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

Example 36 with ResourceMetrics

use of org.apache.storm.scheduler.resource.normalization.ResourceMetrics in project storm by apache.

the class TestGenericResourceAwareStrategy method testGenericResourceAwareStrategyWithoutSettingAckerExecutors.

/**
 * Test if the scheduling logic for the GenericResourceAwareStrategy is correct
 * without setting {@link Config#TOPOLOGY_ACKER_EXECUTORS}.
 *
 * Test details refer to {@link TestDefaultResourceAwareStrategy#testDefaultResourceAwareStrategyWithoutSettingAckerExecutors(int)}
 */
@ParameterizedTest
@ValueSource(ints = { -1, 0, 1, 2 })
public void testGenericResourceAwareStrategyWithoutSettingAckerExecutors(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");
    // Parameterized test on different numOfAckersPerWorker
    if (numOfAckersPerWorker == -1) {
    // Both Config.TOPOLOGY_ACKER_EXECUTORS and Config.TOPOLOGY_RAS_ACKER_EXECUTORS_PER_WORKER are not set
    // Default will be 2 (estimate num of workers) and 1 respectively
    } 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<>();
    if (numOfAckersPerWorker == -1 || numOfAckersPerWorker == 1) {
        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(// bolt-2 - 500 MB, 50% CPU, 1 GPU
        new ExecutorDetails(6, 6), // bolt-1 - 500 MB, 50% CPU, 0 GPU
        new ExecutorDetails(2, 2), // bolt-2 - 500 MB, 50% CPU, 1 GPU
        new ExecutorDetails(5, 5), // acker - 250 MB, 50% CPU, 0 GPU
        new ExecutorDetails(8, 8))));
        // Total 1750 MB, 200% CPU, 2 GPU -> this node has 250 MB, 0% CPU, 0 GPU left
        expectedScheduling.add(new HashSet<>(Arrays.asList(// bolt-3 500 MB, 50% cpu, 2 GPU
        new ExecutorDetails(4, 4), // bolt-1 - 500 MB, 50% CPU, 0 GPU
        new ExecutorDetails(1, 1), // Spout - 500 MB, 50% CPU, 0 GPU
        new ExecutorDetails(0, 0), // acker - 250 MB, 50% CPU, 0 GPU
        new ExecutorDetails(7, 7))));
    // Total 1750 MB, 200% CPU, 2 GPU -> this node has 250 MB, 0% CPU, 0 GPU left
    } else if (numOfAckersPerWorker == 0) {
        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(// bolt-2 - 500 MB, 50% CPU, 1 GPU
        new ExecutorDetails(6, 6), // bolt-1 - 500 MB, 50% CPU, 0 GPU
        new ExecutorDetails(2, 2), // bolt-2 - 500 MB, 50% CPU, 1 GPU
        new ExecutorDetails(5, 5), // bolt-1 - 500 MB, 50% CPU, 0 GPU
        new ExecutorDetails(1, 1))));
        // Total 2000 MB, 200% CPU, 2 GPU -> this node has 0 MB, 0% CPU, 0 GPU left
        expectedScheduling.add(new HashSet<>(Arrays.asList(// Spout - 500 MB, 50% CPU, 0 GPU
        new ExecutorDetails(0, 0), // bolt-3 500 MB, 50% cpu, 2 GPU
        new ExecutorDetails(4, 4))));
    // Total 1000 MB, 100% CPU, 2 GPU -> this node has 1000 MB, 100% CPU, 0 GPU left
    } else if (numOfAckersPerWorker == 2) {
        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);
}

Example 37 with ResourceMetrics

use of org.apache.storm.scheduler.resource.normalization.ResourceMetrics in project storm by apache.

the class TestGenericResourceAwareStrategy method testGenericResourceAwareStrategyInFavorOfShuffle.

