Introduction to the method of completing parallel processing in Python multi-process

This article mainly introduces a summary of the methods of using multiple processes to achieve parallel processing in Python. It has certain reference value. Interested friends can refer to it

Processes and threads are in the field of computer software A very important concept in , process and thread are different and closely related. Let’s first analyze these two concepts:

1. Definition

A process is a program with certain independent functions. Regarding a running activity on a certain data set, a process is an independent unit for resource allocation and scheduling by the system.
A thread is an entity of a process and is the basic basis for CPU scheduling and dispatching. A unit is a basic unit that is smaller than a process and can run independently. The thread itself basically does not own system resources, only a few resources that are essential during operation (such as a program counter, a set of registers and a stack), but it All resources owned by the process can be shared with other threads belonging to the same process.

2. Relationship

One thread can create and destroy another thread ;Multiple threads in the same process can be executed concurrently.

Compared to a process, a thread is a concept closer to an execution body, and it can be shared with other threads in the same process data, but has its own stack space and independent execution sequence.

3. Difference

The main difference between processes and threads is that they are different operating system resource management methods. The process has an independent address space. After a process crashes, it will not affect other processes in protected mode, and threads are just different execution paths in a process. Threads have their own stacks and local variables, but there is no separate address space between threads. The death of one thread is equivalent to the death of the entire process. Therefore, multi-process programs are more robust than multi-thread programs, but when switching processes, it consumes a lot of time. The resources are larger and the efficiency is lower. But for some concurrent operations that require simultaneous execution and sharing of certain variables, only threads, not processes, can be used.

1) In short, a program has at least one process, and a process has at least one thread.

2) The division scale of threads is smaller than that of processes , making the concurrency of multi-threaded programs high.

3) In addition, the process has an independent memory unit during execution, and multiple threads share memory, thus greatly improving the running efficiency of the program.

4) There are differences between threads and processes during execution. Each independent thread has an entry point for program execution, a sequential execution sequence, and an exit point for the program. However, threads cannot execute independently and must exist in the application program, and the application program provides multiple thread execution control.

5) From a logical point of view, the meaning of multi-threading is that in an application, multiple execution parts can be executed at the same time. However, the operating system does not regard multiple threads as multiple independent applications to implement process scheduling and management and resource allocation. This is the important difference between processes and threads.

4. Advantages and Disadvantages

Threads and processes have their own advantages and disadvantages in use: thread execution overhead is small, but it is not conducive to resource management. and protection; and process is the opposite. At the same time, threads are suitable for running on SMP machines, while processes can be migrated across machines.

This article mainly talks about the application of multi-process in Python

The Unix/Linux operating system provides a fork() system call, which is very special. Ordinary function calls call once and return once, but fork() calls once and returns twice, because the operating system automatically makes a copy of the current process (called the parent process) (called the child process), and then, respectively, in Returned within the parent process and child process.

The child process always returns 0, and the parent process returns the ID of the child process. The reason for this is that a parent process can fork out many child processes, so the parent process must record the ID of each child process, and the child process only needs to call getpid() to get the ID of the parent process.

Python's os module encapsulates common system calls, including fork, which can easily create sub-processes in Python programs:

import os

print('Process (%s) start...' % os.getpid())
# Only works on Unix/Linux/Mac:
pid = os.fork()
if pid == 0:
  print('I am child process (%s) and my parent is %s.' % (os.getpid(), os.getppid()))
else:
  print('I (%s) just created a child process (%s).' % (os.getpid(), pid))

Run results As follows:

Process (876) start...
I (876) just created a child process (877).
I am child process (877) and my parent is 876.

Since Windows does not have a fork call, the above code cannot run on Windows.

With the fork call, when a process receives a new task, it can copy a child process to handle the new task. A common Apache server has the parent process listening on the port. Whenever When there is a new http request, the child process is forked to handle the new http request.

multiprocessing

If you plan to write a multi-process service program, Unix/linux is undoubtedly the right choice. Since Windows does not have a fork call, is it impossible to write multi-process programs in Python on Windows?

Since Python is cross-platform, it should naturally provide cross-platform multi-process support. The multiprocessing module is a cross-platform version of the multi-process module.

