前言

虽然异步是我们急需掌握的高阶技术，但是不积跬步无以至千里，同步技术的学习是不能省略的。今天这篇文章主要用Python来介绍Web并发模型，直观地展现同步技术的缺陷以及异步好在哪里。

最简单的并发

import socket

response = 'HTTP/1.1 200 OK\r\nConnection: Close\r\nContent-Length: 11\r\n\r\nHello World'

server = socket.socket()
server.bind(('0.0.0.0', 9527))
server.listen(1024)

while True:
    client, clientaddr = server.accept()  # blocking
    request = client.recv(1024)  # blocking
    client.send(response)  # maybe blocking
    client.close()

上面这个例子太简单了，访问localhost:9527，返回“Hello World”。用ab来测试性能，数据如下：

ab -n 100000 -c 8 http://localhost:9527/
Time taken for tests:   1.568 seconds

多进程

每个请求对应一个进程倒是能解决上面的问题，但是进程太占资源，每个请求的资源都是独立的，无法共享，而且进程的上下文切换成本也很高。

import socket
import signal
import multiprocessing 

response = 'HTTP/1.1 200 OK\r\nConnection: Close\r\nContent-Length: 11\r\n\r\nHello World'

server = socket.socket()
server.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
server.bind(('0.0.0.0', 9527))
server.listen(1024)

def handler(client):
    request = client.recv(1024)
    client.send(response)
    client.close()

#多进程里的子进程执行完后并不会死掉，而是变成僵尸进程，等待主进程挂掉后才会死掉，下面这条语句可以解决这个问题。
signal.signal(signal.SIGCHLD,signal.SIG_IGN)

while True:
    client, addr = server.accept()
    process = multiprocessing.Process(target=handler, args=(client,))
    process.start()

Prefork

这是多进程的改良版，预先分配好和CPU核数一样的进程数，可以控制资源占用，高效处理请求。

import socket
import multiprocessing

response = 'HTTP/1.1 200 OK\r\nConnection: Close\r\nContent-Length: 11\r\n\r\nHello World'

server = socket.socket()
server.bind(('0.0.0.0', 9527))
server.listen(1024)

def handler():
    while True:
        client, addr = server.accept()
        request = client.recv(1024)
        client.send(response)
        client.close()
processors = 8
for i in range(0, processors):
    process = multiprocessing.Process(target=handler, args=())
    process.start()

耗时：1.640秒。

线程池

import Queue
import socket
import threading

response = 'HTTP/1.1 200 OK\r\nConnection: Close\r\nContent-Length: 11\r\n\r\nHello World'

server = socket.socket()
server.bind(('0.0.0.0', 9527))
server.listen(1024)

def handler(queue):
    while True:
        client  = queue.get()
        request = client.recv(1024)
        client.send(response)
        client.close()

queue = Queue.Queue()
processors = 8
for i in range(0, processors):
    thread = threading.Thread(target=handler, args=(queue,))
    thread.daemon = True
    thread.start()

while True:
    client, clientaddr = server.accept()
    queue.put(client)

耗时：3.901秒，大部分时间花在队列上，线程占用资源比进程少（资源可以共享），但是要考虑线程安全问题和锁的性能。

epoll

import select
import socket

response = 'HTTP/1.1 200 OK\r\nConnection: Close\r\nContent-Length: 11\r\n\r\nHello World'
server = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
server.setblocking(False)
server_address = ('localhost', 9527)
server.bind(server_address)
server.listen(1024)
READ_ONLY = select.EPOLLIN | select.EPOLLPRI
epoll = select.epoll()
epoll.register(server, READ_ONLY)
timeout = 60
fd_to_socket = { server.fileno(): server}
while True:
    events = epoll.poll(timeout)
    for fd, flag in events:
        sock = fd_to_socket[fd]
        if flag & READ_ONLY:
            if sock is server:
                conn, client_address = sock.accept()
                conn.setblocking(False)
                fd_to_socket[conn.fileno()] = conn
                epoll.register(conn, READ_ONLY)
            else:
                request = sock.recv(1024)
                sock.send(response)
                sock.close()
                del fd_to_socket[fd]

最后祭出epoll大神，三大异步通信框架Netty、NodeJS、Tornado共同采用的通信技术，耗时1.582秒，但是要注意是单进程单线程哦。epoll真正发挥作用是在长连接应用里，单线程处理上万个长连接玩一样，占用资源极少。