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

Example 11 with NormalizedResourceRequest

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

the class TestNodeSorterHostProximity method testMultipleRacksOrderedByCapacity.

/**
 * Racks should be returned in order of decreasing capacity.
 */
@Test
public void testMultipleRacksOrderedByCapacity() {
    final Map<String, SupervisorDetails> supMap = new HashMap<>();
    final int numRacks = 1;
    final int numSupersPerRack = 10;
    final int numPortsPerSuper = 4;
    final int numZonesPerHost = 1;
    final double numaResourceMultiplier = 1.0;
    int rackStartNum = 0;
    int supStartNum = 0;
    final Map<String, SupervisorDetails> supMapRack0 = genSupervisorsWithRacksAndNuma(numRacks, numSupersPerRack, numZonesPerHost, numPortsPerSuper, rackStartNum++, supStartNum, 600, 8000 - rackStartNum, Collections.emptyMap(), numaResourceMultiplier);
    supStartNum += numSupersPerRack;
    final Map<String, SupervisorDetails> supMapRack1 = genSupervisorsWithRacksAndNuma(numRacks, numSupersPerRack, numZonesPerHost, numPortsPerSuper, rackStartNum++, supStartNum, 500, 8000 - rackStartNum, Collections.emptyMap(), numaResourceMultiplier);
    supStartNum += numSupersPerRack;
    final Map<String, SupervisorDetails> supMapRack2 = genSupervisorsWithRacksAndNuma(numRacks, numSupersPerRack, numZonesPerHost, numPortsPerSuper, rackStartNum++, supStartNum, 400, 8000 - rackStartNum, Collections.emptyMap(), numaResourceMultiplier);
    supStartNum += numSupersPerRack;
    final Map<String, SupervisorDetails> supMapRack3 = genSupervisorsWithRacksAndNuma(numRacks, numSupersPerRack, numZonesPerHost, numPortsPerSuper, rackStartNum++, supStartNum, 300, 8000 - rackStartNum, Collections.emptyMap(), numaResourceMultiplier);
    supStartNum += numSupersPerRack;
    final Map<String, SupervisorDetails> supMapRack4 = genSupervisorsWithRacksAndNuma(numRacks, numSupersPerRack, numZonesPerHost, numPortsPerSuper, rackStartNum++, supStartNum, 200, 8000 - rackStartNum, Collections.emptyMap(), numaResourceMultiplier);
    // too small to hold topology
    supStartNum += numSupersPerRack;
    final Map<String, SupervisorDetails> supMapRack5 = genSupervisorsWithRacksAndNuma(numRacks, numSupersPerRack, numZonesPerHost, numPortsPerSuper, rackStartNum++, supStartNum, 100, 8000 - rackStartNum, Collections.singletonMap("gpu.count", 0.0), numaResourceMultiplier);
    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
    TestDNSToSwitchMapping testDNSToSwitchMapping = 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);
    cluster.setNetworkTopography(testDNSToSwitchMapping.getRackToHosts());
    NodeSorterHostProximity nodeSorter = new NodeSorterHostProximity(cluster, topo1);
    nodeSorter.prepare(null);
    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 = topo1.getApproximateTotalResources();
    String topoRequest = String.format("Topo %s, approx-requested-resources %s", topo1.getId(), topoResourceRequest.toString());
    Iterator<ObjectResourcesItem> it = sortedRacks.iterator();
    Assert.assertEquals(topoRequest + "\n\t" + rackSummaries + "\nRack-000 should be ordered first since it has the largest capacity", "rack-000", it.next().id);
    Assert.assertEquals(topoRequest + "\n\t" + rackSummaries + "\nrack-001 should be ordered second since it smaller than rack-000", "rack-001", it.next().id);
    Assert.assertEquals(topoRequest + "\n\t" + rackSummaries + "\nrack-002 should be ordered third since it is smaller than rack-001", "rack-002", it.next().id);
    Assert.assertEquals(topoRequest + "\n\t" + rackSummaries + "\nrack-003 should be ordered fourth since it since it is smaller than rack-002", "rack-003", it.next().id);
    Assert.assertEquals(topoRequest + "\n\t" + rackSummaries + "\nrack-004 should be ordered fifth since it since it is smaller than rack-003", "rack-004", it.next().id);
    Assert.assertEquals(topoRequest + "\n\t" + rackSummaries + "\nrack-005 should be ordered last since it since it is has smallest capacity", "rack-005", it.next().id);
}

