YARN-6775. CapacityScheduler: Improvements to assignContainers, avoid unnecessary canAssignToUser/Queue calls. (Nathan Roberts via wangda)

Change-Id: Iaf9bb7e5ed3aa1300abdccf87ae6fcbddbd25e3e

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

@@ -63,9 +63,14 @@ public class ActivitiesLogger {
         SchedulerApplicationAttempt application, Priority priority,
         String diagnostic) {
       String type = "app";
-      recordActivity(activitiesManager, node, application.getQueueName(),
-          application.getApplicationId().toString(), priority,
-          ActivityState.REJECTED, diagnostic, type);
+      if (activitiesManager == null) {
+        return;
+      }
+      if (activitiesManager.shouldRecordThisNode(node.getNodeID())) {
+        recordActivity(activitiesManager, node, application.getQueueName(),
+            application.getApplicationId().toString(), priority,
+            ActivityState.REJECTED, diagnostic, type);
+      }
       finishSkippedAppAllocationRecording(activitiesManager,
           application.getApplicationId(), ActivityState.REJECTED, diagnostic);
     }
@@ -203,8 +208,13 @@ public class ActivitiesLogger {
     public static void recordQueueActivity(ActivitiesManager activitiesManager,
         SchedulerNode node, String parentQueueName, String queueName,
         ActivityState state, String diagnostic) {
-      recordActivity(activitiesManager, node, parentQueueName, queueName, null,
-          state, diagnostic, null);
+      if (activitiesManager == null) {
+        return;
+      }
+      if (activitiesManager.shouldRecordThisNode(node.getNodeID())) {
+        recordActivity(activitiesManager, node, parentQueueName, queueName,
+            null, state, diagnostic, null);
+      }
     }
   }
 
@@ -266,13 +276,10 @@ public class ActivitiesLogger {
   private static void recordActivity(ActivitiesManager activitiesManager,
       SchedulerNode node, String parentName, String childName,
       Priority priority, ActivityState state, String diagnostic, String type) {
-    if (activitiesManager == null) {
-      return;
-    }
-    if (activitiesManager.shouldRecordThisNode(node.getNodeID())) {
-      activitiesManager.addSchedulingActivityForNode(node.getNodeID(),
-          parentName, childName, priority != null ? priority.toString() : null,
-          state, diagnostic, type);
-    }
+
+    activitiesManager.addSchedulingActivityForNode(node.getNodeID(), parentName,
+        childName, priority != null ? priority.toString() : null, state,
+        diagnostic, type);
+
   }
 }

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

@@ -1039,6 +1039,8 @@ public class LeafQueue extends AbstractCSQueue {
       return CSAssignment.NULL_ASSIGNMENT;
     }
 
+    Map<String, CachedUserLimit> userLimits = new HashMap<>();
+    boolean needAssignToQueueCheck = true;
     for (Iterator<FiCaSchedulerApp> assignmentIterator =
          orderingPolicy.getAssignmentIterator();
          assignmentIterator.hasNext(); ) {
@@ -1048,24 +1050,50 @@ public class LeafQueue extends AbstractCSQueue {
           node.getNodeID(), SystemClock.getInstance().getTime(), application);
 
