use of com.facebook.presto.execution.scheduler.NodeMap 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);
}
use of com.facebook.presto.execution.scheduler.NodeMap in project presto by prestodb.
the class SimpleTtlNodeSelector method getEligibleNodes.
private List<InternalNode> getEligibleNodes(int limit, NodeMap nodeMap, List<RemoteTask> existingTasks) {
Map<InternalNode, NodeTtl> nodeTtlInfo = nodeTtlFetcherManager.getAllTtls();
Map<InternalNode, Optional<ConfidenceBasedTtlInfo>> ttlInfo = nodeTtlInfo.entrySet().stream().collect(toImmutableMap(Map.Entry::getKey, e -> e.getValue().getTtlInfo().stream().min(Comparator.comparing(ConfidenceBasedTtlInfo::getExpiryInstant))));
Duration estimatedExecutionTimeRemaining = getEstimatedExecutionTimeRemaining();
// Of the nodes on which already have existing tasks, pick only those whose TTL is enough
List<InternalNode> existingEligibleNodes = existingTasks.stream().map(remoteTask -> nodeMap.getActiveNodesByNodeId().get(remoteTask.getNodeId())).filter(Objects::nonNull).filter(ttlInfo::containsKey).filter(node -> ttlInfo.get(node).isPresent()).filter(node -> isTtlEnough(ttlInfo.get(node).get(), estimatedExecutionTimeRemaining)).collect(toList());
int alreadySelectedNodeCount = existingEligibleNodes.size();
List<InternalNode> activeNodes = nodeMap.getActiveNodes();
List<InternalNode> newEligibleNodes = filterNodesByTtl(activeNodes, ImmutableSet.copyOf(existingEligibleNodes), ttlInfo, estimatedExecutionTimeRemaining);
if (alreadySelectedNodeCount < limit && newEligibleNodes.size() > 0) {
List<InternalNode> moreNodes = selectNodes(limit - alreadySelectedNodeCount, new ResettableRandomizedIterator<>(newEligibleNodes));
existingEligibleNodes.addAll(moreNodes);
}
verify(existingEligibleNodes.stream().allMatch(Objects::nonNull), "existingNodes list must not contain any nulls");
return existingEligibleNodes;
}
use of com.facebook.presto.execution.scheduler.NodeMap in project presto by prestodb.
the class SimpleTtlNodeSelector method computeAssignments.
@Override
public SplitPlacementResult computeAssignments(Set<Split> splits, List<RemoteTask> existingTasks) {
boolean isNodeSelectionStrategyNoPreference = splits.stream().allMatch(split -> split.getNodeSelectionStrategy() == NodeSelectionStrategy.NO_PREFERENCE);
// Current NodeSelectionStrategy support is limited to NO_PREFERENCE
if (!isNodeSelectionStrategyNoPreference) {
return simpleNodeSelector.computeAssignments(splits, existingTasks);
}
ImmutableMultimap.Builder<InternalNode, Split> assignment = ImmutableMultimap.builder();
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);
boolean splitWaitingForAnyNode = false;
OptionalInt preferredNodeCount = OptionalInt.empty();
for (Split split : splits) {
if (split.getNodeSelectionStrategy() != NodeSelectionStrategy.NO_PREFERENCE) {
throw new PrestoException(NODE_SELECTION_NOT_SUPPORTED, format("Unsupported node selection strategy for TTL scheduling: %s", split.getNodeSelectionStrategy()));
}
List<InternalNode> candidateNodes = randomNodeSelection.pickNodes(split);
if (candidateNodes.isEmpty()) {
log.warn("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 = simpleNodeSelector.chooseLeastBusyNode(splitWeight, candidateNodes, assignmentStats::getTotalSplitsWeight, preferredNodeCount, maxSplitsWeightPerNode, assignmentStats);
if (!chosenNodeInfo.isPresent()) {
chosenNodeInfo = simpleNodeSelector.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 {
splitWaitingForAnyNode = true;
}
}
ListenableFuture<?> blocked = splitWaitingForAnyNode ? toWhenHasSplitQueueSpaceFuture(existingTasks, calculateLowWatermark(maxPendingSplitsWeightPerTask)) : immediateFuture(null);
return new SplitPlacementResult(blocked, assignment.build());
}
use of com.facebook.presto.execution.scheduler.NodeMap in project presto by prestodb.
the class SimpleTtlNodeSelector method selectRandomNodes.
@Override
public List<InternalNode> selectRandomNodes(int limit, Set<InternalNode> excludedNodes) {
Map<InternalNode, NodeTtl> nodeTtlInfo = nodeTtlFetcherManager.getAllTtls();
Map<InternalNode, Optional<ConfidenceBasedTtlInfo>> ttlInfo = nodeTtlInfo.entrySet().stream().collect(toImmutableMap(Map.Entry::getKey, e -> e.getValue().getTtlInfo().stream().min(Comparator.comparing(ConfidenceBasedTtlInfo::getExpiryInstant))));
NodeMap nodeMap = this.nodeMap.get().get();
List<InternalNode> activeNodes = nodeMap.getActiveNodes();
Duration estimatedExecutionTimeRemaining = getEstimatedExecutionTimeRemaining();
List<InternalNode> eligibleNodes = filterNodesByTtl(activeNodes, excludedNodes, ttlInfo, estimatedExecutionTimeRemaining);
return selectNodes(limit, new ResettableRandomizedIterator<>(eligibleNodes));
}
use of com.facebook.presto.execution.scheduler.NodeMap 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);
}
Aggregations