1 TCP的开销
a 连接协商三次握手,c->syn->s,s->syn ack->c, c->ack->s
b 关闭协商四次握手,c->fin->s, s->ack-c,s->fin->c,c->ack->s
c 保持数据有序,响应确认等计算开销
d 网络拥塞引起的重试开销
2 使用知名端口初始化 serversocket可能需要超级权限。ServerSocket(int port, int backlog)参数backlog用来配置连接队列,在accept之前预先完成连接,加速连接TCP连接阶段,默认为50.
backlog表示ServerSocket可以接受的同时最大连接数量,超过这个连接数量,将会拒绝连接。
如果要提高吞吐量,可以通过设置更大的.setReceiveBufferSize来实现,但是必须在bind之前设置,也就是说要先调用无参构造,然后再调用.bind( endpoint)
3 网络io写操作,提高吞吐量较好的实践有使用java.io.BufferedOutputStream,作为缓冲,减少用户线程和内核线程的切换频率。缓冲区大小一般大于.setReceiveBufferSize。
4 避免对象流死锁,较好的实践是如果要在同一个socket上构建对象输入流和输出流,最好是先构造输出流,再构造输入流。
5 tcp半关闭,shut down output,完成后,对方的read收到eof,结束阻塞。
6 tcp关闭可以用socket.close,socket.getoutputstream.close,socket.getinputstream.close,较好的方式是调用socket.getoutpurtstream.close,它会把未flush的flush掉。三个方法只需调用其中一个即可。isClose方法只会告诉我们本地tcp是否关闭,但是不能告诉我们远程是否关闭。
7 socket read 设置timeout时间,防止无止境阻塞。一般来说,timeout时间会设定为预期时间的两倍。timeout时间设置只对之后的阻塞读有效。
8 每个socket都有send buffer和receive buffer,这个buffer在内核地址空间而非jvm。buffer的size由操作系统实现决定,一般来说是2kb。send buffer可以在tcp关闭前随时设定,通过java.net.Socket.setSendBufferSize(int)设置。但是size的设置只是一种hint,不是绝对值。size设得越大,减少网络写次数,减少拥塞控制,tcp效率、吞吐量越高,类似http://en.wikipedia.org/wiki/Nagle‘s_algorithm 原理。
一般设定为MSS的三倍;至少大于对方receive buffer;receive buffer也要设定大一点,不拖send buffer后腿;
bufferedoutputstream,bytebuffer一般也要设定为匹配的值;
buffersize(bits)=bandwidth(bits/sec)* delay(sec),有点类似于线程数量的控制,不让cpu闲下来。这边的白话是不让buffer空下来,随时处于最大填充状态。
9 nagle算法,为了提高网络传输效率,减少网络拥塞,延迟小包发送,组装为大包一起发送。默认为开,可以通过setTcpnodelay为true来关闭。一般来说,不会关闭,除非是需要实时交互的场景。另外如果真需要关闭,可以采用巧妙的方式,使用bufferedoutputstream,把buffer size设为大于最大请求或响应包,socket send buffer和receive buffer也设为此值,用一次操作写出请求或响应,bufferedoutputstream.flush,充分利用网络。
10 setlinger,用于关闭socket时,进行磨蹭,拖延。
11 keep alive,是个鸡肋。用于检测连接是否处于连接状态,检测对方是否active。它比较有争议,不是tcp协议的标准内容。另外检测需要消耗网络,当检测对方无反应,socket会被置为reset状态,不可读写。一般不推荐使用。
可以考虑用应用层的心跳检测替代。
参考http://hi.baidu.com/tantea/blog/item/580b9d0218f981793812bb7b.html
12 settrafficclass,设置流量类别,只是hint作用,具体效果取决于实现。有这些类别 IPTOS_LOWCOST (0x02),IPTOS_RELIABILITY (0x04),IPTOS_THROUGHPUT (0x08),IPTOS_LOWDELAY (0x10)
13 接口中文翻译http://hi.baidu.com/%EC%C5%BF%E1%D0%A1%B7%E5/blog/item/5d8e0f58aee147471038c29d.html
14 java nio进入新时代,提供非阻塞和多路复用特性,就绪选择器,事件驱动,不再是一个线程处理一个请求,大大节约了线程数量和内存,提高了可伸缩性。
http隧道穿透防火墙,白话为露丝想写情书给他男朋友,但是他爸妈(防火墙)不允许,于是露丝就把情书包装起来写给她的闺蜜莉莉(http 代理服务器,这个代理服务器在防火墙之内),再由莉莉转交给他男朋友。
17 UDP size比较受限(512kb),不可靠,无连接,但是成本低。丢失不重发,重发需要应用控制,要考虑发送消息是否幂等。UDP数据报是个独立传输单位,在java里UDP用java.net.DatagramPacket。适用于发送心跳场景。DatagramSocket的connect,close操作都是针对本地的,并无对连接产生什么效果,毕竟是无连接协议。
如果想提高可靠度,可以在应用实现,clinet维护一个序列号,等待server响应这个序列号,否则进行重发策略。/*
- * ReliableDatagramSocket.java.
