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"
13 #include "xbt/xbt_os_thread.h"
15 #include <boost/optional.hpp>
16 #include <condition_variable>
21 #include <linux/futex.h>
22 #include <sys/syscall.h>
25 XBT_LOG_EXTERNAL_CATEGORY(xbt_parmap);
30 /** @addtogroup XBT_parmap
32 * @brief Parallel map class
35 template <typename T> class Parmap {
37 Parmap(unsigned num_workers, e_xbt_parmap_mode_t mode);
38 Parmap(const Parmap&) = delete;
39 Parmap& operator=(const Parmap&) = delete;
41 void apply(void (*fun)(T), const std::vector<T>& data);
42 boost::optional<T> next();
45 enum Flag { PARMAP_WORK, PARMAP_DESTROY };
48 * @brief Thread data transmission structure
52 ThreadData(Parmap<T>& parmap, int id) : parmap(parmap), worker_id(id) {}
58 * @brief Synchronization object (different specializations).
62 explicit Synchro(Parmap<T>& parmap) : parmap(parmap) {}
63 virtual ~Synchro() = default;
65 * @brief Wakes all workers and waits for them to finish the tasks.
67 * This function is called by the controller thread.
69 virtual void master_signal() = 0;
71 * @brief Starts the parmap: waits for all workers to be ready and returns.
73 * This function is called by the controller thread.
75 virtual void master_wait() = 0;
77 * @brief Ends the parmap: wakes the controller thread when all workers terminate.
79 * This function is called by all worker threads when they end (not including the controller).
81 virtual void worker_signal() = 0;
83 * @brief Waits for some work to process.
85 * This function is called by each worker thread (not including the controller) when it has no more work to do.
87 * @param round the expected round number
89 virtual void worker_wait(unsigned) = 0;
94 class PosixSynchro : public Synchro {
96 explicit PosixSynchro(Parmap<T>& parmap);
100 void worker_signal();
101 void worker_wait(unsigned round);
104 std::condition_variable ready_cond;
105 std::mutex ready_mutex;
106 std::condition_variable done_cond;
107 std::mutex done_mutex;
111 class FutexSynchro : public Synchro {
113 explicit FutexSynchro(Parmap<T>& parmap) : Synchro(parmap) {}
114 void master_signal();
116 void worker_signal();
117 void worker_wait(unsigned);
120 static void futex_wait(unsigned* uaddr, unsigned val);
121 static void futex_wake(unsigned* uaddr, unsigned val);
125 class BusyWaitSynchro : public Synchro {
127 explicit BusyWaitSynchro(Parmap<T>& parmap) : Synchro(parmap) {}
128 void master_signal();
130 void worker_signal();
131 void worker_wait(unsigned);
134 static void* worker_main(void* arg);
135 Synchro* new_synchro(e_xbt_parmap_mode_t mode);
138 Flag status; /**< is the parmap active or being destroyed? */
139 unsigned work_round; /**< index of the current round */
140 xbt_os_thread_t* workers; /**< worker thread handlers */
141 unsigned num_workers; /**< total number of worker threads including the controller */
142 Synchro* synchro; /**< synchronization object */
144 unsigned thread_counter = 0; /**< number of workers that have done the work */
145 void (*fun)(const T) = nullptr; /**< function to run in parallel on each element of data */
146 const std::vector<T>* data = nullptr; /**< parameters to pass to fun in parallel */
147 std::atomic<unsigned> index; /**< index of the next element of data to pick */
151 * @brief Creates a parallel map object
152 * @param num_workers number of worker threads to create
153 * @param mode how to synchronize the worker threads
155 template <typename T> Parmap<T>::Parmap(unsigned num_workers, e_xbt_parmap_mode_t mode)
157 XBT_CDEBUG(xbt_parmap, "Create new parmap (%u workers)", num_workers);
159 /* Initialize the thread pool data structure */
160 this->status = PARMAP_WORK;
161 this->work_round = 0;
162 this->workers = new xbt_os_thread_t[num_workers];
163 this->num_workers = num_workers;
164 this->synchro = new_synchro(mode);
166 /* Create the pool of worker threads */
167 this->workers[0] = nullptr;
168 unsigned int core_bind = 0;
169 for (unsigned i = 1; i < num_workers; i++) {
170 ThreadData* data = new ThreadData(*this, i);
