Hadoop的RPC的通信与其他系统的RPC通信不太一样，作者针对Hadoop的使用特点，专门的设计了一套RPC框架，这套框架个人感觉还是有点小复杂的。所以我打算分成Client客户端和Server服务端2个模块做分析。如果你对RPC的整套流程已经非常了解的前提下，对于Hadoop的RPC，你也一定可以非常迅速的了解的。OK，下面切入正题。
Hadoop的RPC的相关代码都在org.apache.hadoop.ipc的包下，首先RPC的通信必须遵守许多的协议，其中最最基本的协议即使如下；
/**
* Superclass of all protocols that use Hadoop RPC.
* Subclasses of this interface are also supposed to have
* a static final long versionID field.
* Hadoop RPC所有协议的基类，返回协议版本号
*/
public interface VersionedProtocol {
/**
* Return protocol version corresponding to protocol interface.
* @param protocol The classname of the protocol interface
* @param clientVersion The version of the protocol that the client speaks
* @return the version that the server will speak
*/
public long getProtocolVersion(String protocol,
long clientVersion) throws IOException;
}
顾名思义，只有客户端和服务端遵循相同的版本号，才能进行通信。
RPC客户端的所有相关操作都被封装在了一个叫Client.java的文件中:
/** A client for an IPC service. IPC calls take a single {@link Writable} as a
* parameter, and return a {@link Writable} as their value. A service runs on
* a port and is defined by a parameter class and a value class.
* RPC客户端类
* @see Server
*/
public class Client {
public static final Log LOG =
LogFactory.getLog(Client.class);
//客户端到服务端的连接
private Hashtable<ConnectionId, Connection> connections =
new Hashtable<ConnectionId, Connection>();
//回调值类
private Class<? extends Writable> valueClass; // class of call values
//call回调id的计数器
private int counter; // counter for call ids
//原子变量判断客户端是否还在运行
private AtomicBoolean running = new AtomicBoolean(true); // if client runs
final private Configuration conf;
//socket工厂，用来创建socket
private SocketFactory socketFactory; // how to create sockets
private int refCount = 1;
......从代码中明显的看到，这里存在着一个类似于connections连接池的东西，其实这暗示着连接是可以被复用的，在hashtable中，与每个Connecttion连接的对应的是一个ConnectionId，显然这里不是一个Long类似的数值:
/**
* This class holds the address and the user ticket. The client connections
* to servers are uniquely identified by <remoteAddress, protocol, ticket>
* 连接的唯一标识，主要通过<远程地址，协议类型，用户组信息>
*/
static class ConnectionId {
//远程的socket地址
InetSocketAddress address;
//用户组信息
UserGroupInformation ticket;
//协议类型
Class<?> protocol;
private static final int PRIME = 16777619;
private int rpcTimeout;
private String serverPrincipal;
private int maxIdleTime; //connections will be culled if it was idle for
//maxIdleTime msecs
private int maxRetries; //the max. no. of retries for socket connections
private boolean tcpNoDelay; // if T then disable Nagle's Algorithm
private int pingInterval; // how often sends ping to the server in msecs
....这里用了3个属性组成唯一的标识属性，为了保证可以进行ID的复用，所以作者对ConnectionId的equal比较方法和hashCode 进行了重写:
/**
* 作者重写了equal比较方法，只要成员变量都想等也就想到了
*/
@Override
public boolean equals(Object obj) {
if (obj == this) {
return true;
}
if (obj instanceof ConnectionId) {
ConnectionId that = (ConnectionId) obj;
return isEqual(this.address, that.address)
&& this.maxIdleTime == that.maxIdleTime
&& this.maxRetries == that.maxRetries
&& this.pingInterval == that.pingInterval
&& isEqual(this.protocol, that.protocol)
&& this.rpcTimeout == that.rpcTimeout
&& isEqual(this.serverPrincipal, that.serverPrincipal)
&& this.tcpNoDelay == that.tcpNoDelay
&& isEqual(this.ticket, that.ticket);
}
return false;
}
/**
* 重写了hashCode的生成规则，保证不同的对象产生不同的hashCode值
*/
@Override
public int hashCode() {
int result = 1;
result = PRIME * result + ((address == null) ? 0 : address.hashCode());
