Search in sources :

Example 1 with NodeProvider

use of com.facebook.presto.spi.NodeProvider in project presto by prestodb.

the class SimpleNodeSelector method computeAssignments.

@Override
public SplitPlacementResult computeAssignments(Set<Split> splits, List<RemoteTask> existingTasks) {
    Multimap<InternalNode, Split> assignment = HashMultimap.create();
    NodeMap nodeMap = this.nodeMap.get().get();
    NodeAssignmentStats assignmentStats = new NodeAssignmentStats(nodeTaskMap, nodeMap, existingTasks);
    List<InternalNode> eligibleNodes = getEligibleNodes(maxTasksPerStage, nodeMap, existingTasks);
    NodeSelection randomNodeSelection = new RandomNodeSelection(eligibleNodes, minCandidates);
    Set<InternalNode> blockedExactNodes = new HashSet<>();
    boolean splitWaitingForAnyNode = false;
    NodeProvider nodeProvider = nodeMap.getActiveNodeProvider(nodeSelectionHashStrategy);
    OptionalInt preferredNodeCount = OptionalInt.empty();
    for (Split split : splits) {
        List<InternalNode> candidateNodes;
        switch(split.getNodeSelectionStrategy()) {
            case HARD_AFFINITY:
                candidateNodes = selectExactNodes(nodeMap, split.getPreferredNodes(nodeProvider), includeCoordinator);
                preferredNodeCount = OptionalInt.of(candidateNodes.size());
                break;
            case SOFT_AFFINITY:
                // Using all nodes for soft affinity scheduling with modular hashing because otherwise temporarily down nodes would trigger too much rehashing
                if (nodeSelectionHashStrategy == MODULAR_HASHING) {
                    nodeProvider = new ModularHashingNodeProvider(nodeMap.getAllNodes());
                }
                candidateNodes = selectExactNodes(nodeMap, split.getPreferredNodes(nodeProvider), includeCoordinator);
                preferredNodeCount = OptionalInt.of(candidateNodes.size());
                candidateNodes = ImmutableList.<InternalNode>builder().addAll(candidateNodes).addAll(randomNodeSelection.pickNodes(split)).build();
                break;
            case NO_PREFERENCE:
                candidateNodes = randomNodeSelection.pickNodes(split);
                break;
            default:
                throw new PrestoException(NODE_SELECTION_NOT_SUPPORTED, format("Unsupported node selection strategy %s", split.getNodeSelectionStrategy()));
        }
        if (candidateNodes.isEmpty()) {
            log.debug("No nodes available to schedule %s. Available nodes %s", split, nodeMap.getActiveNodes());
            throw new PrestoException(NO_NODES_AVAILABLE, "No nodes available to run query");
        }
        SplitWeight splitWeight = split.getSplitWeight();
        Optional<InternalNodeInfo> chosenNodeInfo = chooseLeastBusyNode(splitWeight, candidateNodes, assignmentStats::getTotalSplitsWeight, preferredNodeCount, maxSplitsWeightPerNode, assignmentStats);
        if (!chosenNodeInfo.isPresent()) {
            chosenNodeInfo = chooseLeastBusyNode(splitWeight, candidateNodes, assignmentStats::getQueuedSplitsWeightForStage, preferredNodeCount, maxPendingSplitsWeightPerTask, assignmentStats);
        }
        if (chosenNodeInfo.isPresent()) {
            split = new Split(split.getConnectorId(), split.getTransactionHandle(), split.getConnectorSplit(), split.getLifespan(), new SplitContext(chosenNodeInfo.get().isCacheable()));
            InternalNode chosenNode = chosenNodeInfo.get().getInternalNode();
            assignment.put(chosenNode, split);
            assignmentStats.addAssignedSplit(chosenNode, splitWeight);
        } else {
            if (split.getNodeSelectionStrategy() != HARD_AFFINITY) {
                splitWaitingForAnyNode = true;
            } else // Exact node set won't matter, if a split is waiting for any node
            if (!splitWaitingForAnyNode) {
                blockedExactNodes.addAll(candidateNodes);
            }
        }
    }
    ListenableFuture<?> blocked;
    if (splitWaitingForAnyNode) {
        blocked = toWhenHasSplitQueueSpaceFuture(existingTasks, calculateLowWatermark(maxPendingSplitsWeightPerTask));
    } else {
        blocked = toWhenHasSplitQueueSpaceFuture(blockedExactNodes, existingTasks, calculateLowWatermark(maxPendingSplitsWeightPerTask));
    }
    return new SplitPlacementResult(blocked, assignment);
}
Also used : NodeAssignmentStats(com.facebook.presto.execution.scheduler.NodeAssignmentStats) InternalNodeInfo(com.facebook.presto.execution.scheduler.InternalNodeInfo) PrestoException(com.facebook.presto.spi.PrestoException) NodeProvider(com.facebook.presto.spi.NodeProvider) ModularHashingNodeProvider(com.facebook.presto.execution.scheduler.ModularHashingNodeProvider) OptionalInt(java.util.OptionalInt) ModularHashingNodeProvider(com.facebook.presto.execution.scheduler.ModularHashingNodeProvider) SplitWeight(com.facebook.presto.spi.SplitWeight) SplitContext(com.facebook.presto.spi.SplitContext) BucketNodeMap(com.facebook.presto.execution.scheduler.BucketNodeMap) NodeMap(com.facebook.presto.execution.scheduler.NodeMap) InternalNode(com.facebook.presto.metadata.InternalNode) Split(com.facebook.presto.metadata.Split) SplitPlacementResult(com.facebook.presto.execution.scheduler.SplitPlacementResult) Sets.newHashSet(com.google.common.collect.Sets.newHashSet) HashSet(java.util.HashSet)