171 this->workers[i] = xbt_os_thread_create(worker_main, data);
172 xbt_os_thread_bind(this->workers[i], core_bind);
173 if (core_bind != std::thread::hardware_concurrency() - 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 kernel::context::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)
310 template <typename T> Parmap<T>::PosixSynchro::~PosixSynchro()
314 template <typename T> void Parmap<T>::PosixSynchro::master_signal()
316 std::unique_lock<std::mutex> lk(ready_mutex);
317 this->parmap.thread_counter = 1;
318 this->parmap.work_round++;
319 /* wake all workers */
320 ready_cond.notify_all();
323 template <typename T> void Parmap<T>::PosixSynchro::master_wait()
325 std::unique_lock<std::mutex> lk(done_mutex);
326 while (this->parmap.thread_counter < this->parmap.num_workers) {
327 /* wait for all workers to be ready */
332 template <typename T> void Parmap<T>::PosixSynchro::worker_signal()
334 std::unique_lock<std::mutex> lk(done_mutex);
335 this->parmap.thread_counter++;
336 if (this->parmap.thread_counter == this->parmap.num_workers) {
337 /* all workers have finished, wake the controller */
338 done_cond.notify_one();
342 template <typename T> void Parmap<T>::PosixSynchro::worker_wait(unsigned round)
344 std::unique_lock<std::mutex> lk(ready_mutex);
345 /* wait for more work */
346 while (this->parmap.work_round != round) {
352 template <typename T> inline void Parmap<T>::FutexSynchro::futex_wait(unsigned* uaddr, unsigned val)
354 XBT_CVERB(xbt_parmap, "Waiting on futex %p", uaddr);
355 syscall(SYS_futex, uaddr, FUTEX_WAIT_PRIVATE, val, nullptr, nullptr, 0);
358 template <typename T> inline void Parmap<T>::FutexSynchro::futex_wake(unsigned* uaddr, unsigned val)
360 XBT_CVERB(xbt_parmap, "Waking futex %p", uaddr);
361 syscall(SYS_futex, uaddr, FUTEX_WAKE_PRIVATE, val, nullptr, nullptr, 0);
364 template <typename T> void Parmap<T>::FutexSynchro::master_signal()
366 __atomic_store_n(&this->parmap.thread_counter, 1, __ATOMIC_SEQ_CST);
367 __atomic_add_fetch(&this->parmap.work_round, 1, __ATOMIC_SEQ_CST);
368 /* wake all workers */
369 futex_wake(&this->parmap.work_round, std::numeric_limits<int>::max());
372 template <typename T> void Parmap<T>::FutexSynchro::master_wait()
374 unsigned count = __atomic_load_n(&this->parmap.thread_counter, __ATOMIC_SEQ_CST);
375 while (count < this->parmap.num_workers) {
376 /* wait for all workers to be ready */
377 futex_wait(&this->parmap.thread_counter, count);
378 count = __atomic_load_n(&this->parmap.thread_counter, __ATOMIC_SEQ_CST);
382 template <typename T> void Parmap<T>::FutexSynchro::worker_signal()
384 unsigned count = __atomic_add_fetch(&this->parmap.thread_counter, 1, __ATOMIC_SEQ_CST);
385 if (count == this->parmap.num_workers) {
386 /* all workers have finished, wake the controller */
387 futex_wake(&this->parmap.thread_counter, std::numeric_limits<int>::max());
391 template <typename T> void Parmap<T>::FutexSynchro::worker_wait(unsigned round)
393 unsigned work_round = __atomic_load_n(&this->parmap.work_round, __ATOMIC_SEQ_CST);
394 /* wait for more work */
395 while (work_round != round) {
396 futex_wait(&this->parmap.work_round, work_round);
397 work_round = __atomic_load_n(&this->parmap.work_round, __ATOMIC_SEQ_CST);
402 template <typename T> void Parmap<T>::BusyWaitSynchro::master_signal()
404 __atomic_store_n(&this->parmap.thread_counter, 1, __ATOMIC_SEQ_CST);
405 __atomic_add_fetch(&this->parmap.work_round, 1, __ATOMIC_SEQ_CST);
408 template <typename T> void Parmap<T>::BusyWaitSynchro::master_wait()
410 while (__atomic_load_n(&this->parmap.thread_counter, __ATOMIC_SEQ_CST) < this->parmap.num_workers) {
411 std::this_thread::yield();
415 template <typename T> void Parmap<T>::BusyWaitSynchro::worker_signal()
417 __atomic_add_fetch(&this->parmap.thread_counter, 1, __ATOMIC_SEQ_CST);
420 template <typename T> void Parmap<T>::BusyWaitSynchro::worker_wait(unsigned round)
422 /* wait for more work */
423 while (__atomic_load_n(&this->parmap.work_round, __ATOMIC_SEQ_CST) != round) {
424 std::this_thread::yield();