- * Copyright ? Esmond Pitt, 1997, 2005. All rights reserved.
- * Permission to use is granted provided this copyright
- * and permission notice is preserved.
- */
- import java.io.*;
- import java.net.*;
- import java.text.*;
- import java.util.*;
- // All times are expressed in seconds.
- // ReliabilityConstants interface, just defines constants.
- interface ReliabilityConstants
- {
- // Timeout minima/maxima
- public static final int MIN_RETRANSMIT_TIMEOUT = 1;
- public static final int MAX_RETRANSMIT_TIMEOUT = 64;
- // Maximum retransmissions per datagram, suggest 3 or 4.
- public static final int MAX_RETRANSMISSIONS = 4;
- }
- The D;; class manages current and smoothed round-trip timers
- and the related timeouts:
- // RoundTripTimer class.
- class RoundTripTimer implements ReliabilityConstants
- {
- float roundTripTime = 0.0f;// most recent RTT
- float smoothedTripTime = 0.0f;// smoothed RTT
- float deviation = 0.75f; // smoothed mean deviation
- short retransmissions = 0;// retransmit count: 0, 1, 2, …
- // current retransmit timeout
- float currentTimeout =
- minmax(calculateRetransmitTimeout());
- /** @return the re-transmission timeout. */
- private int calculateRetransmitTimeout()
- {
- return (int)(smoothedTripTime+4.0*deviation);
- }
- /** @return the bounded retransmission timeout. */
- private float minmax(float rto)
- {
- return Math.min
- (Math.max(rto, MIN_RETRANSMIT_TIMEOUT),
- MAX_RETRANSMIT_TIMEOUT);
- }
- /** Called before each new packet is transmitted. */
- void newPacket()
- {
- retransmissions = 0;
- }
- /**
- * @return the timeout for the packet.
- */
- float currentTimeout()
- {
- return currentTimeout;
- }
- /**
- * Called straight after a successful receive.
- * Calculates the round-trip time, then updates the
- * smoothed round-trip time and the variance (deviation).
- * @param ms time in ms since starting the transmission.
- */
- void stoppedAt(long ms)
- {
- // Calculate the round-trip time for this packet.
- roundTripTime = ms/1000;
- // Update our estimators of round-trip time
- // and its mean deviation.
- double delta = roundTripTime ? smoothedTripTime;
- smoothedTripTime += delta/8.0;
- deviation += (Math.abs(delta)-deviation)/4.0;
- // Recalculate the current timeout.
- currentTimeout = minmax(calculateRetransmitTimeout());
- }
- /**
- * Called after a timeout has occurred.
- * @return true if it‘s time to give up,
- * false if we can retransmit.
- */
- boolean isTimeout()
- {
- currentTimeout *= 2; // next retransmit timeout
- retransmissions++;
- return retransmissions > MAX_RETRANSMISSIONS;
- }
- } // RoundTripTimer class
- The D
-
-
-
- " class exports a D method like the ones
- we have already seen.
