<\/p>\n\n\n\n
Exemples de langages\/concepts :<\/strong><\/p>\n\n\n\n\nJava<\/strong> : Thread, ExecutorService<\/li>\n\n\n\nPython<\/strong> : threading, asyncio<\/li>\n\n\n\nJavaScript<\/strong> : Promise, async\/await<\/li>\n\n\n\nGo<\/strong> : goroutines<\/li>\n\n\n\nRust<\/strong> : tokio, async<\/li>\n<\/ul>\n\n\n\nEn C\/C++<\/strong>, la programmation concurrente peut se faire de plusieurs mani\u00e8res, notamment avec :<\/p>\n\n\n\n\nLes threads (pthreads en C)<\/strong><\/li>\n<\/ol>\n\n\n\nVoici un exemple simple en C<\/strong> avec POSIX threads<\/strong> (pthreads) :<\/p>\n\n\n\n#include <stdio.h>\n#include <pthread.h>\n\nvoid* print_message(void* arg) {\n\u00a0\u00a0\u00a0 char* message = (char*) arg;\n\u00a0\u00a0\u00a0 printf(\"%s\\n\", message);\n\u00a0\u00a0\u00a0 return NULL;\n}\n\nint main() {\n\u00a0\u00a0\u00a0 pthread_t thread1, thread2;\n\u00a0\u00a0\u00a0 pthread_create(&thread1, NULL, print_message, \"Bonjour depuis le thread 1\");\n\u00a0\u00a0\u00a0 pthread_create(&thread2, NULL, print_message, \"Bonjour depuis le thread 2\");\n\u00a0\u00a0\u00a0 pthread_join(thread1, NULL);\n\u00a0\u00a0\u00a0 pthread_join(thread2, NULL);\n\u00a0\u00a0\u00a0 return 0;\n}<\/code><\/pre>\n\n\n\n<\/p>\n\n\n\n
\npthread_create lance une nouvelle fonction dans un thread<\/strong> s\u00e9par\u00e9.<\/li>\n\n\n\npthread_join attend que le thread se termine.<\/li>\n<\/ul>\n\n\n\n2. C++11 et plus : std::thread<\/strong><\/p>\n\n\n\nDepuis C++11<\/strong>, on peut faire plus simple avec la biblioth\u00e8que standard :<\/p>\n\n\n\n<\/p>\n\n\n\n
#include <iostream>\n#include <thread>\n\nvoid say_hello(const std::string& name) {\n\u00a0\u00a0\u00a0 std::cout << \"Bonjour \" << name << \" depuis un thread !\" << std::endl;\n}\n\nint main() {\n\u00a0\u00a0\u00a0 std::thread t1(say_hello, \"Alice\");\n\u00a0\u00a0\u00a0 std::thread t2(say_hello, \"Bob\");\n\u00a0\u00a0\u00a0 t1.join();\n\u00a0\u00a0\u00a0 t2.join();\n\u00a0\u00a0\u00a0 return 0;\n}<\/code><\/pre>\n\n\n\n<\/p>\n\n\n\n
\nstd::thread cr\u00e9e un nouveau thread qui ex\u00e9cute une fonction.<\/li>\n\n\n\n .join() bloque le thread principal jusqu\u2019\u00e0 ce que le thread secondaire soit termin\u00e9.<\/li>\n<\/ul>\n\n\n\n3. Ex\u00e9cution parall\u00e8le vs concurrence<\/strong><\/p>\n\n\n\n\nLes deux exemples ci-dessus montrent de la concurrence<\/strong>.<\/li>\n\n\n\nSi ton ordinateur a plusieurs c\u0153urs, il pourrait<\/strong> ex\u00e9cuter les threads en parall\u00e8le<\/strong>, sinon, ils seront intercal\u00e9s<\/strong> (le syst\u00e8me d\u2019exploitation change de thread r\u00e9guli\u00e8rement).<\/li>\n<\/ul>\n\n\n\nPoints d\u2019attention :<\/strong><\/p>\n\n\n\n\nLes probl\u00e8mes de synchronisation<\/strong> (par exemple : deux threads \u00e9crivant sur la m\u00eame variable en m\u00eame temps)<\/li>\n\n\n\nUtiliser des mutex<\/strong> (std::mutex, pthread_mutex_t) pour prot\u00e9ger les ressources partag\u00e9es<\/li>\n<\/ul>\n\n\n\nLe throughput<\/h1>\n\n\n\n Le throughput<\/strong> (ou d\u00e9bit<\/strong> en fran\u00e7ais) est un terme technique<\/strong> qui d\u00e9signe la quantit\u00e9 de travail accomplie dans un certain laps de temps<\/strong>. En informatique, on l\u2019utilise souvent pour mesurer la performance<\/strong> d\u2019un syst\u00e8me, comme un programme, un r\u00e9seau ou un processeur.<\/p>\n\n\n\nExemples selon le contexte :<\/strong><\/p>\n\n\n\n1. En r\u00e9seau :<\/strong><\/p>\n\n\n\n\nLe throughput est la quantit\u00e9 de donn\u00e9es<\/strong> transmises par seconde<\/strong>.