YARN-10380: Import logic of multi-node allocation in CapacityScheduler (#2494)

Contributed by Qi Zhu.

hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/main/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/ClusterNodeTracker.java

@@ -496,4 +496,15 @@ public List<N> getNodesPerPartition(String partition) {
     }
     return nodesPerPartition;
   }
+
+  public List<String> getPartitions() {
+    List<String> partitions = null;
+    readLock.lock();
+    try {
+      partitions = new ArrayList(nodesPerLabel.keySet());
+    } finally {
+      readLock.unlock();
+    }
+    return partitions;
+  }
 }

hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/main/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/capacity/CapacityScheduler.java

@@ -531,6 +531,8 @@ private static boolean shouldSkipNodeSchedule(FiCaSchedulerNode node,
 
   /**
    * Schedule on all nodes by starting at a random point.
+   * Schedule on all partitions by starting at a random partition
+   * when multiNodePlacementEnabled is true.
    * @param cs
    */
   static void schedule(CapacityScheduler cs) throws InterruptedException{
@@ -544,6 +546,8 @@ static void schedule(CapacityScheduler cs) throws InterruptedException{
     if(nodeSize == 0) {
       return;
     }
+
+    if (!cs.multiNodePlacementEnabled) {
       int start = random.nextInt(nodeSize);
 
       // To avoid too verbose DEBUG logging, only print debug log once for
@@ -581,7 +585,40 @@ static void schedule(CapacityScheduler cs) throws InterruptedException{
       if (printSkipedNodeLogging) {
         printedVerboseLoggingForAsyncScheduling = true;
       }
+    } else {
+      // Get all partitions
+      List<String> partitions = cs.nodeTracker.getPartitions();
+      int partitionSize = partitions.size();
+      // First randomize the start point
+      int start = random.nextInt(partitionSize);
+      // Allocate containers of partition [start, end)
+      for (String partititon : partitions) {
+        if (current++ >= start) {
+          CandidateNodeSet<FiCaSchedulerNode> candidates =
+                  cs.getCandidateNodeSet(partititon);
+          if (candidates == null) {
+            continue;
+          }
+          cs.allocateContainersToNode(candidates, false);
+        }
+      }
 
+      current = 0;
+
+      // Allocate containers of partition [0, start)
+      for (String partititon : partitions) {
+        if (current++ > start) {
+          break;
+        }
+        CandidateNodeSet<FiCaSchedulerNode> candidates =
+                cs.getCandidateNodeSet(partititon);
+        if (candidates == null) {
+          continue;
+        }
+        cs.allocateContainersToNode(candidates, false);
+      }
+
+    }
     Thread.sleep(cs.getAsyncScheduleInterval());
   }
 
@@ -1485,6 +1522,23 @@ private boolean canAllocateMore(CSAssignment assignment, int offswitchCount,
             || assignedContainers < maxAssignPerHeartbeat);
   }
 
+  private CandidateNodeSet<FiCaSchedulerNode> getCandidateNodeSet(
+          String partition) {
+    CandidateNodeSet<FiCaSchedulerNode> candidates = null;
+    Map<NodeId, FiCaSchedulerNode> nodesByPartition = new HashMap<>();
+    List<FiCaSchedulerNode> nodes = nodeTracker
+            .getNodesPerPartition(partition);
+    if (nodes != null && !nodes.isEmpty()) {
+      //Filter for node heartbeat too long
+      nodes.stream()
+              .filter(node -> !shouldSkipNodeSchedule(node, this, true))
+              .forEach(n -> nodesByPartition.put(n.getNodeID(), n));
+      candidates = new SimpleCandidateNodeSet<FiCaSchedulerNode>(
+              nodesByPartition, partition);
+    }
+    return candidates;
+  }
+
   private CandidateNodeSet<FiCaSchedulerNode> getCandidateNodeSet(
           FiCaSchedulerNode node) {
     CandidateNodeSet<FiCaSchedulerNode> candidates = null;
@@ -1513,8 +1567,8 @@ private void allocateContainersToNode(NodeId nodeId,
       int offswitchCount = 0;
       int assignedContainers = 0;
 
-      CandidateNodeSet<FiCaSchedulerNode> candidates = getCandidateNodeSet(
+      CandidateNodeSet<FiCaSchedulerNode> candidates =
-          node);
+              getCandidateNodeSet(node);
       CSAssignment assignment = allocateContainersToNode(candidates,
           withNodeHeartbeat);
       // Only check if we can allocate more container on the same node when

hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/test/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/capacity/TestCapacitySchedulerAsyncScheduling.java

@@ -85,6 +85,10 @@ public class TestCapacitySchedulerAsyncScheduling {
 
   private NMHeartbeatThread nmHeartbeatThread = null;
 
+  private static final String POLICY_CLASS_NAME =
+          "org.apache.hadoop.yarn.server.resourcemanager.scheduler" +
+                  ".placement.ResourceUsageMultiNodeLookupPolicy";
+
   @Before
   public void setUp() throws Exception {
     conf = new YarnConfiguration();
@@ -111,6 +115,21 @@ public void testThreeThreadsAsyncContainerAllocation() throws Exception {
     testAsyncContainerAllocation(3);
   }
 
+  @Test(timeout = 300000)
+  public void testAsyncContainerAllocationWithMultiNode() throws Exception {
+    conf.set(CapacitySchedulerConfiguration.MULTI_NODE_SORTING_POLICIES,
+            "resource-based");
+    conf.set(CapacitySchedulerConfiguration.MULTI_NODE_SORTING_POLICY_NAME,
+            "resource-based");
+    String policyName =
+            CapacitySchedulerConfiguration.MULTI_NODE_SORTING_POLICY_NAME
+                    + ".resource-based" + ".class";
+    conf.set(policyName, POLICY_CLASS_NAME);
+    conf.setBoolean(CapacitySchedulerConfiguration.MULTI_NODE_PLACEMENT_ENABLED,
+            true);
+    testAsyncContainerAllocation(2);
+  }
+
   public void testAsyncContainerAllocation(int numThreads) throws Exception {
     conf.setInt(
         CapacitySchedulerConfiguration.SCHEDULE_ASYNCHRONOUSLY_MAXIMUM_THREAD,