hadoop 2.6 NetworkTopology源代码分析

Node接口定义了一个网络拓扑的一个结点。结点可能是一个代表数据结点的叶子结点，或者代表数据中心或者机架的中间结点。

/** The interface defines a node in a network topology.
 * A node may be a leave representing a data node or an inner
 * node representing a datacenter or rack.
 * Each data has a name and its location in the network is
 * decided by a string with syntax similar to a file name. 
 * For example, a data node's name is hostname:port# and if it's located at
 * rack "orange" in datacenter "dog", the string representation of its
 * network location is /dog/orange
 */

@InterfaceAudience.LimitedPrivate({"HDFS", "MapReduce"})
@InterfaceStability.Unstable
public interface Node {
  /** @return the string representation of this node's network location */
  public String getNetworkLocation();

  /** Set this node's network location
   * @param location the location
   */
  public void setNetworkLocation(String location);

  /** @return this node's name */
  public String getName();

  /** @return this node's parent */
  public Node getParent();

  /** Set this node's parent
   * @param parent the parent
   */
  public void setParent(Node parent);

  /** @return this node's level in the tree.
   * E.g. the root of a tree returns 0 and its children return 1
   */
  public int getLevel();

  /** Set this node's level in the tree
   * @param i the level
   */
  public void setLevel(int i);
}

NodeBase是Node的一个实现类，代码如下：

@InterfaceAudience.LimitedPrivate({"HDFS", "MapReduce"})
@InterfaceStability.Unstable
public class NodeBase implements Node {
  /** Path separator {@value} */
  public final static char PATH_SEPARATOR = '/';
  /** Path separator as a string {@value} */
  public final static String PATH_SEPARATOR_STR = "/";
  /** string representation of root {@value} */
  public final static String ROOT = "";
  
  protected String name; //host:port#
  protected String location; //string representation of this node's location
  protected int level; //which level of the tree the node resides
  protected Node parent; //its parent
  
  /** Default constructor */
  public NodeBase() {
  }
  
  /** Construct a node from its path
   * @param path 
   *   a concatenation of this node's location, the path seperator, and its name 
   */
  public NodeBase(String path) {
    path = normalize(path);
    int index = path.lastIndexOf(PATH_SEPARATOR);
    if (index== -1) {
      set(ROOT, path);
    } else {
      set(path.substring(index+1), path.substring(0, index));
    }
  }
  
  /** Construct a node from its name and its location
   * @param name this node's name (can be null, must not contain {@link #PATH_SEPARATOR})
   * @param location this node's location 
   */
  public NodeBase(String name, String location) {
    set(name, normalize(location));
  }
  
  /** Construct a node from its name and its location
   * @param name this node's name (can be null, must not contain {@link #PATH_SEPARATOR})
   * @param location this node's location 
   * @param parent this node's parent node
   * @param level this node's level in the tree
   */
  public NodeBase(String name, String location, Node parent, int level) {
    set(name, normalize(location));
    this.parent = parent;
    this.level = level;
  }

  /**
   * set this node's name and location
   * @param name the (nullable) name -which cannot contain the {@link #PATH_SEPARATOR}
   * @param location the location
   */
  private void set(String name, String location) {
    if (name != null && name.contains(PATH_SEPARATOR_STR))
      throw new IllegalArgumentException(
                                         "Network location name contains /: "+name);
    this.name = (name==null)?"":name;
    this.location = location;      
  }
  
  /** @return this node's name */
  @Override
  public String getName() { return name; }
  
  /** @return this node's network location */
  @Override
  public String getNetworkLocation() { return location; }
  
  /** Set this node's network location
   * @param location the location
   */
  @Override
  public void setNetworkLocation(String location) { this.location = location; }
  
  /**
   * Get the path of a node
   * @param node a non-null node
   * @return the path of a node
   */
  public static String getPath(Node node) {
    return node.getNetworkLocation() + PATH_SEPARATOR_STR + node.getName();
  }
  
  /** @return this node's path as its string representation */
  @Override
  public String toString() {
    return getPath(this);
  }