The multiprocessing module provides a Process class to represent a process object. The following example demonstrates starting a child process and waiting for it to end:

from multiprocessing import Process
import os

# 子进程要执行的代码
def run_proc(name):
  print('Run child process %s (%s)...' % (name, os.getpid()))

if __name__=='__main__':
  print('Parent process %s.' % os.getpid())
  p = Process(target=run_proc, args=('test',))
  print('Child process will start.')
  p.start()
  p.join()
  print('Child process end.')

创建子进程时，只需要传入一个执行函数和函数的参数，创建一个Process实例，用start()方法启动，这样创建进程比fork()还要简单。

join()方法可以等待子进程结束后再继续往下运行，通常用于进程间的同步。

Pool

如果要启动大量的子进程，可以用进程池的方式批量创建子进程：

from multiprocessing import Pool
import os, time, random

def long_time_task(name):
  print('Run task %s (%s)...' % (name, os.getpid()))
  start = time.time()
  time.sleep(random.random() * 3)
  end = time.time()
  print('Task %s runs %0.2f seconds.' % (name, (end - start)))

if __name__=='__main__':
  print('Parent process %s.' % os.getpid())
  p = Pool(4)
  for i in range(5):
    p.apply_async(long_time_task, args=(i,))
  print('Waiting for all subprocesses done...')
  p.close()
  p.join()
  print('All subprocesses done.')

执行结果如下：

Parent process 669.
Waiting for all subprocesses done...
Run task 0 (671)...
Run task 1 (672)...
Run task 2 (673)...
Run task 3 (674)...
Task 2 runs 0.14 seconds.
Run task 4 (673)...
Task 1 runs 0.27 seconds.
Task 3 runs 0.86 seconds.
Task 0 runs 1.41 seconds.
Task 4 runs 1.91 seconds.
All subprocesses done.

代码解读：

对Pool对象调用join()方法会等待所有子进程执行完毕，调用join()之前必须先调用close()，调用close()之后就不能继续添加新的Process了。

请注意输出的结果，task 0，1，2，3是立刻执行的，而task 4要等待前面某个task完成后才执行，这是因为Pool的默认大小在我的电脑上是4，因此，最多同时执行4个进程。这是Pool有意设计的限制，并不是操作系统的限制。如果改成：

p = Pool(5)

就可以同时跑5个进程。

由于Pool的默认大小是CPU的核数，如果你不幸拥有8核CPU，你要提交至少9个子进程才能看到上面的等待效果。

子进程

很多时候，子进程并不是自身，而是一个外部进程。我们创建了子进程后，还需要控制子进程的输入和输出。

subprocess模块可以让我们非常方便地启动一个子进程，然后控制其输入和输出。

下面的例子演示了如何在Python代码中运行命令nslookup www.python.org，这和命令行直接运行的效果是一样的：

import subprocess

print('$ nslookup www.python.org')
r = subprocess.call(['nslookup', 'www.python.org'])
print('Exit code:', r)

运行结果：

$ nslookup www.python.org
Server:        192.168.19.4
Address:    192.168.19.4#53
Non-authoritative answer:
www.python.org    canonical name = python.map.fastly.net.
Name:    python.map.fastly.net
Address: 199.27.79.223
Exit code: 0

如果子进程还需要输入，则可以通过communicate()方法输入：

import subprocess

print('$ nslookup')
p = subprocess.Popen(['nslookup'], stdin=subprocess.PIPE, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
output, err = p.communicate(b'set q=mx\npython.org\nexit\n')
print(output.decode('utf-8'))
print('Exit code:', p.returncode)

上面的代码相当于在命令行执行命令nslookup，然后手动输入：

set q=mx
python.org
exit

进程间通信

Process之间肯定是需要通信的，操作系统提供了很多机制来实现进程间的通信。Python的multiprocessing模块包装了底层的机制，提供了Queue、Pipes等多种方式来交换数据。

我们以Queue为例，在父进程中创建两个子进程，一个往Queue里写数据，一个从Queue里读数据：

from multiprocessing import Process, Queue
import os, time, random

# 写数据进程执行的代码:
def write(q):
  print('Process to write: %s' % os.getpid())
  for value in ['A', 'B', 'C']:
    print('Put %s to queue...' % value)
    q.put(value)
    time.sleep(random.random())

