Examples with ExecutorDetails org.apache.storm.scheduler.ExecutorDetails used on opensource projects

Example 26 with ExecutorDetails

use of org.apache.storm.scheduler.ExecutorDetails in project storm by apache.

the class TestResourceAwareScheduler method testTopologyWithMultipleSpouts.

@Test
public void testTopologyWithMultipleSpouts() {
    INimbus iNimbus = new INimbusTest();
    Map<String, SupervisorDetails> supMap = genSupervisors(2, 4, 400, 2000);
    // a topology with multiple spouts
    TopologyBuilder builder1 = new TopologyBuilder();
    builder1.setSpout("wordSpout1", new TestWordSpout(), 1);
    builder1.setSpout("wordSpout2", new TestWordSpout(), 1);
    builder1.setBolt("wordCountBolt1", new TestWordCounter(), 1).shuffleGrouping("wordSpout1").shuffleGrouping("wordSpout2");
    builder1.setBolt("wordCountBolt2", new TestWordCounter(), 1).shuffleGrouping("wordCountBolt1");
    builder1.setBolt("wordCountBolt3", new TestWordCounter(), 1).shuffleGrouping("wordCountBolt1");
    builder1.setBolt("wordCountBolt4", new TestWordCounter(), 1).shuffleGrouping("wordCountBolt2");
    builder1.setBolt("wordCountBolt5", new TestWordCounter(), 1).shuffleGrouping("wordSpout2");
    StormTopology stormTopology1 = builder1.createTopology();
    Config config = new Config();
    config.putAll(defaultTopologyConf);
    Map<ExecutorDetails, String> executorMap1 = genExecsAndComps(stormTopology1);
    TopologyDetails topology1 = new TopologyDetails("topology1", config, stormTopology1, 0, executorMap1, 0, "user");
    // a topology with two unconnected partitions
    TopologyBuilder builder2 = new TopologyBuilder();
    builder2.setSpout("wordSpoutX", new TestWordSpout(), 1);
    builder2.setSpout("wordSpoutY", new TestWordSpout(), 1);
    StormTopology stormTopology2 = builder2.createTopology();
    Map<ExecutorDetails, String> executorMap2 = genExecsAndComps(stormTopology2);
    TopologyDetails topology2 = new TopologyDetails("topology2", config, stormTopology2, 0, executorMap2, 0, "user");
    scheduler = new ResourceAwareScheduler();
    Topologies topologies = new Topologies(topology1, topology2);
    Cluster cluster = new Cluster(iNimbus, new ResourceMetrics(new StormMetricsRegistry()), supMap, new HashMap<>(), topologies, config);
    scheduler.prepare(config, new StormMetricsRegistry());
    scheduler.schedule(topologies, cluster);
    SchedulerAssignment assignment1 = cluster.getAssignmentById(topology1.getId());
    Set<WorkerSlot> assignedSlots1 = assignment1.getSlots();
    Set<String> nodesIDs1 = new HashSet<>();
    for (WorkerSlot slot : assignedSlots1) {
        nodesIDs1.add(slot.getNodeId());
    }
    Collection<ExecutorDetails> executors1 = assignment1.getExecutors();
    assertEquals(1, assignedSlots1.size());
    assertEquals(1, nodesIDs1.size());
    assertEquals(7, executors1.size());
    assertFalse(cluster.needsSchedulingRas(topology1));
    assertTrue(cluster.getStatusMap().get(topology1.getId()).startsWith("Running - Fully Scheduled by DefaultResourceAwareStrategy"));
    SchedulerAssignment assignment2 = cluster.getAssignmentById(topology2.getId());
    Set<WorkerSlot> assignedSlots2 = assignment2.getSlots();
    Set<String> nodesIDs2 = new HashSet<>();
    for (WorkerSlot slot : assignedSlots2) {
        nodesIDs2.add(slot.getNodeId());
    }
    Collection<ExecutorDetails> executors2 = assignment2.getExecutors();
    assertEquals(1, assignedSlots2.size());
    assertEquals(1, nodesIDs2.size());
    assertEquals(2, executors2.size());
    assertFalse(cluster.needsSchedulingRas(topology2));
    assertTrue(cluster.getStatusMap().get(topology2.getId()).startsWith("Running - Fully Scheduled by DefaultResourceAwareStrategy"));
}

Example 27 with ExecutorDetails

use of org.apache.storm.scheduler.ExecutorDetails in project storm by apache.

the class TestResourceAwareScheduler method testRASNodeSlotAssign.