  /** Normalize a path by stripping off any trailing {@link #PATH_SEPARATOR}
   * @param path path to normalize.
   * @return the normalised path
   * If <i>path</i>is null or empty {@link #ROOT} is returned
   * @throws IllegalArgumentException if the first character of a non empty path
   * is not {@link #PATH_SEPARATOR}
   */
  public static String normalize(String path) {
    if (path == null || path.length() == 0) return ROOT;
    
    if (path.charAt(0) != PATH_SEPARATOR) {
      throw new IllegalArgumentException(
                                         "Network Location path does not start with "
                                         +PATH_SEPARATOR_STR+ ": "+path);
    }
    
    int len = path.length();
    if (path.charAt(len-1) == PATH_SEPARATOR) {
      return path.substring(0, len-1);
    }
    return path;
  }
  
  /** @return this node's parent */
  @Override
  public Node getParent() { return parent; }
  
  /** Set this node's parent
   * @param parent the parent
   */
  @Override
  public void setParent(Node parent) {
    this.parent = parent;
  }
  
  /** @return this node's level in the tree.
   * E.g. the root of a tree returns 0 and its children return 1
   */
  @Override
  public int getLevel() { return level; }
  
  /** Set this node's level in the tree
   * @param level the level
   */
  @Override
  public void setLevel(int level) {
    this.level = level;
  }
  
  public static int locationToDepth(String location) {
    String normalizedLocation = normalize(location);
    int length = normalizedLocation.length();
    int depth = 0;
    for (int i = 0; i < length; i++) {
      if (normalizedLocation.charAt(i) == PATH_SEPARATOR) {
        depth++;
      }
    }
    return depth;
  }
}

NetworkTopology的InnerNode扩展自NodeBase，代码如下：

有两个变量

protected List<Node> children=new ArrayList<Node>();
    private int numOfLeaves;

有以下方法比较重要：isRack方法判断此结点是否为一个rack结点。如果子结点为空或者子结点的类型不是InnerNode，则是rack结点。

 /** Judge if this node represents a rack 
     * @return true if it has no child or its children are not InnerNodes
     */ 
    boolean isRack() {
      if (children.isEmpty()) {
        return true;
      }
            
      Node firstChild = children.get(0);
      if (firstChild instanceof InnerNode) {
        return false;
      }
            
      return true;
    }

isAncestor方法判断当前结点是否为结点n的祖先。

    /** Judge if this node is an ancestor of node <i>n</i>
     *
     * @param n a node
     * @return true if this node is an ancestor of <i>n</i>
     */
    boolean isAncestor(Node n) {
      return getPath(this).equals(NodeBase.PATH_SEPARATOR_STR) ||
        (n.getNetworkLocation()+NodeBase.PATH_SEPARATOR_STR).
        startsWith(getPath(this)+NodeBase.PATH_SEPARATOR_STR);
    }

isAncestor方法判断当前结点是否为结点n的父结点。

 /** Judge if this node is the parent of node <i>n</i>
     * 
     * @param n a node
     * @return true if this node is the parent of <i>n</i>
     */
    boolean isParent(Node n) {
      return n.getNetworkLocation().equals(getPath(this));
    }

得到当前结点的子结点的名称，这个子结点是node n的祖先结点。

/* Return a child name of this node who is an ancestor of node <i>n</i> */
    private String getNextAncestorName(Node n) {
      if (!isAncestor(n)) {
        throw new IllegalArgumentException(
                                           this + "is not an ancestor of " + n);
      }
      String name = n.getNetworkLocation().substring(getPath(this).length());
      if (name.charAt(0) == PATH_SEPARATOR) {
        name = name.substring(1);
      }
      int index=name.indexOf(PATH_SEPARATOR);
      if (index !=-1)
        name = name.substring(0, index);
      return name;
    }

