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

Example 71 with Topologies

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

the class TestResourceAwareScheduler method testScheduleResilience.

@Test
public void testScheduleResilience() {
    INimbus iNimbus = new INimbusTest();
    Map<String, SupervisorDetails> supMap = genSupervisors(2, 2, 400, 2000);
    TopologyBuilder builder1 = new TopologyBuilder();
    builder1.setSpout("wordSpout1", new TestWordSpout(), 3);
    StormTopology stormTopology1 = builder1.createTopology();
    Config config1 = new Config();
    config1.putAll(defaultTopologyConf);
    Map<ExecutorDetails, String> executorMap1 = genExecsAndComps(stormTopology1);
    TopologyDetails topology1 = new TopologyDetails("topology1", config1, stormTopology1, 3, executorMap1, 0, "user");
    TopologyBuilder builder2 = new TopologyBuilder();
    builder2.setSpout("wordSpout2", new TestWordSpout(), 2);
    StormTopology stormTopology2 = builder2.createTopology();
    Config config2 = new Config();
    config2.putAll(defaultTopologyConf);
    // memory requirement is large enough so that two executors can not be fully assigned to one node
    config2.put(Config.TOPOLOGY_COMPONENT_RESOURCES_ONHEAP_MEMORY_MB, 1280.0);
    Map<ExecutorDetails, String> executorMap2 = genExecsAndComps(stormTopology2);
    TopologyDetails topology2 = new TopologyDetails("topology2", config2, stormTopology2, 2, executorMap2, 0, "user");
    // Test1: When a worker fails, RAS does not alter existing assignments on healthy workers
    scheduler = new ResourceAwareScheduler();
    Topologies topologies = new Topologies(topology2);
    Cluster cluster = new Cluster(iNimbus, new ResourceMetrics(new StormMetricsRegistry()), supMap, new HashMap<>(), topologies, config1);
    scheduler.prepare(config1, new StormMetricsRegistry());
    scheduler.schedule(topologies, cluster);
    SchedulerAssignment assignment = cluster.getAssignmentById(topology2.getId());
    // pick a worker to mock as failed
    WorkerSlot failedWorker = new ArrayList<>(assignment.getSlots()).get(0);
    Map<ExecutorDetails, WorkerSlot> executorToSlot = assignment.getExecutorToSlot();
    List<ExecutorDetails> failedExecutors = new ArrayList<>();
    for (Map.Entry<ExecutorDetails, WorkerSlot> entry : executorToSlot.entrySet()) {
        if (entry.getValue().equals(failedWorker)) {
            failedExecutors.add(entry.getKey());
        }
    }
    for (ExecutorDetails executor : failedExecutors) {
        // remove executor details assigned to the failed worker
        executorToSlot.remove(executor);
    }
    Map<ExecutorDetails, WorkerSlot> copyOfOldMapping = new HashMap<>(executorToSlot);
    Set<ExecutorDetails> healthyExecutors = copyOfOldMapping.keySet();
    scheduler.schedule(topologies, cluster);
    SchedulerAssignment newAssignment = cluster.getAssignmentById(topology2.getId());
    Map<ExecutorDetails, WorkerSlot> newExecutorToSlot = newAssignment.getExecutorToSlot();
    for (ExecutorDetails executor : healthyExecutors) {
        assertEquals(copyOfOldMapping.get(executor), newExecutorToSlot.get(executor));
    }
    assertFalse(cluster.needsSchedulingRas(topology2));
    assertTrue(cluster.getStatusMap().get(topology2.getId()).startsWith("Running - Fully Scheduled by DefaultResourceAwareStrategy"));
    // end of Test1
    // Test2: When a supervisor fails, RAS does not alter existing assignments
    executorToSlot = new HashMap<>();
    executorToSlot.put(new ExecutorDetails(0, 0), new WorkerSlot("r000s000", 0));
    executorToSlot.put(new ExecutorDetails(1, 1), new WorkerSlot("r000s000", 1));
    executorToSlot.put(new ExecutorDetails(2, 2), new WorkerSlot("r000s001", 1));
    Map<String, SchedulerAssignment> existingAssignments = new HashMap<>();
    assignment = new SchedulerAssignmentImpl(topology1.getId(), executorToSlot, null, null);
    existingAssignments.put(topology1.getId(), assignment);
    copyOfOldMapping = new HashMap<>(executorToSlot);