# 读数据进程执行的代码:
def read(q):
  print('Process to read: %s' % os.getpid())
  while True:
    value = q.get(True)
    print('Get %s from queue.' % value)

if __name__=='__main__':
  # 父进程创建Queue，并传给各个子进程：
  q = Queue()
  pw = Process(target=write, args=(q,))
  pr = Process(target=read, args=(q,))
  # 启动子进程pw，写入:
  pw.start()
  # 启动子进程pr，读取:
  pr.start()
  # 等待pw结束:
  pw.join()
  # pr进程里是死循环，无法等待其结束，只能强行终止:
  pr.terminate()

运行结果如下：

Process to write: 50563
Put A to queue...
Process to read: 50564
Get A from queue.
Put B to queue...
Get B from queue.
Put C to queue...
Get C from queue.

在Unix/Linux下，multiprocessing模块封装了fork()调用，使我们不需要关注fork()的细节。由于Windows没有fork调用，因此，multiprocessing需要“模拟”出fork的效果，父进程所有Python对象都必须通过pickle序列化再传到子进程去，所有，如果multiprocessing在Windows下调用失败了，要先考虑是不是pickle失败了。

小结

在Unix/Linux下，可以使用fork()调用实现多进程。

要实现跨平台的多进程，可以使用multiprocessing模块。

进程间通信是通过Queue、Pipes等实现的。

多线程

多任务可以由多进程完成，也可以由一个进程内的多线程完成。进程是由若干线程组成的，一个进程至少有一个线程。

由于线程是操作系统直接支持的执行单元，因此，高级语言通常都内置多线程的支持，Python也不例外，并且，Python的线程是真正的Posix Thread，而不是模拟出来的线程。

Python的标准库提供了两个模块：_thread 和 threading，_thread是低级模块，threading是高级模块，对_thread进行了封装。绝大多数情况下，我们只需要使用threading这个高级模块。

启动一个线程就是把一个函数传入并创建Thread实例，然后调用start()开始执行：

import time, threading

# 新线程执行的代码:
def loop():
  print('thread %s is running...' % threading.current_thread().name)
  n = 0
  while n < 5:
    n = n + 1
    print(&#39;thread %s >>> %s&#39; % (threading.current_thread().name, n))
    time.sleep(1)
  print(&#39;thread %s ended.&#39; % threading.current_thread().name)

print(&#39;thread %s is running...&#39; % threading.current_thread().name)
t = threading.Thread(target=loop, name=&#39;LoopThread&#39;)
t.start()
t.join()
print(&#39;thread %s ended.&#39; % threading.current_thread().name)
thread MainThread is running...
thread LoopThread is running...
thread LoopThread >>> 1
thread LoopThread >>> 2
thread LoopThread >>> 3
thread LoopThread >>> 4
thread LoopThread >>> 5
thread LoopThread ended.
thread MainThread ended.

由于任何进程默认就会启动一个线程，我们把该线程称为主线程，主线程又可以启动新的线程，Python的threading模块有个current_thread()函数，它永远返回当前线程的实例。主线程实例的名字叫MainThread，子线程的名字在创建时指定，我们用LoopThread命名子线程。名字仅仅在打印时用来显示，完全没有其他意义，如果不起名字Python就自动给线程命名为Thread-1，Thread-2……

Lock

多线程和多进程最大的不同在于，多进程中，同一个变量，各自有一份拷贝存在于每个进程中，互不影响，而多线程中，所有变量都由所有线程共享，所以，任何一个变量都可以被任何一个线程修改，因此，线程之间共享数据最大的危险在于多个线程同时改一个变量，把内容给改乱了。

来看看多个线程同时操作一个变量怎么把内容给改乱了：

import time, threading
# 假定这是你的银行存款:
balance = 0
def change_it(n):
  # 先存后取，结果应该为0:
  global balance
  balance = balance + n
  balance = balance - n
def run_thread(n):
  for i in range(100000):
    change_it(n)
t1 = threading.Thread(target=run_thread, args=(5,))
t2 = threading.Thread(target=run_thread, args=(8,))
t1.start()
t2.start()
t1.join()
t2.join()
print(balance)