result = PRIME * result + maxIdleTime;
result = PRIME * result + maxRetries;
result = PRIME * result + pingInterval;
result = PRIME * result + ((protocol == null) ? 0 : protocol.hashCode());
result = PRIME * rpcTimeout;
result = PRIME * result
+ ((serverPrincipal == null) ? 0 : serverPrincipal.hashCode());
result = PRIME * result + (tcpNoDelay ? 1231 : 1237);
result = PRIME * result + ((ticket == null) ? 0 : ticket.hashCode());
return result;
}这样就能保证对应同类型的连接就能够完全复用了，而不是仅仅凭借引用的关系判断对象是否相等，这里就是一个不错的设计了。
与连接Id对应的就是Connection了，它里面维护是一下的一些变量；
/** Thread that reads responses and notifies callers. Each connection owns a
* socket connected to a remote address. Calls are multiplexed through this
* socket: responses may be delivered out of order. */
private class Connection extends Thread {
//所连接的服务器地址
private InetSocketAddress server; // server ip:port
//服务端的krb5的名字，与安全方面相关
private String serverPrincipal; // server's krb5 principal name
//连接头部，内部包含了，所用的协议，客户端用户组信息以及验证的而方法
private ConnectionHeader header; // connection header
//远程连接ID
private final ConnectionId remoteId; // connection id
//连接验证方法
private AuthMethod authMethod; // authentication method
//下面3个变量都是安全方面的
private boolean useSasl;
private Token<? extends TokenIdentifier> token;
private SaslRpcClient saslRpcClient;
//下面是一组socket通信方面的变量
private Socket socket = null; // connected socket
private DataInputStream in;
private DataOutputStream out;
private int rpcTimeout;
private int maxIdleTime; //connections will be culled if it was idle for
//maxIdleTime msecs
private int maxRetries; //the max. no. of retries for socket connections
//tcpNoDelay可设置是否阻塞模式
private boolean tcpNoDelay; // if T then disable Nagle's Algorithm
private int pingInterval; // how often sends ping to the server in msecs
// currently active calls 当前活跃的回调，一个连接 可能会有很多个call回调
private Hashtable<Integer, Call> calls = new Hashtable<Integer, Call>();
//最后一次IO活动通信的时间
private AtomicLong lastActivity = new AtomicLong();// last I/O activity time
//连接关闭标记
private AtomicBoolean shouldCloseConnection = new AtomicBoolean(); // indicate if the connection is closed
private IOException closeException; // close reason
.....里面维护了大量的和连接通信相关的变量，在这里有一个很有意思的东西connectionHeader，连接头部，里面的数据时为了在通信最开始的时候被使用:
class ConnectionHeader implements Writable {
public static final Log LOG = LogFactory.getLog(ConnectionHeader.class);
//客户端和服务端通信的协议名称
private String protocol;
//客户端的用户组信息
private UserGroupInformation ugi = null;
//验证的方式，关系到写入数据的时的格式
private AuthMethod authMethod;
.....起到标识验证的作用。一个Client类的基本结构我们基本可以描绘出来了，下面是完整的类关系图:
在上面这幅图中，你肯定会发现我少了一个很关键的类了，就是Call回调类。Call回调在很多异步通信中是经常出现的。因为在通信过程中，当一个对象通过网络发送请求给另外一个对象的时候，如果采用同步的方式，会一直阻塞在那里，会带来非常不好的效率和体验的，所以很多时候，我们采用的是一种叫回调接口的方式。在这期间，用户可以继续做自己的事情。所以同样的Call这个概念当然也是适用在Hadoop RPC中。在Hadoop的RPC的核心调用原理， 简单的说，就是我把parame参数序列化到一个对象中，通过参数的形式把对象传入，进行RPC通信，最后服务端把处理好的结果值放入call对象，在返回给客户端，也就是说客户端和服务端都是通过Call对象进行操作，Call里面存着，请求的参数，和处理后的结构值2个变量。通过Call对象的封装，客户单实现了完美的无须知道细节的调用。下面是Call类的类按时
/** A call waiting for a value. */
//客户端的一个回调
private class Call {
//回调ID
int id; // call id
//被序列化的参数
Writable param; // parameter
//返回值
Writable value; // value, null if error
//出错时返回的异常
IOException error; // exception, null if value
//回调是否已经被完成
boolean done; // true when call is done
....看到这个Call回调类，也许你慢慢的会明白Hadoop RPC的一个基本原型了，这些Call当然是存在于某个连接中的，一个连接可能会发生多个回调，所以在Connection中维护了calls列表: private class Connection extends Thread {
....