<\/li>\n\n\n\nExemple : 100 Mbps = 100 m\u00e9gabits par seconde.<\/li>\n<\/ul>\n\n\n\n2. En programmation concurrente :<\/strong><\/p>\n\n\n\n\nC\u2019est le nombre de t\u00e2ches trait\u00e9es par seconde<\/strong>.<\/li>\n\n\n\nExemple : un serveur web peut avoir un throughput de 1000 requ\u00eates HTTP par seconde<\/strong>.<\/li>\n<\/ul>\n\n\n\n3. En base de donn\u00e9es :<\/strong><\/p>\n\n\n\n\nNombre de transactions<\/strong> ou de requ\u00eates<\/strong> trait\u00e9es par seconde.<\/li>\n<\/ul>\n\n\n\n4. En syst\u00e8me (OS) :<\/strong><\/p>\n\n\n\n\nCombien de processus ou de threads peuvent \u00eatre ex\u00e9cut\u00e9s dans un temps donn\u00e9.<\/li>\n<\/ul>\n\n\n\nPourquoi c\u2019est important ?<\/strong><\/p>\n\n\n\n\nPlus le throughput est \u00e9lev\u00e9, plus le syst\u00e8me est performant<\/strong>.<\/li>\n\n\n\nIl est souvent en \u00e9quilibre avec la latence<\/strong> :\n\nLatence<\/strong> = temps de r\u00e9ponse pour une requ\u00eate<\/li>\n\n\n\nThroughput<\/strong> = combien de requ\u00eates on peut g\u00e9rer par seconde<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\nOn peut avoir :<\/strong><\/p>\n\n\n\n\nUne faible latence mais un faible throughput (rapide mais peu de volume)<\/li>\n\n\n\n Un haut throughput avec une latence plus grande (beaucoup de volume, mais plus lent par requ\u00eate)<\/li>\n<\/ul>\n\n\n\n1. Exemple : mesurer le throughput<\/strong><\/p>\n\n\n\nImaginons que tu veux lancer 100 000 t\u00e2ches, et mesurer combien sont trait\u00e9es par seconde avec des threads.<\/p>\n\n\n\n
Code C++ avec std::thread et std::chrono :<\/strong><\/p>\n\n\n\n#include <iostream>\n#include <thread>\n#include <vector>\n#include <chrono>\n#include <atomic>\n\nvoid task(std::atomic<int>& counter) {\n\u00a0\u00a0\u00a0 \/\/ Simule une petite t\u00e2che\n\u00a0\u00a0\u00a0 counter++;\n}\n\nint main() {\n\n\u00a0\u00a0\u00a0 const int num_tasks = 100000;\n\u00a0\u00a0\u00a0 std::atomic<int> counter{0};\n\u00a0\u00a0\u00a0 auto start = std::chrono::high_resolution_clock::now();\n\u00a0\u00a0\u00a0 std::vector<std::thread> threads;\n\u00a0\u00a0\u00a0 for (int i = 0; i < num_tasks; ++i) {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 threads.emplace_back(task, std::ref(counter));\n\u00a0\u00a0\u00a0 }\n\u00a0\u00a0\u00a0 for (auto& t : threads) {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 t.join();\n\u00a0\u00a0\u00a0 }\n\u00a0\u00a0\u00a0 auto end = std::chrono::high_resolution_clock::now();\n\u00a0\u00a0\u00a0 std::chrono::duration<double> duration = end - start;\n\u00a0\u00a0\u00a0 double throughput = num_tasks \/ duration.count();\n\u00a0\u00a0\u00a0 std::cout << \"Throughput: \" << throughput << \" tasks\/sec\" << std::endl;\n\u00a0\u00a0\u00a0 return 0;\n}<\/code><\/pre>\n\n\n\n<\/p>\n\n\n\n
Ce que fait ce code :<\/strong><\/p>\n\n\n\n\nLance 100 000 threads qui incr\u00e9mentent un compteur.<\/li>\n\n\n\n Mesure le temps total.<\/li>\n\n\n\n Calcule le throughput = nombre de t\u00e2ches \/ temps en secondes.<\/li>\n<\/ul>\n\n\n\n2. Comment am\u00e9liorer le throughput ?<\/strong><\/p>\n\n\n\nLancer 100 000 threads n\u2019est pas efficace<\/strong>. Voici quelques techniques d\u2019optimisation<\/strong> :<\/p>\n\n\n\na) Utiliser un thread pool<\/strong><\/p>\n\n\n\nAu lieu de cr\u00e9er un thread par t\u00e2che, on cr\u00e9e un petit nombre de threads<\/strong> qui travaillent en boucle.