kangfoo's blog

工作学习笔记，生活掠影。

Hadoop RPC

| 评论

Remote Procedure Call 远程方法调用。不需要了解网络细节，某一程序即可使用该协议请求来自网络内另一台及其程序的服务。它是一个 Client/Server 的结构,提供服务的一方称为Server，消费服务的一方称为Client。

Hadoop 底层的交互都是通过 rpc 进行的。例 如：datanode 和 namenode、tasktracker 和 jobtracker、secondary namenode 和 namenode 之间的通信都是通过 rpc 实现的。

TODO: 此文未写明了。明显需要画 4张图， rpc 原理图，Hadoop rpc 时序图， 客户端 流程图，服端流程图。最好帖几个包图＋ 类图（组件图）。待完善。

要实现远程过程调用，需要有3要素： 1、server 必须发布服务 2、在 client 和 server 两端都需要有模块来处理协议和连接 3、server 发布的服务，需要将接口给到 client

Hadoop RPC

  1. 序列化层。 Client 与 Server 端通讯传递的信息采用实现自 Writable 类型
  2. 函数调用层。 Hadoop RPC 通过动态代理和 java 反射实现函数调用
  3. 网络传输层。Hadoop RPC 采用 TCP/IP socket 机制
  4. 服务器框架层。Hadoop RPC 采用 java NIO 事件驱动模型提高 RPC Server 吞吐量

TODO 缺个 RPC 图

Hadoop RPC 源代码主要在org.apache.hadoop.ipc包下。org.apache.hadoop.ipc.RPC 内部包含5个内部类。

  • Invocation ：用于封装方法名和参数，作为数据传输层，相当于VO（Value Object）。
  • ClientCache ：用于存储client对象，用 socket factory 作为 hash key,存储结构为 hashMap 。
  • Invoker ：是动态代理中的调用实现类，继承了 java.lang.reflect.InvocationHandler。
  • Server ：是ipc.Server的实现类。
  • VersionMismatch : 协议版本。

从客户端开始进行通讯源代码分析

org.apache.hadoop.ipc.Client 有5个内部类

  • Call: A call waiting for a value.
  • Connection: 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.
  • ConnectionId: This class holds the address and the user ticket. The client connections to servers are uniquely identified by
  • ParallelCall: Call implementation used for parallel calls.
  • ParallelResults: Result collector for parallel calls.

客户端和服务端建立连接的大致执行过程为：

  1. 在 Object org.apache.hadoop.ipc.RPC.Invoker.invoke(Object proxy, Method method, Object[] args) 方法中调用
    client.call(new Invocation(method, args), remoteId);

  2. 上述的 new Invocation(method, args) 是 org.apache.hadoop.ipc.RPC 的内部类，它包含被调用的方法名称及其参数。此处主要是设置方法和参数。 client 为 org.apache.hadoop.ipc.Client 的实例对象。

  3. org.apache.hadoop.ipc.Client.call() 方法的具体源代码。在call()方法中 getConnection()内部获取一个 org.apache.hadoop.ipc.Client.Connection 对象并启动 io 流 setupIOstreams()。

    Writable org.apache.hadoop.ipc.Client.call(Writable param, ConnectionId remoteId) throwsInterruptedException, IOException {
Call call = new Call(param); //A call waiting for a value.   
// Get a connection from the pool, or create a new one and add it to the
// pool.  Connections to a given ConnectionId are reused. 
Connection connection = getConnection(remoteId, call);// 主要在 org.apache.hadoop.net 包下。
connection.sendParam(call); //客户端发送数据过程
boolean interrupted = false;
synchronized (call) {
   while (!call.done) {
    try {
      call.wait();                           // wait for the result
    } catch (InterruptedException ie) {
      // save the fact that we were interrupted
      interrupted = true;
    }
  }
… …
}
}
// Get a connection from the pool, or create a new one and add it to the
// pool.  Connections to a given ConnectionId are reused. 
private Connection getConnection(ConnectionId remoteId,
                               Call call)
                               throws IOException, InterruptedException {
if (!running.get()) {
  // the client is stopped
  throw new IOException("The client is stopped");
}
Connection connection;
// 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) {
    connection = connections.get(remoteId);
    if (connection == null) {
      connection = new Connection(remoteId);
      connections.put(remoteId, connection);
    }
  }
} while (!connection.addCall(call)); 
//we don't invoke the method below inside "synchronized (connections)"
//block above. The reason for that is if the server happens to be slow,
//it will take longer to establish a connection and that will slow the
//entire system down.
connection.setupIOstreams(); // 向服务段发送一个 header 并等待结果
return connection;
}

  4. setupIOstreams() 方法。

    void org.apache.hadoop.ipc.Client.Connection.setupIOstreams() throws InterruptedException {
// Connect to the server and set up the I/O streams. It then sends
// a header to the server and starts
// the connection thread that waits for responses.
while (true) {
      setupConnection();//  建立连接
      InputStream inStream = NetUtils.getInputStream(socket); // 输入
      OutputStream outStream = NetUtils.getOutputStream(socket); // 输出
      writeRpcHeader(outStream);
      }
… … 
// update last activity time
  touch();
// start the receiver thread after the socket connection has been set up            start(); 
}