@Test
public void testRASNodeSlotAssign() {
    Config config = new Config();
    config.putAll(defaultTopologyConf);
    INimbus iNimbus = new INimbusTest();
    Map<String, SupervisorDetails> supMap = genSupervisors(5, 4, 400, 2000);
    TopologyDetails topology1 = genTopology("topology1", config, 1, 0, 2, 0, 0, 0, "user");
    TopologyDetails topology2 = genTopology("topology2", config, 1, 0, 2, 0, 0, 0, "user");
    Topologies topologies = new Topologies(topology1, topology2);
    Cluster cluster = new Cluster(iNimbus, new ResourceMetrics(new StormMetricsRegistry()), supMap, new HashMap<>(), topologies, config);
    Map<String, RasNode> nodes = RasNodes.getAllNodesFrom(cluster);
    assertEquals(5, nodes.size());
    RasNode node = nodes.get("r000s000");
    assertEquals("r000s000", node.getId());
    assertTrue(node.isAlive());
    assertEquals(0, node.getRunningTopologies().size());
    assertTrue(node.isTotallyFree());
    assertEquals(4, node.totalSlotsFree());
    assertEquals(0, node.totalSlotsUsed());
    assertEquals(4, node.totalSlots());
    List<ExecutorDetails> executors11 = new ArrayList<>();
    executors11.add(new ExecutorDetails(1, 1));
    node.assign(node.getFreeSlots().iterator().next(), topology1, executors11);
    assertEquals(1, node.getRunningTopologies().size());
    assertFalse(node.isTotallyFree());
    assertEquals(3, node.totalSlotsFree());
    assertEquals(1, node.totalSlotsUsed());
    assertEquals(4, node.totalSlots());
    List<ExecutorDetails> executors12 = new ArrayList<>();
    executors12.add(new ExecutorDetails(2, 2));
    node.assign(node.getFreeSlots().iterator().next(), topology1, executors12);
    assertEquals(1, node.getRunningTopologies().size());
    assertFalse(node.isTotallyFree());
    assertEquals(2, node.totalSlotsFree());
    assertEquals(2, node.totalSlotsUsed());
    assertEquals(4, node.totalSlots());
    List<ExecutorDetails> executors21 = new ArrayList<>();
    executors21.add(new ExecutorDetails(1, 1));
    node.assign(node.getFreeSlots().iterator().next(), topology2, executors21);
    assertEquals(2, node.getRunningTopologies().size());
    assertFalse(node.isTotallyFree());
    assertEquals(1, node.totalSlotsFree());
    assertEquals(3, node.totalSlotsUsed());
    assertEquals(4, node.totalSlots());
    List<ExecutorDetails> executors22 = new ArrayList<>();
    executors22.add(new ExecutorDetails(2, 2));
    node.assign(node.getFreeSlots().iterator().next(), topology2, executors22);
    assertEquals(2, node.getRunningTopologies().size());
    assertFalse(node.isTotallyFree());
    assertEquals(0, node.totalSlotsFree());
    assertEquals(4, node.totalSlotsUsed());
    assertEquals(4, node.totalSlots());
    node.freeAllSlots();
    assertEquals(0, node.getRunningTopologies().size());
    assertTrue(node.isTotallyFree());
    assertEquals(4, node.totalSlotsFree());
    assertEquals(0, node.totalSlotsUsed());
    assertEquals(4, node.totalSlots());
}

Example 28 with ExecutorDetails

use of org.apache.storm.scheduler.ExecutorDetails in project storm by apache.

the class TestNodeSorterHostProximity method testAntiAffinityWithMultipleTopologies.