// currently active calls 当前活跃的回调，一个连接 可能会有很多个call回调
private Hashtable<Integer, Call> calls = new Hashtable<Integer, Call>();作者在设计Call类的时候，比较聪明的考虑一种并发情况下的Call调用，所以为此设计了下面这个Call的子类，就是专门用于短时间内的瞬间Call调用:
/** Call implementation used for parallel calls. */
/** 继承自Call回调类，可以并行的使用，通过加了index下标做Call的区分 */
private class ParallelCall extends Call {
//每个ParallelCall并行的回调就会有对应的结果类
private ParallelResults results;
//index作为Call的区分
private int index;
....如果要查找值，就通过里面的ParallelCall查找，原理是根据index索引:
/** Result collector for parallel calls. */
private static class ParallelResults {
//并行结果类中拥有一组返回值，需要ParallelCall的index索引匹配
private Writable[] values;
//结果值的数量
private int size;
//values中已知的值的个数
private int count;
.....
/** Collect a result. */
public synchronized void callComplete(ParallelCall call) {
//将call中的值赋给result中
values[call.index] = call.value; // store the value
count++; // count it
//如果计数的值等到最终大小，通知caller
if (count == size) // if all values are in
notify(); // then notify waiting caller
}
}因为Call结构集是这些并发Call共有的，所以用的是static变量，都存在在了values数组中了，只有所有的并发Call都把值取出来了，才算回调成功，这个是个非常细小的辅助设计，这个在有些书籍上并没有多少提及。下面我们看看一般Call回调的流程，正如刚刚说的，最终客户端看到的形式就是，传入参数，获得结果，忽略内部一切逻辑，这是怎么做到的呢,答案在下面:
在执行之前，你会先得到ConnectionId：
public Writable call(Writable param, InetSocketAddress addr,
Class<?> protocol, UserGroupInformation ticket,
int rpcTimeout)
throws InterruptedException, IOException {
ConnectionId remoteId = ConnectionId.getConnectionId(addr, protocol,
ticket, rpcTimeout, conf);
return call(param, remoteId);
}接着才是主流程：
public Writable call(Writable param, ConnectionId remoteId)
throws InterruptedException, IOException {
//根据参数构造一个Call回调
Call call = new Call(param);
//根据远程ID获取连接
Connection connection = getConnection(remoteId, call);
//发送参数
connection.sendParam(call); // send the parameter
boolean interrupted = false;
synchronized (call) {
//如果call.done为false，就是Call还没完成
while (!call.done) {
try {
//等待远端程序的执行完毕
call.wait(); // wait for the result
} catch (InterruptedException ie) {
// save the fact that we were interrupted
interrupted = true;
}
}
//如果是异常中断，则终止当前线程
if (interrupted) {
// set the interrupt flag now that we are done waiting
Thread.currentThread().interrupt();
}
//如果call回到出错，则返回call出错信息
if (call.error != null) {
if (call.error instanceof RemoteException) {
call.error.fillInStackTrace();
throw call.error;
} else { // local exception
// use the connection because it will reflect an ip change, unlike
// the remoteId
throw wrapException(connection.getRemoteAddress(), call.error);
}
} else {
//如果是正常情况下，返回回调处理后的值
return call.value;
}
}
}在这上面的操作步骤中，重点关注2个函数，获取连接操作，看看人家是如何保证连接的复用性的:
private Connection getConnection(ConnectionId remoteId,
Call call)
throws IOException, InterruptedException {
.....