Example 12 with NormalizedResourceRequest

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

the class ServerUtils method getAckerExecutorMemoryUsageForTopo.

private static double getAckerExecutorMemoryUsageForTopo(StormTopology topology, Map<String, Object> topologyConf) throws InvalidTopologyException {
    topology = StormCommon.systemTopology(topologyConf, topology);
    Map<String, NormalizedResourceRequest> boltResources = ResourceUtils.getBoltsResources(topology, topologyConf);
    NormalizedResourceRequest entry = boltResources.get(Acker.ACKER_COMPONENT_ID);
    return entry.getTotalMemoryMb();
}

Example 13 with NormalizedResourceRequest

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

the class NodeSorterHostProximity method sortObjectResourcesGeneric.

/**
 * Sort objects by the following two criteria.
 *
 * <li>the number executors of the topology that needs to be scheduled is already on the
 * object (node or rack) in descending order. The reasoning to sort based on criterion 1 is so we schedule the rest
 * of a topology on the same object (node or rack) as the existing executors of the topology.</li>
 *
 * <li>the subordinate/subservient resource availability percentage of a rack in descending order We calculate the
 * resource availability percentage by dividing the resource availability of the object (node or rack) by the
 * resource availability of the entire rack or cluster depending on if object references a node or a rack.
 * How this differs from the DefaultResourceAwareStrategy is that the percentage boosts the node or rack if it is
 * requested by the executor that the sorting is being done for and pulls it down if it is not.
 * By doing this calculation, objects (node or rack) that have exhausted or little of one of the resources mentioned
 * above will be ranked after racks that have more balanced resource availability and nodes or racks that have
 * resources that are not requested will be ranked below . So we will be less likely to pick a rack that
 * have a lot of one resource but a low amount of another and have a lot of resources that are not requested by the executor.</li>
 *
 * @param allResources         contains all individual ObjectResources as well as cumulative stats
 * @param exec                 executor for which the sorting is done
 * @param existingScheduleFunc a function to get existing executors already scheduled on this object
 * @return an {@link Iterable} of sorted {@link ObjectResourcesItem}
 */
@Deprecated
private Iterable<ObjectResourcesItem> sortObjectResourcesGeneric(final ObjectResourcesSummary allResources, ExecutorDetails exec, final ExistingScheduleFunc existingScheduleFunc) {
    ObjectResourcesSummary affinityBasedAllResources = new ObjectResourcesSummary(allResources);
    final NormalizedResourceOffer availableResourcesOverall = allResources.getAvailableResourcesOverall();
    final NormalizedResourceRequest requestedResources = (exec != null) ? topologyDetails.getTotalResources(exec) : null;
    affinityBasedAllResources.getObjectResources().forEach(x -> {
        if (requestedResources != null) {
            // negate unrequested resources
            x.availableResources.updateForRareResourceAffinity(requestedResources);
        }
        x.minResourcePercent = availableResourcesOverall.calculateMinPercentageUsedBy(x.availableResources);
        x.avgResourcePercent = availableResourcesOverall.calculateAveragePercentageUsedBy(x.availableResources);
        LOG.trace("for {}: minResourcePercent={}, avgResourcePercent={}, numExistingSchedule={}", x.id, x.minResourcePercent, x.avgResourcePercent, existingScheduleFunc.getNumExistingSchedule(x.id));
    });
    Comparator<ObjectResourcesItem> comparator = (o1, o2) -> {
        int execsScheduled1 = existingScheduleFunc.getNumExistingSchedule(o1.id);
        int execsScheduled2 = existingScheduleFunc.getNumExistingSchedule(o2.id);
        if (execsScheduled1 > execsScheduled2) {
            return -1;
        } else if (execsScheduled1 < execsScheduled2) {
            return 1;
        }
        double o1Avg = o1.avgResourcePercent;
        double o2Avg = o2.avgResourcePercent;
        if (o1Avg > o2Avg) {
            return -1;
        } else if (o1Avg < o2Avg) {
            return 1;
        }
        return o1.id.compareTo(o2.id);
    };
    TreeSet<ObjectResourcesItem> sortedObjectResources = new TreeSet<>(comparator);
    sortedObjectResources.addAll(affinityBasedAllResources.getObjectResources());
    LOG.debug("Sorted Object Resources: {}", sortedObjectResources);
    return sortedObjectResources;
}