    Set<ExecutorDetails> existingExecutors = copyOfOldMapping.keySet();
    Map<String, SupervisorDetails> supMap1 = new HashMap<>(supMap);
    // mock the supervisor r000s000 as a failed supervisor
    supMap1.remove("r000s000");
    topologies = new Topologies(topology1);
    Cluster cluster1 = new Cluster(iNimbus, new ResourceMetrics(new StormMetricsRegistry()), supMap1, existingAssignments, topologies, config1);
    scheduler.schedule(topologies, cluster1);
    newAssignment = cluster1.getAssignmentById(topology1.getId());
    newExecutorToSlot = newAssignment.getExecutorToSlot();
    for (ExecutorDetails executor : existingExecutors) {
        assertEquals(copyOfOldMapping.get(executor), newExecutorToSlot.get(executor));
    }
    assertEquals("Fully Scheduled", cluster1.getStatusMap().get(topology1.getId()));
    // end of Test2
    // Test3: When a supervisor and a worker on it fails, RAS does not alter existing assignments
    executorToSlot = new HashMap<>();
    // the worker to orphan
    executorToSlot.put(new ExecutorDetails(0, 0), new WorkerSlot("r000s000", 1));
    // the worker that fails
    executorToSlot.put(new ExecutorDetails(1, 1), new WorkerSlot("r000s000", 2));
    // the healthy worker
    executorToSlot.put(new ExecutorDetails(2, 2), new WorkerSlot("r000s001", 1));
    existingAssignments = new HashMap<>();
    assignment = new SchedulerAssignmentImpl(topology1.getId(), executorToSlot, null, null);
    existingAssignments.put(topology1.getId(), assignment);
    // delete one worker of r000s000 (failed) from topo1 assignment to enable actual schedule for testing
    executorToSlot.remove(new ExecutorDetails(1, 1));
    copyOfOldMapping = new HashMap<>(executorToSlot);
    // namely the two eds on the orphaned worker and the healthy worker
    existingExecutors = copyOfOldMapping.keySet();
    supMap1 = new HashMap<>(supMap);
    // mock the supervisor r000s000 as a failed supervisor
    supMap1.remove("r000s000");
    topologies = new Topologies(topology1);
    cluster1 = new Cluster(iNimbus, new ResourceMetrics(new StormMetricsRegistry()), supMap1, existingAssignments, topologies, config1);
    scheduler.schedule(topologies, cluster1);
    newAssignment = cluster1.getAssignmentById(topology1.getId());
    newExecutorToSlot = newAssignment.getExecutorToSlot();
    for (ExecutorDetails executor : existingExecutors) {
        assertEquals(copyOfOldMapping.get(executor), newExecutorToSlot.get(executor));
    }
    assertFalse(cluster1.needsSchedulingRas(topology1));
    assertEquals("Fully Scheduled", cluster1.getStatusMap().get(topology1.getId()));
    // end of Test3
    // Test4: Scheduling a new topology does not disturb other assignments unnecessarily
    topologies = new Topologies(topology1);
    cluster1 = new Cluster(iNimbus, new ResourceMetrics(new StormMetricsRegistry()), supMap, new HashMap<>(), topologies, config1);
    scheduler.schedule(topologies, cluster1);
    assignment = cluster1.getAssignmentById(topology1.getId());
    executorToSlot = assignment.getExecutorToSlot();
    copyOfOldMapping = new HashMap<>(executorToSlot);
    topologies = addTopologies(topologies, topology2);
    cluster1 = new Cluster(iNimbus, new ResourceMetrics(new StormMetricsRegistry()), supMap, new HashMap<>(), topologies, config1);
    scheduler.schedule(topologies, cluster1);
    newAssignment = cluster1.getAssignmentById(topology1.getId());
    newExecutorToSlot = newAssignment.getExecutorToSlot();
    for (ExecutorDetails executor : copyOfOldMapping.keySet()) {
        assertEquals(copyOfOldMapping.get(executor), newExecutorToSlot.get(executor));
    }
    assertFalse(cluster1.needsSchedulingRas(topology1));
    assertFalse(cluster1.needsSchedulingRas(topology2));
    String expectedStatusPrefix = "Running - Fully Scheduled by DefaultResourceAwareStrategy";
    assertTrue(cluster1.getStatusMap().get(topology1.getId()).startsWith(expectedStatusPrefix));
    assertTrue(cluster1.getStatusMap().get(topology2.getId()).startsWith(expectedStatusPrefix));
}