  5. 启动org.apache.hadoop.ipc.Client.Connection 客户端获取服务器端放回数据过程

    void org.apache.hadoop.ipc.Client.Connection.run()
while (waitForWork()) {//wait here for work - read or close connection
    receiveResponse();
  }

ipc.Server源码分析

ipc.Server 有6个内部类：

  • Call ：用于存储客户端发来的请求
  • Listener ： 监听类，用于监听客户端发来的请求，同时Listener内部还有一个静态类，Listener.Reader，当监听器监听到用户请求，便让Reader读取用户请求。
  • ExceptionsHandler: 异常管理
  • Responder ：响应RPC请求类，请求处理完毕，由Responder发送给请求客户端。
  • Connection ：连接类，真正的客户端请求读取逻辑在这个类中。
  • Handler ：请求处理类，会循环阻塞读取callQueue中的call对象，并对其进行操作。

大致过程为：

  1. Namenode的初始化时，RPC的server对象是通过ipc.RPC类的getServer()方法获得的。

    void org.apache.hadoop.hdfs.server.namenode.NameNode.initialize(Configuration conf) throwsIOException
// create rpc server
InetSocketAddress dnSocketAddr = getServiceRpcServerAddress(conf);
if (dnSocketAddr != null) {
  int serviceHandlerCount =
    conf.getInt(DFSConfigKeys.DFS_NAMENODE_SERVICE_HANDLER_COUNT_KEY,
                DFSConfigKeys.DFS_NAMENODE_SERVICE_HANDLER_COUNT_DEFAULT);
  this.serviceRpcServer = RPC.getServer(this, dnSocketAddr.getHostName(), 
      dnSocketAddr.getPort(), serviceHandlerCount,
      false, conf, namesystem.getDelegationTokenSecretManager());
  this.serviceRPCAddress = this.serviceRpcServer.getListenerAddress();
  setRpcServiceServerAddress(conf);
}
… …
this.server.start();  //start RPC server

  2. 启动 server

    void org.apache.hadoop.ipc.Server.start()
// Starts the service.  Must be called before any calls will be handled.
public synchronized void start() {
responder.start();
listener.start();
handlers = new Handler[handlerCount];
for (int i = 0; i < handlerCount; i++) {
  handlers[i] = new Handler(i);
  handlers[i].start(); //处理call
}
}

  3. Server处理请求, server 同样使用非阻塞 nio 以提高吞吐量

    org.apache.hadoop.ipc.Server.Listener.Listener(Server) throws IOException
public Listener() throws IOException {
  address = new InetSocketAddress(bindAddress, port);
  // Create a new server socket and set to non blocking mode
  acceptChannel = ServerSocketChannel.open();
  acceptChannel.configureBlocking(false);
… … }

  4. 真正建立连接

    void org.apache.hadoop.ipc.Server.Listener.doAccept(SelectionKey key) throws IOException,OutOfMemoryError

    Reader 读数据接收请求

    void org.apache.hadoop.ipc.Server.Listener.doRead(SelectionKey key) throws InterruptedException
try {
    count = c.readAndProcess();
  } catch (InterruptedException ieo) {
    LOG.info(getName() + ": readAndProcess caught InterruptedException", ieo);
    throw ieo;
  }

    int org.apache.hadoop.ipc.Server.Connection.readAndProcess() throws IOException,InterruptedException
if (!rpcHeaderRead) {
      //Every connection is expected to send the header.
      if (rpcHeaderBuffer == null) {
        rpcHeaderBuffer = ByteBuffer.allocate(2);
      }
      count = channelRead(channel, rpcHeaderBuffer);
      if (count < 0 || rpcHeaderBuffer.remaining() > 0) {
        return count;
      }
      int version = rpcHeaderBuffer.get(0);
… … 
processOneRpc(data.array()); // 数据处理

  5. 下面贴出Server.Connection类中的processOneRpc()方法和processData()方法的源码。

    void org.apache.hadoop.ipc.Server.Connection.processOneRpc(byte[] buf) throws IOException,InterruptedException
private void processOneRpc(byte[] buf) throws IOException,
    InterruptedException {
  if (headerRead) {
    processData(buf);
  } else {
    processHeader(buf);
    headerRead = true;
    if (!authorizeConnection()) {
      throw new AccessControlException("Connection from " + this
          + " for protocol " + header.getProtocol()
          + " is unauthorized for user " + user);
    }
  }
}

  6. 处理call

    void org.apache.hadoop.ipc.Server.Handler.run()
while (running) {
    try {
      final Call call = callQueue.take(); // pop the queue; maybe blocked here
      … … 
      CurCall.set(call);
      try {
        // Make the call as the user via Subject.doAs, thus associating
        // the call with the Subject
        if (call.connection.user == null) {
          value = call(call.connection.protocol, call.param, 
                       call.timestamp);
        } else {
… …}

  7. 返回请求

下面贴出Server.Responder类中的doRespond()方法源码：

void org.apache.hadoop.ipc.Server.Responder.doRespond(Call call) throws IOException
    //
    // Enqueue a response from the application.
    //
    void doRespond(Call call) throws IOException {
      synchronized (call.connection.responseQueue) {
        call.connection.responseQueue.addLast(call);
        if (call.connection.responseQueue.size() == 1) {
          processResponse(call.connection.responseQueue, true);
        }
      }
    }

补充： notify()让因wait()进入阻塞队列里的线程（blocked状态）变为runnable，然后发出notify()动作的线程继续执行完，待其完成后，进行调度时，调用wait()的线程可能会被再次调度而进入running状态。

参考资源：

评论