Python的并行（持续更新）

0. 参考：

1. 《Python并行编程 中文版》https://python-parallel-programmning-cookbook.readthedocs.io/zh_CN/latest/index.html

1. 线程和进程：

1. 进程可以包含多个并行运行的线程；
2. 通常，操作系统创建和管理线程比进程更省CPU资源；
3. 线程用于一些小任务，进程用于繁重的任务；
4. 同一个进程下的线程共享地址空间和其他资源，进程之间相互独立；

2. 在Python中使用线程:

2.1 多线程简介：

1. 线程是独立的处理流程，可以和系统的其他线程并行或并发地执行。
2. 多线程可以共享数据和资源，利用所谓的共享内存空间。
3. 每一个线程基本上包含3个元素：程序计数器，寄存器和栈。
4. 线程的状态大体上可以分为ready, running, blocked。
5. 多线程编程一般使用共享内容空间进行线程间的通讯，这就使管理内容空间成为多线程编程的重点和难点。
6. 线程的典型应用是应用软件的并行化。
7. 相比于进程，使用线程的优势主要是性能。

2.2 threading 库实现多线程:

1. 如何定义一个线程 threading.Thread():

class threading.Thread(group=None,   ## 一般设置为 None ，这是为以后的一些特性预留的target=None,  ## 当线程启动的时候要执行的函数name=None,    ## 线程的名字，默认会分配一个唯一名字 Thread-Nargs=(),      ## 使用 tuple 类型给 target 传递参数kwargs={})    ## 使用 dict 类型给 target 传递参数

举例：

import threadingdef function(i):print("function called by thread %i\n" % i)return#threads = []
for i in range(5):t = threading.Thread(target=function, args=(i,)) ## 用 function 函数初始化一个 Thread 对象 t，并将参数 i 传入；#threads.append(t) t.start() ## 线程被创建后不会马上执行，需要手动调用 .start() 方法执行线程t.join() ## 阻塞调用 t 线程的主线程，t 线程执行结束，主线程才会继续执行

运行结果：

function called by thread 0function called by thread 1function called by thread 2function called by thread 3function called by thread 4

2. 如何确定当前线程 threading.current_thread().name：

1. 通常一个服务进程中有多个线程服务，负责不同的操作，所以对于线程的命名是很重要的；
2. Python中每一个线程在被 Thread 被创建时都有一个默认的名字（可以修改）；

举例:

import threading
import timedef first_func():print(threading.current_thread().name + str(" is Starting"))time.sleep(2)print(threading.current_thread().name + str("is Exiting"))returndef second_func():print(threading.current_thread().name + str(" is Starting"))time.sleep(2)print(threading.current_thread().name + str("is Exiting"))returndef third_func():print(threading.current_thread().name + str(" is Starting"))time.sleep(2)print(threading.current_thread().name + str("is Exiting"))returnif __name__ == "__main__":t1 = threading.Thread(name="first_func", target=first_func)t2 = threading.Thread(name="second_func", target=second_func)t3 = threading.Thread(target=third_func)t1.start()t2.start()t3.start()t1.join()t2.join()t3.join()

运行结果：

first_func is Starting
second_func is Starting
Thread-36 (third_func) is Starting
first_funcis Exiting
second_funcis Exiting
Thread-36 (third_func)is Exiting

从上面运行结果可以看出，如果不用 name= 参数指定线程名称的话，那么线程名称将使用默认值。

3. 如何实现一个线程 threading:

使用 threading 模块实现一个线程，需要3步：

1. 定义一个 Thread 类的子类；
2. 重写 __init__(self, [,args]) 方法；
3. 重写 run(self, [,args]) 方法实现一个线程；

举例：

import threading
#import _thread
import time#exitFlag = 0class myThread(threading.Thread): ## 定义一个 threading 子类，继承 threading.Thread 父类def __init__(self, threadID, name, counter):  ## 重写 __init__() 方法，并添加额外的参数threading.Thread.__init__(self) ## 初始化继承自Thread类的属性，使子类对象能够正确地继承和使用父类的属性和方法self.threadID = threadID ## 子类额外的属性self.name = nameself.counter = counterdef run(self):print("Starting " + self.name)print_time(self.name, self.counter, 5)print("Exiting " + self.name)def print_time(threadName, delay, counter):while counter:##if exitFlag:  ## 当 exitFlag != 0时，执行 _thread.exit()，线程退出 (但是在该段代码中，exitFlag的值没有被改变，所以不会执行 _thread.exit()，所以可以直接注释掉)##    _thread.exit()time.sleep(delay)print("%s: %s" % (threadName, time.ctime(time.time())))counter -= 1## 创建线程
thread1 = myThread(1, "Thread-1", 1)
thread2 = myThread(2, "Thread-2", 2)
## 开启线程
thread1.start()
thread2.start()
## .join()
thread1.join()
thread2.join()
print("Exiting Main Thread")