/**
 * Schedule two topologies, once with special resources and another without.
 * There are enough special resources to hold one topology with special resource ("my.gpu").
 * If the sort order is incorrect, scheduling will not succeed.
 */
@Test
public void testAntiAffinityWithMultipleTopologies() {
    INimbus iNimbus = new INimbusTest();
    Map<String, SupervisorDetails> supMap = genSupervisorsWithRacks(1, 40, 66, 0, 0, 4700, 226200, new HashMap<>());
    HashMap<String, Double> extraResources = new HashMap<>();
    extraResources.put("my.gpu", 1.0);
    supMap.putAll(genSupervisorsWithRacks(1, 40, 66, 1, 0, 4700, 226200, extraResources));
    Config config = new Config();
    config.putAll(createGrasClusterConfig(88, 775, 25, null, null));
    IScheduler scheduler = new ResourceAwareScheduler();
    scheduler.prepare(config, new StormMetricsRegistry());
    TopologyDetails tdSimple = genTopology("topology-simple", config, 1, 5, 100, 300, 0, 0, "user", 8192);
    // Schedule the simple topology first
    Topologies topologies = new Topologies(tdSimple);
    Cluster cluster = new Cluster(iNimbus, new ResourceMetrics(new StormMetricsRegistry()), supMap, new HashMap<>(), topologies, config);
    {
        NodeSorterHostProximity nodeSorter = new NodeSorterHostProximity(cluster, tdSimple);
        for (ExecutorDetails exec : tdSimple.getExecutors()) {
            nodeSorter.prepare(exec);
            List<ObjectResourcesItem> sortedRacks = StreamSupport.stream(nodeSorter.getSortedRacks().spliterator(), false).collect(Collectors.toList());
            String rackSummaries = StreamSupport.stream(sortedRacks.spliterator(), false).map(x -> String.format("Rack %s -> scheduled-cnt %d, min-avail %f, avg-avail %f, cpu %f, mem %f", x.id, nodeSorter.getScheduledExecCntByRackId().getOrDefault(x.id, new AtomicInteger(-1)).get(), x.minResourcePercent, x.avgResourcePercent, x.availableResources.getTotalCpu(), x.availableResources.getTotalMemoryMb())).collect(Collectors.joining("\n\t"));
            NormalizedResourceRequest topoResourceRequest = tdSimple.getApproximateTotalResources();
            String topoRequest = String.format("Topo %s, approx-requested-resources %s", tdSimple.getId(), topoResourceRequest.toString());
            Assert.assertEquals(rackSummaries + "\n# of racks sorted", 2, sortedRacks.size());
            Assert.assertEquals(rackSummaries + "\nFirst rack sorted", "rack-000", sortedRacks.get(0).id);
            Assert.assertEquals(rackSummaries + "\nSecond rack sorted", "rack-001", sortedRacks.get(1).id);
        }
    }
    scheduler.schedule(topologies, cluster);
    TopologyBuilder builder = topologyBuilder(1, 5, 100, 300);
    builder.setBolt("gpu-bolt", new TestBolt(), 40).addResource("my.gpu", 1.0).shuffleGrouping("spout-0");
    TopologyDetails tdGpu = topoToTopologyDetails("topology-gpu", config, builder.createTopology(), 0, 0, "user", 8192);
    // Now schedule GPU but with the simple topology in place.
    topologies = new Topologies(tdSimple, tdGpu);
    cluster = new Cluster(cluster, topologies);
    {