Example 14 with NormalizedResourceRequest

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

the class NodeSorter method sortObjectResourcesGeneric.

/**
 * Sort objects by the following two criteria.
 *
 * <li>the number executors of the topology that needs to be scheduled is already on the
 * object (node or rack) in descending order. The reasoning to sort based on criterion 1 is so we schedule the rest
 * of a topology on the same object (node or rack) as the existing executors of the topology.</li>
 *
 * <li>the subordinate/subservient resource availability percentage of a rack in descending order We calculate the
 * resource availability percentage by dividing the resource availability of the object (node or rack) by the
 * resource availability of the entire rack or cluster depending on if object references a node or a rack.
 * How this differs from the DefaultResourceAwareStrategy is that the percentage boosts the node or rack if it is
 * requested by the executor that the sorting is being done for and pulls it down if it is not.
 * By doing this calculation, objects (node or rack) that have exhausted or little of one of the resources mentioned
 * above will be ranked after racks that have more balanced resource availability and nodes or racks that have
 * resources that are not requested will be ranked below . So we will be less likely to pick a rack that
 * have a lot of one resource but a low amount of another and have a lot of resources that are not requested by the executor.</li>
 *
 * @param allResources         contains all individual ObjectResources as well as cumulative stats
 * @param exec                 executor for which the sorting is done
 * @param existingScheduleFunc a function to get existing executors already scheduled on this object
 * @return a sorted list of ObjectResources
 */
@Deprecated
private List<ObjectResourcesItem> sortObjectResourcesGeneric(final ObjectResourcesSummary allResources, ExecutorDetails exec, final ExistingScheduleFunc existingScheduleFunc) {
    ObjectResourcesSummary affinityBasedAllResources = new ObjectResourcesSummary(allResources);
    NormalizedResourceRequest requestedResources = topologyDetails.getTotalResources(exec);
    affinityBasedAllResources.getObjectResources().forEach(x -> x.availableResources.updateForRareResourceAffinity(requestedResources));
    final NormalizedResourceOffer availableResourcesOverall = allResources.getAvailableResourcesOverall();
    List<ObjectResourcesItem> sortedObjectResources = new ArrayList<>();
    Comparator<ObjectResourcesItem> comparator = (o1, o2) -> {
        int execsScheduled1 = existingScheduleFunc.getNumExistingSchedule(o1.id);
        int execsScheduled2 = existingScheduleFunc.getNumExistingSchedule(o2.id);
        if (execsScheduled1 > execsScheduled2) {
            return -1;
        } else if (execsScheduled1 < execsScheduled2) {
            return 1;
        }
        double o1Avg = availableResourcesOverall.calculateAveragePercentageUsedBy(o1.availableResources);
        double o2Avg = availableResourcesOverall.calculateAveragePercentageUsedBy(o2.availableResources);
        if (o1Avg > o2Avg) {
            return -1;
        } else if (o1Avg < o2Avg) {
            return 1;
        }
        return o1.id.compareTo(o2.id);
    };
    sortedObjectResources.addAll(affinityBasedAllResources.getObjectResources());
    sortedObjectResources.sort(comparator);
    LOG.debug("Sorted Object Resources: {}", sortedObjectResources);
    return sortedObjectResources;
}

Example 15 with NormalizedResourceRequest

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

the class RasBlacklistStrategy method releaseBlacklistWhenNeeded.