Example 2 with NodeProvider

use of com.facebook.presto.spi.NodeProvider in project presto by prestodb.

the class TopologyAwareNodeSelector method computeAssignments.

@Override
public SplitPlacementResult computeAssignments(Set<Split> splits, List<RemoteTask> existingTasks) {
    NodeMap nodeMap = this.nodeMap.get().get();
    Multimap<InternalNode, Split> assignment = HashMultimap.create();
    NodeAssignmentStats assignmentStats = new NodeAssignmentStats(nodeTaskMap, nodeMap, existingTasks);
    int[] topologicCounters = new int[topologicalSplitCounters.size()];
    Set<NetworkLocation> filledLocations = new HashSet<>();
    Set<InternalNode> blockedExactNodes = new HashSet<>();
    boolean splitWaitingForAnyNode = false;
    NodeProvider nodeProvider = nodeMap.getActiveNodeProvider(nodeSelectionHashStrategy);
    for (Split split : splits) {
        SplitWeight splitWeight = split.getSplitWeight();
        if (split.getNodeSelectionStrategy() == HARD_AFFINITY) {
            List<InternalNode> candidateNodes = selectExactNodes(nodeMap, split.getPreferredNodes(nodeProvider), includeCoordinator);
            if (candidateNodes.isEmpty()) {
                log.debug("No nodes available to schedule %s. Available nodes %s", split, nodeMap.getActiveNodes());
                throw new PrestoException(NO_NODES_AVAILABLE, "No nodes available to run query");
            }
            InternalNode chosenNode = bestNodeSplitCount(splitWeight, candidateNodes.iterator(), minCandidates, maxPendingSplitsWeightPerTask, assignmentStats);
            if (chosenNode != null) {
                assignment.put(chosenNode, split);
                assignmentStats.addAssignedSplit(chosenNode, splitWeight);
            } else // Exact node set won't matter, if a split is waiting for any node
            if (!splitWaitingForAnyNode) {
                blockedExactNodes.addAll(candidateNodes);
            }
            continue;
        }
        InternalNode chosenNode = null;
        int depth = networkLocationSegmentNames.size();
        int chosenDepth = 0;
        Set<NetworkLocation> locations = new HashSet<>();
        for (HostAddress host : split.getPreferredNodes(nodeProvider)) {
            locations.add(networkLocationCache.get(host));
        }
        if (locations.isEmpty()) {
            // Add the root location
            locations.add(ROOT_LOCATION);
            depth = 0;
        }
        // Try each address at progressively shallower network locations
        for (int i = depth; i >= 0 && chosenNode == null; i--) {
            for (NetworkLocation location : locations) {
                // For example, locations which couldn't be located will be at the "root" location
                if (location.getSegments().size() < i) {
                    continue;
                }
                location = location.subLocation(0, i);
                if (filledLocations.contains(location)) {
                    continue;
                }
                Set<InternalNode> nodes = nodeMap.getActiveWorkersByNetworkPath().get(location);
                chosenNode = bestNodeSplitCount(splitWeight, new ResettableRandomizedIterator<>(nodes), minCandidates, calculateMaxPendingSplitsWeightPerTask(i, depth), assignmentStats);
