@Test
public void testDefaultSingleQueueMetrics() {
String queueName = "single";
String user = "******";
QueueMetrics metrics = QueueMetrics.forQueue(ms, queueName, null, false, conf);
MetricsSource queueSource = queueSource(ms, queueName);
AppSchedulingInfo app = mockApp(user);
metrics.submitApp(user, 1);
MetricsSource userSource = userSource(ms, queueName, user);
checkApps(queueSource, 1, 1, 0, 0, 0, 0);
metrics.setAvailableResourcesToQueue(Resources.createResource(100 * GB));
metrics.incrPendingResources(user, 5, Resources.createResource(15 * GB));
// Available resources is set externally, as it depends on dynamic
// configurable cluster/queue resources
checkResources(queueSource, 0, 0, 0, 0, 100 * GB, 15 * GB, 5, 0, 0);
metrics.incrAppsRunning(app, user);
checkApps(queueSource, 1, 0, 1, 0, 0, 0);
metrics.allocateResources(user, 3, Resources.createResource(2 * GB));
checkResources(queueSource, 6 * GB, 3, 3, 0, 100 * GB, 9 * GB, 2, 0, 0);
metrics.releaseResources(user, 1, Resources.createResource(2 * GB));
checkResources(queueSource, 4 * GB, 2, 3, 1, 100 * GB, 9 * GB, 2, 0, 0);
metrics.finishApp(app, RMAppAttemptState.FINISHED);
checkApps(queueSource, 1, 0, 0, 1, 0, 0);
assertNull(userSource);
}
@Test
public void testTwoLevelWithUserMetrics() {
String parentQueueName = "root";
String leafQueueName = "root.leaf";
String user = "******";
QueueMetrics parentMetrics = QueueMetrics.forQueue(ms, parentQueueName, null, true, conf);
Queue parentQueue = make(stub(Queue.class).returning(parentMetrics).from.getMetrics());
QueueMetrics metrics = QueueMetrics.forQueue(ms, leafQueueName, parentQueue, true, conf);
MetricsSource parentQueueSource = queueSource(ms, parentQueueName);
MetricsSource queueSource = queueSource(ms, leafQueueName);
AppSchedulingInfo app = mockApp(user);
metrics.submitApp(user, 1);
MetricsSource userSource = userSource(ms, leafQueueName, user);
MetricsSource parentUserSource = userSource(ms, parentQueueName, user);
checkApps(queueSource, 1, 1, 0, 0, 0, 0);
checkApps(parentQueueSource, 1, 1, 0, 0, 0, 0);
checkApps(userSource, 1, 1, 0, 0, 0, 0);
checkApps(parentUserSource, 1, 1, 0, 0, 0, 0);
parentMetrics.setAvailableResourcesToQueue(Resources.createResource(100 * GB));
metrics.setAvailableResourcesToQueue(Resources.createResource(100 * GB));
parentMetrics.setAvailableResourcesToUser(user, Resources.createResource(10 * GB));
metrics.setAvailableResourcesToUser(user, Resources.createResource(10 * GB));
metrics.incrPendingResources(user, 5, Resources.createResource(15 * GB));
checkResources(queueSource, 0, 0, 0, 0, 100 * GB, 15 * GB, 5, 0, 0);
checkResources(parentQueueSource, 0, 0, 0, 0, 100 * GB, 15 * GB, 5, 0, 0);
checkResources(userSource, 0, 0, 0, 0, 10 * GB, 15 * GB, 5, 0, 0);
checkResources(parentUserSource, 0, 0, 0, 0, 10 * GB, 15 * GB, 5, 0, 0);
metrics.incrAppsRunning(app, user);
checkApps(queueSource, 1, 0, 1, 0, 0, 0);
checkApps(userSource, 1, 0, 1, 0, 0, 0);
metrics.allocateResources(user, 3, Resources.createResource(2 * GB));
metrics.reserveResource(user, Resources.createResource(3 * GB));
// Available resources is set externally, as it depends on dynamic
// configurable cluster/queue resources
checkResources(queueSource, 6 * GB, 3, 3, 0, 100 * GB, 9 * GB, 2, 3 * GB, 1);
checkResources(parentQueueSource, 6 * GB, 3, 3, 0, 100 * GB, 9 * GB, 2, 3 * GB, 1);
checkResources(userSource, 6 * GB, 3, 3, 0, 10 * GB, 9 * GB, 2, 3 * GB, 1);
checkResources(parentUserSource, 6 * GB, 3, 3, 0, 10 * GB, 9 * GB, 2, 3 * GB, 1);
metrics.releaseResources(user, 1, Resources.createResource(2 * GB));