       // Check queue max-capacity limit
-      if (!super.canAssignToThisQueue(clusterResource, ps.getPartition(),
-          currentResourceLimits, application.getCurrentReservation(),
-          schedulingMode)) {
-        ActivitiesLogger.APP.recordRejectedAppActivityFromLeafQueue(
-            activitiesManager, node, application, application.getPriority(),
-            ActivityDiagnosticConstant.QUEUE_MAX_CAPACITY_LIMIT);
-        ActivitiesLogger.QUEUE.recordQueueActivity(activitiesManager, node,
-            getParent().getQueueName(), getQueueName(), ActivityState.SKIPPED,
-            ActivityDiagnosticConstant.EMPTY);
-        return CSAssignment.NULL_ASSIGNMENT;
+      Resource appReserved = application.getCurrentReservation();
+      if (needAssignToQueueCheck) {
+        if (!super.canAssignToThisQueue(clusterResource, node.getPartition(),
+            currentResourceLimits, appReserved, schedulingMode)) {
+          ActivitiesLogger.APP.recordRejectedAppActivityFromLeafQueue(
+              activitiesManager, node, application, application.getPriority(),
+              ActivityDiagnosticConstant.QUEUE_MAX_CAPACITY_LIMIT);
+          ActivitiesLogger.QUEUE.recordQueueActivity(activitiesManager, node,
+              getParent().getQueueName(), getQueueName(), ActivityState.SKIPPED,
+              ActivityDiagnosticConstant.EMPTY);
+          return CSAssignment.NULL_ASSIGNMENT;
+        }
+        // If there was no reservation and canAssignToThisQueue returned
+        // true, there is no reason to check further.
+        if (!this.reservationsContinueLooking
+            || appReserved.equals(Resources.none())) {
+          needAssignToQueueCheck = false;
+        }
       }
 
+      CachedUserLimit cul = userLimits.get(application.getUser());
+      Resource cachedUserLimit = null;
+      if (cul != null) {
+        cachedUserLimit = cul.userLimit;
+      }
       Resource userLimit = computeUserLimitAndSetHeadroom(application,
-          clusterResource, ps.getPartition(), schedulingMode);
-
+          clusterResource, ps.getPartition(), schedulingMode, cachedUserLimit);
+      if (cul == null) {
+        cul = new CachedUserLimit(userLimit);
+        userLimits.put(application.getUser(), cul);
+      }
       // Check user limit
-      if (!canAssignToUser(clusterResource, application.getUser(), userLimit,
-          application, ps.getPartition(), currentResourceLimits)) {
+      boolean userAssignable = true;
+      if (!cul.canAssign && Resources.fitsIn(appReserved, cul.reservation)) {
+        userAssignable = false;
+      } else {
+        userAssignable = canAssignToUser(clusterResource, application.getUser(),
+            userLimit, application, node.getPartition(), currentResourceLimits);
+        if (!userAssignable && Resources.fitsIn(cul.reservation, appReserved)) {
+          cul.canAssign = false;
+          cul.reservation = appReserved;
+        }
+      }
+      if (!userAssignable) {
         application.updateAMContainerDiagnostics(AMState.ACTIVATED,
             "User capacity has reached its maximum limit.");
         ActivitiesLogger.APP.recordRejectedAppActivityFromLeafQueue(
@@ -1140,7 +1168,7 @@ public class LeafQueue extends AbstractCSQueue {
 
         // check user-limit
         Resource userLimit = computeUserLimitAndSetHeadroom(app, cluster, p,
-            allocation.getSchedulingMode());
+            allocation.getSchedulingMode(), null);
 
         // Deduct resources that we can release
         Resource usedResource = Resources.clone(getUser(username).getUsed(p));
@@ -1345,19 +1373,20 @@ public class LeafQueue extends AbstractCSQueue {
   @Lock({LeafQueue.class})
   Resource computeUserLimitAndSetHeadroom(FiCaSchedulerApp application,
       Resource clusterResource, String nodePartition,
-      SchedulingMode schedulingMode) {
+      SchedulingMode schedulingMode, Resource userLimit) {
     String user = application.getUser();
     User queueUser = getUser(user);
 
     // Compute user limit respect requested labels,
     // TODO, need consider headroom respect labels also
-    Resource userLimit =
-        getResourceLimitForActiveUsers(application.getUser(), clusterResource,
-            nodePartition, schedulingMode);
-
+    if (userLimit == null) {
+      userLimit = getResourceLimitForActiveUsers(application.getUser(),
+          clusterResource, nodePartition, schedulingMode);
+    }
     setQueueResourceLimitsInfo(clusterResource);
 
     Resource headroom =
+        metrics.getUserMetrics(user) == null ? Resources.none() :
         getHeadroom(queueUser, cachedResourceLimitsForHeadroom.getLimit(),
             clusterResource, userLimit, nodePartition);
     