我们定义了一个共享变量balance，初始值为0，并且启动两个线程，先存后取，理论上结果应该为0，但是，由于线程的调度是由操作系统决定的，当t1、t2交替执行时，只要循环次数足够多，balance的结果就不一定是0了。

原因是因为高级语言的一条语句在CPU执行时是若干条语句，即使一个简单的计算：

balance = balance + n

也分两步：

1. 计算balance + n，存入临时变量中；

2. 将临时变量的值赋给balance。

也就是可以看成：

x = balance + n
balance = x

数据错误的原因：是因为修改balance需要多条语句，而执行这几条语句时，线程可能中断，从而导致多个线程把同一个对象的内容改乱了。

两个线程同时一存一取，就可能导致余额不对，你肯定不希望你的银行存款莫名其妙地变成了负数，所以，我们必须确保一个线程在修改balance的时候，别的线程一定不能改。

如果我们要确保balance计算正确，就要给change_it()上一把锁，当某个线程开始执行change_it()时，我们说，该线程因为获得了锁，因此其他线程不能同时执行change_it()，只能等待，直到锁被释放后，获得该锁以后才能改。由于锁只有一个，无论多少线程，同一时刻最多只有一个线程持有该锁，所以，不会造成修改的冲突。创建一个锁就是通过threading.Lock()来实现：

balance = 0
lock = threading.Lock()

def run_thread(n):
  for i in range(100000):
    # 先要获取锁:
    lock.acquire()
    try:
      # 放心地改吧:
      change_it(n)
    finally:
      # 改完了一定要释放锁:
      lock.release()

当多个线程同时执行lock.acquire()时，只有一个线程能成功地获取锁，然后继续执行代码，其他线程就继续等待直到获得锁为止。

获得锁的线程用完后一定要释放锁，否则那些苦苦等待锁的线程将永远等待下去，成为死线程。所以我们用try...finally来确保锁一定会被释放。

锁的好处就是确保了某段关键代码只能由一个线程从头到尾完整地执行，坏处当然也很多，首先是阻止了多线程并发执行，包含锁的某段代码实际上只能以单线程模式执行，效率就大大地下降了。其次，由于可以存在多个锁，不同的线程持有不同的锁，并试图获取对方持有的锁时，可能会造成死锁，导致多个线程全部挂起，既不能执行，也无法结束，只能靠操作系统强制终止。

多核CPU

如果你不幸拥有一个多核CPU，你肯定在想，多核应该可以同时执行多个线程。

如果写一个死循环的话，会出现什么情况呢？

打开Mac OS X的Activity Monitor，或者Windows的Task Manager，都可以监控某个进程的CPU使用率。

我们可以监控到一个死循环线程会100%占用一个CPU。如果有两个死循环线程，在多核CPU中，可以监控到会占用200%的CPU，也就是占用两个CPU核心。要想把N核CPU的核心全部跑满，就必须启动N个死循环线程。

试试用Python写个死循环：

import threading, multiprocessing

def loop():
  x = 0
  while True:
    x = x ^ 1

for i in range(multiprocessing.cpu_count()):
  t = threading.Thread(target=loop)
  t.start()

启动与CPU核心数量相同的N个线程，在4核CPU上可以监控到CPU占用率仅有102%，也就是仅使用了一核。

但是用C、C++或Java来改写相同的死循环，直接可以把全部核心跑满，4核就跑到400%，8核就跑到800%，为什么Python不行呢？

因为Python的线程虽然是真正的线程，但解释器执行代码时，有一个GIL锁：Global Interpreter Lock，任何Python线程执行前，必须先获得GIL锁，然后，每执行100条字节码，解释器就自动释放GIL锁，让别的线程有机会执行。这个GIL全局锁实际上把所有线程的执行代码都给上了锁，所以，多线程在Python中只能交替执行，即使100个线程跑在100核CPU上，也只能用到1个核。

GIL是Python解释器设计的历史遗留问题，通常我们用的解释器是官方实现的CPython，要真正利用多核，除非重写一个不带GIL的解释器。

所以，在Python中，可以使用多线程，但不要指望能有效利用多核。如果一定要通过多线程利用多核，那只能通过C扩展来实现，不过这样就失去了Python简单易用的特点。