运行结果：

Starting Thread-1
Starting Thread-2Thread-1: Wed Jun 21 11:12:09 2023
Thread-2: Wed Jun 21 11:12:10 2023
Thread-1: Wed Jun 21 11:12:10 2023
Thread-1: Wed Jun 21 11:12:11 2023
Thread-2: Wed Jun 21 11:12:12 2023
Thread-1: Wed Jun 21 11:12:12 2023
Thread-1: Wed Jun 21 11:12:13 2023
Exiting Thread-1
Thread-2: Wed Jun 21 11:12:14 2023
Thread-2: Wed Jun 21 11:12:16 2023
Thread-2: Wed Jun 21 11:12:18 2023
Exiting Thread-2
Exiting Main Thread

由于 thread1 的 sleep 时间比 thread2 的时间短，所以 thread2 会执行更久一些，退出也就更晚一些。

4. 使用 Lock 进行线程同步 (lock()):

1. 并发线程中，多个线程对共享内存进行操作，并且至少有一个可以改变数据。这种情况下如果没有同步机制，那么多个线程之间就会产生竞争，从而导致代码无效或出错。
2. 解决多线程竞争问题的最简单的方法就是用锁 (Lock)。当一个线程需要访问共享内存时，它必须先获得 Lock 之后才能访问；当该线程对共享资源使用完成后，必须释放 Lock，然后其他线程在拿到 Lock 进行访问资源。因此，为了避免多线程竞争的出现，必须保证：同一时刻只能允许一个线程访问共享内存。
3. 在实际使用中，该方法经常会导致一种 死锁 现象，原因是不同线程互相拿着对方需要的 Lock，导致死锁的发生。
   详见：https://python-parallel-programmning-cookbook.readthedocs.io/zh_CN/latest/chapter2/06_Thread_synchronization_with_Lock_and_Rlock.html

举例:

import threadingshared_resource_with_lock = 0
shared_resource_with_no_lock = 0
COUNT = 100000
shared_resource_lock = threading.Lock() ## 锁## 有锁的情况
def increment_with_lock():global shared_resource_with_lock ## shared_resource_with_lock 即最外面的 shared_resource_with_lock，这样写就不需要再通过函数的参数引入 shared_resource_with_lock 了for _ in range(COUNT):shared_resource_lock.acquire() ## 获取 锁shared_resource_with_lock += 1shared_resource_lock.release() ## 释放 锁def decrement_with_lock():global shared_resource_with_lockfor _ in range(COUNT):shared_resource_lock.acquire()shared_resource_with_lock -= 1shared_resource_lock.release()## 没有锁的情况
def increment_without_lock():global shared_resource_with_no_lockfor _ in range(COUNT):shared_resource_with_no_lock += 1def decrement_without_lock():global shared_resource_with_no_lockfor _ in range(COUNT):shared_resource_with_no_lock -= 1if __name__ == "__main__":t1 = threading.Thread(target=increment_with_lock)t2 = threading.Thread(target=decrement_with_lock)t3 = threading.Thread(target=increment_without_lock)t4 = threading.Thread(target=decrement_without_lock)## 开启线程t1.start()t2.start()t3.start()t4.start()## .join()t1.join()t2.join()t3.join()t4.join()print ("the value of shared variable with lock management is %s" % shared_resource_with_lock)print ("the value of shared variable with race condition is %s" % shared_resource_with_no_lock)

运行结果：

the value of shared variable with lock management is 0
the value of shared variable with race condition is 0

尽管在上面的结果中，没锁的情况下得到的结果也是正确的，但是执行多次，总会出现错误的结果；而有锁的情况下，执行多次，结果一定是正确的。

尽管理论上用锁的策略可以避免多线程中的竞争问题，但是可能会对程序的其他方面产生负面影响。此外，锁的策略经常会导致不必要的开销，也会限制程序的可扩展性和可读性。更重要的是，有时候需要对多进程共享的内存分配优先级，使用锁可能和这种优先级冲突。从实践的经验来看，使用锁的应用将对debug带来不小的麻烦。所以，最好使用其他可选的方法确保同步读取共享内存，避免竞争条件。

