Examples with SchedulingResult org.apache.storm.scheduler.resource.SchedulingResult used on opensource projects

Example 1 with SchedulingResult

use of org.apache.storm.scheduler.resource.SchedulingResult in project storm by apache.

the class DefaultResourceAwareStrategy method schedule.

public SchedulingResult schedule(TopologyDetails td) {
    if (_nodes.getNodes().size() <= 0) {
        LOG.warn("No available nodes to schedule tasks on!");
        return SchedulingResult.failure(SchedulingStatus.FAIL_NOT_ENOUGH_RESOURCES, "No available nodes to schedule tasks on!");
    }
    Collection<ExecutorDetails> unassignedExecutors = new HashSet<ExecutorDetails>(_cluster.getUnassignedExecutors(td));
    Map<WorkerSlot, Collection<ExecutorDetails>> schedulerAssignmentMap = new HashMap<>();
    LOG.debug("ExecutorsNeedScheduling: {}", unassignedExecutors);
    Collection<ExecutorDetails> scheduledTasks = new ArrayList<>();
    List<Component> spouts = this.getSpouts(td);
    if (spouts.size() == 0) {
        LOG.error("Cannot find a Spout!");
        return SchedulingResult.failure(SchedulingStatus.FAIL_INVALID_TOPOLOGY, "Cannot find a Spout!");
    }
    //order executors to be scheduled
    List<ExecutorDetails> orderedExecutors = orderExecutors(td, unassignedExecutors);
    Collection<ExecutorDetails> executorsNotScheduled = new HashSet<>(unassignedExecutors);
    for (ExecutorDetails exec : orderedExecutors) {
        LOG.debug("\n\nAttempting to schedule: {} of component {}[ REQ {} ]", exec, td.getExecutorToComponent().get(exec), td.getTaskResourceReqList(exec));
        scheduleExecutor(exec, td, schedulerAssignmentMap, scheduledTasks);
    }
    executorsNotScheduled.removeAll(scheduledTasks);
    LOG.debug("/* Scheduling left over task (most likely sys tasks) */");
    // schedule left over system tasks
    for (ExecutorDetails exec : executorsNotScheduled) {
        scheduleExecutor(exec, td, schedulerAssignmentMap, scheduledTasks);
    }
    SchedulingResult result;
    executorsNotScheduled.removeAll(scheduledTasks);
    if (executorsNotScheduled.size() > 0) {
        LOG.error("Not all executors successfully scheduled: {}", executorsNotScheduled);
        schedulerAssignmentMap = null;
        result = SchedulingResult.failure(SchedulingStatus.FAIL_NOT_ENOUGH_RESOURCES, (td.getExecutors().size() - unassignedExecutors.size()) + "/" + td.getExecutors().size() + " executors scheduled");
    } else {
        LOG.debug("All resources successfully scheduled!");
        result = SchedulingResult.successWithMsg(schedulerAssignmentMap, "Fully Scheduled by DefaultResourceAwareStrategy");
    }
    if (schedulerAssignmentMap == null) {
        LOG.error("Topology {} not successfully scheduled!", td.getId());
    }
    return result;
}

Example 2 with SchedulingResult

use of org.apache.storm.scheduler.resource.SchedulingResult in project storm by apache.

the class BaseResourceAwareStrategy method schedule.

