Examples with NodeSorterHostProximity org.apache.storm.scheduler.resource.strategies.scheduling.sorter.NodeSorterHostProximity used on opensource projects

Example 1 with NodeSorterHostProximity

use of org.apache.storm.scheduler.resource.strategies.scheduling.sorter.NodeSorterHostProximity in project storm by apache.

the class BaseResourceAwareStrategy method prepareForScheduling.

/**
 * Initialize instance variables as the first step in {@link #schedule(Cluster, TopologyDetails)}.
 * This method may be extended by subclasses to initialize additional variables as in
 * {@link ConstraintSolverStrategy#prepareForScheduling(Cluster, TopologyDetails)}.
 *
 * @param cluster on which executors will be scheduled.
 * @param topologyDetails to be scheduled.
 */
protected void prepareForScheduling(Cluster cluster, TopologyDetails topologyDetails) {
    this.cluster = cluster;
    this.topologyDetails = topologyDetails;
    // from Cluster
    this.nodes = new RasNodes(cluster);
    networkTopography = cluster.getNetworkTopography();
    hostnameToNodes = this.nodes.getHostnameToNodes();
    // from TopologyDetails
    topoName = topologyDetails.getName();
    execToComp = topologyDetails.getExecutorToComponent();
    compToExecs = topologyDetails.getComponentToExecutors();
    Map<String, Object> topoConf = topologyDetails.getConf();
    orderExecutorsByProximity = isOrderByProximity(topoConf);
    maxSchedulingTimeMs = computeMaxSchedulingTimeMs(topoConf);
    // From Cluster and TopologyDetails - and cleaned-up
    // all unassigned executors including system components execs
    unassignedExecutors = Collections.unmodifiableSet(new HashSet<>(cluster.getUnassignedExecutors(topologyDetails)));
    int confMaxStateSearch = getMaxStateSearchFromTopoConf(topologyDetails.getConf());
    int daemonMaxStateSearch = ObjectReader.getInt(cluster.getConf().get(DaemonConfig.RESOURCE_AWARE_SCHEDULER_MAX_STATE_SEARCH));
    maxStateSearch = Math.min(daemonMaxStateSearch, confMaxStateSearch);
    LOG.debug("The max state search configured by topology {} is {}", topologyDetails.getId(), confMaxStateSearch);
    LOG.debug("The max state search that will be used by topology {} is {}", topologyDetails.getId(), maxStateSearch);
    searcherState = createSearcherState();
    setNodeSorter(new NodeSorterHostProximity(cluster, topologyDetails, nodeSortType));
    setExecSorter(orderExecutorsByProximity ? new ExecSorterByProximity(topologyDetails) : new ExecSorterByConnectionCount(topologyDetails));
    logClusterInfo();
}

Example 2 with NodeSorterHostProximity

use of org.apache.storm.scheduler.resource.strategies.scheduling.sorter.NodeSorterHostProximity 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());
}

Example 3 with NodeSorterHostProximity

use of org.apache.storm.scheduler.resource.strategies.scheduling.sorter.NodeSorterHostProximity in project storm by apache.

the class TestDefaultResourceAwareStrategy method testMultipleRacksWithFavoritism.

/**
 * Test whether strategy will choose correct rack
 */
@Test
public void testMultipleRacksWithFavoritism() {
    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);
    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
    DNSToSwitchMapping TestNetworkTopographyPlugin = 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);
    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();
        List<String> nodesForRack = rackToNodes.get(rack);
        if (nodesForRack == null) {
            nodesForRack = new ArrayList<>();
            rackToNodes.put(rack, nodesForRack);
        }
        nodesForRack.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 last since rack-2 has not cpu resources
    Assert.assertEquals("rack-2 should be ordered fifth", "rack-2", it.next().id);
    SchedulingResult schedulingResult = rs.schedule(cluster, topo1);
    assert (schedulingResult.isSuccess());
    SchedulerAssignment assignment = cluster.getAssignmentById(topo1.getId());
    for (WorkerSlot ws : assignment.getSlotToExecutors().keySet()) {
        String hostName = rs.idToNode(ws.getNodeId()).getHostname();
        String rackId = resolvedSuperVisors.get(hostName);
        Assert.assertTrue(ws + " is neither on a favored node " + t1FavoredHostNames + " nor the highest priority rack (rack-0)", t1FavoredHostNames.contains(hostName) || "rack-0".equals(rackId));
        Assert.assertFalse(ws + " is a part of an unfavored node " + t1UnfavoredHostIds, t1UnfavoredHostIds.contains(hostName));
    }
    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
        // The favored nodes would have put it on a different rack, but because that rack does not have free space to run the
        // topology it falls back to this rack
        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());
}