- // ReliableDatagramSocket class
- public class ReliableDatagramSocket
- extends DatagramSocket
- implements ReliabilityConstants
- {
- RoundTripTimer roundTripTimer = new RoundTripTimer();
- private boolean reinit = false;
- private long sendSequenceNo = 0; // send sequence #
- private long recvSequenceNo = 0; // recv sequence #
- /* anonymous initialization for all constructors */
- {
- init();
- }
- /**
- * Construct a ReliableDatagramSocket
- * @param port Local port: reeive on any interface/address
- * @exception SocketException can‘t create the socket
- */
- public ReliableDatagramSocket(int port)
- throws SocketException
- {
- super(port);
- }
- /**
- * Construct a ReliableDatagramSocket
- * @param port Local port
- * @param localAddr local interface address to use
- * @exception SocketException can‘t create the socket
- */
- public ReliableDatagramSocket
- (int port, InetAddress localAddr) throws SocketException
- {
- super(port, localAddr);
- }
- /**
- * Construct a ReliableDatagramSocket, JDK >= 1.4.
- * @param localAddr local socket address to use
- * @exception SocketException can‘t create the socket
- */
- public ReliableDatagramSocket(SocketAddress localAddr)
- throws SocketException
- {
- super(localAddr);
- }
- /**
- * Overrides DatagramSocket.connect():
- * Does the connect, then (re-)initializes
- * the statistics for the connection.
- * @param dest Destination address
- * @param port Destination port
- */
- public void connect(InetAddress dest, int port)
- {
- super.connect(dest, port);
- init();
- }
- /**
- * Overrides JDK 1.4 DatagramSocket.connect().
- * Does the connect, then (re-)initializes
- * the statistics for the connection.
- * @param dest Destination address
- */
- public void connect(SocketAddress dest)
- {
- super.connect(dest);
- init();
- }
- /** Initialize */
- private void init()
- {
- this.roundTripTimer = new RoundTripTimer();
- }
- /**
- * Send and receive reliably,
- * retrying adaptively with exponential backoff
- * until the response is received or timeout occurs.
- * @param sendPacket outgoing request datagram
- * @param recvPacket incoming reply datagram
- * @exception IOException on any error
- * @exception InterruptedIOException on timeout
- */
- public synchronized void sendReceive
- (DatagramPacket sendPacket, DatagramPacket recvPacket)
- throws IOException, InterruptedIOException
- {
- // re-initialize after timeout
- if (reinit)
- {
- init();
- reinit = false;
- }
- roundTripTimer.newPacket();
- long start = System.currentTimeMillis();
- long sequenceNumber = getSendSequenceNo();
- // Loop until final timeout or some unexpected exception
- for (;;)
- {
- // keep using the same sequenceNumber while retrying
- setSendSequenceNo(sequenceNumber);
- send(sendPacket);// may throw
- int timeout =
- (int)(roundTripTimer.currentTimeout()*1000.0+0.5);
- long soTimeoutStart = System.currentTimeMillis();
- try
- {
- for (;;)
- {
- // Adjust socket timeout for time already elapsed
- int soTimeout = timeout?(int)
- (System.currentTimeMillis()?soTimeoutStart);
- setSoTimeout(soTimeout);
- receive(recvPacket);
- long recvSequenceNumber = getRecvSequenceNo();
- if (recvSequenceNumber == sequenceNumber)
- {
- // Got the correct reply:
- // stop timer, calculate new RTT values
- long ms = System.currentTimeMillis()-start;
- roundTripTimer.stoppedAt(ms);
- return;
- }
- }
- }
- catch (InterruptedIOException exc)
- {
- // timeout: retry?
- if (roundTripTimer.isTimeout())
- {
- reinit = true;
- // rethrow InterruptedIOException to caller
- throw exc;
- }
- // else continue
- }
- // may throw other SocketException or IOException
- } // end re-transmit loop
- } // sendReceive()
- /**
- * @return the last received sequence number;
- * used by servers to obtain the reply sequenceNumber.