不过，也不用过于担心，Python虽然不能利用多线程实现多核任务，但可以通过多进程实现多核任务。多个Python进程有各自独立的GIL锁，互不影响。

多线程编程，模型复杂，容易发生冲突，必须用锁加以隔离，同时，又要小心死锁的发生。

Python解释器由于设计时有GIL全局锁，导致了多线程无法利用多核。

ThreadLocal

在多线程环境下，每个线程都有自己的数据。一个线程使用自己的局部变量比使用全局变量好，因为局部变量只有线程自己能看见，不会影响其他线程，而全局变量的修改必须加锁。但是局部变量也有问题，就是在函数调用的时候，传递起来很麻烦：

import threading

# 创建全局ThreadLocal对象:
local_school = threading.local()

def process_student():
  # 获取当前线程关联的student:
  std = local_school.student
  print('Hello, %s (in %s)' % (std, threading.current_thread().name))

def process_thread(name):
  # 绑定ThreadLocal的student:
  local_school.student = name
  process_student()

t1 = threading.Thread(target= process_thread, args=('Alice',), name='Thread-A')
t2 = threading.Thread(target= process_thread, args=('Bob',), name='Thread-B')
t1.start()
t2.start()
t1.join()
t2.join()

全局变量local_school就是一个ThreadLocal对象，每个Thread对它都可以读写student属性，但互不影响。你可以把local_school看成全局变量，但每个属性如local_school.student都是线程的局部变量，可以任意读写而互不干扰，也不用管理锁的问题，ThreadLocal内部会处理。

可以理解为全局变量local_school是一个dict，不但可以用local_school.student，还可以绑定其他变量，如local_school.teacher等等。

ThreadLocal最常用的地方就是为每个线程绑定一个数据库连接，HTTP请求，用户身份信息等，这样一个线程的所有调用到的处理函数都可以非常方便地访问这些资源。

一个ThreadLocal变量虽然是全局变量，但每个线程都只能读写自己线程的独立副本，互不干扰。ThreadLocal解决了参数在一个线程中各个函数之间互相传递的问题。

进程 vs. 线程

我们介绍了多进程和多线程，这是实现多任务最常用的两种方式。现在，我们来讨论一下这两种方式的优缺点。

首先，要实现多任务，通常我们会设计Master-Worker模式，Master负责分配任务，Worker负责执行任务，因此，多任务环境下，通常是一个Master，多个Worker。

如果用多进程实现Master-Worker，主进程就是Master，其他进程就是Worker。

如果用多线程实现Master-Worker，主线程就是Master，其他线程就是Worker。

多进程模式最大的优点就是稳定性高，因为一个子进程崩溃了，不会影响主进程和其他子进程。（当然主进程挂了所有进程就全挂了，但是Master进程只负责分配任务，挂掉的概率低）著名的Apache最早就是采用多进程模式。

多进程模式的缺点是创建进程的代价大，在Unix/Linux系统下，用fork调用还行，在Windows下创建进程开销巨大。另外，操作系统能同时运行的进程数也是有限的，在内存和CPU的限制下，如果有几千个进程同时运行，操作系统连调度都会成问题。

多线程模式通常比多进程快一点，但是也快不到哪去，而且，多线程模式致命的缺点就是任何一个线程挂掉都可能直接造成整个进程崩溃，因为所有线程共享进程的内存。在Windows上，如果一个线程执行的代码出了问题，你经常可以看到这样的提示：“该程序执行了非法操作，即将关闭”，其实往往是某个线程出了问题，但是操作系统会强制结束整个进程。

在Windows下，多线程的效率比多进程要高，所以微软的IIS服务器默认采用多线程模式。由于多线程存在稳定性的问题，IIS的稳定性就不如Apache。为了缓解这个问题，IIS和Apache现在又有多进程+多线程的混合模式，真是把问题越搞越复杂。

线程切换

无论是多进程还是多线程，只要数量一多，效率肯定上不去，为什么呢？

我们打个比方，假设你不幸正在准备中考，每天晚上需要做语文、数学、英语、物理、化学这5科的作业，每项作业耗时1小时。

如果你先花1小时做语文作业，做完了，再花1小时做数学作业，这样，依次全部做完，一共花5小时，这种方式称为单任务模型，或者批处理任务模型。

Suppose you plan to switch to the multi-tasking model, you can first do Chinese for 1 minute, then switch to math homework, do 1 minute, then switch to English, and so on. As long as the switching speed is fast enough, this method is the same as single-tasking. It is the same for the core CPU to perform multitasking. From the perspective of a kindergarten child, you are doing homework for 5 subjects at the same time.

