java socket参数详解:TcpNoDelay

TcpNoDelay=false，为启用nagle算法，也是默认值。 Nagle算法的立意是良好的，避免网络中充塞小封包，提高网络的利用率。但是当Nagle算法遇到delayed ACK悲剧就发生了。Delayed ACK的本意也是为了提高TCP性能，跟应答数据捎带上ACK，同时避免糊涂窗口综合症，也可以一个ack确认多个段来节省开销。悲剧发生在这种情况，假设一端发送数据并等待另一端应答，协议上分为头部和数据，发送的时候不幸地选择了write-write，然后再read，也就是先发送头部，再发送数据，最后等待应答。
实验模型：
发送端（客户端）
write(head);
write(body);
read(response);
接收端（服务端）
read(request);
process(request);
write(response);
这里假设head和body都比较小，当默认启用nagle算法，并且是第一次发送的时候，根据nagle算法，第一个段head可以立即发送，因为没有等待确认的段；接收端（服务端）收到head，但是包不完整，继续等待body达到并延迟ACK；发送端（客户端）继续写入body，这时候nagle算法起作用了，因为head还没有被ACK，所以body要延迟发送。这就造成了发送端（客户端）和接收端（服务端）都在等待对方发送数据的现象：
发送端（客户端）等待接收端ACK head以便继续发送body；
接收端（服务端）在等待发送方发送body并延迟ACK，悲剧的无以言语。
这种时候只有等待一端超时并发送数据才能继续往下走。
代码：
发送端代码
[java] view plaincopyprint?

package socket.nagle;

import java.io.*;
import java.net.*;
import org.apache.log4j.Logger;

public class Client {
private static Logger logger = Logger.getLogger(Client.class);
public static void main(String[] args) throws Exception {
// 是否分开写head和body
boolean writeSplit = true;
String host = "localhost";
logger.debug("WriteSplit:" + writeSplit);

Socket socket = new Socket();
socket.setTcpNoDelay(false);
socket.connect(new InetSocketAddress(host, 10000));

InputStream in = socket.getInputStream();
OutputStream out = socket.getOutputStream();
BufferedReader reader = new BufferedReader(new InputStreamReader(in));

String head = "hello ";
String body = "world\r\n";
for (int i = 0; i < 10; i++) {
long label = System.currentTimeMillis();
if (writeSplit) {
out.write(head.getBytes());
out.write(body.getBytes());
} else {
out.write((head + body).getBytes());
}
String line = reader.readLine();
logger.debug("RTT:" + (System.currentTimeMillis() - label) + ", receive: " + line);
}
in.close();
out.close();
socket.close();
}
}

接收端代码
[java] view plaincopyprint?

package socket.nagle;

import java.io.*;
import java.net.*;

import org.apache.log4j.Logger;

public class Server {
private static Logger logger = Logger.getLogger(Server.class);

public static void main(String[] args) throws Exception {
ServerSocket serverSocket = new ServerSocket();
serverSocket.bind(new InetSocketAddress(10000));
logger.debug(serverSocket);
logger.debug("Server startup at 10000");
while (true) {
Socket socket = serverSocket.accept();
InputStream in = socket.getInputStream();
OutputStream out = socket.getOutputStream();

while (true) {
try {
BufferedReader reader = new BufferedReader(new InputStreamReader(in));
String line = reader.readLine();
logger.debug(line);
out.write((line + "\r\n").getBytes());
} catch (Exception e) {
break;
}
}
}
}
}

实验结果：
[plain] view plaincopyprint?

[test5@cent4 ~]$ java socket.nagle.Server
1 [main] DEBUG socket.nagle.Server - ServerSocket[addr=0.0.0.0/0.0.0.0,port=0,localport=10000]
6 [main] DEBUG socket.nagle.Server - Server startup at 10000
4012 [main] DEBUG socket.nagle.Server - hello world
4062 [main] DEBUG socket.nagle.Server - hello world
4105 [main] DEBUG socket.nagle.Server - hello world
4146 [main] DEBUG socket.nagle.Server - hello world
4187 [main] DEBUG socket.nagle.Server - hello world
4228 [main] DEBUG socket.nagle.Server - hello world
4269 [main] DEBUG socket.nagle.Server - hello world
4310 [main] DEBUG socket.nagle.Server - hello world
4350 [main] DEBUG socket.nagle.Server - hello world
4390 [main] DEBUG socket.nagle.Server - hello world
4392 [main] DEBUG socket.nagle.Server -
4392 [main] DEBUG socket.nagle.Server -

实验1：当WriteSplit=true and TcpNoDelay=false 启用nagle算法
[plain] view plaincopyprint?