metrics.unreserveResource(user, Resources.createResource(3 * GB));
checkResources(queueSource, 4 * GB, 2, 3, 1, 100 * GB, 9 * GB, 2, 0, 0);
checkResources(parentQueueSource, 4 * GB, 2, 3, 1, 100 * GB, 9 * GB, 2, 0, 0);
checkResources(userSource, 4 * GB, 2, 3, 1, 10 * GB, 9 * GB, 2, 0, 0);
checkResources(parentUserSource, 4 * GB, 2, 3, 1, 10 * GB, 9 * GB, 2, 0, 0);
metrics.finishApp(app, RMAppAttemptState.FINISHED);
checkApps(queueSource, 1, 0, 0, 1, 0, 0);
checkApps(parentQueueSource, 1, 0, 0, 1, 0, 0);
checkApps(userSource, 1, 0, 0, 1, 0, 0);
checkApps(parentUserSource, 1, 0, 0, 1, 0, 0);
}
/**
* A Schedulable represents an entity that can launch tasks, such as a job or a queue. It provides a
* common interface so that algorithms such as fair sharing can be applied both within a queue and
* across queues. There are currently two types of Schedulables: JobSchedulables, which represent a
* single job, and QueueSchedulables, which allocate among jobs in their queue.
*
* <p>Separate sets of Schedulables are used for maps and reduces. Each queue has both a
* mapSchedulable and a reduceSchedulable, and so does each job.
*
* <p>A Schedulable is responsible for three roles: 1) It can launch tasks through assignTask(). 2)
* It provides information about the job/queue to the scheduler, including: - Demand (maximum number
* of tasks required) - Number of currently running tasks - Minimum share (for queues) - Job/queue
* weight (for fair sharing) - Start time and priority (for FIFO) 3) It can be assigned a fair
* share, for use with fair scheduling.
*
* <p>Schedulable also contains two methods for performing scheduling computations: - updateDemand()
* is called periodically to compute the demand of the various jobs and queues, which may be
* expensive (e.g. jobs must iterate through all their tasks to count failed tasks, tasks that can
* be speculated, etc). - redistributeShare() is called after demands are updated and a
* Schedulable's fair share has been set by its parent to let it distribute its share among the
* other Schedulables within it (e.g. for queues that want to perform fair sharing among their
* jobs).
*/
@Private
@Unstable
abstract class Schedulable {
/** Fair share assigned to this Schedulable */
private Resource fairShare = Resources.createResource(0);
/**
* Name of job/queue, used for debugging as well as for breaking ties in scheduling order
* deterministically.
*/
public abstract String getName();
/**
* Maximum number of resources required by this Schedulable. This is defined as number of
* currently utilized resources + number of unlaunched resources (that are either not yet launched
* or need to be speculated).
*/
public abstract Resource getDemand();
/** Get the aggregate amount of resources consumed by the schedulable. */
public abstract Resource getResourceUsage();
/** Minimum Resource share assigned to the schedulable. */
public abstract Resource getMinShare();
/** Job/queue weight in fair sharing. */
public abstract double getWeight();
/** Start time for jobs in FIFO queues; meaningless for QueueSchedulables. */
public abstract long getStartTime();
/** Job priority for jobs in FIFO queues; meaningless for QueueSchedulables. */
public abstract Priority getPriority();
/** Refresh the Schedulable's demand and those of its children if any. */
public abstract void updateDemand();
/**
* Assign a container on this node if possible, and return the amount of resources assigned. If
* {@code reserved} is true, it means a reservation already exists on this node, and the
* schedulable should fulfill that reservation if possible.