@@ -1365,7 +1394,7 @@ public class LeafQueue extends AbstractCSQueue {
       LOG.debug("Headroom calculation for user " + user + ": " + " userLimit="
           + userLimit + " queueMaxAvailRes="
           + cachedResourceLimitsForHeadroom.getLimit() + " consumed="
-          + queueUser.getUsed() + " headroom=" + headroom + " partition="
+          + queueUser.getUsed() + " partition="
           + nodePartition);
     }
     
@@ -1726,7 +1755,7 @@ public class LeafQueue extends AbstractCSQueue {
           .getSchedulableEntities()) {
         computeUserLimitAndSetHeadroom(application, clusterResource,
             RMNodeLabelsManager.NO_LABEL,
-            SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
+            SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY, null);
       }
     } finally {
       writeLock.unlock();
@@ -2065,4 +2094,14 @@ public class LeafQueue extends AbstractCSQueue {
   public Set<String> getAllUsers() {
     return this.getUsersManager().getUsers().keySet();
   }
+
+  static class CachedUserLimit {
+    final Resource userLimit;
+    boolean canAssign = true;
+    Resource reservation = Resources.none();
+
+    CachedUserLimit(Resource userLimit) {
+      this.userLimit = userLimit;
+    }
+  }
 }

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

@@ -31,6 +31,7 @@ import org.apache.hadoop.security.UserGroupInformation;
 import org.apache.hadoop.security.token.Token;
 import org.apache.hadoop.security.token.TokenIdentifier;
 import org.apache.hadoop.test.GenericTestUtils;
+import org.apache.hadoop.util.Time;
 import org.apache.hadoop.yarn.LocalConfigurationProvider;
 import org.apache.hadoop.yarn.api.ApplicationMasterProtocol;
 import org.apache.hadoop.yarn.api.protocolrecords.AllocateRequest;
@@ -135,8 +136,12 @@ import org.apache.hadoop.yarn.server.utils.BuilderUtils;
 import org.apache.hadoop.yarn.util.resource.DefaultResourceCalculator;
 import org.apache.hadoop.yarn.util.resource.DominantResourceCalculator;
 import org.apache.hadoop.yarn.util.resource.Resources;
+import org.apache.log4j.Level;
+import org.apache.log4j.LogManager;
+import org.apache.log4j.Logger;
 import org.junit.After;
 import org.junit.Assert;
+import org.junit.Assume;
 import org.junit.Before;
 import org.junit.Test;
 import org.mockito.Mockito;
@@ -148,9 +153,11 @@ import java.util.ArrayList;
 import java.util.Arrays;
 import java.util.Collection;
 import java.util.Collections;
+import java.util.Enumeration;
 import java.util.HashMap;
 import java.util.List;
 import java.util.Map;
+import java.util.PriorityQueue;
 import java.util.Set;
 import java.util.concurrent.BrokenBarrierException;
 import java.util.concurrent.CyclicBarrier;
@@ -3491,6 +3498,7 @@ public class TestCapacityScheduler {
     rm.stop();
   }
 
+
   @Test
   public void testHeadRoomCalculationWithDRC() throws Exception {
     // test with total cluster resource of 20GB memory and 20 vcores.
@@ -4073,6 +4081,143 @@ public class TestCapacityScheduler {
     rm.stop();
   }
 