Example 72 with Topologies

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

the class TestResourceAwareScheduler method testNodeFreeSlot.

/**
 * test if free slots on nodes work correctly
 */
@Test
public void testNodeFreeSlot() {
    INimbus iNimbus = new INimbusTest();
    Map<String, SupervisorDetails> supMap = genSupervisors(4, 4, 100, 1000);
    Config config = createClusterConfig(100, 500, 500, null);
    Topologies topologies = new Topologies(genTopology("topo-1", config, 1, 0, 2, 0, currentTime - 2, 29, "user"), genTopology("topo-2", config, 1, 0, 2, 0, currentTime - 2, 10, "user"));
    Cluster cluster = new Cluster(iNimbus, new ResourceMetrics(new StormMetricsRegistry()), supMap, new HashMap<>(), topologies, config);
    scheduler = new ResourceAwareScheduler();
    scheduler.prepare(config, new StormMetricsRegistry());
    scheduler.schedule(topologies, cluster);
    Map<String, RasNode> nodes = RasNodes.getAllNodesFrom(cluster);
    for (SchedulerAssignment assignment : cluster.getAssignments().values()) {
        for (Entry<WorkerSlot, WorkerResources> entry : new HashMap<>(assignment.getScheduledResources()).entrySet()) {
            WorkerSlot ws = entry.getKey();
            WorkerResources wr = entry.getValue();
            double memoryBefore = nodes.get(ws.getNodeId()).getAvailableMemoryResources();
            double cpuBefore = nodes.get(ws.getNodeId()).getAvailableCpuResources();
            double memoryUsedByWorker = wr.get_mem_on_heap() + wr.get_mem_off_heap();
            assertEquals("Check if memory used by worker is calculated correctly", 1000.0, memoryUsedByWorker, 0.001);
            double cpuUsedByWorker = wr.get_cpu();
            assertEquals("Check if CPU used by worker is calculated correctly", 100.0, cpuUsedByWorker, 0.001);
            nodes.get(ws.getNodeId()).free(ws);
            double memoryAfter = nodes.get(ws.getNodeId()).getAvailableMemoryResources();
            double cpuAfter = nodes.get(ws.getNodeId()).getAvailableCpuResources();
            assertEquals("Check if free correctly frees amount of memory", memoryBefore + memoryUsedByWorker, memoryAfter, 0.001);
            assertEquals("Check if free correctly frees amount of memory", cpuBefore + cpuUsedByWorker, cpuAfter, 0.001);
            assertFalse("Check if worker was removed from assignments", assignment.getSlotToExecutors().containsKey(ws));
        }
    }
}

Example 73 with Topologies

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

the class TestResourceAwareScheduler method testLargeTopologiesCommon.

public void testLargeTopologiesCommon(final String strategy, final boolean includeGpu, final int multiplier) {
    INimbus iNimbus = new INimbusTest();
    Map<String, SupervisorDetails> supMap = genSupervisorsWithRacks(25 * multiplier, 40, 66, 3 * multiplier, 0, 4700, 226200, new HashMap<>());
    if (includeGpu) {
        HashMap<String, Double> extraResources = new HashMap<>();
        extraResources.put("my.gpu", 1.0);
        supMap.putAll(genSupervisorsWithRacks(3 * multiplier, 40, 66, 0, 0, 4700, 226200, extraResources));
    }
    Config config = new Config();
    config.putAll(createClusterConfig(88, 775, 25, null));
    config.put(Config.TOPOLOGY_SCHEDULER_STRATEGY, strategy);
    scheduler = new ResourceAwareScheduler();
    Map<String, TopologyDetails> topologyDetailsMap = new HashMap<>();
    for (int i = 0; i < 11 * multiplier; i++) {
        TopologyDetails td = genTopology(String.format("topology-%05d", i), config, 5, 40, 30, 114, 0, 0, "user", 8192);
        topologyDetailsMap.put(td.getId(), td);
    }
    if (includeGpu) {
        for (int i = 0; i < multiplier; i++) {
            TopologyBuilder builder = topologyBuilder(5, 40, 30, 114);
            builder.setBolt("gpu-bolt", new TestBolt(), 40).addResource("my.gpu", 1.0).shuffleGrouping("spout-0");
            TopologyDetails td = topoToTopologyDetails(String.format("topology-gpu-%05d", i), config, builder.createTopology(), 0, 0, "user", 8192);
            topologyDetailsMap.put(td.getId(), td);
        }
    }
    Topologies topologies = new Topologies(topologyDetailsMap);
    Cluster cluster = new Cluster(iNimbus, new ResourceMetrics(new StormMetricsRegistry()), supMap, new HashMap<>(), topologies, config);
    long startTime = Time.currentTimeMillis();
    scheduler.prepare(config, new StormMetricsRegistry());
    scheduler.schedule(topologies, cluster);
    long schedulingDuration = Time.currentTimeMillis() - startTime;
    LOG.info("Scheduling took " + schedulingDuration + " ms");
    LOG.info("HAS {} SLOTS USED", cluster.getUsedSlots().size());
    Map<String, SchedulerAssignment> assignments = new TreeMap<>(cluster.getAssignments());
    for (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);
    }
}

Example 74 with Topologies

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

the class TestNodeSorterHostProximity method testMultipleRacksWithHostProximity.

