YARN-7149. Cross-queue preemption sometimes starves an underserved queue. (Eric Payne via wangda)

Change-Id: Ib269991dbebce160378e8372ee6d24849c4a5ed6
(cherry picked from commit 3dfa937a1fadfc62947755872515f549b3b15e6a)

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

@@ -731,7 +731,9 @@ public class UsersManager implements AbstractUsersManager {
      * should be higher than queue-hard-limit * ulMin
      */
     float usersSummedByWeight = activeUsersTimesWeights;
-    Resource resourceUsed = totalResUsageForActiveUsers.getUsed(nodePartition);
+    Resource resourceUsed = Resources.add(
+                            totalResUsageForActiveUsers.getUsed(nodePartition),
+                            required);
 
     // For non-activeUser calculation, consider all users count.
     if (!activeUser) {

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

@@ -24,6 +24,7 @@ import java.util.List;
 import org.apache.commons.logging.Log;
 import org.apache.commons.logging.LogFactory;
 import org.apache.hadoop.conf.Configuration;
+import org.apache.hadoop.net.NetworkTopology;
 import org.apache.hadoop.security.SecurityUtilTestHelper;
 import org.apache.hadoop.yarn.api.protocolrecords.AllocateResponse;
 import org.apache.hadoop.yarn.api.records.Container;
@@ -887,4 +888,53 @@ public class TestContainerAllocation {
 
     rm1.close();
   }
+
+
+
+  @Test(timeout = 60000)
+  public void testUserLimitAllocationMultipleContainers() throws Exception {
+    CapacitySchedulerConfiguration newConf =
+        (CapacitySchedulerConfiguration) TestUtils
+            .getConfigurationWithMultipleQueues(conf);
+    newConf.setUserLimit("root.c", 50);
+    MockRM rm1 = new MockRM(newConf);
+
+    rm1.getRMContext().setNodeLabelManager(mgr);
+    rm1.start();
+    MockNM nm1 = rm1.registerNode("h1:1234", 1000 * GB);
+
+    // launch app from 1st user to queue C, AM container should be launched in nm1
+    RMApp app1 = rm1.submitApp(2 * GB, "app", "user1", null, "c");
+    MockAM am1 = MockRM.launchAndRegisterAM(app1, rm1, nm1);
+
+    // launch app from 2nd user to queue C, AM container should be launched in nm1
+    RMApp app2 = rm1.submitApp(2 * GB, "app", "user2", null, "c");
+    MockAM am2 = MockRM.launchAndRegisterAM(app2, rm1, nm1);
+
+    // Each application asks 1000 * 5GB containers
+    am1.allocate("*", 5 * GB, 1000, null);
+    am1.allocate("h1", 5 * GB, 1000, null);
+    am1.allocate(NetworkTopology.DEFAULT_RACK, 5 * GB, 1000, null);
+
+    // Each application asks 1000 * 5GB containers
+    am2.allocate("*", 5 * GB, 1000, null);
+    am2.allocate("h1", 5 * GB, 1000, null);
+    am2.allocate(NetworkTopology.DEFAULT_RACK, 5 * GB, 1000, null);
+
+    CapacityScheduler cs = (CapacityScheduler) rm1.getResourceScheduler();
+    RMNode rmNode1 = rm1.getRMContext().getRMNodes().get(nm1.getNodeId());
+
+    FiCaSchedulerApp schedulerApp1 =
+        cs.getApplicationAttempt(am1.getApplicationAttemptId());
+    FiCaSchedulerApp schedulerApp2 =
+        cs.getApplicationAttempt(am2.getApplicationAttemptId());
+
+    // container will be allocated to am1
+    // App1 will get 2 container allocated (plus AM container)
+    cs.handle(new NodeUpdateSchedulerEvent(rmNode1));
+    Assert.assertEquals(101, schedulerApp1.getLiveContainers().size());
+    Assert.assertEquals(100, schedulerApp2.getLiveContainers().size());
+
+    rm1.close();
+  }
 }

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

@@ -1252,7 +1252,7 @@ public class TestLeafQueue {
     //app4 is user 0
     //maxqueue 16G, userlimit 7G, used 8G, headroom 5G
     //(8G used is 6G from this test case - app4, 2 from last test case, app_1)
-    assertEquals(0*GB, app_4.getHeadroom().getMemorySize());
+    assertEquals(1*GB, app_4.getHeadroom().getMemorySize());
   }
 
   @Test
@@ -1436,7 +1436,7 @@ public class TestLeafQueue {
     assertEquals(2*GB, app_0.getCurrentConsumption().getMemorySize());
     assertEquals(0*GB, app_1.getCurrentConsumption().getMemorySize());
     // TODO, fix headroom in the future patch
-    assertEquals(0*GB, app_0.getHeadroom().getMemorySize());
+    assertEquals(1*GB, app_0.getHeadroom().getMemorySize());
       // User limit = 2G, 2 in use
     assertEquals(0*GB, app_1.getHeadroom().getMemorySize());
       // the application is not yet active
@@ -1449,8 +1449,8 @@ public class TestLeafQueue {
     assertEquals(3*GB, a.getUsedResources().getMemorySize());
     assertEquals(2*GB, app_0.getCurrentConsumption().getMemorySize());
     assertEquals(1*GB, app_1.getCurrentConsumption().getMemorySize());
-    assertEquals(0*GB, app_0.getHeadroom().getMemorySize()); // 4G - 3G
-    assertEquals(0*GB, app_1.getHeadroom().getMemorySize()); // 4G - 3G
+    assertEquals(1*GB, app_0.getHeadroom().getMemorySize()); // 4G - 3G
+    assertEquals(1*GB, app_1.getHeadroom().getMemorySize()); // 4G - 3G
     
     // Submit requests for app_1 and set max-cap
     a.setMaxCapacity(.1f);