/**
 * Note that this method is not thread-safe.
 * Several instance variables are generated from supplied
 * parameters. In addition, the following instance variables are set to complete scheduling:
 *  <li>{@link #searcherState}</li>
 *  <li>{@link #execSorter} to sort executors</li>
 *  <li>{@link #nodeSorter} to sort nodes</li>
 * <p>
 * Scheduling consists of three main steps:
 *  <li>{@link #prepareForScheduling(Cluster, TopologyDetails)}</li>
 *  <li>{@link #checkSchedulingFeasibility()}, and</li>
 *  <li>{@link #scheduleExecutorsOnNodes(List, Iterable)}</li>
 * </p><p>
 * The executors and nodes are sorted in the order most conducive to scheduling for the strategy.
 * Those interfaces may be overridden by subclasses using mutators:
 *  <li>{@link #setExecSorter(IExecSorter)} and</li>
 *  <li>{@link #setNodeSorter(INodeSorter)}</li>
 *</p>
 *
 * @param cluster on which executors will be scheduled.
 * @param td the topology to schedule for.
 * @return result of scheduling (success, failure, or null when interrupted).
 */
@Override
public SchedulingResult schedule(Cluster cluster, TopologyDetails td) {
    prepareForScheduling(cluster, td);
    // early detection of success or failure
    SchedulingResult earlyResult = checkSchedulingFeasibility();
    if (earlyResult != null) {
        return earlyResult;
    }
    LOG.debug("Topology {} {} Number of ExecutorsNeedScheduling: {}", topoName, topologyDetails.getId(), unassignedExecutors.size());
    // order executors to be scheduled
    List<ExecutorDetails> orderedExecutors = execSorter.sortExecutors(unassignedExecutors);
    Iterable<String> sortedNodes = null;
    if (!this.sortNodesForEachExecutor) {
        nodeSorter.prepare(null);
        sortedNodes = nodeSorter.sortAllNodes();
    }
    return scheduleExecutorsOnNodes(orderedExecutors, sortedNodes);
}

Example 3 with SchedulingResult

use of org.apache.storm.scheduler.resource.SchedulingResult in project storm by apache.

the class TestConstraintSolverStrategy method basicFailureTest.

public void basicFailureTest(String confKey, Object confValue, ConstraintSolverStrategy cs) {
    Map<String, Object> config = makeTestTopoConf();
    config.put(confKey, confValue);
    cs.prepare(config);
    TopologyDetails topo = makeTopology(config, NORMAL_BOLT_PARALLEL);
    Topologies topologies = new Topologies(topo);
    Cluster cluster = makeCluster(topologies);
    LOG.info("Scheduling...");
    SchedulingResult result = cs.schedule(cluster, topo);
    LOG.info("Done scheduling {}...", result);
    Assert.assertTrue("Assert scheduling topology success " + result, !result.isSuccess());
}

Example 4 with SchedulingResult

use of org.apache.storm.scheduler.resource.SchedulingResult in project storm by apache.

the class TestConstraintSolverStrategy method basicUnitTestWithKillAndRecover.

public void basicUnitTestWithKillAndRecover(ConstraintSolverStrategy cs, int boltParallel, int coLocationCnt) {
    Map<String, Object> config = makeTestTopoConf(coLocationCnt);
    cs.prepare(config);
    TopologyDetails topo = makeTopology(config, boltParallel);
    Topologies topologies = new Topologies(topo);
    Cluster cluster = makeCluster(topologies);
    LOG.info("Scheduling...");
    SchedulingResult result = cs.schedule(cluster, topo);
    LOG.info("Done scheduling {}...", result);
    Assert.assertTrue("Assert scheduling topology success " + result, result.isSuccess());
    Assert.assertEquals("Assert no unassigned executors, found unassigned: " + cluster.getUnassignedExecutors(topo), 0, cluster.getUnassignedExecutors(topo).size());
    Assert.assertTrue("Valid Scheduling?", ConstraintSolverStrategy.validateSolution(cluster, topo));
    LOG.info("Slots Used {}", cluster.getAssignmentById(topo.getId()).getSlots());
    LOG.info("Assignment {}", cluster.getAssignmentById(topo.getId()).getSlotToExecutors());
    // simulate worker loss
    SchedulerAssignment assignment = cluster.getAssignmentById(topo.getId());
    Set<WorkerSlot> slotsToDelete = new HashSet<>();
    Set<WorkerSlot> slots = assignment.getSlots();
    int i = 0;
    for (WorkerSlot slot : slots) {
        if (i % 2 == 0) {
            slotsToDelete.add(slot);
        }
        i++;
    }
    LOG.info("KILL WORKER(s) {}", slotsToDelete);
    for (WorkerSlot slot : slotsToDelete) {
        cluster.freeSlot(slot);
    }
    cs = new ConstraintSolverStrategy();
    cs.prepare(config);
    LOG.info("Scheduling again...");
    result = cs.schedule(cluster, topo);
    LOG.info("Done scheduling {}...", result);
    Assert.assertTrue("Assert scheduling topology success " + result, result.isSuccess());
    Assert.assertEquals("topo all executors scheduled?", 0, cluster.getUnassignedExecutors(topo).size());
    Assert.assertTrue("Valid Scheduling?", ConstraintSolverStrategy.validateSolution(cluster, topo));
}