增加结点

/** Add node <i>n</i> to the subtree of this node 
     * @param n node to be added
     * @return true if the node is added; false otherwise
     */
    boolean add(Node n) {
      if (!isAncestor(n))
        throw new IllegalArgumentException(n.getName()+", which is located at "
                +n.getNetworkLocation()+", is not a decendent of "
                +getPath(this));
      if (isParent(n)) {
        // this node is the parent of n; add n directly
        n.setParent(this);
        n.setLevel(this.level+1);
        for(int i=0; i<children.size(); i++) {
          if (children.get(i).getName().equals(n.getName())) {
            children.set(i, n);
            return false;
          }
        }
        children.add(n);
        numOfLeaves++;
        return true;
      } else {
        // find the next ancestor node
        String parentName = getNextAncestorName(n);
        InnerNode parentNode = null;
        for(int i=0; i<children.size(); i++) {
          if (children.get(i).getName().equals(parentName)) {
            parentNode = (InnerNode)children.get(i);
            break;
          }
        }
        if (parentNode == null) {
          // create a new InnerNode
          parentNode = createParentNode(parentName);
          children.add(parentNode);
        }
        // add n to the subtree of the next ancestor node
        if (parentNode.add(n)) {
          numOfLeaves++;
          return true;
        } else {
          return false;
        }
      }
    }

其中createParentNode的代码如下：

    /**
     * Creates a parent node to be added to the list of children.  
     * Creates a node using the InnerNode four argument constructor specifying 
     * the name, location, parent, and level of this node.
     * 
     * <p>To be overridden in subclasses for specific InnerNode implementations,
     * as alternative to overriding the full {@link #add(Node)} method.
     * 
     * @param parentName The name of the parent node
     * @return A new inner node
     * @see InnerNode#InnerNode(String, String, InnerNode, int)
     */
    protected InnerNode createParentNode(String parentName) {
      return new InnerNode(parentName, getPath(this), this, this.getLevel()+1);
    }

remove一个结点的代码如下：

 /** Remove node <i>n</i> from the subtree of this node
     * @param n node to be deleted 
     * @return true if the node is deleted; false otherwise
     */
    boolean remove(Node n) {
      String parent = n.getNetworkLocation();
      String currentPath = getPath(this);
      if (!isAncestor(n))
        throw new IllegalArgumentException(n.getName()
                                           +", which is located at "
                                           +parent+", is not a descendent of "+currentPath);
      if (isParent(n)) {
        // this node is the parent of n; remove n directly
        for(int i=0; i<children.size(); i++) {
          if (children.get(i).getName().equals(n.getName())) {
            children.remove(i);
            numOfLeaves--;
            n.setParent(null);
            return true;
          }
        }
        return false;
      } else {
        // find the next ancestor node: the parent node
        String parentName = getNextAncestorName(n);
        InnerNode parentNode = null;
        int i;
        for(i=0; i<children.size(); i++) {
          if (children.get(i).getName().equals(parentName)) {
            parentNode = (InnerNode)children.get(i);
            break;
          }
        }
        if (parentNode==null) {
          return false;
        }
        // remove n from the parent node
        boolean isRemoved = parentNode.remove(n);
        // if the parent node has no children, remove the parent node too
        if (isRemoved) {
          if (parentNode.getNumOfChildren() == 0) {
            children.remove(i);
          }
          numOfLeaves--;
        }
        return isRemoved;
      }
    } // end of remove

getLoc 根据一个字符串，返回代表这个字符串的Node类型的引用。

/** Given a node's string representation, return a reference to the node
     * @param loc string location of the form /rack/node
     * @return null if the node is not found or the childnode is there but
     * not an instance of {@link InnerNode}
     */
    private Node getLoc(String loc) {
      if (loc == null || loc.length() == 0) return this;
            
      String[] path = loc.split(PATH_SEPARATOR_STR, 2);
      Node childnode = null;
      for(int i=0; i<children.size(); i++) {
        if (children.get(i).getName().equals(path[0])) {
          childnode = children.get(i);
        }
      }
      if (childnode == null) return null; // non-existing node
      if (path.length == 1) return childnode;
      if (childnode instanceof InnerNode) {
        return ((InnerNode)childnode).getLoc(path[1]);
      } else {
        return null;
      }
    }