5. 使用RLock进行线程同步:

1. 为了保证 “只有拿到锁的线程才能释放锁”，那么应该使用 RLock() 对象；
2. 和 Lock()一样，RLock()也有acquire()和release()两种方法；
3. RLock() 有三个特点：
     1). 谁拿到谁释放。如果线程A拿到锁，线程B无法释放这个锁，只有A可以释放；
     2). 同一线程可以多次拿到该锁，即可以acquire多次；
     3). acquire多少次就必须release多少次，只有最后一次release才能改变RLock的状态为unlocked）;

举例：

import threading
import timeclass Box(object):lock = threading.RLock()def __init__(self):self.total_items = 0def execute(self, n):Box.lock.acquire()self.total_items += nBox.lock.release()def add(self):Box.lock.acquire()self.execute(1)Box.lock.release()def remove(self):Box.lock.acquire()self.execute(-1)Box.lock.release()def adder(box, items):while items > 0:print("adding 1 item in the box")box.add()time.sleep(1)items -= 1def remover(box, items):while items > 0:print("removing 1 item in the box")box.remove()time.sleep(1)items -= 1if __name__ == "__main__":items = 5print("putting %s items in the box"% items)box = Box()t1 = threading.Thread(target=adder, args=(box, items))t2 = threading.Thread(target=remover, args=(box, items))t1.start()t2.start()t1.join()t2.join()print("%s items still remain in the box " % box.total_items)

运行结果：

putting 5 items in the box
adding 1 item in the box
removing 1 item in the box
adding 1 item in the box
removing 1 item in the boxremoving 1 item in the box
adding 1 item in the boxremoving 1 item in the box
adding 1 item in the box
adding 1 item in the box
removing 1 item in the box0 items still remain in the box 

Box类的execute()方法包含RLock，adder()和remover()方法也包含RLock，就是说无论是调用Box还是adder()或者remover()，每个线程的每一步都有拿到资源、释放资源的过程。

6. 使用信号量进行线程同步：

1. 信号量是由操作系统管理的一种抽象数据类型，用于多线程中同步对共享资源的使用；
2. 信号量是一个内部数据，用于表明当前共享资源可以有多少并发读取；
3. 在 Threading 中，信号量的操作有两个函数：acquire()和release()；

举例：

import threading
import time
import randomsemaphore = threading.Semaphore(0) ## 可以理解为一个内置的计数器，当调用 acquire 方法时候内置计数器 -1，对应着申请资源；调用 release 方法时候内置计数器+1，对应着释放可用资源。
print("init semaphore %s" % semaphore._value)def consumer():print("consumer is waiting.")semaphore.acquire()print("consumer notify: consumed item number %s" % item)print("consumer semaphore %s" % semaphore._value)def producer():global itemtime.sleep(10)item = random.randint(0, 1000)print("producer notify : produced item number %s" % item)semaphore.release()print("producer semaphore %s" % semaphore._value)if __name__ == "__main__":for _ in range(0, 5):t1 = threading.Thread(target=producer)t2 = threading.Thread(target=consumer)t1.start()t2.start()t1.join()t2.join()print("program terminated")

运行结果：

init semaphore 0
consumer is waiting.
producer notify : produced item number 756
producer semaphore 1
consumer notify: consumed item number 756
consumer semaphore 0
consumer is waiting.
producer notify : produced item number 948
producer semaphore 1
consumer notify: consumed item number 948
consumer semaphore 0
consumer is waiting.
producer notify : produced item number 597
producer semaphore 1
consumer notify: consumed item number 597
consumer semaphore 0
consumer is waiting.
producer notify : produced item number 239
producer semaphore 1
consumer notify: consumed item number 239
consumer semaphore 0
consumer is waiting.
producer notify : produced item number 141
producer semaphore 1
consumer notify: consumed item number 141
consumer semaphore 0
program terminated

根据semaphore = threading.Semaphore(0)将信号量初始化为0，其目的在于同步两个或多个线程。
producer()执行完后，通过seaphore.release()释放资源，之后consumer()通过semaphore.acquire()拿到资源；
相应的信号量的计数器也会从初始化的0 --> 1 --> 多次重复。

