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>
22 #include <linux/futex.h>
23 #include <sys/syscall.h>
27 #include <pthread_np.h>
30 XBT_LOG_EXTERNAL_CATEGORY(xbt_parmap);
35 /** @addtogroup XBT_parmap
37 * @brief Parallel map class
40 template <typename T> class Parmap {
42 Parmap(unsigned num_workers, e_xbt_parmap_mode_t mode);
43 Parmap(const Parmap&) = delete;
44 Parmap& operator=(const Parmap&) = delete;
46 void apply(std::function<void(T)>&& fun, const std::vector<T>& data);
47 boost::optional<T> next();
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() override;
102 void master_wait() override;
103 void worker_signal() override;
104 void worker_wait(unsigned round) override;
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() override;
118 void master_wait() override;
119 void worker_signal() override;
120 void worker_wait(unsigned) override;
123 static void futex_wait(std::atomic_uint* uaddr, unsigned val);
124 static void futex_wake(std::atomic_uint* uaddr, unsigned val);
128 class BusyWaitSynchro : public Synchro {
130 explicit BusyWaitSynchro(Parmap<T>& parmap) : Synchro(parmap) {}
131 void master_signal() override;
132 void master_wait() override;
133 void worker_signal() override;
134 void worker_wait(unsigned) override;
137 static void worker_main(ThreadData* data);
138 Synchro* new_synchro(e_xbt_parmap_mode_t mode);
141 bool destroying; /**< is the parmap being destroyed? */
142 std::atomic_uint 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 std::atomic_uint thread_counter{0}; /**< number of workers that have done the work */
148 std::function<void(T)> fun; /**< 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_uint 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->destroying = false;
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()
203 synchro->master_signal();
205 for (unsigned i = 1; i < num_workers; i++) {
213 * @brief Applies a list of tasks in parallel.
214 * @param fun the function to call in parallel
215 * @param data each element of this vector will be passed as an argument to fun
217 template <typename T> void Parmap<T>::apply(std::function<void(T)>&& fun, const std::vector<T>& data)
219 /* Assign resources to worker threads (we are maestro here)*/
220 this->fun = std::move(fun);
223 this->synchro->master_signal(); // maestro runs futex_wake to wake all the minions (the working threads)
224 this->work(); // maestro works with its minions
225 this->synchro->master_wait(); // When there is no more work to do, then maestro waits for the last minion to stop
226 XBT_CDEBUG(xbt_parmap, "Job done"); // ... and proceeds
230 * @brief Returns a next task to process.
232 * Worker threads call this function to get more work.
234 * @return the next task to process, or throws a std::out_of_range exception if there is no more work
236 template <typename T> boost::optional<T> Parmap<T>::next()
238 unsigned index = this->index.fetch_add(1, std::memory_order_relaxed);
239 if (index < this->data->size())
240 return (*this->data)[index];
246 * @brief Main work loop: applies fun to elements in turn.
248 template <typename T> void Parmap<T>::work()
250 unsigned length = this->data->size();
251 unsigned index = this->index.fetch_add(1, std::memory_order_relaxed);
252 while (index < length) {
253 this->fun((*this->data)[index]);
254 index = this->index.fetch_add(1, std::memory_order_relaxed);
259 * Get a synchronization object for given mode.