getLeaf方法有点难度，作用是得到当前子结点去除excludedNode之外的第leftIndex个结点。

/** get <i>leafIndex</i> leaf of this subtree 
     * if it is not in the <i>excludedNode</i>
     *
     * @param leafIndex an indexed leaf of the node
     * @param excludedNode an excluded node (can be null)
     * @return
     */
    Node getLeaf(int leafIndex, Node excludedNode) {
      int count=0;
      // check if the excluded node a leaf
      boolean isLeaf =
        excludedNode == null || !(excludedNode instanceof InnerNode);
      // calculate the total number of excluded leaf nodes
      int numOfExcludedLeaves =
        isLeaf ? 1 : ((InnerNode)excludedNode).getNumOfLeaves();
      if (isLeafParent()) { // children are leaves
        if (isLeaf) { // excluded node is a leaf node
          int excludedIndex = children.indexOf(excludedNode);
          if (excludedIndex != -1 && leafIndex >= 0) {
            // excluded node is one of the children so adjust the leaf index
            leafIndex = leafIndex>=excludedIndex ? leafIndex+1 : leafIndex;
          }
        }
        // range check
        if (leafIndex<0 || leafIndex>=this.getNumOfChildren()) {
          return null;
        }
        return children.get(leafIndex);
      } else {
        for(int i=0; i<children.size(); i++) {
          InnerNode child = (InnerNode)children.get(i);
          if (excludedNode == null || excludedNode != child) {
            // not the excludedNode
            int numOfLeaves = child.getNumOfLeaves();
            if (excludedNode != null && child.isAncestor(excludedNode)) {
              numOfLeaves -= numOfExcludedLeaves;
            }
            if (count+numOfLeaves > leafIndex) {
              // the leaf is in the child subtree
              return child.getLeaf(leafIndex-count, excludedNode);
            } else {
              // go to the next child
              count = count+numOfLeaves;
            }
          } else { // it is the excluededNode
            // skip it and set the excludedNode to be null
            excludedNode = null;
          }
        }
        return null;
      }
    }

到此InnerNode方法完毕，现在介绍NetworkTopology的方法：

/** Add a leaf node
   * Update node counter & rack counter if necessary
   * @param node node to be added; can be null
   * @exception IllegalArgumentException if add a node to a leave
                                         or node to be added is not a leaf
   */
  public void add(Node node) {
    if (node==null) return;
    String oldTopoStr = this.toString();
    if( node instanceof InnerNode ) {
      throw new IllegalArgumentException(
        "Not allow to add an inner node: "+NodeBase.getPath(node));
    }
    int newDepth = NodeBase.locationToDepth(node.getNetworkLocation()) + 1;
    netlock.writeLock().lock();
    try {
      if ((depthOfAllLeaves != -1) && (depthOfAllLeaves != newDepth)) {
        LOG.error("Error: can't add leaf node " + NodeBase.getPath(node) +
            " at depth " + newDepth + " to topology:\n" + oldTopoStr);
        throw new InvalidTopologyException("Failed to add " + NodeBase.getPath(node) +
            ": You cannot have a rack and a non-rack node at the same " +
            "level of the network topology.");
      }
      Node rack = getNodeForNetworkLocation(node);
      if (rack != null && !(rack instanceof InnerNode)) {
        throw new IllegalArgumentException("Unexpected data node "
                                           + node.toString()
                                           + " at an illegal network location");
      }
      if (clusterMap.add(node)) {
        LOG.info("Adding a new node: "+NodeBase.getPath(node));
        if (rack == null) {
          numOfRacks++;
        }
        if (!(node instanceof InnerNode)) {
          if (depthOfAllLeaves == -1) {
            depthOfAllLeaves = node.getLevel();
          }
        }
      }
      if(LOG.isDebugEnabled()) {
        LOG.debug("NetworkTopology became:\n" + this.toString());
      }
    } finally {
      netlock.writeLock().unlock();
    }
  }

以下方法得到一个机架下的所有datanode.

/**
   * Given a string representation of a rack, return its children
   * @param loc a path-like string representation of a rack
   * @return a newly allocated list with all the node's children
   */
  public List<Node> getDatanodesInRack(String loc) {
    netlock.readLock().lock();
    try {
      loc = NodeBase.normalize(loc);
      if (!NodeBase.ROOT.equals(loc)) {
        loc = loc.substring(1);
      }
      InnerNode rack = (InnerNode) clusterMap.getLoc(loc);
      if (rack == null) {
        return null;
      }
      return new ArrayList<Node>(rack.getChildren());
    } finally {
      netlock.readLock().unlock();
    }
  }