/**
 * Test if hosts are presented together regardless of resource availability.
 * Supervisors are created with multiple Numa zones in such a manner that resources on two numa zones on the same host
 * differ widely in resource availability.
 */
@Test
public void testMultipleRacksWithHostProximity() {
    final Map<String, SupervisorDetails> supMap = new HashMap<>();
    final int numRacks = 1;
    final int numSupersPerRack = 12;
    final int numPortsPerSuper = 4;
    final int numZonesPerHost = 3;
    final double numaResourceMultiplier = 0.4;
    int rackStartNum = 0;
    int supStartNum = 0;
    final Map<String, SupervisorDetails> supMapRack0 = genSupervisorsWithRacksAndNuma(numRacks, numSupersPerRack, numZonesPerHost, numPortsPerSuper, rackStartNum++, supStartNum, 400, 8000, Collections.emptyMap(), numaResourceMultiplier);
    // generate another rack of supervisors with less resources
    supStartNum += numSupersPerRack;
    final Map<String, SupervisorDetails> supMapRack1 = genSupervisorsWithRacksAndNuma(numRacks, numSupersPerRack, numZonesPerHost, numPortsPerSuper, rackStartNum++, supStartNum, 200, 4000, Collections.emptyMap(), numaResourceMultiplier);
    // generate some supervisors that are depleted of one resource
    supStartNum += numSupersPerRack;
    final Map<String, SupervisorDetails> supMapRack2 = genSupervisorsWithRacksAndNuma(numRacks, numSupersPerRack, numZonesPerHost, numPortsPerSuper, rackStartNum++, supStartNum, 0, 8000, Collections.emptyMap(), numaResourceMultiplier);
    // generate some that has a lot of memory but little of cpu
    supStartNum += numSupersPerRack;
    final Map<String, SupervisorDetails> supMapRack3 = genSupervisorsWithRacksAndNuma(numRacks, numSupersPerRack, numZonesPerHost, numPortsPerSuper, rackStartNum++, supStartNum, 10, 8000 * 2 + 4000, Collections.emptyMap(), numaResourceMultiplier);
    // generate some that has a lot of cpu but little of memory
    supStartNum += numSupersPerRack;
    final Map<String, SupervisorDetails> supMapRack4 = genSupervisorsWithRacksAndNuma(numRacks, numSupersPerRack, numZonesPerHost, numPortsPerSuper, rackStartNum++, supStartNum, 400 + 200 + 10, 1000, Collections.emptyMap(), numaResourceMultiplier);
    supMap.putAll(supMapRack0);
    supMap.putAll(supMapRack1);
    supMap.putAll(supMapRack2);
    supMap.putAll(supMapRack3);
    supMap.putAll(supMapRack4);
    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
    TestDNSToSwitchMapping testDNSToSwitchMapping = new TestDNSToSwitchMapping(supMapRack0, supMapRack1, supMapRack2, supMapRack3, supMapRack4);
    Config t1Conf = new Config();
    t1Conf.putAll(config);
    final List<String> t1FavoredHostNames = Arrays.asList("host-41", "host-42", "host-43");
    t1Conf.put(Config.TOPOLOGY_SCHEDULER_FAVORED_NODES, t1FavoredHostNames);
    final List<String> t1UnfavoredHostIds = Arrays.asList("host-1", "host-2", "host-3");
    t1Conf.put(Config.TOPOLOGY_SCHEDULER_UNFAVORED_NODES, t1UnfavoredHostIds);
    // generate topologies
    TopologyDetails topo1 = genTopology("topo-1", t1Conf, 8, 0, 2, 0, CURRENT_TIME - 2, 10, "user");
    Config t2Conf = new Config();
    t2Conf.putAll(config);
    t2Conf.put(Config.TOPOLOGY_SCHEDULER_FAVORED_NODES, Arrays.asList("host-31", "host-32", "host-33"));
    t2Conf.put(Config.TOPOLOGY_SCHEDULER_UNFAVORED_NODES, Arrays.asList("host-11", "host-12", "host-13"));
    TopologyDetails topo2 = genTopology("topo-2", t2Conf, 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);
    cluster.setNetworkTopography(testDNSToSwitchMapping.getRackToHosts());
    INodeSorter nodeSorter = new NodeSorterHostProximity(cluster, topo1);
    nodeSorter.prepare(null);
    Set<String> seenHosts = new HashSet<>();
    String prevHost = null;
    List<String> errLines = new ArrayList();
    Map<String, String> nodeToHost = new RasNodes(cluster).getNodeIdToHostname();
    for (String nodeId : nodeSorter.sortAllNodes()) {
        String host = nodeToHost.getOrDefault(nodeId, "no-host-for-node-" + nodeId);
        errLines.add(String.format("\tnodeId:%s, host:%s", nodeId, host));
        if (!host.equals(prevHost) && seenHosts.contains(host)) {
            String err = String.format("Host %s for node %s is out of order:\n\t%s", host, nodeId, String.join("\n\t", errLines));
            Assert.fail(err);
        }
        seenHosts.add(host);
        prevHost = host;
    }
}