@Override
protected Set<String> releaseBlacklistWhenNeeded(Cluster cluster, final List<String> blacklistedNodeIds) {
    LOG.info("RAS We have {} nodes blacklisted...", blacklistedNodeIds.size());
    Set<String> readyToRemove = new HashSet<>();
    if (blacklistedNodeIds.size() > 0) {
        int availableSlots = cluster.getNonBlacklistedAvailableSlots(blacklistedNodeIds).size();
        int neededSlots = 0;
        NormalizedResourceOffer available = cluster.getNonBlacklistedClusterAvailableResources(blacklistedNodeIds);
        NormalizedResourceOffer needed = new NormalizedResourceOffer();
        for (TopologyDetails td : cluster.getTopologies()) {
            if (cluster.needsSchedulingRas(td)) {
                int slots = 0;
                try {
                    slots = ServerUtils.getEstimatedWorkerCountForRasTopo(td.getConf(), td.getTopology());
                } catch (InvalidTopologyException e) {
                    LOG.warn("Could not guess the number of slots needed for {}", td.getName(), e);
                }
                int assignedSlots = cluster.getAssignedNumWorkers(td);
                int tdSlotsNeeded = slots - assignedSlots;
                neededSlots += tdSlotsNeeded;
                NormalizedResourceRequest resources = td.getApproximateTotalResources();
                needed.add(resources);
                LOG.warn("{} needs to be scheduled with {} and {} slots", td.getName(), resources, tdSlotsNeeded);
            }
        }
        // Now we need to free up some resources...
        Map<String, SupervisorDetails> availableSupervisors = cluster.getSupervisors();
        NormalizedResourceOffer shortage = new NormalizedResourceOffer(needed);
        shortage.remove(available, cluster.getResourceMetrics());
        int shortageSlots = neededSlots - availableSlots;
        LOG.debug("Need {} and {} slots.", needed, neededSlots);
        LOG.debug("Available {} and {} slots.", available, availableSlots);
        LOG.debug("Shortage {} and {} slots.", shortage, shortageSlots);
        if (shortage.areAnyOverZero() || shortageSlots > 0) {
            LOG.info("Need {} and {} slots more. Releasing some blacklisted nodes to cover it.", shortage, shortageSlots);
            // release earliest blacklist - but release all supervisors on a given blacklisted host.
            Map<String, Set<String>> hostToSupervisorIds = createHostToSupervisorMap(blacklistedNodeIds, cluster);
            for (Set<String> supervisorIds : hostToSupervisorIds.values()) {
                for (String supervisorId : supervisorIds) {
                    SupervisorDetails sd = availableSupervisors.get(supervisorId);
                    if (sd != null) {
                        NormalizedResourcesWithMemory sdAvailable = cluster.getAvailableResources(sd);
                        int sdAvailableSlots = cluster.getAvailablePorts(sd).size();
                        readyToRemove.add(supervisorId);
                        shortage.remove(sdAvailable, cluster.getResourceMetrics());
                        shortageSlots -= sdAvailableSlots;
                        LOG.info("Releasing {} with {} and {} slots leaving {} and {} slots to go", supervisorId, sdAvailable, sdAvailableSlots, shortage, shortageSlots);
                    }
                }
                // make sure we've handled all supervisors on the host before we break
                if (!shortage.areAnyOverZero() && shortageSlots <= 0) {
                    // we have enough resources now...
                    break;
                }
            }
        }
    }
    return readyToRemove;
}