- */
- public long getRecvSequenceNo()
- {
- return recvSequenceNo;
- }
- /** @return the last sent sequence number */
- private long getSendSequenceNo()
- {
- return sendSequenceNo;
- }
- /**
- * Set the next send sequence number.
- * Used by servers to set the reply
- * sequenceNumber from the received packet:
- *
- . * socket.setSendSequenceNo(socket.getRecvSequenceNo());
- *
- * @param sendSequenceNo Next sequence number to send.
- */
- public void setSendSequenceNo(long sendSequenceNo)
- {
- this.sendSequenceNo = sendSequenceNo;
- }
- /**
- * override for DatagramSocket.receive:
- * handles the sequence number.
- * @param packet DatagramPacket
- * @exception IOException I/O error
- */
- public void receive(DatagramPacket packet)
- throws IOException
- {
- super.receive(packet);
- // read sequence number and remove it from the packet
- ByteArrayInputStream bais = new ByteArrayInputStream
- (packet.getData(), packet.getOffset(),
- packet.getLength());
- DataInputStream dis = new DataInputStream(bais);
- recvSequenceNo = dis.readLong();
- byte[] buffer = new byte[dis.available()];
- dis.read(buffer);
- packet.setData(buffer,0,buffer.length);
- }
- /**
- * override for DatagramSocket.send:
- * handles the sequence number.
- * @param packet DatagramPacket
- * @exception IOException I/O error
- */
- public void send(DatagramPacket packet)
- throws IOException
- {
- ByteArrayOutputStreambaos = new ByteArrayOutputStream();
- DataOutputStreamdos = new DataOutputStream(baos);
- // Write the sequence number, then the user data.
- dos.writeLong(sendSequenceNo++);
- dos.write
- (packet.getData(), packet.getOffset(),
- packet.getLength());
- dos.flush();
- // Construct a new packet with this new data and send it.
- byte[]data = baos.toByteArray();
- packet = new DatagramPacket
- (data, baos.size(), packet.getAddress(),
- packet.getPort());
- super.send(packet);
- }
- } // end of ReliableDatagramSocket class
Java代码
- public class ReliableEchoServer implements Runnable
- {
- ReliableDatagramSocket
- socket;
- byte[] buffer = new byte[1024];
- DatagramPacket recvPacket =
- new DatagramPacket(buffer, buffer.length);
- ReliableEchoServer(int port) throws IOException
- {
- this.socket = new ReliableDatagramSocket(port);
- }
- public void run()
- {
- for (;;)
- {
- try
- {
- // Restore the receive length to the maximum
- recvPacket.setLength(buffer.length);
- socket.receive(recvPacket);
- // Reply must have same seqno as request
- long seqno = socket.getRecvSequenceNo();
- socket.setSendSequenceNo(seqno);
- // Echo the request back as the response
- socket.send(recvPacket);
- }
- catch (IOException exc)
- {
- exc.printStackTrace();
- }
- } // for (;;)
- } // run()
- } // class
UDP支持多播和广播(广播是一种特殊的多播,尽量不使用广播,广播产生更多没必要的网络流量),而TCP只支持单播。一般多播用于服务发现,如jini look up。多播与多次单播相比,好处是减少开销、减小网络流量、减少服务器负载,而且速度更快,并且接受者接收到消息的时间更接近,对于某些场景来说很重要。
多播的缺点是继承了udp,不可靠网络,依赖路由器,安全问题更加复杂。并且多播并不知道多播消息会被哪些接受者接收,也不知道接受者是否接收到,设计协议的时候需要考虑这点。
发送多播消息,发送端可以用MulticastSocket和DatagramSocket,而接收端只能用MulticastSocket。
多播使用场景
(a) Software distribution
(b) Time services
(c) Naming services like
(d) Stock-market tickers, race results, and the like
(e) Database replication
(f) Video and audio streaming: video conferencing, movie shows, etc
(g) Multi-player gaming
(h) Distributed resource allocation
(i) Service discovery.