remove一个结点，注意在clusterMap.remove(node)方法，可以把没有子结点的上级结点删除。

/** Remove a node
   * Update node counter and rack counter if necessary
   * @param node node to be removed; can be null
   */ 
  public void remove(Node node) {
    if (node==null) return;
    if( node instanceof InnerNode ) {
      throw new IllegalArgumentException(
        "Not allow to remove an inner node: "+NodeBase.getPath(node));
    }
    LOG.info("Removing a node: "+NodeBase.getPath(node));
    netlock.writeLock().lock();
    try {
      if (clusterMap.remove(node)) {
        InnerNode rack = (InnerNode)getNode(node.getNetworkLocation());
        if (rack == null) {
          numOfRacks--;
        }
      }
      if(LOG.isDebugEnabled()) {
        LOG.debug("NetworkTopology became:\n" + this.toString());
      }
    } finally {
      netlock.writeLock().unlock();
    }
  }

getDistance方法，返回两个结点的距离。

/** Return the distance between two nodes
   * It is assumed that the distance from one node to its parent is 1
   * The distance between two nodes is calculated by summing up their distances
   * to their closest common ancestor.
   * @param node1 one node
   * @param node2 another node
   * @return the distance between node1 and node2 which is zero if they are the same
   *  or {@link Integer#MAX_VALUE} if node1 or node2 do not belong to the cluster
   */
  public int getDistance(Node node1, Node node2) {
    if (node1 == node2) {
      return 0;
    }
    Node n1=node1, n2=node2;
    int dis = 0;
    netlock.readLock().lock();
    try {
      int level1=node1.getLevel(), level2=node2.getLevel();
      while(n1!=null && level1>level2) {
        n1 = n1.getParent();
        level1--;
        dis++;
      }
      while(n2!=null && level2>level1) {
        n2 = n2.getParent();
        level2--;
        dis++;
      }
      while(n1!=null && n2!=null && n1.getParent()!=n2.getParent()) {
        n1=n1.getParent();
        n2=n2.getParent();
        dis+=2;
      }
    } finally {
      netlock.readLock().unlock();
    }
    if (n1==null) {
      LOG.warn("The cluster does not contain node: "+NodeBase.getPath(node1));
      return Integer.MAX_VALUE;
    }
    if (n2==null) {
      LOG.warn("The cluster does not contain node: "+NodeBase.getPath(node2));
      return Integer.MAX_VALUE;
    }
    return dis+2;
  }

chooseRandom方法从scope范围内随机选取一个结点，如果scope 以 ~开头，那么代表除了scope之外，随机选取一个结点。

/** randomly choose one node from <i>scope</i>
   * if scope starts with ~, choose one from the all nodes except for the
   * ones in <i>scope</i>; otherwise, choose one from <i>scope</i>
   * @param scope range of nodes from which a node will be chosen
   * @return the chosen node
   */
  public Node chooseRandom(String scope) {
    netlock.readLock().lock();
    try {
      if (scope.startsWith("~")) {
        return chooseRandom(NodeBase.ROOT, scope.substring(1));
      } else {
        return chooseRandom(scope, null);
      }
    } finally {
      netlock.readLock().unlock();
    }
  }

  private Node chooseRandom(String scope, String excludedScope){
    if (excludedScope != null) {
      if (scope.startsWith(excludedScope)) {
        return null;
      }
      if (!excludedScope.startsWith(scope)) {
        excludedScope = null;
      }
    }
    Node node = getNode(scope);
    if (!(node instanceof InnerNode)) {
      return node;
    }
    InnerNode innerNode = (InnerNode)node;
    int numOfDatanodes = innerNode.getNumOfLeaves();
    if (excludedScope == null) {
      node = null;
    } else {
      node = getNode(excludedScope);
      if (!(node instanceof InnerNode)) {
        numOfDatanodes -= 1;
      } else {
        numOfDatanodes -= ((InnerNode)node).getNumOfLeaves();
      }
    }
    if (numOfDatanodes == 0) {
      throw new InvalidTopologyException(
          "Failed to find datanode (scope=\"" + String.valueOf(scope) +
          "\" excludedScope=\"" + String.valueOf(excludedScope) + "\").");
    }
    int leaveIndex = r.nextInt(numOfDatanodes);
    return innerNode.getLeaf(leaveIndex, node);
  }

方法countNumOfAvailableNodes方法返回在scope范围内，但不是excludedNodes列表中。如果scope以~开头，那么返回不在scope范围的结点的数量。

