HDFS-3601. Add BlockPlacementPolicyWithNodeGroup to support block placement with 4-layer network topology. Contributed by Junping Du

git-svn-id: https://svn.apache.org/repos/asf/hadoop/common/trunk@1357442 13f79535-47bb-0310-9956-ffa450edef68

Author: szetszwo

hadoop-common-project/hadoop-common/src/main/java/org/apache/hadoop/net/NetworkTopology.java

@@ -170,8 +170,7 @@ boolean add(Node n) {
         }
         if (parentNode == null) {
           // create a new InnerNode
-          parentNode = new InnerNode(parentName, getPath(this),
-                                     this, this.getLevel()+1);
+          parentNode = createParentNode(parentName);
           children.add(parentNode);
         }
         // add n to the subtree of the next ancestor node
@@ -288,7 +287,7 @@ Node getLeaf(int leafIndex, Node excludedNode) {
       // calculate the total number of excluded leaf nodes
       int numOfExcludedLeaves =
         isLeaf ? 1 : ((InnerNode)excludedNode).getNumOfLeaves();
-      if (isRack()) { // children are leaves
+      if (isLeafParent()) { // children are leaves
         if (isLeaf) { // excluded node is a leaf node
           int excludedIndex = children.indexOf(excludedNode);
           if (excludedIndex != -1 && leafIndex >= 0) {
@@ -326,6 +325,10 @@ Node getLeaf(int leafIndex, Node excludedNode) {
       }
     }
 
+    protected boolean isLeafParent() {
+      return isRack();
+    }
+
     /**
       * Determine if children a leaves, default implementation calls {@link #isRack()}
       * <p>To be overridden in subclasses for specific InnerNode implementations,
@@ -752,6 +755,30 @@ public String toString() {
     }
     return tree.toString();
   }
+  
+  /**
+   * Divide networklocation string into two parts by last separator, and get 
+   * the first part here.
+   * 
+   * @param networkLocation
+   * @return
+   */
+  public static String getFirstHalf(String networkLocation) {
+    int index = networkLocation.lastIndexOf(NodeBase.PATH_SEPARATOR_STR);
+    return networkLocation.substring(0, index);
+  }
+
+  /**
+   * Divide networklocation string into two parts by last separator, and get 
+   * the second part here.
+   * 
+   * @param networkLocation
+   * @return
+   */
+  public static String getLastHalf(String networkLocation) {
+    int index = networkLocation.lastIndexOf(NodeBase.PATH_SEPARATOR_STR);
+    return networkLocation.substring(index);
+  }
 
   /** swap two array items */
   static protected void swap(Node[] nodes, int i, int j) {

hadoop-common-project/hadoop-common/src/main/java/org/apache/hadoop/net/NetworkTopologyWithNodeGroup.java

@@ -49,7 +49,7 @@ protected Node getNodeForNetworkLocation(Node node) {
     }
     Node nodeGroup = getNode(node.getNetworkLocation());
     if (nodeGroup == null) {
-      nodeGroup = new InnerNode(node.getNetworkLocation());
+      nodeGroup = new InnerNodeWithNodeGroup(node.getNetworkLocation());
     }
     return getNode(nodeGroup.getNetworkLocation());
   }
@@ -383,6 +383,11 @@ boolean isNodeGroup() {
       }
       return true;
     }
+    
+    @Override
+    protected boolean isLeafParent() {
+      return isNodeGroup();
+    }
 
     @Override
     protected InnerNode createParentNode(String parentName) {

hadoop-hdfs-project/hadoop-hdfs/CHANGES.txt

@@ -13,6 +13,9 @@ Trunk (unreleased changes)
 
     HDFS-3125. Add JournalService to enable Journal Daemon. (suresh)
 
+    HDFS-3601. Add BlockPlacementPolicyWithNodeGroup to support block placement
+    with 4-layer network topology.  (Junping Du via szetszwo)
+
   IMPROVEMENTS
 
     HDFS-1620. Rename HdfsConstants -> HdfsServerConstants, FSConstants ->

hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/BlockPlacementPolicyDefault.java

@@ -148,6 +148,7 @@ DatanodeDescriptor[] chooseTarget(int numOfReplicas,
       new ArrayList<DatanodeDescriptor>(chosenNodes);
     for (Node node:chosenNodes) {
       excludedNodes.put(node, node);
+      adjustExcludedNodes(excludedNodes, node);
     }
 
     if (!clusterMap.contains(writer)) {
@@ -359,10 +360,11 @@ protected DatanodeDescriptor chooseRandom(
         (DatanodeDescriptor)(clusterMap.chooseRandom(nodes));
 
       Node oldNode = excludedNodes.put(chosenNode, chosenNode);
-      if (oldNode == null) { // choosendNode was not in the excluded list
+      if (oldNode == null) { // chosenNode was not in the excluded list
         numOfAvailableNodes--;
         if (isGoodTarget(chosenNode, blocksize, maxNodesPerRack, results)) {
           results.add(chosenNode);
+          adjustExcludedNodes(excludedNodes, chosenNode);
           return chosenNode;
         } else {
           badTarget = true;
@@ -409,6 +411,7 @@ protected void chooseRandom(int numOfReplicas,
         if (isGoodTarget(chosenNode, blocksize, maxNodesPerRack, results)) {
           numOfReplicas--;
           results.add(chosenNode);
+          adjustExcludedNodes(excludedNodes, chosenNode);
         } else {
           badTarget = true;
         }
@@ -426,7 +429,21 @@ protected void chooseRandom(int numOfReplicas,
       throw new NotEnoughReplicasException(detail);
     }
   }
-    
+  
+  /**
+   * After choosing a node to place replica, adjust excluded nodes accordingly.
+   * It should do nothing here as chosenNode is already put into exlcudeNodes, 
+   * but it can be overridden in subclass to put more related nodes into 
+   * excludedNodes.
+   * 
+   * @param excludedNodes
+   * @param chosenNode
+   */
+  protected void adjustExcludedNodes(HashMap<Node, Node> excludedNodes,
+      Node chosenNode) {
+    // do nothing here.
+  }
+
   /* judge if a node is a good target.
    * return true if <i>node</i> has enough space, 
    * does not have too much load, and the rack does not have too many nodes