[test5@cent4 ~]$ java socket.nagle.Client
0 [main] DEBUG socket.nagle.Client - WriteSplit:true
52 [main] DEBUG socket.nagle.Client - RTT:12, receive: hello world
95 [main] DEBUG socket.nagle.Client - RTT:42, receive: hello world
137 [main] DEBUG socket.nagle.Client - RTT:41, receive: hello world
178 [main] DEBUG socket.nagle.Client - RTT:41, receive: hello world
218 [main] DEBUG socket.nagle.Client - RTT:40, receive: hello world
259 [main] DEBUG socket.nagle.Client - RTT:40, receive: hello world
300 [main] DEBUG socket.nagle.Client - RTT:41, receive: hello world
341 [main] DEBUG socket.nagle.Client - RTT:41, receive: hello world
382 [main] DEBUG socket.nagle.Client - RTT:41, receive: hello world
422 [main] DEBUG socket.nagle.Client - RTT:40, receive: hello world

可以看到，每次请求到应答的时间间隔都在40ms，除了第一次。linux的delayed ack是40ms，而不是原来以为的200ms。第一次立即ACK，似乎跟linux的quickack mode有关，这里我不是特别清楚，
其实问题不是出在nagle算法身上的，问题是出在write-write-read这种应用编程上。禁用nagle算法可以暂时解决问题，但是禁用 nagle算法也带来很大坏处，网络中充塞着小封包，网络的利用率上不去，在极端情况下，大量小封包导致网络拥塞甚至崩溃。在这种情况下，其实你只要避免write-write-read形式的调用就可以避免延迟现象，如下面这种情况发送的数据不要再分割成两部分。
实验2：当WriteSplit=false and TcpNoDelay=false 启用nagle算法
[plain] view plaincopyprint?

[test5@cent4 ~]$ java socket.nagle.Client
0 [main] DEBUG socket.nagle.Client - WriteSplit:false
27 [main] DEBUG socket.nagle.Client - RTT:4, receive: hello world
31 [main] DEBUG socket.nagle.Client - RTT:3, receive: hello world
34 [main] DEBUG socket.nagle.Client - RTT:2, receive: hello world
38 [main] DEBUG socket.nagle.Client - RTT:3, receive: hello world
42 [main] DEBUG socket.nagle.Client - RTT:2, receive: hello world
44 [main] DEBUG socket.nagle.Client - RTT:2, receive: hello world
47 [main] DEBUG socket.nagle.Client - RTT:1, receive: hello world
50 [main] DEBUG socket.nagle.Client - RTT:2, receive: hello world
53 [main] DEBUG socket.nagle.Client - RTT:1, receive: hello world
54 [main] DEBUG socket.nagle.Client - RTT:1, receive: hello world

实验3：当WriteSplit=true and TcpNoDelay=true 禁用nagle算法
[plain] view plaincopyprint?

[test5@cent4 ~]$ java socket.nagle.Client
0 [main] DEBUG socket.nagle.Client - WriteSplit:true
25 [main] DEBUG socket.nagle.Client - RTT:6, receive: hello world
28 [main] DEBUG socket.nagle.Client - RTT:2, receive: hello world
31 [main] DEBUG socket.nagle.Client - RTT:3, receive: hello world
33 [main] DEBUG socket.nagle.Client - RTT:1, receive: hello world
35 [main] DEBUG socket.nagle.Client - RTT:2, receive: hello world
41 [main] DEBUG socket.nagle.Client - RTT:6, receive: hello world
49 [main] DEBUG socket.nagle.Client - RTT:3, receive: hello world
52 [main] DEBUG socket.nagle.Client - RTT:2, receive: hello world
56 [main] DEBUG socket.nagle.Client - RTT:3, receive: hello world
59 [main] DEBUG socket.nagle.Client - RTT:3, receive: hello world

实验4：当WriteSplit=false and TcpNoDelay=true 禁用nagle算法
[plain] view plaincopyprint?

[test5@cent4 ~]$ java socket.nagle.Client
0 [main] DEBUG socket.nagle.Client - WriteSplit:false
21 [main] DEBUG socket.nagle.Client - RTT:3, receive: hello world
23 [main] DEBUG socket.nagle.Client - RTT:1, receive: hello world
27 [main] DEBUG socket.nagle.Client - RTT:3, receive: hello world
30 [main] DEBUG socket.nagle.Client - RTT:2, receive: hello world
32 [main] DEBUG socket.nagle.Client - RTT:1, receive: hello world
35 [main] DEBUG socket.nagle.Client - RTT:2, receive: hello world
38 [main] DEBUG socket.nagle.Client - RTT:2, receive: hello world
41 [main] DEBUG socket.nagle.Client - RTT:2, receive: hello world
43 [main] DEBUG socket.nagle.Client - RTT:1, receive: hello world
46 [main] DEBUG socket.nagle.Client - RTT:2, receive: hello world

实验2到4，都没有出现延时的情况。
注意：以上实验在windows上测试下面的代码，客户端和服务器必须分在两台机器上，似乎winsock对loopback连接的处理不一样。下面的我的做法是：服务端与客户端都在一台Linux机上。