/**
 * test if the scheduling logic for the GenericResourceAwareStrategy (when in favor of shuffle) is correct.
 */
@Test
public void testGenericResourceAwareStrategyInFavorOfShuffle() 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);
    StormTopology stormToplogy = builder.createTopology();
    INimbus iNimbus = new INimbusTest();
    Config conf = createGrasClusterConfig(50, 250, 250, 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, "testTopology");
    conf.put(Config.TOPOLOGY_WORKER_MAX_HEAP_SIZE_MB, Double.MAX_VALUE);
    conf.put(Config.TOPOLOGY_SUBMITTER_USER, "user");
    conf.put(Config.TOPOLOGY_RAS_ORDER_EXECUTORS_BY_PROXIMITY_NEEDS, true);
    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);
    ResourceAwareScheduler rs = new ResourceAwareScheduler();
    rs.prepare(conf, new StormMetricsRegistry());
    rs.schedule(topologies, cluster);
    // Sorted execs: [[0, 0], [2, 2], [6, 6], [4, 4], [1, 1], [5, 5], [3, 3], [7, 7]]
    // Ackers: [[7, 7]]]
    HashSet<HashSet<ExecutorDetails>> expectedScheduling = new HashSet<>();
    expectedScheduling.add(new HashSet<>(Arrays.asList(// spout
    new ExecutorDetails(0, 0), // bolt-1
    new ExecutorDetails(2, 2), // bolt-2
    new ExecutorDetails(6, 6), // acker
    new ExecutorDetails(7, 7))));
    expectedScheduling.add(new HashSet<>(Arrays.asList(// bolt-3
    new ExecutorDetails(4, 4), // bolt-1
    new ExecutorDetails(1, 1))));
    // bolt-2
    expectedScheduling.add(new HashSet<>(Arrays.asList(new ExecutorDetails(5, 5))));
    // bolt-3
    expectedScheduling.add(new HashSet<>(Arrays.asList(new ExecutorDetails(3, 3))));
    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);
}

Example 38 with ResourceMetrics

use of org.apache.storm.scheduler.resource.normalization.ResourceMetrics in project storm by apache.

the class TestGenericResourceAwareStrategy method testGrasRequiringEviction.

/*
     * test requiring eviction until Generic Resource (gpu) is evicted.
     */
@Test
public void testGrasRequiringEviction() {
    int spoutParallelism = 3;
    double cpuPercent = 10;
    double memoryOnHeap = 10;
    double memoryOffHeap = 10;
    // Sufficient Cpu/Memory. But insufficient gpu to schedule all topologies (gpu1, noGpu, gpu2).
    // gpu topology (requires 3 gpu's in total)
    TopologyBuilder builder = new TopologyBuilder();
    builder.setSpout("spout", new TestSpout(), spoutParallelism).addResource("gpu.count", 1.0);
    StormTopology stormTopologyWithGpu = builder.createTopology();
    // non-gpu topology
    builder = new TopologyBuilder();
    builder.setSpout("spout", new TestSpout(), spoutParallelism);
    StormTopology stormTopologyNoGpu = builder.createTopology();
    Config conf = createGrasClusterConfig(cpuPercent, memoryOnHeap, memoryOffHeap, null, Collections.emptyMap());
    // allow 1 round of evictions
    conf.put(DaemonConfig.RESOURCE_AWARE_SCHEDULER_MAX_TOPOLOGY_SCHEDULING_ATTEMPTS, 2);
    String gpu1 = "hasGpu1";
    String noGpu = "hasNoGpu";
    String gpu2 = "hasGpu2";
    TopologyDetails[] topo = { createTestStormTopology(stormTopologyWithGpu, 10, gpu1, conf), createTestStormTopology(stormTopologyNoGpu, 10, noGpu, conf), createTestStormTopology(stormTopologyWithGpu, 9, gpu2, conf) };
    Topologies topologies = new Topologies(topo[0], topo[1]);
    Map<String, Double> genericResourcesMap = new HashMap<>();
    genericResourcesMap.put("gpu.count", 1.0);
    Map<String, SupervisorDetails> supMap = genSupervisors(4, 4, 500, 2000, genericResourcesMap);
    Cluster cluster = new Cluster(new INimbusTest(), new ResourceMetrics(new StormMetricsRegistry()), supMap, new HashMap<>(), topologies, conf);
    // should schedule gpu1 and noGpu successfully
    scheduler = new ResourceAwareScheduler();
    scheduler.prepare(conf, new StormMetricsRegistry());
    scheduler.schedule(topologies, cluster);
    assertTopologiesFullyScheduled(cluster, gpu1);
    assertTopologiesFullyScheduled(cluster, noGpu);
    // should evict gpu1 and noGpu topologies in order to schedule gpu2 topology; then fail to reschedule gpu1 topology;
    // then schedule noGpu topology.
    // Scheduling used to ignore gpu resource when deciding when to stop evicting, and gpu2 would fail to schedule.
    topologies = new Topologies(topo[0], topo[1], topo[2]);
    cluster = new Cluster(cluster, topologies);
    scheduler.schedule(topologies, cluster);
    assertTopologiesNotScheduled(cluster, gpu1);
    assertTopologiesFullyScheduled(cluster, noGpu);
    assertTopologiesFullyScheduled(cluster, gpu2);
}

Example 39 with ResourceMetrics

use of org.apache.storm.scheduler.resource.normalization.ResourceMetrics in project storm by apache.

the class TestGenericResourceAwareStrategy method testAntiAffinityWithMultipleTopologies.