信号量的一个特殊用法是互斥量，互斥量是初始值为1的信号量，可以实现数据、资源的互斥访问；

7. 使用条件进行线程同步：

1. 条件：指的是程序状态的改变；
2. 某些线程在等待某一条件发生，其他的线程会在该条件发生的时候进行通知。一旦条件发生，线程会拿到共享资源的唯一权限。

举例：

from threading import Thread, Condition
import timeitems = []
condition = Condition() ## A condition variable allows one or more threads to wait until they are notified by another thread.class consumer(Thread):def __init__(self):Thread.__init__(self)def consume(self):global conditionglobal itemscondition.acquire()if len(items) == 0:condition.wait() ## Wait until notified or until a timeout occurs.print("Consumer notify : no item to consume")items.pop()print("Consumer notify : consumed 1 item")print("Consumer notify : items to consume are " + str(len(items)))condition.notify() ## Wake up one or more threads waiting on this condition, if any.condition.release()def run(self):for _ in range(0, 20):time.sleep(2)self.consume()class producer(Thread):def __init__(self):Thread.__init__(self)def produce(self):global conditionglobal itemscondition.acquire()if len(items) == 10:condition.wait()print("Producer notify : items producted are " + str(len(items)))print("Producer notify : stop the production!!")items.append(1)print("Producer notify : total items producted " + str(len(items)))condition.notify()condition.release()def run(self):for _ in range(0, 20):time.sleep(1)self.produce()if __name__ == "__main__":producer = producer()consumer = consumer()producer.start()consumer.start()producer.join()consumer.join()

运行结果：

Producer notify : total items producted 1
Consumer notify : consumed 1 item
Consumer notify : items to consume are 0
Producer notify : total items producted 1
Producer notify : total items producted 2
Consumer notify : consumed 1 item
Consumer notify : items to consume are 1
Producer notify : total items producted 2
Producer notify : total items producted 3
Consumer notify : consumed 1 item
Consumer notify : items to consume are 2
Producer notify : total items producted 3
Producer notify : total items producted 4
Consumer notify : consumed 1 item
Consumer notify : items to consume are 3
Producer notify : total items producted 4
Producer notify : total items producted 5
Consumer notify : consumed 1 item
Consumer notify : items to consume are 4
Producer notify : total items producted 5
Producer notify : total items producted 6
Consumer notify : consumed 1 item
Consumer notify : items to consume are 5
Producer notify : total items producted 6
Producer notify : total items producted 7
Consumer notify : consumed 1 item
Consumer notify : items to consume are 6
Producer notify : total items producted 7
Producer notify : total items producted 8
Consumer notify : consumed 1 item
Consumer notify : items to consume are 7
Producer notify : total items producted 8
Producer notify : total items producted 9
Consumer notify : consumed 1 item
Consumer notify : items to consume are 8
Producer notify : total items producted 9
Producer notify : total items producted 10
Consumer notify : consumed 1 item
Consumer notify : items to consume are 9
Producer notify : total items producted 10
Consumer notify : consumed 1 item
Consumer notify : items to consume are 9
Consumer notify : consumed 1 item
Consumer notify : items to consume are 8
Consumer notify : consumed 1 item
Consumer notify : items to consume are 7
Consumer notify : consumed 1 item
Consumer notify : items to consume are 6
Consumer notify : consumed 1 item
Consumer notify : items to consume are 5
Consumer notify : consumed 1 item
Consumer notify : items to consume are 4
Consumer notify : consumed 1 item
Consumer notify : items to consume are 3
Consumer notify : consumed 1 item
Consumer notify : items to consume are 2
Consumer notify : consumed 1 item
Consumer notify : items to consume are 1
Consumer notify : consumed 1 item
Consumer notify : items to consume are 0

整个过程有点绕，可以通过这个例子简单理解 使用 condition 进行线程同步 (https://blog.csdn.net/lzanze/article/details/105351064)