+  @Test (timeout = 300000)
+  public void testUserLimitThroughput() throws Exception {
+    // Since this is more of a performance unit test, only run if
+    // RunUserLimitThroughput is set (-DRunUserLimitThroughput=true)
+    Assume.assumeTrue(Boolean.valueOf(
+        System.getProperty("RunUserLimitThroughput")));
+
+    CapacitySchedulerConfiguration csconf =
+        new CapacitySchedulerConfiguration();
+    csconf.setMaximumApplicationMasterResourcePerQueuePercent("root", 100.0f);
+    csconf.setMaximumAMResourcePercentPerPartition("root", "", 100.0f);
+    csconf.setMaximumApplicationMasterResourcePerQueuePercent("root.default",
+        100.0f);
+    csconf.setMaximumAMResourcePercentPerPartition("root.default", "", 100.0f);
+    csconf.setResourceComparator(DominantResourceCalculator.class);
+
+    YarnConfiguration conf = new YarnConfiguration(csconf);
+      conf.setClass(YarnConfiguration.RM_SCHEDULER, CapacityScheduler.class,
+          ResourceScheduler.class);
+
+    MockRM rm = new MockRM(conf);
+    rm.start();
+
+    CapacityScheduler cs = (CapacityScheduler) rm.getResourceScheduler();
+    LeafQueue qb = (LeafQueue)cs.getQueue("default");
+
+    // For now make user limit large so we can activate all applications
+    qb.setUserLimitFactor((float)100.0);
+    qb.setupConfigurableCapacities();
+
+    SchedulerEvent addAppEvent;
+    SchedulerEvent addAttemptEvent;
+    Container container = mock(Container.class);
+    ApplicationSubmissionContext submissionContext =
+        mock(ApplicationSubmissionContext.class);
+
+    final int appCount = 100;
+    ApplicationId[] appids = new ApplicationId[appCount];
+    RMAppAttemptImpl[] attempts = new RMAppAttemptImpl[appCount];
+    ApplicationAttemptId[] appAttemptIds = new ApplicationAttemptId[appCount];
+    RMAppImpl[] apps = new RMAppImpl[appCount];
+    RMAppAttemptMetrics[] attemptMetrics = new RMAppAttemptMetrics[appCount];
+    for (int i=0; i<appCount; i++) {
+      appids[i] = BuilderUtils.newApplicationId(100, i);
+      appAttemptIds[i] =
+      BuilderUtils.newApplicationAttemptId(appids[i], 1);
+
+      attemptMetrics[i] =
+          new RMAppAttemptMetrics(appAttemptIds[i], rm.getRMContext());
+      apps[i] = mock(RMAppImpl.class);
+      when(apps[i].getApplicationId()).thenReturn(appids[i]);
+      attempts[i] = mock(RMAppAttemptImpl.class);
+      when(attempts[i].getMasterContainer()).thenReturn(container);
+      when(attempts[i].getSubmissionContext()).thenReturn(submissionContext);
+      when(attempts[i].getAppAttemptId()).thenReturn(appAttemptIds[i]);
+      when(attempts[i].getRMAppAttemptMetrics()).thenReturn(attemptMetrics[i]);
+      when(apps[i].getCurrentAppAttempt()).thenReturn(attempts[i]);
+
+      rm.getRMContext().getRMApps().put(appids[i], apps[i]);
+      addAppEvent =
+          new AppAddedSchedulerEvent(appids[i], "default", "user1");
+      cs.handle(addAppEvent);
+      addAttemptEvent =
+          new AppAttemptAddedSchedulerEvent(appAttemptIds[i], false);
+      cs.handle(addAttemptEvent);
+    }
+
+    // add nodes  to cluster, so cluster has 20GB and 20 vcores
+    Resource newResource = Resource.newInstance(10 * GB, 10);
+    RMNode node = MockNodes.newNodeInfo(0, newResource, 1, "127.0.0.1");
+    cs.handle(new NodeAddedSchedulerEvent(node));
+
+    Resource newResource2 = Resource.newInstance(10 * GB, 10);
+    RMNode node2 = MockNodes.newNodeInfo(0, newResource2, 1, "127.0.0.2");
+    cs.handle(new NodeAddedSchedulerEvent(node2));
+
+    Priority u0Priority = TestUtils.createMockPriority(1);
+    RecordFactory recordFactory =
+        RecordFactoryProvider.getRecordFactory(null);
+
+    FiCaSchedulerApp[] fiCaApps = new FiCaSchedulerApp[appCount];
+    for (int i=0;i<appCount;i++) {
+      fiCaApps[i] =
+          cs.getSchedulerApplications().get(apps[i].getApplicationId())
+              .getCurrentAppAttempt();
+      // allocate container for app2 with 1GB memory and 1 vcore
+      fiCaApps[i].updateResourceRequests(Collections.singletonList(
+          TestUtils.createResourceRequest(ResourceRequest.ANY, 1*GB, 1, true,
+              u0Priority, recordFactory)));
+    }
+    // Now force everything to be over user limit
+    qb.setUserLimitFactor((float)0.0);
+
+    // Quiet the loggers while measuring throughput
+    for (Enumeration<?> loggers=LogManager.getCurrentLoggers();
+        loggers.hasMoreElements(); )  {
+      Logger logger = (Logger) loggers.nextElement();
+      logger.setLevel(Level.WARN);
+    }
+    final int topn = 20;
+    final int iterations = 2000000;
+    final int printInterval = 20000;
+    final float numerator = 1000.0f * printInterval;
+    PriorityQueue<Long> queue = new PriorityQueue<>(topn,
+        Collections.reverseOrder());
+
+    long n = Time.monotonicNow();
+    long timespent = 0;
+    for (int i = 0; i < iterations; i+=2) {
+      if (i > 0  && i % printInterval == 0){
+        long ts = (Time.monotonicNow() - n);
+        if (queue.size() < topn) {
+          queue.offer(ts);
+        } else {
+          Long last = queue.peek();
+          if (last > ts) {
+            queue.poll();
+            queue.offer(ts);
+          }
+        }
+        System.out.println(i + " " + (numerator / ts));
+        n= Time.monotonicNow();
+      }
+    cs.handle(new NodeUpdateSchedulerEvent(node));
+    cs.handle(new NodeUpdateSchedulerEvent(node2));
+    }
+    timespent=0;
+    int entries = queue.size();
+    while(queue.size() > 0){
+      long l = queue.poll();
+      timespent += l;
+    }
+    System.out.println("Avg of fastest " + entries + ": "
+        + numerator / (timespent / entries));
+    rm.stop();
+  }
+
   @Test
   public void testCSQueueBlocked() throws Exception {
     CapacitySchedulerConfiguration conf = new CapacitySchedulerConfiguration();

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

@@ -1146,7 +1146,7 @@ public class TestLeafQueue {
         new ResourceLimits(clusterResource),
         SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY), qb, nodes, apps);
     qb.computeUserLimitAndSetHeadroom(app_0, clusterResource,
-        "", SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
+        "", SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY, null);
 
     //maxqueue 16G, userlimit 13G, - 4G used = 9G
     assertEquals(9*GB,app_0.getHeadroom().getMemorySize());
@@ -1169,7 +1169,7 @@ public class TestLeafQueue {
         new ResourceLimits(clusterResource),
         SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY), qb, nodes, apps);
     qb.computeUserLimitAndSetHeadroom(app_0, clusterResource,
-        "", SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
+        "", SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY, null);
 
     assertEquals(8*GB, qb.getUsedResources().getMemorySize());
     assertEquals(4*GB, app_0.getCurrentConsumption().getMemorySize());
@@ -1219,7 +1219,7 @@ public class TestLeafQueue {
         new ResourceLimits(clusterResource),
         SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY), qb, nodes, apps);
     qb.computeUserLimitAndSetHeadroom(app_3, clusterResource,
-        "", SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
+        "", SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY, null);
     assertEquals(4*GB, qb.getUsedResources().getMemorySize());
     //maxqueue 16G, userlimit 7G, used (by each user) 2G, headroom 5G (both)
     assertEquals(5*GB, app_3.getHeadroom().getMemorySize());
@@ -1240,9 +1240,9 @@ public class TestLeafQueue {
         new ResourceLimits(clusterResource),
         SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY), qb, nodes, apps);
     qb.computeUserLimitAndSetHeadroom(app_4, clusterResource,
-        "", SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
+        "", SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY, null);
     qb.computeUserLimitAndSetHeadroom(app_3, clusterResource,
-        "", SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY);
+        "", SchedulingMode.RESPECT_PARTITION_EXCLUSIVITY, null);
     
     
     //app3 is user1, active from last test case