However, there is a price for switching homework. For example, when switching from Chinese to mathematics, you must first clear away the Chinese books and pens on the table (this is called saving the scene), then open the mathematics textbook and find the compass and ruler. (This is called preparing for the new environment) before you can start doing math homework. The operating system is the same when switching processes or threads. It needs to first save the current execution environment (CPU register state, memory pages, etc.), and then prepare the execution environment for the new task (restore the last register state, switch memory page, etc.) before execution can begin. Although this switching process is fast, it also takes time. If there are thousands of tasks running at the same time, the operating system may be mainly busy switching tasks, and there is not much time to perform tasks. The most common situation in this situation is that the hard disk beeps wildly, there is no response when clicking on the window, and the system is in a state of suspended animation.

Therefore, once the number of multitasking reaches a limit, all the resources of the system will be consumed. As a result, the efficiency will drop sharply and all tasks will not be completed well.

Computation-intensive vs. IO-intensive

The second consideration in whether to use multitasking is the type of task. We can divide tasks into computing-intensive and IO-intensive.

Computing-intensive tasks are characterized by a large amount of calculations that consume CPU resources, such as calculating pi, high-definition decoding of videos, etc., all relying on the computing power of the CPU. Although this kind of computing-intensive task can also be completed with multi-tasking, the more tasks there are, the more time spent on task switching, and the lower the efficiency of the CPU in executing tasks. Therefore, to make the most efficient use of the CPU, computing-intensive tasks The number of simultaneous tasks should be equal to the number of CPU cores.

Computing-intensive tasks mainly consume CPU resources, so code running efficiency is crucial. Scripting languages ​​like Python run very inefficiently and are completely unsuitable for computationally intensive tasks. For computationally intensive tasks, it is better to write in C language.

The second type of task is IO-intensive. Tasks involving network and disk IO are all IO-intensive tasks. The characteristic of this type of task is that the CPU consumption is very small, and most of the time of the task is Wait for the IO operation to complete (because the speed of IO is much slower than the speed of CPU and memory). For IO-intensive tasks, the more tasks, the higher the CPU efficiency, but there is a limit. Most common tasks are IO-intensive tasks, such as web applications.

During the execution of IO-intensive tasks, 99% of the time is spent on IO, and very little time is spent on the CPU. Therefore, use the extremely fast C language to replace Python with extremely slow running speed. The scripting language cannot improve operating efficiency at all. For IO-intensive tasks, the most suitable language is the language with the highest development efficiency (the least amount of code). Scripting language is the first choice, and C language is the worst.

Asynchronous IO

Considering the huge speed difference between CPU and IO, a task spends most of its execution waiting for IO operations. , the single-process single-thread model will cause other tasks to be unable to be executed in parallel. Therefore, we need a multi-process model or a multi-thread model to support the concurrent execution of multi-tasks.

Modern operating systems have made huge improvements to IO operations. The biggest feature is that they support asynchronous IO. If you make full use of the asynchronous IO support provided by the operating system, you can use a single-process single-thread model to perform multiple tasks. This new model is called an event-driven model. Nginx is a web server that supports asynchronous IO. It runs on a single-core CPU. Using a single-process model can efficiently support multitasking. On a multi-core CPU, you can run multiple processes (the number is the same as the number of CPU cores), taking full advantage of the multi-core CPU. Since the total number of processes in the system is very limited, operating system scheduling is very efficient. Using the asynchronous IO programming model to implement multitasking is a major trend.

Corresponding to the Python language, the single-process asynchronous programming model is called coroutine. With the support of coroutine, efficient multi-task programs can be written based on event-driven. We will discuss how to write coroutines later.