Example 5 with SchedulingResult

use of org.apache.storm.scheduler.resource.SchedulingResult in project storm by apache.

the class TestDefaultResourceAwareStrategy method testMultipleRacks.

/**
 * Test whether strategy will choose correct rack
 */
@Test
public void testMultipleRacks() {
    final Map<String, SupervisorDetails> supMap = new HashMap<>();
    final Map<String, SupervisorDetails> supMapRack0 = genSupervisors(10, 4, 0, 400, 8000);
    // generate another rack of supervisors with less resources
    final Map<String, SupervisorDetails> supMapRack1 = genSupervisors(10, 4, 10, 200, 4000);
    // generate some supervisors that are depleted of one resource
    final Map<String, SupervisorDetails> supMapRack2 = genSupervisors(10, 4, 20, 0, 8000);
    // generate some that has alot of memory but little of cpu
    final Map<String, SupervisorDetails> supMapRack3 = genSupervisors(10, 4, 30, 10, 8000 * 2 + 4000);
    // generate some that has alot of cpu but little of memory
    final Map<String, SupervisorDetails> supMapRack4 = genSupervisors(10, 4, 40, 400 + 200 + 10, 1000);
    // Generate some that have neither resource, to verify that the strategy will prioritize this last
    // Also put a generic resource with 0 value in the resources list, to verify that it doesn't affect the sorting
    final Map<String, SupervisorDetails> supMapRack5 = genSupervisors(10, 4, 50, 0.0, 0.0, Collections.singletonMap("gpu.count", 0.0));
    supMap.putAll(supMapRack0);
    supMap.putAll(supMapRack1);
    supMap.putAll(supMapRack2);
    supMap.putAll(supMapRack3);
    supMap.putAll(supMapRack4);
    supMap.putAll(supMapRack5);
    Config config = createClusterConfig(100, 500, 500, null);
    config.put(Config.TOPOLOGY_WORKER_MAX_HEAP_SIZE_MB, Double.MAX_VALUE);
    INimbus iNimbus = new INimbusTest();
    // create test DNSToSwitchMapping plugin
    DNSToSwitchMapping TestNetworkTopographyPlugin = new TestDNSToSwitchMapping(supMapRack0, supMapRack1, supMapRack2, supMapRack3, supMapRack4, supMapRack5);
    // generate topologies
    TopologyDetails topo1 = genTopology("topo-1", config, 8, 0, 2, 0, CURRENT_TIME - 2, 10, "user");
    TopologyDetails topo2 = genTopology("topo-2", config, 8, 0, 2, 0, CURRENT_TIME - 2, 10, "user");
    Topologies topologies = new Topologies(topo1, topo2);
    Cluster cluster = new Cluster(iNimbus, new ResourceMetrics(new StormMetricsRegistry()), supMap, new HashMap<>(), topologies, config);
    List<String> supHostnames = new LinkedList<>();
    for (SupervisorDetails sup : supMap.values()) {
        supHostnames.add(sup.getHost());
    }
    Map<String, List<String>> rackToNodes = new HashMap<>();
    Map<String, String> resolvedSuperVisors = TestNetworkTopographyPlugin.resolve(supHostnames);
    for (Map.Entry<String, String> entry : resolvedSuperVisors.entrySet()) {
        String hostName = entry.getKey();
        String rack = entry.getValue();
        rackToNodes.computeIfAbsent(rack, rid -> new ArrayList<>()).add(hostName);
    }
    cluster.setNetworkTopography(rackToNodes);
    DefaultResourceAwareStrategyOld rs = new DefaultResourceAwareStrategyOld();
