From 51ccb87031eda6a2b75be098a88f1d89ea82c610 Mon Sep 17 00:00:00 2001 From: Alejandro Abdelnur Date: Thu, 9 May 2013 22:15:58 +0000 Subject: [PATCH] YARN-655. Fair scheduler metrics should subtract allocated memory from available memory. (sandyr via tucu) git-svn-id: https://svn.apache.org/repos/asf/hadoop/common/trunk@1480809 13f79535-47bb-0310-9956-ffa450edef68 --- hadoop-yarn-project/CHANGES.txt | 3 ++ .../scheduler/QueueMetrics.java | 5 +++ .../scheduler/fair/FairScheduler.java | 20 +++++++++--- .../scheduler/fair/TestFairScheduler.java | 32 +++++++++++++++++++ 4 files changed, 56 insertions(+), 4 deletions(-) diff --git a/hadoop-yarn-project/CHANGES.txt b/hadoop-yarn-project/CHANGES.txt index fb9e6c3cde..80d004a701 100644 --- a/hadoop-yarn-project/CHANGES.txt +++ b/hadoop-yarn-project/CHANGES.txt @@ -367,6 +367,9 @@ Release 2.0.5-beta - UNRELEASED YARN-637. FS: maxAssign is not honored. (kkambatl via tucu) + YARN-655. Fair scheduler metrics should subtract allocated memory from + available memory. (sandyr via tucu) + Release 2.0.4-alpha - 2013-04-25 INCOMPATIBLE CHANGES diff --git a/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/main/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/QueueMetrics.java b/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/main/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/QueueMetrics.java index ef1f78b48e..13def49c20 100644 --- a/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/main/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/QueueMetrics.java +++ b/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/main/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/QueueMetrics.java @@ -43,6 +43,7 @@ import org.apache.hadoop.yarn.conf.YarnConfiguration; import org.apache.hadoop.yarn.server.resourcemanager.resource.Resources; import org.apache.hadoop.yarn.server.resourcemanager.rmapp.attempt.RMAppAttemptState; +import org.apache.hadoop.yarn.util.BuilderUtils; import org.slf4j.Logger; import org.slf4j.LoggerFactory; @@ -425,6 +426,10 @@ public int getAppsKilled() { public int getAppsFailed() { return appsFailed.value(); } + + public Resource getAllocatedResources() { + return BuilderUtils.newResource(allocatedMB.value(), 0); + } public int getAllocatedMB() { return allocatedMB.value(); diff --git a/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/main/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/fair/FairScheduler.java b/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/main/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/fair/FairScheduler.java index 5252b811c3..5904ae8b22 100644 --- a/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/main/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/fair/FairScheduler.java +++ b/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/main/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/fair/FairScheduler.java @@ -234,10 +234,6 @@ protected synchronized void update() { // Recursively compute fair shares for all queues // and update metrics rootQueue.recomputeShares(); - - // Update recorded capacity of root queue (child queues are updated - // when fair share is calculated). - rootMetrics.setAvailableResourcesToQueue(clusterCapacity); } /** @@ -670,6 +666,7 @@ private synchronized void completedContainer(RMContainer rmContainer, } else { application.containerCompleted(rmContainer, containerStatus, event); node.releaseContainer(container); + updateRootQueueMetrics(); } LOG.info("Application " + applicationAttemptId + @@ -681,6 +678,7 @@ private synchronized void completedContainer(RMContainer rmContainer, private synchronized void addNode(RMNode node) { nodes.put(node.getNodeID(), new FSSchedulerNode(node)); Resources.addTo(clusterCapacity, node.getTotalCapability()); + updateRootQueueMetrics(); LOG.info("Added node " + node.getNodeAddress() + " cluster capacity: " + clusterCapacity); @@ -689,6 +687,7 @@ private synchronized void addNode(RMNode node) { private synchronized void removeNode(RMNode rmNode) { FSSchedulerNode node = nodes.get(rmNode.getNodeID()); Resources.subtractFrom(clusterCapacity, rmNode.getTotalCapability()); + updateRootQueueMetrics(); // Remove running containers List runningContainers = node.getRunningContainers(); @@ -901,6 +900,7 @@ private synchronized void nodeUpdate(RMNode nm) { if ((assignedContainers >= maxAssign) && (maxAssign > 0)) { break; } } } + updateRootQueueMetrics(); } @Override @@ -922,6 +922,18 @@ public SchedulerAppReport getSchedulerAppInfo( } return new SchedulerAppReport(applications.get(appAttemptId)); } + + /** + * Subqueue metrics might be a little out of date because fair shares are + * recalculated at the update interval, but the root queue metrics needs to + * be updated synchronously with allocations and completions so that cluster + * metrics will be consistent. + */ + private void updateRootQueueMetrics() { + rootMetrics.setAvailableResourcesToQueue( + Resources.subtract( + clusterCapacity, rootMetrics.getAllocatedResources())); + } @Override public QueueMetrics getRootQueueMetrics() { diff --git a/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/test/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/fair/TestFairScheduler.java b/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/test/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/fair/TestFairScheduler.java index 88b8a63df7..f5c0d5512f 100644 --- a/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/test/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/fair/TestFairScheduler.java +++ b/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/test/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/fair/TestFairScheduler.java @@ -68,6 +68,7 @@ import org.apache.hadoop.yarn.server.resourcemanager.rmapp.attempt.RMAppAttemptState; import org.apache.hadoop.yarn.server.resourcemanager.rmcontainer.RMContainer; import org.apache.hadoop.yarn.server.resourcemanager.rmnode.RMNode; +import org.apache.hadoop.yarn.server.resourcemanager.scheduler.QueueMetrics; import org.apache.hadoop.yarn.server.resourcemanager.scheduler.ResourceScheduler; import org.apache.hadoop.yarn.server.resourcemanager.scheduler.event.AppAddedSchedulerEvent; import org.apache.hadoop.yarn.server.resourcemanager.scheduler.event.AppRemovedSchedulerEvent; @@ -128,6 +129,7 @@ public void setUp() throws IOException { public void tearDown() { scheduler = null; resourceManager = null; + QueueMetrics.clearQueueMetrics(); } private Configuration createConfiguration() { @@ -337,6 +339,13 @@ public void testSimpleContainerAllocation() { assertEquals(1024, scheduler.getQueueManager().getQueue("queue1"). getResourceUsage().getMemory()); + + // verify metrics + QueueMetrics queue1Metrics = scheduler.getQueueManager().getQueue("queue1") + .getMetrics(); + assertEquals(1024, queue1Metrics.getAllocatedMB()); + assertEquals(1024, scheduler.getRootQueueMetrics().getAllocatedMB()); + assertEquals(512, scheduler.getRootQueueMetrics().getAvailableMB()); } @Test (timeout = 5000) @@ -1276,6 +1285,7 @@ public void testReservationWhileMultiplePriorities() { scheduler.handle(updateEvent); assertEquals(1, app.getLiveContainers().size()); + assertEquals(0, scheduler.getRootQueueMetrics().getAvailableMB()); // Create request at higher priority createSchedulingRequestExistingApplication(1024, 1, attId); @@ -1291,6 +1301,7 @@ public void testReservationWhileMultiplePriorities() { // Complete container scheduler.allocate(attId, new ArrayList(), Arrays.asList(containerId)); + assertEquals(1024, scheduler.getRootQueueMetrics().getAvailableMB()); // Schedule at opening scheduler.update(); @@ -1302,6 +1313,7 @@ public void testReservationWhileMultiplePriorities() { for (RMContainer liveContainer : liveContainers) { Assert.assertEquals(2, liveContainer.getContainer().getPriority().getPriority()); } + assertEquals(0, scheduler.getRootQueueMetrics().getAvailableMB()); } @Test @@ -1606,4 +1618,24 @@ public void testReservationThatDoesntFit() { assertEquals(1, app.getLiveContainers().size()); assertEquals(0, app.getReservedContainers().size()); } + + @Test + public void testRemoveNodeUpdatesRootQueueMetrics() { + assertEquals(0, scheduler.getRootQueueMetrics().getAvailableMB()); + + RMNode node1 = MockNodes.newNodeInfo(1, Resources.createResource(1024)); + NodeAddedSchedulerEvent addEvent = new NodeAddedSchedulerEvent(node1); + scheduler.handle(addEvent); + + assertEquals(1024, scheduler.getRootQueueMetrics().getAvailableMB()); + scheduler.update(); // update shouldn't change things + assertEquals(1024, scheduler.getRootQueueMetrics().getAvailableMB()); + + NodeRemovedSchedulerEvent removeEvent = new NodeRemovedSchedulerEvent(node1); + scheduler.handle(removeEvent); + + assertEquals(0, scheduler.getRootQueueMetrics().getAvailableMB()); + scheduler.update(); // update shouldn't change things + assertEquals(0, scheduler.getRootQueueMetrics().getAvailableMB()); + } }