Example 75 with Topologies

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

the class TestNodeSorterHostProximity method testWithBlackListedHosts.

/**
 * Black list all nodes for a rack before sorting nodes.
 * Confirm that {@link NodeSorterHostProximity#sortAllNodes()} still works.
 */
@Test
void testWithBlackListedHosts() {
    INimbus iNimbus = new INimbusTest();
    double compPcore = 100;
    double compOnHeap = 775;
    double compOffHeap = 25;
    int topo1NumSpouts = 1;
    int topo1NumBolts = 5;
    int topo1SpoutParallelism = 100;
    int topo1BoltParallelism = 200;
    final int numSupersPerRack = 10;
    final int numPortsPerSuper = 66;
    long compPerRack = (topo1NumSpouts * topo1SpoutParallelism + topo1NumBolts * topo1BoltParallelism + 10);
    long compPerSuper = compPerRack / numSupersPerRack;
    double cpuPerSuper = compPcore * compPerSuper;
    double memPerSuper = (compOnHeap + compOffHeap) * compPerSuper;
    double topo1MaxHeapSize = memPerSuper;
    final String topoName1 = "topology1";
    int numRacks = 3;
    Map<String, SupervisorDetails> supMap = genSupervisorsWithRacks(numRacks, numSupersPerRack, numPortsPerSuper, 0, 0, cpuPerSuper, memPerSuper, new HashMap<>());
    TestDNSToSwitchMapping testDNSToSwitchMapping = new TestDNSToSwitchMapping(supMap.values());
    Config config = new Config();
    config.putAll(createGrasClusterConfig(compPcore, compOnHeap, compOffHeap, null, null));
    config.put(Config.TOPOLOGY_SCHEDULER_STRATEGY, GenericResourceAwareStrategy.class.getName());
    IScheduler scheduler = new ResourceAwareScheduler();
    scheduler.prepare(config, new StormMetricsRegistry());
    TopologyDetails td1 = genTopology(topoName1, config, topo1NumSpouts, topo1NumBolts, topo1SpoutParallelism, topo1BoltParallelism, 0, 0, "user", topo1MaxHeapSize);
    Topologies topologies = new Topologies(td1);
    Cluster cluster = new Cluster(iNimbus, new ResourceMetrics(new StormMetricsRegistry()), supMap, new HashMap<>(), topologies, config);
    cluster.setNetworkTopography(testDNSToSwitchMapping.getRackToHosts());
    Map<String, List<String>> networkTopography = cluster.getNetworkTopography();
    assertEquals("Expecting " + numRacks + " racks found " + networkTopography.size(), numRacks, networkTopography.size());
    assertTrue("Expecting racks count to be >= 3, found " + networkTopography.size(), networkTopography.size() >= 3);
    Set<String> blackListedHosts = new HashSet<>();
    List<SupervisorDetails> supArray = new ArrayList<>(supMap.values());
    for (int i = 0; i < numSupersPerRack; i++) {
        blackListedHosts.add(supArray.get(i).getHost());
    }
    blacklistHostsAndSortNodes(blackListedHosts, supMap.values(), cluster, td1);
    String rackToClear = cluster.getNetworkTopography().keySet().stream().findFirst().get();
    blackListedHosts = new HashSet<>(cluster.getNetworkTopography().get(rackToClear));
    blacklistHostsAndSortNodes(blackListedHosts, supMap.values(), cluster, td1);
}