18 设计server需要考虑两点:同时连接的客户数量,每个连接的持续时间。当客户超过一个的时候,我们就要考虑用多线程,这个时候就涉及到线程如何创建、线程运行、线程销毁。服务器端由等待连接的线程和处理连接的线程组成。
服务器模型进化趋势:单线程接收连接、处理连接,无法同时处理多个客户,淘汰;每接收一个请求,创建一个线程对请求处理,可以并发,但是会耗尽服务器资源;采用线程池方式,并进行阀值控制,保护服务器,并进行优雅降级。
关于线程池线程数量的控制,一般是预创建N个线程,当峰值访问来临时,临时创建M个动态线程,一旦访问峰值降下来,再释放动态线程。
连接模型可以分为一个连接一个对话(请求-响应);一个连接多次对话。不同的模型,连接释放的方式不一样。
代码如下
- public void processSession(Socket socket)
- {
- receive(request);
- // process request and construct reply, not shown …
- send(reply);
- // close connection
- socket.close();// exception handling not shown
- }
- void processSession(Socket socket)
- {
- while (receive(request)) // i.e. while not end-of-stream
- {
- // process request and construct reply, not shown …
- send(reply);
- }
- // close connection
- socket.close();// exception handling not shown
- }
多次对话的连接释放方式,可以根据输入流的返回结果,或者遇到eof来关闭连接。
归结点
(a) On receipt of an end-of-stream when reading the connection.
(b) If the request or the client is deemed invalid.
(c) On detection of a read timeout or idle timeout on the connection.
(d) After writing a reply
19 设计客户端,一般需要考虑连接失败和读数据超时。为了减少创建连接的开销,一般还会使用线程池,如rmi。
在请求-响应事务中,一般会采取header_body_trailler的结构。结合使用gathering、scattering io来较少内存和cpu开销
- // Initialization - common to both ends
- static final int HEADER_LENGTH = 16;
- static final int BODY_LENGTH = 480;
- static final int TRAILER_LENGTH = 16;
- ByteBuffer header = ByteBuffer.allocate(HEADER_LENGTH);
- ByteBuffer body = ByteBuffer.allocate(BODY_LENGTH);
- ByteBuffer trailer = ByteBuffer.allocate(TRAILER_LENGTH);
- ByteBuffer[]
- buffers = new ByteBuffer[]
- { header, body, trailer };
- // sending end - populate the buffers, not shown
- long count = channel.write(buffers);
- // repeat until all data sent
- // receiving end
- long count = channel.read(buffers);
- // repeat until all data read
对于浏览器加载页面的过程,由于加载对交互顺序不敏感,所以client可以同时并发多个连接、多个线程并行从服务端获取数据
20 jdk为编写并发服务器提供了很好的支持。如Executors提供了线程池,java.util.concurrent.ThreadPoolExecutor.DiscardPolicy提供了阀值控制,ThreadFactory提供了创建线程的方式。
21 客户端技术一般来用连接池,如memcache client每个连接某时刻只在一个request-reply事务中。或者多个事务公用一个连接,比如tair client,需要在协议上维护request-reply的匹配关系。
22 网络编程的八个谬论
a 网络是可靠的
b 网络没有延迟
c 带宽是无限的
d 网络是安全的
e 网络拓扑不会变
f 只有一个管理员
g 传输开销为0
h 网络是均匀的,网络由不同带宽的节点组成,木桶理论,以最小的那个为带宽。
i 网络io如同磁盘io。网络io更容易出错,不如磁盘稳定
j 和peer的状态是同步的。除非在应用层接收到ack,否则不要假定对方收到你的数据。
k 所有的网络失败都是可以检测的。
l 资源是无限的。其实网络编程涉及的资源包括端口、缓冲都是有限的
m 应用可以无限等待远程服务。任何远程调用都应该设定超时时间。
n 远程服务的响应是及时的
o 有单点失败。在分布式系统中,一般一个host的失败不会引发整个系统的崩溃。除非有一个中心节点。
p 只有一个资源分配器。每个host的资源都可以独立分配。
q 时间是完全统一的
参考: