1 /* A thread pool (C++ version). */
3 /* Copyright (c) 2004-2020 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 = false; /**< is the parmap being destroyed? */
142 std::atomic_uint work_round{0}; /**< 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{0}; /**< 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->workers.resize(num_workers);
164 this->num_workers = num_workers;
165 this->synchro = new_synchro(mode);
167 /* Create the pool of worker threads (the caller of apply() will be worker[0]) */
168 this->workers[0] = nullptr;
170 for (unsigned i = 1; i < num_workers; i++) {
171 ThreadData* data = new ThreadData(*this, i);
172 this->workers[i] = new std::thread(worker_main, data);
174 /* Bind the worker to a core if possible */
175 #if HAVE_PTHREAD_SETAFFINITY
176 #if HAVE_PTHREAD_NP_H /* FreeBSD ? */
178 size_t size = sizeof(cpuset_t);
181 size_t size = sizeof(cpu_set_t);
183 pthread_t pthread = this->workers[i]->native_handle();
184 int core_bind = (i - 1) % std::thread::hardware_concurrency();
186 CPU_SET(core_bind, &cpuset);
187 pthread_setaffinity_np(pthread, size, &cpuset);
193 * @brief Destroys a parmap
195 template <typename T> Parmap<T>::~Parmap()
198 synchro->master_signal();
200 for (unsigned i = 1; i < num_workers; i++) {
208 * @brief Applies a list of tasks in parallel.
209 * @param fun the function to call in parallel
210 * @param data each element of this vector will be passed as an argument to fun
212 template <typename T> void Parmap<T>::apply(std::function<void(T)>&& fun, const std::vector<T>& data)
214 /* Assign resources to worker threads (we are maestro here)*/
215 this->fun = std::move(fun);
218 this->synchro->master_signal(); // maestro runs futex_wake to wake all the minions (the working threads)
219 this->work(); // maestro works with its minions
220 this->synchro->master_wait(); // When there is no more work to do, then maestro waits for the last minion to stop
221 XBT_CDEBUG(xbt_parmap, "Job done"); // ... and proceeds
225 * @brief Returns a next task to process.
227 * Worker threads call this function to get more work.
229 * @return the next task to process, or throws a std::out_of_range exception if there is no more work
231 template <typename T> boost::optional<T> Parmap<T>::next()
233 unsigned index = this->index.fetch_add(1, std::memory_order_relaxed);
234 if (index < this->data->size())
235 return (*this->data)[index];
241 * @brief Main work loop: applies fun to elements in turn.
243 template <typename T> void Parmap<T>::work()
245 unsigned length = this->data->size();
246 unsigned index = this->index.fetch_add(1, std::memory_order_relaxed);
247 while (index < length) {
248 this->fun((*this->data)[index]);
249 index = this->index.fetch_add(1, std::memory_order_relaxed);
254 * Get a synchronization object for given mode.
255 * @param mode the synchronization mode
257 template <typename T> typename Parmap<T>::Synchro* Parmap<T>::new_synchro(e_xbt_parmap_mode_t mode)
259 if (mode == XBT_PARMAP_DEFAULT) {
261 mode = XBT_PARMAP_FUTEX;
263 mode = XBT_PARMAP_POSIX;
268 case XBT_PARMAP_POSIX:
269 res = new PosixSynchro(*this);
271 case XBT_PARMAP_FUTEX:
273 res = new FutexSynchro(*this);
275 xbt_die("Futex is not available on this OS.");
278 case XBT_PARMAP_BUSY_WAIT:
279 res = new BusyWaitSynchro(*this);
287 /** @brief Main function of a worker thread */
288 template <typename T> void Parmap<T>::worker_main(ThreadData* data)
290 Parmap<T>& parmap = data->parmap;
292 kernel::context::Context* context = simix_global->context_factory->create_context(std::function<void()>(), nullptr);
293 kernel::context::Context::set_current(context);
295 XBT_CDEBUG(xbt_parmap, "New worker thread created");
297 /* Worker's main loop */
299 round++; // New scheduling round
300 parmap.synchro->worker_wait(round);
301 if (parmap.destroying)
304 XBT_CDEBUG(xbt_parmap, "Worker %d got a job", data->worker_id);
306 parmap.synchro->worker_signal();
307 XBT_CDEBUG(xbt_parmap, "Worker %d has finished", data->worker_id);
309 /* We are destroying the parmap */
314 template <typename T> Parmap<T>::PosixSynchro::PosixSynchro(Parmap<T>& parmap) : Synchro(parmap)
318 template <typename T> Parmap<T>::PosixSynchro::~PosixSynchro()
322 template <typename T> void Parmap<T>::PosixSynchro::master_signal()
324 std::unique_lock<std::mutex> lk(ready_mutex);
325 this->parmap.thread_counter = 1;
326 this->parmap.work_round++;
327 /* wake all workers */
328 ready_cond.notify_all();
331 template <typename T> void Parmap<T>::PosixSynchro::master_wait()
333 std::unique_lock<std::mutex> lk(done_mutex);
334 while (this->parmap.thread_counter < this->parmap.num_workers) {
335 /* wait for all workers to be ready */
340 template <typename T> void Parmap<T>::PosixSynchro::worker_signal()
342 std::unique_lock<std::mutex> lk(done_mutex);
343 this->parmap.thread_counter++;
344 if (this->parmap.thread_counter == this->parmap.num_workers) {
345 /* all workers have finished, wake the controller */
346 done_cond.notify_one();
350 template <typename T> void Parmap<T>::PosixSynchro::worker_wait(unsigned round)
352 std::unique_lock<std::mutex> lk(ready_mutex);
353 /* wait for more work */
354 while (this->parmap.work_round != round) {
360 template <typename T> inline void Parmap<T>::FutexSynchro::futex_wait(std::atomic_uint* uaddr, unsigned val)
362 XBT_CVERB(xbt_parmap, "Waiting on futex %p", uaddr);
363 syscall(SYS_futex, uaddr, FUTEX_WAIT_PRIVATE, val, nullptr, nullptr, 0);
366 template <typename T> inline void Parmap<T>::FutexSynchro::futex_wake(std::atomic_uint* uaddr, unsigned val)
368 XBT_CVERB(xbt_parmap, "Waking futex %p", uaddr);
369 syscall(SYS_futex, uaddr, FUTEX_WAKE_PRIVATE, val, nullptr, nullptr, 0);
372 template <typename T> void Parmap<T>::FutexSynchro::master_signal()
374 this->parmap.thread_counter.store(1);
375 this->parmap.work_round.fetch_add(1);
376 /* wake all workers */
377 futex_wake(&this->parmap.work_round, std::numeric_limits<int>::max());
380 template <typename T> void Parmap<T>::FutexSynchro::master_wait()
382 unsigned count = this->parmap.thread_counter.load();
383 while (count < this->parmap.num_workers) {
384 /* wait for all workers to be ready */
385 futex_wait(&this->parmap.thread_counter, count);
386 count = this->parmap.thread_counter.load();
390 template <typename T> void Parmap<T>::FutexSynchro::worker_signal()
392 unsigned count = this->parmap.thread_counter.fetch_add(1) + 1;
393 if (count == this->parmap.num_workers) {
394 /* all workers have finished, wake the controller */
395 futex_wake(&this->parmap.thread_counter, std::numeric_limits<int>::max());
399 template <typename T> void Parmap<T>::FutexSynchro::worker_wait(unsigned round)
401 unsigned work_round = this->parmap.work_round.load();
402 /* wait for more work */
403 while (work_round != round) {
404 futex_wait(&this->parmap.work_round, work_round);
405 work_round = this->parmap.work_round.load();
410 template <typename T> void Parmap<T>::BusyWaitSynchro::master_signal()
412 this->parmap.thread_counter.store(1);
413 this->parmap.work_round.fetch_add(1);
416 template <typename T> void Parmap<T>::BusyWaitSynchro::master_wait()
418 while (this->parmap.thread_counter.load() < this->parmap.num_workers) {
419 std::this_thread::yield();
423 template <typename T> void Parmap<T>::BusyWaitSynchro::worker_signal()
425 this->parmap.thread_counter.fetch_add(1);
428 template <typename T> void Parmap<T>::BusyWaitSynchro::worker_wait(unsigned round)
430 /* wait for more work */
431 while (this->parmap.work_round.load() != round) {
432 std::this_thread::yield();