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"
13 #include "src/simix/smx_private.hpp" /* simix_global */
15 #include <boost/optional.hpp>
16 #include <condition_variable>
21 #include <linux/futex.h>
22 #include <sys/syscall.h>
26 #include <pthread_np.h>
29 XBT_LOG_EXTERNAL_CATEGORY(xbt_parmap);
34 /** @addtogroup XBT_parmap
36 * @brief Parallel map class
39 template <typename T> class Parmap {
41 Parmap(unsigned num_workers, e_xbt_parmap_mode_t mode);
42 Parmap(const Parmap&) = delete;
43 Parmap& operator=(const Parmap&) = delete;
45 void apply(void (*fun)(T), const std::vector<T>& data);
46 boost::optional<T> next();
49 enum Flag { PARMAP_WORK, PARMAP_DESTROY };
52 * @brief Thread data transmission structure
56 ThreadData(Parmap<T>& parmap, int id) : parmap(parmap), worker_id(id) {}
62 * @brief Synchronization object (different specializations).
66 explicit Synchro(Parmap<T>& parmap) : parmap(parmap) {}
67 virtual ~Synchro() = default;
69 * @brief Wakes all workers and waits for them to finish the tasks.
71 * This function is called by the controller thread.
73 virtual void master_signal() = 0;
75 * @brief Starts the parmap: waits for all workers to be ready and returns.
77 * This function is called by the controller thread.
79 virtual void master_wait() = 0;
81 * @brief Ends the parmap: wakes the controller thread when all workers terminate.
83 * This function is called by all worker threads when they end (not including the controller).
85 virtual void worker_signal() = 0;
87 * @brief Waits for some work to process.
89 * This function is called by each worker thread (not including the controller) when it has no more work to do.
91 * @param round the expected round number
93 virtual void worker_wait(unsigned) = 0;
98 class PosixSynchro : public Synchro {
100 explicit PosixSynchro(Parmap<T>& parmap);
102 void master_signal() override;
103 void master_wait() override;
104 void worker_signal() override;
105 void worker_wait(unsigned round) override;
108 std::condition_variable ready_cond;
109 std::mutex ready_mutex;
110 std::condition_variable done_cond;
111 std::mutex done_mutex;
115 class FutexSynchro : public Synchro {
117 explicit FutexSynchro(Parmap<T>& parmap) : Synchro(parmap) {}
118 void master_signal() override;
119 void master_wait() override;
120 void worker_signal() override;
121 void worker_wait(unsigned) override;
124 static void futex_wait(std::atomic_uint* uaddr, unsigned val);
125 static void futex_wake(std::atomic_uint* uaddr, unsigned val);
129 class BusyWaitSynchro : public Synchro {
131 explicit BusyWaitSynchro(Parmap<T>& parmap) : Synchro(parmap) {}
132 void master_signal() override;
133 void master_wait() override;
134 void worker_signal() override;
135 void worker_wait(unsigned) override;
138 static void* worker_main(void* arg);
139 Synchro* new_synchro(e_xbt_parmap_mode_t mode);
142 Flag status; /**< is the parmap active or being destroyed? */
143 std::atomic_uint work_round; /**< index of the current round */
144 std::vector<std::thread*> workers; /**< worker thread handlers */
145 unsigned num_workers; /**< total number of worker threads including the controller */
146 Synchro* synchro; /**< synchronization object */
148 std::atomic_uint thread_counter{0}; /**< number of workers that have done the work */
149 void (*fun)(const T) = nullptr; /**< function to run in parallel on each element of data */
150 const std::vector<T>* data = nullptr; /**< parameters to pass to fun in parallel */
151 std::atomic_uint index; /**< index of the next element of data to pick */
155 * @brief Creates a parallel map object
156 * @param num_workers number of worker threads to create
157 * @param mode how to synchronize the worker threads