<\/p>\n\n\n\nPourquoi ?<\/strong>co\u00fbteux<\/strong>. Les thread pools r\u00e9utilisent les threads.<\/p>\n\n\n\nb) Limiter les acc\u00e8s partag\u00e9s<\/strong><\/p>\n\n\n\nLes acc\u00e8s concurrents \u00e0 des ressources partag\u00e9es (comme counter) n\u00e9cessitent des verrous ou des op\u00e9rations atomiques<\/strong>, ce qui ralentit<\/strong>.<\/p>\n\n\n\nc) Batching (traitement par lots)<\/strong><\/p>\n\n\n\nSi possible, regrouper plusieurs petites t\u00e2ches en une plus grosse r\u00e9duit les appels aux threads.<\/p>\n\n\n\n
d) Asynchrone \/ futures \/ coroutines<\/strong><\/p>\n\n\n\nUtiliser des techniques non bloquantes (ex: std::async, std::future, coroutines en C++20) peut am\u00e9liorer le d\u00e9bit.<\/p>\n\n\n\n
En r\u00e9sum\u00e9 :<\/strong><\/p>\n\n\n\n\nThroughput = nombre de t\u00e2ches \/ temps<\/strong><\/li>\n\n\n\nPour le mesurer : chronom\u00e8tre + compteur<\/li>\n\n\n\n Pour l\u2019am\u00e9liorer : \u00e9viter la surcharge (cr\u00e9ation de threads, contention), utiliser des pools<\/strong>, faire du batch<\/strong>, etc.<\/li>\n<\/ul>\n\n\n\nExemple : Thread Pool + mesure du throughput<\/strong><\/p>\n\n\n\n#include <iostream>\n#include <thread>\n#include <vector>\n#include <queue>\n#include <functional>\n#include <mutex>\n#include <condition_variable>\n#include <atomic>\n#include <chrono>\n\nclass ThreadPool {\npublic:\n\u00a0\u00a0\u00a0 ThreadPool(size_t num_threads) {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 for (size_t i = 0; i < num_threads; ++i) {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 workers.emplace_back([this]() {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 while (true) {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 std::function<void()> task;\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 {\u00a0\u00a0 \/\/ zone prot\u00e9g\u00e9e\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 std::unique_lock<std::mutex> lock(queue_mutex);\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 condition.wait(lock, [this]() {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 return stop || !tasks.empty();\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 });\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 if (stop && tasks.empty()) return;\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 task = std::move(tasks.front());\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 tasks.pop();\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 }\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 task();\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 }\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 });\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 }\n\u00a0\u00a0\u00a0 }\n\u00a0\u00a0\u00a0 void enqueue(std::function<void()> task) {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 std::lock_guard<std::mutex> lock(queue_mutex);\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 tasks.push(task);\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 }\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 condition.notify_one();\n\u00a0\u00a0\u00a0 }\n\u00a0\u00a0\u00a0 void shutdown() {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 std::lock_guard<std::mutex> lock(queue_mutex);\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 stop = true;\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 }\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 condition.notify_all();\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 for (auto& worker : workers) worker.join();\n\u00a0\u00a0\u00a0 }\n\u00a0\u00a0\u00a0 ~ThreadPool() {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 if (!stop) shutdown();\n\u00a0\u00a0\u00a0 }\nprivate:\n\u00a0\u00a0\u00a0 std::vector<std::thread> workers;\n\u00a0\u00a0\u00a0 std::queue<std::function<void()>> tasks;\n\u00a0\u00a0\u00a0 std::mutex queue_mutex;\n\u00a0\u00a0\u00a0 std::condition_variable condition;\n\u00a0\u00a0\u00a0 bool stop = false;\n};\n\n\/\/ Mesure du throughput\nint main() {\n\u00a0\u00a0\u00a0 const int num_tasks = 100000;\n\u00a0\u00a0\u00a0 const int num_threads = std::thread::hardware_concurrency(); \/\/ ex: 4, 8, etc.\n\u00a0\u00a0\u00a0 ThreadPool pool(num_threads);\n\u00a0\u00a0\u00a0 std::atomic<int> counter{0};\n\u00a0\u00a0\u00a0 auto start = std::chrono::high_resolution_clock::now();\n\u00a0\u00a0\u00a0 for (int i = 0; i < num_tasks; ++i) {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 pool.enqueue([&counter]() {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 counter++;\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 });\n\u00a0\u00a0\u00a0 }\n\u00a0\u00a0\u00a0 \/\/ Attendre que tout soit termin\u00e9 (simple mais pas parfait)\n\u00a0\u00a0\u00a0 while (counter.load() < num_tasks) {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 std::this_thread::sleep_for(std::chrono::milliseconds(1));\n\u00a0\u00a0\u00a0 }\n\u00a0\u00a0\u00a0 auto end = std::chrono::high_resolution_clock::now();\n\u00a0\u00a0\u00a0 std::chrono::duration<double> duration = end - start;\n\u00a0\u00a0\u00a0 double throughput = num_tasks \/ duration.count();\n\u00a0\u00a0\u00a0 std::cout << \"Throughput avec thread pool (\" << num_threads\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 << \" threads) : \" << throughput << \" tasks\/sec\" << std::endl;\n\u00a0\u00a0\u00a0 pool.shutdown();\n\u00a0\u00a0\u00a0 return 0;\n}<\/code><\/pre>\n\n\n\n<\/p>\n\n\n\n
Ce que fait ce programme :<\/strong><\/p>\n\n\n\n\nCr\u00e9e un thread pool<\/strong> avec n threads (selon ton CPU).<\/li>\n\n\n\nLance 100 000 t\u00e2ches<\/strong> dans la file du pool.<\/li>\n\n\n\nAttend que toutes soient termin\u00e9es.<\/li>\n\n\n\n Affiche le throughput en t\u00e2ches\/sec<\/strong>.<\/li>\n<\/ul>\n\n\n\nComparaison :<\/strong><\/p>\n\n\n\nM\u00e9thode<\/strong><\/td>Threads cr\u00e9\u00e9s<\/strong><\/td>\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 Performance<\/strong><\/td><\/tr><\/thead>1 thread par t\u00e2che<\/td> 100 000<\/td> \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 Tr\u00e8s lent, surcharge<\/td><\/tr> Thread pool<\/td> 4-16<\/td> \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 Rapide, scalable<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\nExemple : Thread Pool + mesure du throughput<\/strong><\/p>\n\n\n\n#include <iostream>\n#include <thread>\n#include <vector>\n#include <queue>\n#include <functional>\n#include <mutex>\n#include <condition_variable>\n#include <atomic>\n#include <chrono>\nclass ThreadPool {\npublic:\n\u00a0\u00a0\u00a0 ThreadPool(size_t num_threads) {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 for (size_t i = 0; i < num_threads; ++i) {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 workers.emplace_back([this]() {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 while (true) {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 std::function<void()> task;\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 {\u00a0\u00a0 \/\/ zone prot\u00e9g\u00e9e\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 std::unique_lock<std::mutex> lock(queue_mutex);\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 condition.wait(lock, [this]() {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 return stop || !tasks.empty();\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 });\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 if (stop && tasks.empty()) return;\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 task = std::move(tasks.front());\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 tasks.pop();\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 }\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 task();\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 }\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 });\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 }\n\u00a0\u00a0\u00a0 }\n\n\u00a0\u00a0\u00a0 void enqueue(std::function<void()> task) {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 std::lock_guard<std::mutex> lock(queue_mutex);\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 tasks.push(task);\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 }\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 condition.notify_one();\n\u00a0\u00a0\u00a0 }\n\u00a0\u00a0\u00a0 void shutdown() {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 std::lock_guard<std::mutex> lock(queue_mutex);\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 stop = true;\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 }\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 condition.notify_all();\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 for (auto& worker : workers) worker.join();\n\u00a0\u00a0\u00a0 }\n\n\u00a0\u00a0\u00a0 ~ThreadPool() {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 if (!stop)\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 shutdown();\n\u00a0\u00a0\u00a0 }\n\nprivate:\n\u00a0\u00a0\u00a0 std::vector<std::thread> workers;\n\u00a0\u00a0\u00a0 std::queue<std::function<void()>> tasks;\n\u00a0\u00a0\u00a0 std::mutex queue_mutex;\n\u00a0\u00a0\u00a0 std::condition_variable condition;\n\u00a0\u00a0\u00a0 bool stop = false;\n};\n\n\/\/ Mesure du throughput\nint main() {\n\u00a0\u00a0\u00a0 const int num_tasks = 100000;\n\u00a0\u00a0\u00a0 const int num_threads = std::thread::hardware_concurrency(); \/\/ ex: 4, 8, etc.\n\u00a0\u00a0\u00a0 ThreadPool pool(num_threads);\n\u00a0\u00a0\u00a0 std::atomic<int> counter{0};\n\u00a0\u00a0\u00a0 auto start = std::chrono::high_resolution_clock::now();\n\u00a0\u00a0\u00a0 for (int i = 0; i < num_tasks; ++i) {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 pool.enqueue([&counter]() {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 counter++;\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 });\n\u00a0\u00a0\u00a0 }\n\u00a0\u00a0\u00a0 \/\/ Attendre que tout soit termin\u00e9 (simple mais pas parfait)\n\u00a0\u00a0\u00a0 while (counter.load() < num_tasks) {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 std::this_thread::sleep_for(std::chrono::milliseconds(1));\n\u00a0\u00a0\u00a0 }\n\u00a0\u00a0\u00a0 auto end = std::chrono::high_resolution_clock::now();\n\u00a0\u00a0\u00a0 std::chrono::duration<double> duration = end - start;\n\u00a0\u00a0\u00a0 double throughput = num_tasks \/ duration.count();\n \u00a0\u00a0 std::cout << \"Throughput avec thread pool (\" << num_threads\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 << \" threads) : \" << throughput << \" tasks\/sec\" << std::endl;\n\u00a0\u00a0\u00a0 pool.shutdown();\n\u00a0\u00a0\u00a0 return 0;\n}<\/code><\/pre>\n\n\n\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0<\/p>\n\n\n\n