        NodeSorterHostProximity nodeSorter = new NodeSorterHostProximity(cluster, tdGpu);
        for (ExecutorDetails exec : tdGpu.getExecutors()) {
            String comp = tdGpu.getComponentFromExecutor(exec);
            nodeSorter.prepare(exec);
            List<ObjectResourcesItem> sortedRacks = StreamSupport.stream(nodeSorter.getSortedRacks().spliterator(), false).collect(Collectors.toList());
            String rackSummaries = sortedRacks.stream().map(x -> String.format("Rack %s -> scheduled-cnt %d, min-avail %f, avg-avail %f, cpu %f, mem %f", x.id, nodeSorter.getScheduledExecCntByRackId().getOrDefault(x.id, new AtomicInteger(-1)).get(), x.minResourcePercent, x.avgResourcePercent, x.availableResources.getTotalCpu(), x.availableResources.getTotalMemoryMb())).collect(Collectors.joining("\n\t"));
            NormalizedResourceRequest topoResourceRequest = tdSimple.getApproximateTotalResources();
            String topoRequest = String.format("Topo %s, approx-requested-resources %s", tdSimple.getId(), topoResourceRequest.toString());
            Assert.assertEquals(rackSummaries + "\n# of racks sorted", 2, sortedRacks.size());
            if (comp.equals("gpu-bolt")) {
                Assert.assertEquals(rackSummaries + "\nFirst rack sorted for " + comp, "rack-001", sortedRacks.get(0).id);
                Assert.assertEquals(rackSummaries + "\nSecond rack sorted for " + comp, "rack-000", sortedRacks.get(1).id);
            } else {
                Assert.assertEquals(rackSummaries + "\nFirst rack sorted for " + comp, "rack-000", sortedRacks.get(0).id);
                Assert.assertEquals(rackSummaries + "\nSecond rack sorted for " + comp, "rack-001", sortedRacks.get(1).id);
            }
        }
    }
    scheduler.schedule(topologies, cluster);
    Map<String, SchedulerAssignment> assignments = new TreeMap<>(cluster.getAssignments());
    assertEquals(2, assignments.size());
    Map<String, Map<String, AtomicLong>> topoPerRackCount = new HashMap<>();
    for (Map.Entry<String, SchedulerAssignment> entry : assignments.entrySet()) {
        SchedulerAssignment sa = entry.getValue();
        Map<String, AtomicLong> slotsPerRack = new TreeMap<>();
        for (WorkerSlot slot : sa.getSlots()) {
            String nodeId = slot.getNodeId();
            String rack = supervisorIdToRackName(nodeId);
            slotsPerRack.computeIfAbsent(rack, (r) -> new AtomicLong(0)).incrementAndGet();
        }
        LOG.info("{} => {}", entry.getKey(), slotsPerRack);
        topoPerRackCount.put(entry.getKey(), slotsPerRack);
    }
    Map<String, AtomicLong> simpleCount = topoPerRackCount.get("topology-simple-0");
    assertNotNull(simpleCount);
    // Because the simple topology was scheduled first we want to be sure that it didn't put anything on
    // the GPU nodes.
    // Only 1 rack is in use
    assertEquals(1, simpleCount.size());
    // r001 is the second rack with GPUs
    assertFalse(simpleCount.containsKey("r001"));
    // r000 is the first rack with no GPUs
    assertTrue(simpleCount.containsKey("r000"));
// We don't really care too much about the scheduling of topology-gpu-0, because it was scheduled.
}