/** return the number of leaves in <i>scope</i> but not in <i>excludedNodes</i>
   * if scope starts with ~, return the number of nodes that are not
   * in <i>scope</i> and <i>excludedNodes</i>;
   * @param scope a path string that may start with ~
   * @param excludedNodes a list of nodes
   * @return number of available nodes
   */
  public int countNumOfAvailableNodes(String scope,
                                      Collection<Node> excludedNodes) {
    boolean isExcluded=false;
    if (scope.startsWith("~")) {
      isExcluded=true;
      scope=scope.substring(1);
    }
    scope = NodeBase.normalize(scope);
    int excludedCountInScope = 0; // the number of nodes in both scope & excludedNodes
    int excludedCountOffScope = 0; // the number of nodes outside scope & excludedNodes
    netlock.readLock().lock();
    try {
      for (Node node : excludedNodes) {
        node = getNode(NodeBase.getPath(node));
        if (node == null) {
          continue;
        }
        if ((NodeBase.getPath(node) + NodeBase.PATH_SEPARATOR_STR)
            .startsWith(scope + NodeBase.PATH_SEPARATOR_STR)) {
          excludedCountInScope++;
        } else {
          excludedCountOffScope++;
        }
      }
      Node n = getNode(scope);
      int scopeNodeCount = 0;
      if (n != null) {
        scopeNodeCount++;
      }
      if (n instanceof InnerNode) {
        scopeNodeCount=((InnerNode)n).getNumOfLeaves();
      }
      if (isExcluded) {
        return clusterMap.getNumOfLeaves() - scopeNodeCount
            - excludedCountOffScope;
      } else {
        return scopeNodeCount - excludedCountInScope;
      }
    } finally {
      netlock.readLock().unlock();
    }
  }

getWeight方法返回结点node和结点reader之间的距离，返回的值越低，代表两个结点的距离越近。

/**
   * Returns an integer weight which specifies how far away {node} is away from
   * {reader}. A lower value signifies that a node is closer.
   * 
   * @param reader Node where data will be read
   * @param node Replica of data
   * @return weight
   */
  protected int getWeight(Node reader, Node node) {
    // 0 is local, 1 is same rack, 2 is off rack
    // Start off by initializing to off rack
    int weight = 2;
    if (reader != null) {
      if (reader == node) {
        weight = 0;
      } else if (isOnSameRack(reader, node)) {
        weight = 1;
      }
    }
    return weight;
  }

sortByDistance方法，按和reader的距离，对nodes进行排序。在三层架构中，对于reader，结点node要么是本机，要么在同一个机架，要么在不同机架。

/**
   * Sort nodes array by network distance to <i>reader</i>.
   * <p/>
   * In a three-level topology, a node can be either local, on the same rack,
   * or on a different rack from the reader. Sorting the nodes based on network
   * distance from the reader reduces network traffic and improves
   * performance.
   * <p/>
   * As an additional twist, we also randomize the nodes at each network
   * distance. This helps with load balancing when there is data skew.
   *
   * @param reader    Node where data will be read
   * @param nodes     Available replicas with the requested data
   * @param activeLen Number of active nodes at the front of the array
   */
  public void sortByDistance(Node reader, Node[] nodes, int activeLen) {
    /** Sort weights for the nodes array */
    int[] weights = new int[activeLen];
    for (int i=0; i<activeLen; i++) {
      weights[i] = getWeight(reader, nodes[i]);
    }
    // Add weight/node pairs to a TreeMap to sort
    TreeMap<Integer, List<Node>> tree = new TreeMap<Integer, List<Node>>();
    for (int i=0; i<activeLen; i++) {
      int weight = weights[i];
      Node node = nodes[i];
      List<Node> list = tree.get(weight);
      if (list == null) {
        list = Lists.newArrayListWithExpectedSize(1);
        tree.put(weight, list);
      }
      list.add(node);
    }

    int idx = 0;
    for (List<Node> list: tree.values()) {
      if (list != null) {
        Collections.shuffle(list, r);
        for (Node n: list) {
          nodes[idx] = n;
          idx++;
        }
      }
    }
    Preconditions.checkState(idx == activeLen,
        "Sorted the wrong number of nodes!");
  }