\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0Ce que fait ce programme :<\/strong><\/p>\n\n\n\n\nCr\u00e9e un thread pool<\/strong> avec n threads (selon ton CPU).<\/li>\n\n\n\nLance 100 000 t\u00e2ches<\/strong> dans la file du pool.<\/li>\n\n\n\nAttend que toutes soient termin\u00e9es.<\/li>\n\n\n\n Affiche le throughput en t\u00e2ches\/sec<\/strong>.<\/li>\n<\/ul>\n\n\n\nComparaison :<\/strong><\/p>\n\n\n\nM\u00e9thode<\/strong><\/td>\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 Threads cr\u00e9\u00e9s<\/strong><\/td>\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 Performance<\/strong><\/td><\/tr><\/thead>1 thread par t\u00e2che<\/td> \u00a0\u00a0\u00a0\u00a0 \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0100 000<\/td> \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 Tr\u00e8s lent, surcharge<\/td><\/tr> Thread pool<\/td> \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 4-16<\/td> \u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 Rapide, scalable<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\nExemple complet : throughput avec coroutines asynchrones<\/strong><\/p>\n\n\n\n#include <iostream>\n#include <coroutine>\n#include <chrono>\n#include <thread>\n#include <atomic>\n#include <vector>\n\n\/\/ Awaiter pour simuler un sleep non bloquant\nstruct SleepAwaiter {\n\u00a0\u00a0\u00a0 std::chrono::milliseconds duration;\n\u00a0\u00a0\u00a0 bool await_ready() const noexcept { return false; }\n\u00a0\u00a0\u00a0 void await_suspend(std::coroutine_handle<> handle) {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 std::jthread([handle, d = duration]() {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 std::this_thread::sleep_for(d);\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 handle.resume();\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 });\n\u00a0\u00a0\u00a0 }\n\u00a0\u00a0\u00a0 void await_resume() const noexcept {}\n};\n\n\/\/ T\u00e2che de base retourn\u00e9e par la coroutine\nstruct Task {\n\u00a0\u00a0\u00a0 struct promise_type {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 Task get_return_object() { return {}; }\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 std::suspend_never initial_suspend() noexcept { return {}; }\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 std::suspend_never final_suspend() noexcept { return {}; }\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 void return_void() {}\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 void unhandled_exception() { std::terminate(); }\n\u00a0\u00a0\u00a0 };\n};\n\n\/\/ Coroutine asynchrone qui incr\u00e9mente un compteur apr\u00e8s un d\u00e9lai\nTask delayed_task(std::atomic<int>& counter, int delay_ms) {\n\u00a0\u00a0\u00a0 co_await SleepAwaiter{std::chrono::milliseconds(delay_ms)};\n\u00a0\u00a0\u00a0 counter++;\n}\n\nint main() {\n\u00a0\u00a0\u00a0 const int num_tasks = 10000; \/\/ exemple raisonnable\n\u00a0\u00a0\u00a0 std::atomic<int> counter{0};\n\u00a0\u00a0\u00a0 auto start = std::chrono::high_resolution_clock::now();\n\u00a0\u00a0\u00a0 \/\/ Lancer toutes les coroutines\n\u00a0\u00a0\u00a0 for (int i = 0; i < num_tasks; ++i) {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 delayed_task(counter, 1);\u00a0 \/\/ d\u00e9lai minimal pour simuler non blocage\n\u00a0\u00a0\u00a0 }\n\u00a0\u00a0\u00a0 \/\/ Attendre que toutes les t\u00e2ches soient termin\u00e9es\n\u00a0\u00a0\u00a0 while (counter.load() < num_tasks) {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 std::this_thread::sleep_for(std::chrono::milliseconds(1));\n\u00a0\u00a0\u00a0 }\n\u00a0\u00a0\u00a0 auto end = std::chrono::high_resolution_clock::now();\n\u00a0\u00a0\u00a0 std::chrono::duration<double> duration = end - start;\n\u00a0\u00a0\u00a0 double throughput = num_tasks \/ duration.count();\n\u00a0\u00a0\u00a0 std::cout << \"Throughput avec coroutines asynchrones : \"\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 << throughput << \" tasks\/sec\" << std::endl;\n\u00a0\u00a0\u00a0 return 0;\n}<\/code><\/pre>\n\n\n\nCe que tu vas voir :<\/strong><\/p>\n\n\n\n\nLe programme lance 10 000 coroutines asynchrones.<\/li>\n\n\n\n Chacune attend 1 ms (sans bloquer un thread).<\/li>\n\n\n\n Le compteur s\u2019incr\u00e9mente quand la t\u00e2che est termin\u00e9e.<\/li>\n\n\n\n Le throughput (t\u00e2ches par seconde) est affich\u00e9.<\/li>\n<\/ul>\n\n\n\nNotes importantes :<\/strong><\/p>\n\n\n\n\nCette simulation utilise std::jthread pour g\u00e9rer le timer dans SleepAwaiter.<\/li>\n\n\n\n En vrai, les coroutines s\u2019int\u00e8grent avec des boucles d\u2019\u00e9v\u00e9nements (event loops) plus efficaces.<\/li>\n\n\n\n Ici, on a un tr\u00e8s l\u00e9ger d\u00e9lai pour mieux simuler un travail \u00ab r\u00e9el \u00bb.<\/li>\n<\/ul>\n\n\n\nOpenMP<\/h1>\n\n\n\n OpenMP<\/strong> peut aider \u00e0 optimiser la performance, surtout quand tu veux parall\u00e9liser des t\u00e2ches CPU-bound<\/strong> (qui utilisent beaucoup le processeur) de fa\u00e7on simple et efficace.<\/p>\n\n\n\nQu\u2019est-ce que OpenMP ?<\/strong><\/p>\n\n\n\n\nC\u2019est une API<\/strong> pour faire de la programmation parall\u00e8le sur CPU.<\/li>\n\n\n\nTr\u00e8s utilis\u00e9e en C\/C++ (et Fortran).<\/li>\n\n\n\n Elle te permet de parall\u00e9liser des boucles ou des sections de code avec des directives<\/strong> (#pragma omp), sans g\u00e9rer explicitement les threads.<\/li>\n\n\n\nOpenMP g\u00e8re la cr\u00e9ation, la synchronisation et la r\u00e9partition des threads automatiquement.<\/li>\n<\/ul>\n\n\n\nExemple simple en OpenMP pour parall\u00e9liser une boucle de t\u00e2ches :<\/strong><\/p>\n\n\n\n#include <iostream>\n#include <atomic>\n#include <omp.h><\/strong>\n\nint main() {\n\u00a0\u00a0\u00a0 const int num_tasks = 100000;\n\u00a0\u00a0\u00a0 std::atomic<int> counter{0};\n\u00a0\u00a0\u00a0 #pragma omp parallel for<\/strong>\n\u00a0\u00a0\u00a0 for (int i = 0; i < num_tasks; ++i) {\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 \/\/ T\u00e2che CPU-bound simul\u00e9e\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 counter++;\n\u00a0\u00a0\u00a0 }\n\u00a0\u00a0\u00a0 std::cout << \"Counter = \" << counter << std::endl;\n\u00a0\u00a0\u00a0 return 0;\n}<\/code><\/pre>\n\n\n\nComment OpenMP peut aider dans ton contexte ?<\/strong><\/p>\n\n\n\n\nSi tes t\u00e2ches sont
![\"\"](\"https:\/\/i0.wp.com\/imalogic.com\/blog\/wp-content\/uploads\/2025\/09\/Concurrent-Programming-1600x944.jp_.webp?ssl=1\")
<\/a><\/figure>\n<\/div>\n\n\n