@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));
    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);
    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);
    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 (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 40 with ResourceMetrics

use of org.apache.storm.scheduler.resource.normalization.ResourceMetrics in project storm by apache.

the class TestLargeCluster method testLargeCluster.

/**
 * Create a large cluster, read topologies and configuration from resource directory and schedule.
 *
 * @throws Exception upon error.
 */
@Test
public void testLargeCluster() throws Exception {
    for (TEST_CLUSTER_NAME testClusterName : TEST_CLUSTER_NAME.values()) {
        LOG.info("********************************************");
        LOG.info("testLargeCluster: Start Processing cluster {}", testClusterName.getClusterName());
        String resourcePath = testClusterName.getResourcePath();
        Map<String, SupervisorDetails> supervisors = createSupervisors(testClusterName, 0);
        TopologyDetails[] topoDetailsArray = createTopoDetailsArray(resourcePath, false);
        Assert.assertTrue("No topologies found for cluster " + testClusterName.getClusterName(), topoDetailsArray.length > 0);
        Topologies topologies = new Topologies(topoDetailsArray);
        Config confWithDefaultStrategy = new Config();
        confWithDefaultStrategy.putAll(topoDetailsArray[0].getConf());
        confWithDefaultStrategy.put(Config.TOPOLOGY_SCHEDULER_STRATEGY, DefaultResourceAwareStrategy.class.getName());
        confWithDefaultStrategy.put(Config.STORM_NETWORK_TOPOGRAPHY_PLUGIN, TestUtilsForResourceAwareScheduler.GenSupervisorsDnsToSwitchMapping.class.getName());
        INimbus iNimbus = new INimbusTest();
        Cluster cluster = new Cluster(iNimbus, new ResourceMetrics(new StormMetricsRegistry()), supervisors, new HashMap<>(), topologies, confWithDefaultStrategy);
        scheduler = new ResourceAwareScheduler();
        List<Class> classesToDebug = Arrays.asList(DefaultResourceAwareStrategy.class, GenericResourceAwareStrategy.class, ResourceAwareScheduler.class, Cluster.class);
        // switch to Level.DEBUG for verbose otherwise Level.INFO
        Level logLevel = Level.INFO;
        classesToDebug.forEach(x -> Configurator.setLevel(x.getName(), logLevel));
        long startTime = System.currentTimeMillis();
        scheduler.prepare(confWithDefaultStrategy, new StormMetricsRegistry());
        scheduler.schedule(topologies, cluster);
        long endTime = System.currentTimeMillis();
        LOG.info("Cluster={} Scheduling Time: {} topologies in {} seconds", testClusterName.getClusterName(), topoDetailsArray.length, (endTime - startTime) / 1000.0);
        for (TopologyDetails td : topoDetailsArray) {
            TestUtilsForResourceAwareScheduler.assertTopologiesFullyScheduled(cluster, td.getName());
        }
        // Remove topology and reschedule it
        for (int i = 0; i < topoDetailsArray.length; i++) {
            startTime = System.currentTimeMillis();
            TopologyDetails topoDetails = topoDetailsArray[i];
            cluster.unassign(topoDetails.getId());
            LOG.info("Cluster={},  ({}) Removed topology {}", testClusterName.getClusterName(), i, topoDetails.getName());
            IScheduler rescheduler = new ResourceAwareScheduler();
            rescheduler.prepare(confWithDefaultStrategy, new StormMetricsRegistry());
            rescheduler.schedule(topologies, cluster);
            TestUtilsForResourceAwareScheduler.assertTopologiesFullyScheduled(cluster, topoDetails.getName());
            endTime = System.currentTimeMillis();
            LOG.info("Cluster={}, ({}) Scheduling Time: Removed topology {} and rescheduled in {} seconds", testClusterName.getClusterName(), i, topoDetails.getName(), (endTime - startTime) / 1000.0);
        }
        classesToDebug.forEach(x -> Configurator.setLevel(x.getName(), Level.INFO));
        LOG.info("testLargeCluster: End Processing cluster {}", testClusterName.getClusterName());
        LOG.info("********************************************");
    }
}