1、为什么要使用线程池?
某类任务非常耗时,严重影响该线程执行其他任务;任务主要分两类,一类是磁盘io,一类是网络io,在处理这些io任务的时候非常耗时,会长期占用cpu可能会影响其他任务的处理;
2、使用线程池原理
在其他线程中异步处理耗时任务,当有耗时任务需要处理的时候,创建一个线程去异步执行这些耗时任务,但是开启线程数量和cpu核数有关,所以开启处理耗时任务的线程数量是有限的;需要在cpu核心之间做出平衡选择;从而通过创建和销毁线程来固定线程数量;但是创建和销毁线程非常消耗线程资源,所以将创建的线程放入线程池中来复用线程资源;充分利用系统资源来异步执行耗时任务;
3、线程池具体实现
线程池是生产者消费者模式:生产者线程发布耗时任务,队列存放任务(有任务的上下文和任务的执行函数),任务的上下文本来实在生产者中执行的,现在要放在消费者线程中执行,线程池是一定线程数量的集合,消费者线程取出并执行任务,通过线程调度唤起休眠线程,唤起线程需要加锁mutex和条件变量condition;线程状态有从无到有及从有到无两种状态,通过条件变量唤醒线程或休眠线程,条件即使队列的状态,如果队列中有任务则唤醒线程,如果没有任务,则让线程睡眠,生产者发布任务后通知休眠线程唤醒来开展取出并执行任务的工作;
线程需要平衡选择,平衡选择根据具体的耗时任务进行,耗时任务分为两种,io密集和cpu密集;io密集型会阻塞线程,cpu密集型则cpu长期被一个线程占用,不能处理其他任务;线程池中线程的个数如果是cpu密集型则和cpu核数相等,如果是io密集型则线程个数通常是2*cpu核数+2;
如何确定线程池线程数量:
(io等待时间+cpu运算时间)*核数/cpu运算时间
4、实现线程池
接口设计:1)产生线程池接口,参数包括线程池数量,队列长度(因为线程的栈空间是固定的)
2)销毁线程池接口,参数包括标记线程池推出,通知所有线程
3)生产者抛出任务的接口,构造任务,放入队列,通知线程唤醒
4)线程池使用对象主要是生产者线程,是生产者线程使用线程池;
代码实现:
#ifndef _THREAD_POOL_H #define _THREAD_POOL_H typedef struct thread_pool_t thread_pool_t; typedef void (*handler_pt) (void *); thread_pool_t *thread_pool_create(int thrd_count, int queue_size); int thread_pool_post(thread_pool_t *pool, handler_pt func, void *arg); int thread_pool_destroy(thread_pool_t *pool); int wait_all_done(thread_pool_t *pool); #endif #include <pthread.h> #include <stdint.h> #include <stddef.h> #include <stdlib.h> #include "thrd_pool.h" typedef struct task_t { handler_pt func; void * arg; } task_t; typedef struct task_queue_t { uint32_t head; uint32_t tail; uint32_t count; task_t *queue; } task_queue_t; struct thread_pool_t { pthread_mutex_t mutex; pthread_cond_t condition; pthread_t *threads; task_queue_t task_queue; int closed; int started; // 当前运行的线程数 int thrd_count; int queue_size; }; static void * thread_worker(void *thrd_pool); static void thread_pool_free(thread_pool_t *pool); thread_pool_t * thread_pool_create(int thrd_count, int queue_size) { thread_pool_t *pool; if (thrd_count <= 0 || queue_size <= 0) { return NULL; } pool = (thread_pool_t*) malloc(sizeof(*pool)); if (pool == NULL) { return NULL; } pool->thrd_count = 0; pool->queue_size = queue_size; pool->task_queue.head = 0; pool->task_queue.tail = 0; pool->task_queue.count = 0; pool->started = pool->closed = 0; pool->task_queue.queue = (task_t*)malloc(sizeof(task_t)*queue_size); if (pool->task_queue.queue == NULL) { // TODO: free pool return NULL; } pool->threads = (pthread_t*) malloc(sizeof(pthread_t) * thrd_count); if (pool->threads == NULL) { // TODO: free pool return NULL; } int i = 0; for (; i < thrd_count; i++) { if (pthread_create(&(pool->threads[i]), NULL, thread_worker, (void*)pool) != 0) { // TODO: free pool return NULL; } pool->thrd_count++; pool->started++; } return pool; } int thread_pool_post(thread_pool_t *pool, handler_pt func, void *arg) { if (pool == NULL || func == NULL) { return -1; } task_queue_t *task_queue = &(pool->task_queue); if (pthread_mutex_lock(&(pool->mutex)) != 0) { return -2; } if (pool->closed) { pthread_mutex_unlock(&(pool->mutex)); return -3; } if (task_queue->count == pool->queue_size) { pthread_mutex_unlock(&(pool->mutex)); return -4; } task_queue->queue[task_queue->tail].func = func; task_queue->queue[task_queue->tail].arg = arg; task_queue->tail = (task_queue->tail + 1) % pool->queue_size; task_queue->count++; if (pthread_cond_signal(&(pool->condition)) != 0) { pthread_mutex_unlock(&(pool->mutex)); return -5; } pthread_mutex_unlock(&(pool->mutex)); return 0; } static void thread_pool_free(thread_pool_t *pool) { if (pool == NULL || pool->started > 0) { return; } if (pool->threads) { free(pool->threads); pool->threads = NULL; pthread_mutex_lock(&(pool->mutex)); pthread_mutex_destroy(&pool->mutex); pthread_cond_destroy(&pool->condition); } if (pool->task_queue.queue) { free(pool->task_queue.queue); pool->task_queue.queue = NULL; } free(pool); } int wait_all_done(thread_pool_t *pool) { int i, ret=0; for (i=0; i < pool->thrd_count; i++) { if (pthread_join(pool->threads[i], NULL) != 0) { ret=1; } } return ret; } int thread_pool_destroy(thread_pool_t *pool) { if (pool == NULL) { return -1; } if (pthread_mutex_lock(&(pool->mutex)) != 0) { return -2; } if (pool->closed) { thread_pool_free(pool); return -3; } pool->closed = 1; if (pthread_cond_broadcast(&(pool->condition)) != 0 || pthread_mutex_unlock(&(pool->mutex)) != 0) { thread_pool_free(pool); return -4; } wait_all_done(pool); thread_pool_free(pool); return 0; } static void * thread_worker(void *thrd_pool) { thread_pool_t *pool = (thread_pool_t*)thrd_pool; task_queue_t *que; task_t task; for (;;) { pthread_mutex_lock(&(pool->mutex)); que = &pool->task_queue; // 虚假唤醒 linux pthread_cond_signal // linux 可能被信号唤醒 // 业务逻辑不严谨,被其他线程抢了该任务 while (que->count == 0 && pool->closed == 0) { // pthread_mutex_unlock(&(pool->mutex)) // 阻塞在 condition // =================================== // 解除阻塞 // pthread_mutex_lock(&(pool->mutex)); pthread_cond_wait(&(pool->condition), &(pool->mutex)); } if (pool->closed == 1) break; task = que->queue[que->head]; que->head = (que->head + 1) % pool->queue_size; que->count--; pthread_mutex_unlock(&(pool->mutex)); (*(task.func))(task.arg); } pool->started--; pthread_mutex_unlock(&(pool->mutex)); pthread_exit(NULL); return NULL; }
5、reactor模型中使用线程池
可以用于处理decode,read,compute,encode,send任务,以为这些任务非常耗时;
6、nginx中的线程池
read读取出数据后,通过decode解码,由于compute的任务非常耗时,线程池可以用于执行compute任务;