Example 29 with ExecutorDetails

use of org.apache.storm.scheduler.ExecutorDetails in project storm by apache.

the class TestConstraintSolverStrategy method testIntegrationWithRAS.

@Test
public void testIntegrationWithRAS() {
    if (!consolidatedConfigFlag) {
        LOG.info("Skipping test since bolt-1 maxCoLocationCnt=10 requires consolidatedConfigFlag=true, current={}", consolidatedConfigFlag);
        return;
    }
    Map<String, Object> config = Utils.readDefaultConfig();
    config.put(Config.TOPOLOGY_SCHEDULER_STRATEGY, ConstraintSolverStrategy.class.getName());
    config.put(Config.TOPOLOGY_RAS_CONSTRAINT_MAX_STATE_SEARCH, MAX_TRAVERSAL_DEPTH);
    config.put(Config.TOPOLOGY_WORKER_MAX_HEAP_SIZE_MB, 100_000);
    config.put(Config.TOPOLOGY_PRIORITY, 1);
    config.put(Config.TOPOLOGY_COMPONENT_CPU_PCORE_PERCENT, 10);
    config.put(Config.TOPOLOGY_COMPONENT_RESOURCES_ONHEAP_MEMORY_MB, 100);
    config.put(Config.TOPOLOGY_COMPONENT_RESOURCES_OFFHEAP_MEMORY_MB, 0.0);
    List<List<String>> constraints = new LinkedList<>();
    addConstraints("spout-0", "bolt-0", constraints);
    addConstraints("bolt-1", "bolt-1", constraints);
    addConstraints("bolt-1", "bolt-2", constraints);
    Map<String, Integer> spreads = new HashMap<String, Integer>();
    spreads.put("spout-0", 1);
    spreads.put("bolt-1", 10);
    setConstraintConfig(constraints, spreads, config);
    TopologyDetails topo = genTopology("testTopo", config, 2, 3, 30, 300, 0, 0, "user");
    Map<String, TopologyDetails> topoMap = new HashMap<>();
    topoMap.put(topo.getId(), topo);
    Topologies topologies = new Topologies(topoMap);
    // Fails with 36 supervisors, works with 37
    Map<String, SupervisorDetails> supMap = genSupervisors(37, 16, 400, 1024 * 4);
    Cluster cluster = makeCluster(topologies, supMap);
    ResourceAwareScheduler rs = new ResourceAwareScheduler();
    rs.prepare(config, new StormMetricsRegistry());
    try {
        rs.schedule(topologies, cluster);
        assertStatusSuccess(cluster, topo.getId());
        Assert.assertEquals("topo all executors scheduled?", 0, cluster.getUnassignedExecutors(topo).size());
    } finally {
        rs.cleanup();
    }
    // simulate worker loss
    Map<ExecutorDetails, WorkerSlot> newExecToSlot = new HashMap<>();
    Map<ExecutorDetails, WorkerSlot> execToSlot = cluster.getAssignmentById(topo.getId()).getExecutorToSlot();
    Iterator<Map.Entry<ExecutorDetails, WorkerSlot>> it = execToSlot.entrySet().iterator();
    for (int i = 0; i < execToSlot.size() / 2; i++) {
        ExecutorDetails exec = it.next().getKey();
        WorkerSlot ws = it.next().getValue();
        newExecToSlot.put(exec, ws);
    }
    Map<String, SchedulerAssignment> newAssignments = new HashMap<>();
    newAssignments.put(topo.getId(), new SchedulerAssignmentImpl(topo.getId(), newExecToSlot, null, null));
    cluster.setAssignments(newAssignments, false);
    rs.prepare(config, new StormMetricsRegistry());
    try {
        rs.schedule(topologies, cluster);
        assertStatusSuccess(cluster, topo.getId());
        Assert.assertEquals("topo all executors scheduled?", 0, cluster.getUnassignedExecutors(topo).size());
    } finally {
        rs.cleanup();
    }
}

