void* Parmap<T>::worker_main(void* arg)
280 ThreadData* data = static_cast<ThreadData*>(arg);
281 Parmap<T>& parmap = data->parmap;
283 smx_context_t context = SIMIX_context_new(std::function<void()>(), nullptr, nullptr);
284 SIMIX_context_set_current(context);
286 XBT_CDEBUG(xbt_parmap, "New worker thread created");
288 /* Worker's main loop */
291 parmap.synchro->worker_wait(round);
292 if (parmap.status == PARMAP_DESTROY)
295 XBT_CDEBUG(xbt_parmap, "Worker %d got a job", data->worker_id);
297 parmap.synchro->worker_signal();
298 XBT_CDEBUG(xbt_parmap, "Worker %d has finished", data->worker_id);
300 /* We are destroying the parmap */
306 template <typename T> Parmap<T>::PosixSynchro::PosixSynchro(Parmap<T>& parmap) : Synchro(parmap)
308 ready_cond = xbt_os_cond_init();
309 ready_mutex = xbt_os_mutex_init();
310 done_cond = xbt_os_cond_init();
311 done_mutex = xbt_os_mutex_init();
314 template <typename T> Parmap<T>::PosixSynchro::~PosixSynchro()
316 xbt_os_cond_destroy(ready_cond);
317 xbt_os_mutex_destroy(ready_mutex);
318 xbt_os_cond_destroy(done_cond);
319 xbt_os_mutex_destroy(done_mutex);
322 template <typename T> void Parmap<T>::PosixSynchro::master_signal()
324 xbt_os_mutex_acquire(ready_mutex);
325 this->parmap.thread_counter = 1;
326 this->parmap.work_round++;
327 /* wake all workers */
328 xbt_os_cond_broadcast(ready_cond);
329 xbt_os_mutex_release(ready_mutex);
332 template <typename T> void Parmap<T>::PosixSynchro::master_wait()
334 xbt_os_mutex_acquire(done_mutex);
335 while (this->parmap.thread_counter < this->parmap.num_workers) {
336 /* wait for all workers to be ready */
337 xbt_os_cond_wait(done_cond, done_mutex);
339 xbt_os_mutex_release(done_mutex);
342 template <typename T> void Parmap<T>::PosixSynchro::worker_signal()
344 xbt_os_mutex_acquire(done_mutex);
345 this->parmap.thread_counter++;
346 if (this->parmap.thread_counter == this->parmap.num_workers) {
347 /* all workers have finished, wake the controller */
348 xbt_os_cond_signal(done_cond);
350 xbt_os_mutex_release(done_mutex);
353 template <typename T> void Parmap<T>::PosixSynchro::worker_wait(unsigned round)
355 xbt_os_mutex_acquire(ready_mutex);
356 /* wait for more work */
357 while (this->parmap.work_round != round) {
358 xbt_os_cond_wait(ready_cond, ready_mutex);
360 xbt_os_mutex_release(ready_mutex);
364 template <typename T> inline void Parmap<T>::FutexSynchro::futex_wait(unsigned* uaddr, unsigned val)
366 XBT_CVERB(xbt_parmap, "Waiting on futex %p", uaddr);
367 syscall(SYS_futex, uaddr, FUTEX_WAIT_PRIVATE, val, nullptr, nullptr, 0);
370 template <typename T> inline void Parmap<T>::FutexSynchro::futex_wake(unsigned* uaddr, unsigned val)
372 XBT_CVERB(xbt_parmap, "Waking futex %p", uaddr);
373 syscall(SYS_futex, uaddr, FUTEX_WAKE_PRIVATE, val, nullptr, nullptr, 0);
376 template <typename T> void Parmap<T>::FutexSynchro::master_signal()
378 __atomic_store_n(&this->parmap.thread_counter, 1, __ATOMIC_SEQ_CST);
379 __atomic_add_fetch(&this->parmap.work_round, 1, __ATOMIC_SEQ_CST);
380 /* wake all workers */
381 futex_wake(&this->parmap.work_round, std::numeric_limits<int>::max());
384 template <typename T> void Parmap<T>::FutexSynchro::master_wait()
386 unsigned count = __atomic_load_n(&this->parmap.thread_counter, __ATOMIC_SEQ_CST);
387 while (count < this->parmap.num_workers) {
388 /* wait for all workers to be ready */
389 futex_wait(&this->parmap.thread_counter, count);
390 count = __atomic_load_n(&this->parmap.thread_counter, __ATOMIC_SEQ_CST);
394 template <typename T> void Parmap<T>::FutexSynchro::worker_signal()
396 unsigned count = __atomic_add_fetch(&this->parmap.thread_counter, 1, __ATOMIC_SEQ_CST);
397 if (count == this->parmap.num_workers) {
398 /* all workers have finished, wake the controller */
399 futex_wake(&this->parmap.thread_counter, std::numeric_limits<int>::max());
403 template <typename T> void Parmap<T>::FutexSynchro::worker_wait(unsigned round)
405 unsigned work_round = __atomic_load_n(&this->parmap.work_round, __ATOMIC_SEQ_CST);
406 /* wait for more work */
407 while (work_round != round) {
408 futex_wait(&this->parmap.work_round, work_round);
409 work_round = __atomic_load_n(&this->parmap.work_round, __ATOMIC_SEQ_CST);
414 template <typename T> void Parmap<T>::BusyWaitSynchro::master_signal()
416 __atomic_store_n(&this->parmap.thread_counter, 1, __ATOMIC_SEQ_CST);
417 __atomic_add_fetch(&this->parmap.work_round, 1, __ATOMIC_SEQ_CST);
420 template <typename T> void Parmap<T>::BusyWaitSynchro::master_wait()
422 while (__atomic_load_n(&this->parmap.thread_counter, __ATOMIC_SEQ_CST) < this->parmap.num_workers) {
423 xbt_os_thread_yield();
427 template <typename T> void Parmap<T>::BusyWaitSynchro::worker_signal()
429 __atomic_add_fetch(&this->parmap.thread_counter, 1, __ATOMIC_SEQ_CST);
432 template <typename T> void Parmap<T>::BusyWaitSynchro::worker_wait(unsigned round)
434 /* wait for more work */
435 while (__atomic_load_n(&this->parmap.work_round, __ATOMIC_SEQ_CST) != round) {
436 xbt_os_thread_yield();