                if (chosenNode != null) {
                    chosenDepth = i;
                    break;
                }
                filledLocations.add(location);
            }
        }
        if (chosenNode != null) {
            assignment.put(chosenNode, split);
            assignmentStats.addAssignedSplit(chosenNode, splitWeight);
            topologicCounters[chosenDepth]++;
        } else {
            splitWaitingForAnyNode = true;
        }
    }
    for (int i = 0; i < topologicCounters.length; i++) {
        if (topologicCounters[i] > 0) {
            topologicalSplitCounters.get(i).update(topologicCounters[i]);
        }
    }
    ListenableFuture<?> blocked;
    long maxPendingForWildcardNetworkAffinity = calculateMaxPendingSplitsWeightPerTask(0, networkLocationSegmentNames.size());
    if (splitWaitingForAnyNode) {
        blocked = toWhenHasSplitQueueSpaceFuture(existingTasks, calculateLowWatermark(maxPendingForWildcardNetworkAffinity));
    } else {
        blocked = toWhenHasSplitQueueSpaceFuture(blockedExactNodes, existingTasks, calculateLowWatermark(maxPendingForWildcardNetworkAffinity));
    }
    return new SplitPlacementResult(blocked, assignment);
}
Also used : NodeAssignmentStats(com.facebook.presto.execution.scheduler.NodeAssignmentStats) PrestoException(com.facebook.presto.spi.PrestoException) NodeProvider(com.facebook.presto.spi.NodeProvider) HostAddress(com.facebook.presto.spi.HostAddress) NetworkLocation(com.facebook.presto.execution.scheduler.NetworkLocation) SplitWeight(com.facebook.presto.spi.SplitWeight) ResettableRandomizedIterator(com.facebook.presto.execution.scheduler.ResettableRandomizedIterator) NodeMap(com.facebook.presto.execution.scheduler.NodeMap) BucketNodeMap(com.facebook.presto.execution.scheduler.BucketNodeMap) InternalNode(com.facebook.presto.metadata.InternalNode) Split(com.facebook.presto.metadata.Split) SplitPlacementResult(com.facebook.presto.execution.scheduler.SplitPlacementResult) HashSet(java.util.HashSet)

Aggregations

BucketNodeMap (com.facebook.presto.execution.scheduler.BucketNodeMap)2 NodeAssignmentStats (com.facebook.presto.execution.scheduler.NodeAssignmentStats)2 NodeMap (com.facebook.presto.execution.scheduler.NodeMap)2 SplitPlacementResult (com.facebook.presto.execution.scheduler.SplitPlacementResult)2 InternalNode (com.facebook.presto.metadata.InternalNode)2 Split (com.facebook.presto.metadata.Split)2 NodeProvider (com.facebook.presto.spi.NodeProvider)2 PrestoException (com.facebook.presto.spi.PrestoException)2 SplitWeight (com.facebook.presto.spi.SplitWeight)2 HashSet (java.util.HashSet)2 InternalNodeInfo (com.facebook.presto.execution.scheduler.InternalNodeInfo)1 ModularHashingNodeProvider (com.facebook.presto.execution.scheduler.ModularHashingNodeProvider)1 NetworkLocation (com.facebook.presto.execution.scheduler.NetworkLocation)1 ResettableRandomizedIterator (com.facebook.presto.execution.scheduler.ResettableRandomizedIterator)1 HostAddress (com.facebook.presto.spi.HostAddress)1 SplitContext (com.facebook.presto.spi.SplitContext)1 Sets.newHashSet (com.google.common.collect.Sets.newHashSet)1 OptionalInt (java.util.OptionalInt)1