Example 30 with ExecutorDetails

use of org.apache.storm.scheduler.ExecutorDetails in project storm by apache.

the class TestGenericResourceAwareStrategy method testGenericResourceAwareStrategyWithSettingAckerExecutors.

/**
 * Test if the scheduling logic for the GenericResourceAwareStrategy is correct
 * with setting {@link Config#TOPOLOGY_ACKER_EXECUTORS}.
 *
 * Test details refer to {@link TestDefaultResourceAwareStrategy#testDefaultResourceAwareStrategyWithSettingAckerExecutors(int)}
 */
@ParameterizedTest
@ValueSource(ints = { -1, 0, 2, 200 })
public void testGenericResourceAwareStrategyWithSettingAckerExecutors(int numOfAckersPerWorker) throws InvalidTopologyException {
    int spoutParallelism = 1;
    int boltParallelism = 2;
    TopologyBuilder builder = new TopologyBuilder();
    builder.setSpout("spout", new TestSpout(), spoutParallelism);
    builder.setBolt("bolt-1", new TestBolt(), boltParallelism).shuffleGrouping("spout");
    builder.setBolt("bolt-2", new TestBolt(), boltParallelism).shuffleGrouping("bolt-1").addResource("gpu.count", 1.0);
    builder.setBolt("bolt-3", new TestBolt(), boltParallelism).shuffleGrouping("bolt-2").addResource("gpu.count", 2.0);
    String topoName = "testTopology";
    StormTopology stormToplogy = builder.createTopology();
    INimbus iNimbus = new INimbusTest();
    Config conf = createGrasClusterConfig(50, 500, 0, null, Collections.emptyMap());
    Map<String, Double> genericResourcesMap = new HashMap<>();
    genericResourcesMap.put("gpu.count", 2.0);
    Map<String, SupervisorDetails> supMap = genSupervisors(4, 4, 200, 2000, genericResourcesMap);
    conf.put(Config.TOPOLOGY_PRIORITY, 0);
    conf.put(Config.TOPOLOGY_NAME, topoName);
    conf.put(Config.TOPOLOGY_WORKER_MAX_HEAP_SIZE_MB, 2000);
    conf.put(Config.TOPOLOGY_SUBMITTER_USER, "user");
    conf.put(Config.TOPOLOGY_ACKER_EXECUTORS, 4);
    if (numOfAckersPerWorker == -1) {
    // Leave topology.acker.executors.per.worker unset
    } else {
        conf.put(Config.TOPOLOGY_RAS_ACKER_EXECUTORS_PER_WORKER, numOfAckersPerWorker);
    }
    int estimatedNumWorker = ServerUtils.getEstimatedWorkerCountForRasTopo(conf, stormToplogy);
    Nimbus.setUpAckerExecutorConfigs(topoName, conf, conf, estimatedNumWorker);
    conf.put(Config.TOPOLOGY_ACKER_RESOURCES_ONHEAP_MEMORY_MB, 250);
    conf.put(Config.TOPOLOGY_ACKER_CPU_PCORE_PERCENT, 50);
    TopologyDetails topo = new TopologyDetails("testTopology-id", conf, stormToplogy, 0, genExecsAndComps(StormCommon.systemTopology(conf, stormToplogy)), currentTime, "user");
    Topologies topologies = new Topologies(topo);
    Cluster cluster = new Cluster(iNimbus, new ResourceMetrics(new StormMetricsRegistry()), supMap, new HashMap<>(), topologies, conf);
    scheduler = new ResourceAwareScheduler();
    scheduler.prepare(conf, new StormMetricsRegistry());
    scheduler.schedule(topologies, cluster);
    // We need to have 3 slots on 3 separate hosts. The topology needs 6 GPUs 3500 MB memory and 350% CPU
    // The bolt-3 instances must be on separate nodes because they each need 2 GPUs.
    // The bolt-2 instances must be on the same node as they each need 1 GPU
    // (this assumes that we are packing the components to avoid fragmentation).
    // The bolt-1 and spout instances fill in the rest.
    // Ordered execs: [[6, 6], [2, 2], [4, 4], [5, 5], [1, 1], [3, 3], [0, 0]]
    // Ackers: [[8, 8], [7, 7]] (+ [[9, 9], [10, 10]] when numOfAckersPerWorker=2)
    HashSet<HashSet<ExecutorDetails>> expectedScheduling = new HashSet<>();
    expectedScheduling.add(new HashSet<>(Arrays.asList(// bolt-3 - 500 MB, 50% CPU, 2 GPU
    new ExecutorDetails(3, 3))));
    // Total 500 MB, 50% CPU, 2 - GPU -> this node has 1500 MB, 150% cpu, 0 GPU left
    expectedScheduling.add(new HashSet<>(Arrays.asList(// acker - 250 MB, 50% CPU, 0 GPU
    new ExecutorDetails(7, 7), // acker - 250 MB, 50% CPU, 0 GPU
    new ExecutorDetails(8, 8), // bolt-2 - 500 MB, 50% CPU, 1 GPU
    new ExecutorDetails(6, 6), // bolt-1 - 500 MB, 50% CPU, 0 GPU
    new ExecutorDetails(2, 2))));
    // Total 1500 MB, 200% CPU, 2 GPU -> this node has 500 MB, 0% CPU, 0 GPU left
    expectedScheduling.add(new HashSet<>(Arrays.asList(// acker- 250 MB, 50% CPU, 0 GPU
    new ExecutorDetails(9, 9), // acker- 250 MB, 50% CPU, 0 GPU
    new ExecutorDetails(10, 10), // bolt-1 - 500 MB, 50% CPU, 0 GPU
    new ExecutorDetails(1, 1), // bolt-3 500 MB, 50% cpu, 2 GPU
    new ExecutorDetails(4, 4))));
    // Total 1500 MB, 200% CPU, 2 GPU -> this node has 500 MB, 0% CPU, 0 GPU left
    expectedScheduling.add(new HashSet<>(Arrays.asList(// Spout - 500 MB, 50% CPU, 0 GPU
    new ExecutorDetails(0, 0), // bolt-2 - 500 MB, 50% CPU, 1 GPU
    new ExecutorDetails(5, 5))));
    // Total 1000 MB, 100% CPU, 2 GPU -> this node has 1000 MB, 100% CPU, 0 GPU left
    HashSet<HashSet<ExecutorDetails>> foundScheduling = new HashSet<>();
    SchedulerAssignment assignment = cluster.getAssignmentById("testTopology-id");
    for (Collection<ExecutorDetails> execs : assignment.getSlotToExecutors().values()) {
        foundScheduling.add(new HashSet<>(execs));
    }
    assertEquals(expectedScheduling, foundScheduling);
}