    rs.prepareForScheduling(cluster, topo1);
    INodeSorter nodeSorter = new NodeSorterHostProximity(cluster, topo1, BaseResourceAwareStrategy.NodeSortType.DEFAULT_RAS);
    nodeSorter.prepare(null);
    Iterable<ObjectResourcesItem> sortedRacks = nodeSorter.getSortedRacks();
    Iterator<ObjectResourcesItem> it = sortedRacks.iterator();
    // Ranked first since rack-0 has the most balanced set of resources
    Assert.assertEquals("rack-0 should be ordered first", "rack-0", it.next().id);
    // Ranked second since rack-1 has a balanced set of resources but less than rack-0
    Assert.assertEquals("rack-1 should be ordered second", "rack-1", it.next().id);
    // Ranked third since rack-4 has a lot of cpu but not a lot of memory
    Assert.assertEquals("rack-4 should be ordered third", "rack-4", it.next().id);
    // Ranked fourth since rack-3 has alot of memory but not cpu
    Assert.assertEquals("rack-3 should be ordered fourth", "rack-3", it.next().id);
    // Ranked fifth since rack-2 has not cpu resources
    Assert.assertEquals("rack-2 should be ordered fifth", "rack-2", it.next().id);
    // Ranked last since rack-5 has neither CPU nor memory available
    assertEquals("Rack-5 should be ordered sixth", "rack-5", it.next().id);
    SchedulingResult schedulingResult = rs.schedule(cluster, topo1);
    assert (schedulingResult.isSuccess());
    SchedulerAssignment assignment = cluster.getAssignmentById(topo1.getId());
    for (WorkerSlot ws : assignment.getSlotToExecutors().keySet()) {
        // make sure all workers on scheduled in rack-0
        Assert.assertEquals("assert worker scheduled on rack-0", "rack-0", resolvedSuperVisors.get(rs.idToNode(ws.getNodeId()).getHostname()));
    }
    Assert.assertEquals("All executors in topo-1 scheduled", 0, cluster.getUnassignedExecutors(topo1).size());
    // Test if topology is already partially scheduled on one rack
    Iterator<ExecutorDetails> executorIterator = topo2.getExecutors().iterator();
    List<String> nodeHostnames = rackToNodes.get("rack-1");
    for (int i = 0; i < topo2.getExecutors().size() / 2; i++) {
        String nodeHostname = nodeHostnames.get(i % nodeHostnames.size());
        RasNode node = rs.hostnameToNodes(nodeHostname).get(0);
        WorkerSlot targetSlot = node.getFreeSlots().iterator().next();
        ExecutorDetails targetExec = executorIterator.next();
        // to keep track of free slots
        node.assign(targetSlot, topo2, Arrays.asList(targetExec));
    }
    rs = new DefaultResourceAwareStrategyOld();
    // schedule topo2
    schedulingResult = rs.schedule(cluster, topo2);
    assert (schedulingResult.isSuccess());
    assignment = cluster.getAssignmentById(topo2.getId());
    for (WorkerSlot ws : assignment.getSlotToExecutors().keySet()) {
        // make sure all workers on scheduled in rack-1
        Assert.assertEquals("assert worker scheduled on rack-1", "rack-1", resolvedSuperVisors.get(rs.idToNode(ws.getNodeId()).getHostname()));
    }
    Assert.assertEquals("All executors in topo-2 scheduled", 0, cluster.getUnassignedExecutors(topo2).size());
}