159 template <typename T> Parmap<T>::Parmap(unsigned num_workers, e_xbt_parmap_mode_t mode)
161 XBT_CDEBUG(xbt_parmap, "Create new parmap (%u workers)", num_workers);
163 /* Initialize the thread pool data structure */
164 this->status = PARMAP_WORK;
165 this->work_round = 0;
166 this->workers.resize(num_workers);
167 this->num_workers = num_workers;
168 this->synchro = new_synchro(mode);
170 /* Create the pool of worker threads (the caller of apply() will be worker[0]) */
171 this->workers[0] = nullptr;
172 XBT_ATTRIB_UNUSED unsigned int core_bind = 0;
174 for (unsigned i = 1; i < num_workers; i++) {
175 this->workers[i] = new std::thread(worker_main, new ThreadData(*this, i));
177 /* Bind the worker to a core if possible */
178 #if HAVE_PTHREAD_SETAFFINITY
179 #if HAVE_PTHREAD_NP_H /* FreeBSD ? */
181 size_t size = sizeof(cpuset_t);
184 size_t size = sizeof(cpu_set_t);
186 pthread_t pthread = this->workers[i]->native_handle();
188 CPU_SET(core_bind, &cpuset);
189 pthread_setaffinity_np(pthread, size, &cpuset);
190 if (core_bind != std::thread::hardware_concurrency() - 1)
199 * @brief Destroys a parmap
201 template <typename T> Parmap<T>::~Parmap()
203 status = PARMAP_DESTROY;
204 synchro->master_signal();
206 for (unsigned i = 1; i < num_workers; i++) {
214 * @brief Applies a list of tasks in parallel.
215 * @param fun the function to call in parallel
216 * @param data each element of this vector will be passed as an argument to fun
218 template <typename T> void Parmap<T>::apply(void (*fun)(T), const std::vector<T>& data)
220 /* Assign resources to worker threads (we are maestro here)*/
224 this->synchro->master_signal(); // maestro runs futex_wake to wake all the minions (the working threads)
225 this->work(); // maestro works with its minions
226 this->synchro->master_wait(); // When there is no more work to do, then maestro waits for the last minion to stop
227 XBT_CDEBUG(xbt_parmap, "Job done"); // ... and proceeds
231 * @brief Returns a next task to process.
233 * Worker threads call this function to get more work.
235 * @return the next task to process, or throws a std::out_of_range exception if there is no more work
237 template <typename T> boost::optional<T> Parmap<T>::next()
239 unsigned index = this->index.fetch_add(1, std::memory_order_relaxed);
240 if (index < this->data->size())
241 return (*this->data)[index];
247 * @brief Main work loop: applies fun to elements in turn.
249 template <typename T> void Parmap<T>::work()
251 unsigned length = this->data->size();
252 unsigned index = this->index.fetch_add(1, std::memory_order_relaxed);
253 while (index < length) {
254 this->fun((*this->data)[index]);
255 index = this->index.fetch_add(1, std::memory_order_relaxed);
260 * Get a synchronization object for given mode.
261 * @param mode the synchronization mode
263 template <typename T> typename Parmap<T>::Synchro* Parmap<T>::new_synchro(e_xbt_parmap_mode_t mode)
265 if (mode == XBT_PARMAP_DEFAULT) {
267 mode = XBT_PARMAP_FUTEX;
269 mode = XBT_PARMAP_POSIX;
274 case XBT_PARMAP_POSIX:
275 res = new PosixSynchro(*this);
277 case XBT_PARMAP_FUTEX:
279 res = new FutexSynchro(*this);
281 xbt_die("Futex is not available on this OS.");
284 case XBT_PARMAP_BUSY_WAIT:
285 res = new BusyWaitSynchro(*this);
293 /** @brief Main function of a worker thread */
294 template <typename T> void* Parmap<T>::worker_main(void* arg)
296 ThreadData* data = static_cast<ThreadData*>(arg);
297 Parmap<T>& parmap = data->parmap;
299 smx_context_t context = simix_global->context_factory->create_context(std::function<void()>(), nullptr);
300 kernel::context::Context::set_current(context);
302 XBT_CDEBUG(xbt_parmap, "New worker thread created");
304 /* Worker's main loop */
306 round++; // New scheduling round
307 parmap.synchro->worker_wait(round);