260 * @param mode the synchronization mode
262 template <typename T> typename Parmap<T>::Synchro* Parmap<T>::new_synchro(e_xbt_parmap_mode_t mode)
264 if (mode == XBT_PARMAP_DEFAULT) {
266 mode = XBT_PARMAP_FUTEX;
268 mode = XBT_PARMAP_POSIX;
273 case XBT_PARMAP_POSIX:
274 res = new PosixSynchro(*this);
276 case XBT_PARMAP_FUTEX:
278 res = new FutexSynchro(*this);
280 xbt_die("Futex is not available on this OS.");
283 case XBT_PARMAP_BUSY_WAIT:
284 res = new BusyWaitSynchro(*this);
292 /** @brief Main function of a worker thread */
293 template <typename T> void Parmap<T>::worker_main(ThreadData* data)
295 Parmap<T>& parmap = data->parmap;
297 kernel::context::Context* context = simix_global->context_factory->create_context(std::function<void()>(), 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.destroying)
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 */
319 template <typename T> Parmap<T>::PosixSynchro::PosixSynchro(Parmap<T>& parmap) : Synchro(parmap)
323 template <typename T> Parmap<T>::PosixSynchro::~PosixSynchro()
327 template <typename T> void Parmap<T>::PosixSynchro::master_signal()
329 std::unique_lock<std::mutex> lk(ready_mutex);
330 this->parmap.thread_counter = 1;
331 this->parmap.work_round++;
332 /* wake all workers */
333 ready_cond.notify_all();
336 template <typename T> void Parmap<T>::PosixSynchro::master_wait()
338 std::unique_lock<std::mutex> lk(done_mutex);
339 while (this->parmap.thread_counter < this->parmap.num_workers) {
340 /* wait for all workers to be ready */
345 template <typename T> void Parmap<T>::PosixSynchro::worker_signal()
347 std::unique_lock<std::mutex> lk(done_mutex);
348 this->parmap.thread_counter++;
349 if (this->parmap.thread_counter == this->parmap.num_workers) {
350 /* all workers have finished, wake the controller */
351 done_cond.notify_one();
355 template <typename T> void Parmap<T>::PosixSynchro::worker_wait(unsigned round)
357 std::unique_lock<std::mutex> lk(ready_mutex);
358 /* wait for more work */
359 while (this->parmap.work_round != round) {
365 template <typename T> inline void Parmap<T>::FutexSynchro::futex_wait(std::atomic_uint* uaddr, unsigned val)
367 XBT_CVERB(xbt_parmap, "Waiting on futex %p", uaddr);
368 syscall(SYS_futex, uaddr, FUTEX_WAIT_PRIVATE, val, nullptr, nullptr, 0);
371 template <typename T> inline void Parmap<T>::FutexSynchro::futex_wake(std::atomic_uint* uaddr, unsigned val)
373 XBT_CVERB(xbt_parmap, "Waking futex %p", uaddr);
374 syscall(SYS_futex, uaddr, FUTEX_WAKE_PRIVATE, val, nullptr, nullptr, 0);
377 template <typename T> void Parmap<T>::FutexSynchro::master_signal()
379 this->parmap.thread_counter.store(1);
380 this->parmap.work_round.fetch_add(1);
381 /* wake all workers */
382 futex_wake(&this->parmap.work_round, std::numeric_limits<int>::max());
385 template <typename T> void Parmap<T>::FutexSynchro::master_wait()
387 unsigned count = this->parmap.thread_counter.load();
388 while (count < this->parmap.num_workers) {
389 /* wait for all workers to be ready */
390 futex_wait(&this->parmap.thread_counter, count);
391 count = this->parmap.thread_counter.load();
395 template <typename T> void Parmap<T>::FutexSynchro::worker_signal()
397 unsigned count = this->parmap.thread_counter.fetch_add(1) + 1;
398 if (count == this->parmap.num_workers) {
399 /* all workers have finished, wake the controller */
400 futex_wake(&this->parmap.thread_counter, std::numeric_limits<int>::max());
404 template <typename T> void Parmap<T>::FutexSynchro::worker_wait(unsigned round)
406 unsigned work_round = this->parmap.work_round.load();
407 /* wait for more work */
408 while (work_round != round) {
409 futex_wait(&this->parmap.work_round, work_round);
410 work_round = this->parmap.work_round.load();
415 template <typename T> void Parmap<T>::BusyWaitSynchro::master_signal()
417 this->parmap.thread_counter.store(1);
418 this->parmap.work_round.fetch_add(1);
421 template <typename T> void Parmap<T>::BusyWaitSynchro::master_wait()
423 while (this->parmap.thread_counter.load() < this->parmap.num_workers) {
424 std::this_thread::yield();
428 template <typename T> void Parmap<T>::BusyWaitSynchro::worker_signal()
430 this->parmap.thread_counter.fetch_add(1);
433 template <typename T> void Parmap<T>::BusyWaitSynchro::worker_wait(unsigned round)
435 /* wait for more work */
436 while (this->parmap.work_round.load() != round) {
437 std::this_thread::yield();