*/
public abstract Resource assignContainer(FSSchedulerNode node, boolean reserved);
/** Assign a fair share to this Schedulable. */
public void setFairShare(Resource fairShare) {
this.fairShare = fairShare;
}
/** Get the fair share assigned to this Schedulable. */
public Resource getFairShare() {
return fairShare;
}
/** Convenient toString implementation for debugging. */
@Override
public String toString() {
return String.format(
"[%s, demand=%s, running=%s, share=%s,], w=%.1f]",
getName(), getDemand(), getResourceUsage(), fairShare, getWeight());
}
}
// @Test
public void testFifoScheduler() throws Exception {
LOG.info("--- START: testFifoScheduler ---");
final int GB = 1024;
// Register node1
String host_0 = "host_0";
org.apache.hadoop.yarn.server.resourcemanager.NodeManager nm_0 =
registerNode(host_0, 1234, 2345, NetworkTopology.DEFAULT_RACK, 4 * GB);
nm_0.heartbeat();
// Register node2
String host_1 = "host_1";
org.apache.hadoop.yarn.server.resourcemanager.NodeManager nm_1 =
registerNode(host_1, 1234, 2345, NetworkTopology.DEFAULT_RACK, 2 * GB);
nm_1.heartbeat();
// ResourceRequest priorities
Priority priority_0 = org.apache.hadoop.yarn.server.resourcemanager.resource.Priority.create(0);
Priority priority_1 = org.apache.hadoop.yarn.server.resourcemanager.resource.Priority.create(1);
// Submit an application
Application application_0 = new Application("user_0", resourceManager);
application_0.submit();
application_0.addNodeManager(host_0, 1234, nm_0);
application_0.addNodeManager(host_1, 1234, nm_1);
Resource capability_0_0 = Resources.createResource(GB);
application_0.addResourceRequestSpec(priority_1, capability_0_0);
Resource capability_0_1 = Resources.createResource(2 * GB);
application_0.addResourceRequestSpec(priority_0, capability_0_1);
Task task_0_0 = new Task(application_0, priority_1, new String[] {host_0, host_1});
application_0.addTask(task_0_0);
// Submit another application
Application application_1 = new Application("user_1", resourceManager);
application_1.submit();
application_1.addNodeManager(host_0, 1234, nm_0);
application_1.addNodeManager(host_1, 1234, nm_1);
Resource capability_1_0 = Resources.createResource(3 * GB);
application_1.addResourceRequestSpec(priority_1, capability_1_0);
Resource capability_1_1 = Resources.createResource(4 * GB);
application_1.addResourceRequestSpec(priority_0, capability_1_1);
Task task_1_0 = new Task(application_1, priority_1, new String[] {host_0, host_1});
application_1.addTask(task_1_0);
// Send resource requests to the scheduler
LOG.info("Send resource requests to the scheduler");
application_0.schedule();
application_1.schedule();
// Send a heartbeat to kick the tires on the Scheduler
LOG.info(
"Send a heartbeat to kick the tires on the Scheduler... "
+ "nm0 -> task_0_0 and task_1_0 allocated, used=4G "
+ "nm1 -> nothing allocated");
nm_0.heartbeat(); // task_0_0 and task_1_0 allocated, used=4G
nm_1.heartbeat(); // nothing allocated
// Get allocations from the scheduler
application_0.schedule(); // task_0_0
checkApplicationResourceUsage(GB, application_0);
application_1.schedule(); // task_1_0
checkApplicationResourceUsage(3 * GB, application_1);
nm_0.heartbeat();
nm_1.heartbeat();
checkNodeResourceUsage(4 * GB, nm_0); // task_0_0 (1G) and task_1_0 (3G)
checkNodeResourceUsage(0 * GB, nm_1); // no tasks, 2G available
LOG.info("Adding new tasks...");
Task task_1_1 = new Task(application_1, priority_1, new String[] {RMNode.ANY});
application_1.addTask(task_1_1);
Task task_1_2 = new Task(application_1, priority_1, new String[] {RMNode.ANY});