308 if (parmap.status == PARMAP_DESTROY)
311 XBT_CDEBUG(xbt_parmap, "Worker %d got a job", data->worker_id);
313 parmap.synchro->worker_signal();
314 XBT_CDEBUG(xbt_parmap, "Worker %d has finished", data->worker_id);
316 /* We are destroying the parmap */
322 template <typename T> Parmap<T>::PosixSynchro::PosixSynchro(Parmap<T>& parmap) : Synchro(parmap)
326 template <typename T> Parmap<T>::PosixSynchro::~PosixSynchro()
330 template <typename T> void Parmap<T>::PosixSynchro::master_signal()
332 std::unique_lock<std::mutex> lk(ready_mutex);
333 this->parmap.thread_counter = 1;
334 this->parmap.work_round++;
335 /* wake all workers */
336 ready_cond.notify_all();
339 template <typename T> void Parmap<T>::PosixSynchro::master_wait()
341 std::unique_lock<std::mutex> lk(done_mutex);
342 while (this->parmap.thread_counter < this->parmap.num_workers) {
343 /* wait for all workers to be ready */
348 template <typename T> void Parmap<T>::PosixSynchro::worker_signal()
350 std::unique_lock<std::mutex> lk(done_mutex);
351 this->parmap.thread_counter++;
352 if (this->parmap.thread_counter == this->parmap.num_workers) {
353 /* all workers have finished, wake the controller */
354 done_cond.notify_one();
358 template <typename T> void Parmap<T>::PosixSynchro::worker_wait(unsigned round)
360 std::unique_lock<std::mutex> lk(ready_mutex);
361 /* wait for more work */
362 while (this->parmap.work_round != round) {
368 template <typename T> inline void Parmap<T>::FutexSynchro::futex_wait(std::atomic_uint* uaddr, unsigned val)
370 XBT_CVERB(xbt_parmap, "Waiting on futex %p", uaddr);
371 syscall(SYS_futex, uaddr, FUTEX_WAIT_PRIVATE, val, nullptr, nullptr, 0);
374 template <typename T> inline void Parmap<T>::FutexSynchro::futex_wake(std::atomic_uint* uaddr, unsigned val)
376 XBT_CVERB(xbt_parmap, "Waking futex %p", uaddr);
377 syscall(SYS_futex, uaddr, FUTEX_WAKE_PRIVATE, val, nullptr, nullptr, 0);
380 template <typename T> void Parmap<T>::FutexSynchro::master_signal()
382 this->parmap.thread_counter.store(1);
383 this->parmap.work_round.fetch_add(1);
384 /* wake all workers */
385 futex_wake(&this->parmap.work_round, std::numeric_limits<int>::max());
388 template <typename T> void Parmap<T>::FutexSynchro::master_wait()
390 unsigned count = this->parmap.thread_counter.load();
391 while (count < this->parmap.num_workers) {
392 /* wait for all workers to be ready */
393 futex_wait(&this->parmap.thread_counter, count);
394 count = this->parmap.thread_counter.load();
398 template <typename T> void Parmap<T>::FutexSynchro::worker_signal()
400 unsigned count = this->parmap.thread_counter.fetch_add(1) + 1;
401 if (count == this->parmap.num_workers) {
402 /* all workers have finished, wake the controller */
403 futex_wake(&this->parmap.thread_counter, std::numeric_limits<int>::max());
407 template <typename T> void Parmap<T>::FutexSynchro::worker_wait(unsigned round)
409 unsigned work_round = this->parmap.work_round.load();
410 /* wait for more work */
411 while (work_round != round) {
412 futex_wait(&this->parmap.work_round, work_round);
413 work_round = this->parmap.work_round.load();
418 template <typename T> void Parmap<T>::BusyWaitSynchro::master_signal()
420 this->parmap.thread_counter.store(1);
421 this->parmap.work_round.fetch_add(1);
424 template <typename T> void Parmap<T>::BusyWaitSynchro::master_wait()
426 while (this->parmap.thread_counter.load() < this->parmap.num_workers) {
427 std::this_thread::yield();
431 template <typename T> void Parmap<T>::BusyWaitSynchro::worker_signal()
433 this->parmap.thread_counter.fetch_add(1);
436 template <typename T> void Parmap<T>::BusyWaitSynchro::worker_wait(unsigned round)
438 /* wait for more work */
439 while (this->parmap.work_round.load() != round) {
440 std::this_thread::yield();