/* we could avoid this allocation for each RPC by having a
* connectionsId object and with set() method. We need to manage the
* refs for keys in HashMap properly. For now its ok.
*/
do {
synchronized (connections) {
//从connection连接池中获取连接，可以保证相同的连接ID可以复用
connection = connections.get(remoteId);
if (connection == null) {
connection = new Connection(remoteId);
connections.put(remoteId, connection);
}
}
} while (!connection.addCall(call));
有点单例模式的味道哦，还有一个方法叫sendParam发送参数方法:
public void sendParam(Call call) {
if (shouldCloseConnection.get()) {
return;
}
DataOutputBuffer d=null;
try {
synchronized (this.out) {
if (LOG.isDebugEnabled())
LOG.debug(getName() + " sending #" + call.id);
//for serializing the
//data to be written
//将call回调中的参数写入到输出流中，传向服务端
d = new DataOutputBuffer();
d.writeInt(call.id);
call.param.write(d);
byte[] data = d.getData();
int dataLength = d.getLength();
out.writeInt(dataLength); //first put the data length
out.write(data, 0, dataLength);//write the data
out.flush();
}
....代码只发送了Call的id，和请求参数，并没有把所有的Call的内容都扔出去了，一定是为了减少数据量的传输，这里还把数据的长度写入了，这是为了方便服务端准确的读取到不定长的数据。这服务端中间的处理操作不是今天讨论的重点。Call的执行过程就是这样。那么Call是如何被调用的呢，这又要重新回到了Client客户端上去了，Client有一个run()函数，所有的操作都是始于此的；
public void run() {
if (LOG.isDebugEnabled())
LOG.debug(getName() + ": starting, having connections "
+ connections.size());
//等待工作，等待请求调用
while (waitForWork()) {//wait here for work - read or close connection
//调用完请求，则立即获取回复
receiveResponse();
}
close();
if (LOG.isDebugEnabled())
LOG.debug(getName() + ": stopped, remaining connections "
+ connections.size());
}操作很简单，程序一直跑着，有请求，处理请求，获取请求，没有请求，就死等。
private synchronized boolean waitForWork() {
if (calls.isEmpty() && !shouldCloseConnection.get() && running.get()) {
long timeout = maxIdleTime-
(System.currentTimeMillis()-lastActivity.get());
if (timeout>0) {
try {
wait(timeout);
} catch (InterruptedException e) {}
}
}
....获取回复的操作如下:
/* Receive a response.
* Because only one receiver, so no synchronization on in.
* 获取回复值
*/
private void receiveResponse() {
if (shouldCloseConnection.get()) {
return;
}
//更新最近一次的call活动时间
touch();
try {
int id = in.readInt(); // try to read an id
if (LOG.isDebugEnabled())
LOG.debug(getName() + " got value #" + id);
//从获取call中取得相应的call
Call call = calls.get(id);
//判断该结果状态
int state = in.readInt(); // read call status
if (state == Status.SUCCESS.state) {
Writable value = ReflectionUtils.newInstance(valueClass, conf);
value.readFields(in); // read value
call.setValue(value);
calls.remove(id);
} else if (state == Status.ERROR.state) {
call.setException(new RemoteException(WritableUtils.readString(in),
WritableUtils.readString(in)));
calls.remove(id);
} else if (state == Status.FATAL.state) {
// Close the connection
markClosed(new RemoteException(WritableUtils.readString(in),
WritableUtils.readString(in)));
}
.....
} catch (IOException e) {
markClosed(e);
}
}从之前维护的Call列表中取出，做判断。Client本身的执行流程比较的简单:
Hadoop RPC客户端的通信模块的部分大致就是我上面的这个流程，中间其实还忽略了很多的细节，大家学习的时候，针对源码会有助于更好的理解，Hadoop RPC的服务端的实现更加复杂，所以建议采用分模块的学习或许会更好一点。
原文地址：http://blog.csdn.net/androidlushangderen/article/details/41751133