application_1.addTask(task_1_2);
Task task_1_3 = new Task(application_1, priority_0, new String[] {RMNode.ANY});
application_1.addTask(task_1_3);
application_1.schedule();
Task task_0_1 = new Task(application_0, priority_1, new String[] {host_0, host_1});
application_0.addTask(task_0_1);
Task task_0_2 = new Task(application_0, priority_1, new String[] {host_0, host_1});
application_0.addTask(task_0_2);
Task task_0_3 = new Task(application_0, priority_0, new String[] {RMNode.ANY});
application_0.addTask(task_0_3);
application_0.schedule();
// Send a heartbeat to kick the tires on the Scheduler
LOG.info("Sending hb from " + nm_0.getHostName());
nm_0.heartbeat(); // nothing new, used=4G
LOG.info("Sending hb from " + nm_1.getHostName());
nm_1.heartbeat(); // task_0_3, used=2G
// Get allocations from the scheduler
LOG.info("Trying to allocate...");
application_0.schedule();
checkApplicationResourceUsage(3 * GB, application_0);
application_1.schedule();
checkApplicationResourceUsage(3 * GB, application_1);
nm_0.heartbeat();
nm_1.heartbeat();
checkNodeResourceUsage(4 * GB, nm_0);
checkNodeResourceUsage(2 * GB, nm_1);
// Complete tasks
LOG.info("Finishing up task_0_0");
application_0.finishTask(task_0_0); // Now task_0_1
application_0.schedule();
application_1.schedule();
nm_0.heartbeat();
nm_1.heartbeat();
checkApplicationResourceUsage(3 * GB, application_0);
checkApplicationResourceUsage(3 * GB, application_1);
checkNodeResourceUsage(4 * GB, nm_0);
checkNodeResourceUsage(2 * GB, nm_1);
LOG.info("Finishing up task_1_0");
application_1.finishTask(task_1_0); // Now task_0_2
application_0.schedule(); // final overcommit for app0 caused here
application_1.schedule();
nm_0.heartbeat(); // final overcommit for app0 occurs here
nm_1.heartbeat();
checkApplicationResourceUsage(4 * GB, application_0);
checkApplicationResourceUsage(0 * GB, application_1);
// checkNodeResourceUsage(1*GB, nm_0); // final over-commit -> rm.node->1G, test.node=2G
checkNodeResourceUsage(2 * GB, nm_1);
LOG.info("Finishing up task_0_3");
application_0.finishTask(task_0_3); // No more
application_0.schedule();
application_1.schedule();
nm_0.heartbeat();
nm_1.heartbeat();
checkApplicationResourceUsage(2 * GB, application_0);
checkApplicationResourceUsage(0 * GB, application_1);
// checkNodeResourceUsage(2*GB, nm_0); // final over-commit, rm.node->1G, test.node->2G
checkNodeResourceUsage(0 * GB, nm_1);
LOG.info("Finishing up task_0_1");
application_0.finishTask(task_0_1);
application_0.schedule();
application_1.schedule();
nm_0.heartbeat();
nm_1.heartbeat();
checkApplicationResourceUsage(1 * GB, application_0);
checkApplicationResourceUsage(0 * GB, application_1);
LOG.info("Finishing up task_0_2");
application_0.finishTask(task_0_2); // now task_1_3 can go!
application_0.schedule();
application_1.schedule();
nm_0.heartbeat();
nm_1.heartbeat();
checkApplicationResourceUsage(0 * GB, application_0);
checkApplicationResourceUsage(4 * GB, application_1);
LOG.info("Finishing up task_1_3");
application_1.finishTask(task_1_3); // now task_1_1
application_0.schedule();
application_1.schedule();
nm_0.heartbeat();
nm_1.heartbeat();
checkApplicationResourceUsage(0 * GB, application_0);
checkApplicationResourceUsage(3 * GB, application_1);
LOG.info("Finishing up task_1_1");
application_1.finishTask(task_1_1);
application_0.schedule();
application_1.schedule();
nm_0.heartbeat();
nm_1.heartbeat();
checkApplicationResourceUsage(0 * GB, application_0);
checkApplicationResourceUsage(3 * GB, application_1);
LOG.info("--- END: testFifoScheduler ---");
}