8. 使用事件进行线程同步：

事件：线程之间用于通讯的对象。

举例：

import time
from threading import Thread, Event
import randomitems = []
event = Event()class consumer(Thread):def __init__(self, items, event):Thread.__init__(self)self.items = itemsself.event = eventdef run(self):while True:time.sleep(2)self.event.wait() ## 线程会阻塞在这里，直到事件 event 被触发(set.evet.set())才会继续执行。item = self.items.pop()print("Consumer notify : %d popped from list by %s" % (item, self.name))class producer(Thread):def __init__(self, items, event):Thread.__init__(self)self.items = itemsself.event = eventdef run(self):global itemfor _ in range(10):time.sleep(2)item = random.randint(0, 256)self.items.append(item) ## 将 item 添加到list末尾然后通过 self.event.set() 和 self.event.clear() 发出事件通知print('Producer notify : item %d appended to list by %s' % (item, self.name))print('Producer notify : event set by %s' % self.name)self.event.set() ## set() 方法将内部变量设置为 True (is_set() == True)print("Produce notify : event cleared by %s" % self.name)#print("Produce set event label : ", self.event.is_set())self.event.clear() ## clear() 方法将内部变量设置为 False (is_set() == False)#print("Produce clear event label : ", self.event.is_set())if __name__ == "__main__":t1 = producer(items, event)t2 = consumer(items, event)t1.start()t2.start()t1.join()t2.join()

运行结果（部分结果）：

Producer notify : item 140 appended to list by Thread-64
Producer notify : event set by Thread-64
Produce notify : event cleared by Thread-64
Consumer notify : 140 popped from list by Thread-65
Producer notify : item 42 appended to list by Thread-64
Producer notify : event set by Thread-64
Produce notify : event cleared by Thread-64
Producer notify : item 101 appended to list by Thread-64
Producer notify : event set by Thread-64
Produce notify : event cleared by Thread-64
Consumer notify : 101 popped from list by Thread-65
Producer notify : item 213 appended to list by Thread-64
Producer notify : event set by Thread-64
Produce notify : event cleared by Thread-64
Producer notify : item 31 appended to list by Thread-64
Producer notify : event set by Thread-64
Produce notify : event cleared by Thread-64
Consumer notify : 31 popped from list by Thread-65
Producer notify : item 235 appended to list by Thread-64
Producer notify : event set by Thread-64
Produce notify : event cleared by Thread-64

该脚本我自己在运行的时候，运行了20多分钟，没结束，所以直接停了（部分结果如上所示）。

9. 使用 with 语法：

1. 当两个相关的操作需要在一部分代码块前后分别执行的时候，可以使用 with 语法；
2. 使用 with 语法可以在特定的地方分配和释放资源，因此， with 语法也叫做“上下文管理器”；
3. 在threading模块中，所有带有 acquire() 方法和 release() 方法的对象（包括Lock, RLock, Condition, Semaphore）都可以使用 with 语法；

举例：

import threading
import logging
logging.basicConfig(level=logging.DEBUG, format="(%(threadName)-10s) %(message)s", )def threading_with(statement):with statement:logging.debug("%s acquired via with" % statement)def threading_not_with(statement):statement.acquire()try:logging.debug("%s acquired directly" % statement)finally:statement.release()if __name__ == "__main__":lock = threading.Lock()rlock = threading.RLock()condition = threading.Condition()mutex = threading.Semaphore(1)threading_synchronization_list = [lock, rlock, condition, mutex] ## 包含要测试的线程同步使用的对象for statement in threading_synchronization_list:t1 = threading.Thread(target=threading_with, args=(statement,))t2 = threading.Thread(target=threading_not_with, args=(statement,))t1.start()t2.start()t1.join()t2.join()

运行结果：

(Thread-68 (threading_with)) <locked _thread.lock object at 0x10ba04f00> acquired via with
(Thread-69 (threading_not_with)) <locked _thread.lock object at 0x10ba04f00> acquired directly
(Thread-70 (threading_with)) <locked _thread.RLock object owner=6223966208 count=1 at 0x10bca8740> acquired via with
(Thread-71 (threading_not_with)) <locked _thread.RLock object owner=6308098048 count=1 at 0x10bca8740> acquired directly
(Thread-72 (threading_with)) <Condition(<locked _thread.RLock object owner=6223966208 count=1 at 0x10bcf9080>, 0)> acquired via with
(Thread-73 (threading_not_with)) <Condition(<locked _thread.RLock object owner=6308098048 count=1 at 0x10bcf9080>, 0)> acquired directly
(Thread-74 (threading_with)) <threading.Semaphore object at 0x10760bf70> acquired via with
(Thread-75 (threading_not_with)) <threading.Semaphore object at 0x10760bf70> acquired directly

上述结果展示了使用with和不用with的每一个函数以及用在了哪些地方(Lock, RLock, Condition, Semaphore)

10. 使用 queue 进行线程通信：

与 threading 模块相比，queue 操作更简单、更安全。

---

更新中……