Distributed process

Among Thread and Process, Process should be preferred because Process is more stable and Process can be distributed to multiple machines. Thread can only be distributed to multiple CPUs on the same machine at most.

Python's multiprocessing module not only supports multiple processes, but the managers submodule also supports distributing multiple processes to multiple machines. A service process can act as a scheduler, distributing tasks to multiple other processes, relying on network communication. Since the managers module is well encapsulated, you can easily write distributed multi-process programs without knowing the details of network communication.

For example: If we already have a multi-process program that communicates through Queue running on the same machine, now, because the process that handles the task has a heavy workload, we want to separate the process that sends the task and the process that handles the task. Distributed to two machines. How to implement it using distributed process?

原有的Queue可以继续使用，但是，通过managers模块把Queue通过网络暴露出去，就可以让其他机器的进程访问Queue了。

我们先看服务进程，服务进程负责启动Queue，把Queue注册到网络上，然后往Queue里面写入任务：

import random, time, queue
from multiprocessing.managers import BaseManager

# 发送任务的队列:
task_queue = queue.Queue()
# 接收结果的队列:
result_queue = queue.Queue()

# 从BaseManager继承的QueueManager:
class QueueManager(BaseManager):
  pass

# 把两个Queue都注册到网络上, callable参数关联了Queue对象:
QueueManager.register('get_task_queue', callable=lambda: task_queue)
QueueManager.register('get_result_queue', callable=lambda: result_queue)
# 绑定端口5000, 设置验证码'abc':
manager = QueueManager(address=('', 5000), authkey=b'abc')
# 启动Queue:
manager.start()
# 获得通过网络访问的Queue对象:
task = manager.get_task_queue()
result = manager.get_result_queue()
# 放几个任务进去:
for i in range(10):
  n = random.randint(0, 10000)
  print('Put task %d...' % n)
  task.put(n)
# 从result队列读取结果:
print('Try get results...')
for i in range(10):
  r = result.get(timeout=10)
  print('Result: %s' % r)
# 关闭:
manager.shutdown()
print('master exit.')

当我们在一台机器上写多进程程序时，创建的Queue可以直接拿来用，但是，在分布式多进程环境下，添加任务到Queue不可以直接对原始的task_queue进行操作，那样就绕过了QueueManager的封装，必须通过manager.get_task_queue()获得的Queue接口添加。

然后，在另一台机器上启动任务进程（本机上启动也可以）：

import time, sys, queue
from multiprocessing.managers import BaseManager

# 创建类似的QueueManager:
class QueueManager(BaseManager):
  pass

# 由于这个QueueManager只从网络上获取Queue，所以注册时只提供名字:
QueueManager.register('get_task_queue')
QueueManager.register('get_result_queue')

# 连接到服务器，也就是运行task_master.py的机器:
server_addr = '127.0.0.1'
print('Connect to server %s...' % server_addr)
# 端口和验证码注意保持与task_master.py设置的完全一致:
m = QueueManager(address=(server_addr, 5000), authkey=b'abc')
# 从网络连接:
m.connect()
# 获取Queue的对象:
task = m.get_task_queue()
result = m.get_result_queue()
# 从task队列取任务,并把结果写入result队列:
for i in range(10):
  try:
    n = task.get(timeout=1)
    print('run task %d * %d...' % (n, n))
    r = '%d * %d = %d' % (n, n, n*n)
    time.sleep(1)
    result.put(r)
  except Queue.Empty:
    print('task queue is empty.')
# 处理结束:
print('worker exit.')

任务进程要通过网络连接到服务进程，所以要指定服务进程的IP。http://www.jb51.net/article/65112.htm

小结

Python的分布式进程接口简单，封装良好，适合需要把繁重任务分布到多台机器的环境下。

注意Queue的作用是用来传递任务和接收结果，每个任务的描述数据量要尽量小。比如发送一个处理日志文件的任务，就不要发送几百兆的日志文件本身，而是发送日志文件存放的完整路径，由Worker进程再去共享的磁盘上读取文件。

The above is the detailed content of Introduction to the method of completing parallel processing in Python multi-process. For more information, please follow other related articles on the PHP Chinese website!