use of com.hazelcast.jet.impl.execution.ConveyorCollector in project hazelcast-jet by hazelcast.
the class ExecutionPlan method createIfAbsentReceiverTasklet.
private void createIfAbsentReceiverTasklet(EdgeDef edge, int[][] ptionsPerProcessor, int totalPtionCount) {
final ConcurrentConveyor<Object>[] localConveyors = localConveyorMap.get(edge.edgeId());
receiverMap.computeIfAbsent(edge.destVertex().vertexId(), x -> new HashMap<>()).computeIfAbsent(edge.destOrdinal(), x -> {
Map<Address, ReceiverTasklet> addrToTasklet = new HashMap<>();
// create a receiver per address
int offset = 0;
for (Address addr : ptionArrgmt.remotePartitionAssignment.get().keySet()) {
final OutboundCollector[] collectors = new OutboundCollector[ptionsPerProcessor.length];
// assign the queues starting from end
final int queueOffset = --offset;
Arrays.setAll(collectors, n -> new ConveyorCollector(localConveyors[n], localConveyors[n].queueCount() + queueOffset, ptionsPerProcessor[n]));
final OutboundCollector collector = compositeCollector(collectors, edge, totalPtionCount);
ReceiverTasklet receiverTasklet = new ReceiverTasklet(collector, edge.getConfig().getReceiveWindowMultiplier(), getConfig().getInstanceConfig().getFlowControlPeriodMs());
addrToTasklet.put(addr, receiverTasklet);
}
return addrToTasklet;
});
}
use of com.hazelcast.jet.impl.execution.ConveyorCollector in project hazelcast-jet by hazelcast.
the class ExecutionPlan method createOutboundCollectors.
private OutboundCollector[] createOutboundCollectors(EdgeDef edge, int processorIndex, Map<Address, ConcurrentConveyor<Object>> senderConveyorMap) {
final int upstreamParallelism = edge.sourceVertex().localParallelism();
final int downstreamParallelism = edge.destVertex().localParallelism();
final int numRemoteMembers = ptionArrgmt.remotePartitionAssignment.get().size();
final int queueSize = edge.getConfig().getQueueSize();
final int[][] ptionsPerProcessor = ptionArrgmt.assignPartitionsToProcessors(downstreamParallelism, edge.isDistributed());
if (edge.routingPolicy() == RoutingPolicy.ISOLATED) {
if (downstreamParallelism < upstreamParallelism) {
throw new IllegalArgumentException(String.format("The edge %s specifies the %s routing policy, but the downstream vertex" + " parallelism (%d) is less than the upstream vertex parallelism (%d)", edge, RoutingPolicy.ISOLATED.name(), downstreamParallelism, upstreamParallelism));
}
if (edge.isDistributed()) {
throw new IllegalArgumentException("Isolated edges must be local: " + edge);
}
// there is only one producer per consumer for a one to many edge, so queueCount is always 1
ConcurrentConveyor<Object>[] localConveyors = localConveyorMap.computeIfAbsent(edge.edgeId(), e -> createConveyorArray(downstreamParallelism, 1, queueSize));
return IntStream.range(0, downstreamParallelism).filter(i -> i % upstreamParallelism == processorIndex).mapToObj(i -> new ConveyorCollector(localConveyors[i], 0, ptionsPerProcessor[i])).toArray(OutboundCollector[]::new);
}
/*
* Each edge is represented by an array of conveyors between the producers and consumers
* There are as many conveyors as there are consumers.
* Each conveyor has one queue per producer.
*
* For a distributed edge, there is one additional producer per member represented
* by the ReceiverTasklet.
*/
final ConcurrentConveyor<Object>[] localConveyors = localConveyorMap.computeIfAbsent(edge.edgeId(), e -> {
int queueCount = upstreamParallelism + (edge.isDistributed() ? numRemoteMembers : 0);
return createConveyorArray(downstreamParallelism, queueCount, queueSize);
});
final OutboundCollector[] localCollectors = new OutboundCollector[downstreamParallelism];
Arrays.setAll(localCollectors, n -> new ConveyorCollector(localConveyors[n], processorIndex, ptionsPerProcessor[n]));
// in a local edge, we only have the local collectors.
if (!edge.isDistributed()) {
return localCollectors;
}
// in a distributed edge, allCollectors[0] is the composite of local collectors, and
// allCollectors[n] where n > 0 is a collector pointing to a remote member _n_.
final int totalPtionCount = nodeEngine.getPartitionService().getPartitionCount();
final OutboundCollector[] allCollectors;
createIfAbsentReceiverTasklet(edge, ptionsPerProcessor, totalPtionCount);
// assign remote partitions to outbound data collectors
final Map<Address, int[]> memberToPartitions = ptionArrgmt.remotePartitionAssignment.get();
allCollectors = new OutboundCollector[memberToPartitions.size() + 1];
allCollectors[0] = compositeCollector(localCollectors, edge, totalPtionCount);
int index = 1;
for (Map.Entry<Address, int[]> entry : memberToPartitions.entrySet()) {
allCollectors[index++] = new ConveyorCollectorWithPartition(senderConveyorMap.get(entry.getKey()), processorIndex, entry.getValue());
}
return allCollectors;
}
use of com.hazelcast.jet.impl.execution.ConveyorCollector in project hazelcast by hazelcast.
the class ExecutionPlan method initialize.
/**
* A method called on the members as part of the InitExecutionOperation.
* Creates tasklets, inboxes/outboxes and connects these to make them ready
* for a later StartExecutionOperation.
*/
public void initialize(NodeEngineImpl nodeEngine, long jobId, long executionId, @Nonnull SnapshotContext snapshotContext, ConcurrentHashMap<String, File> tempDirectories, InternalSerializationService jobSerializationService) {
this.nodeEngine = nodeEngine;
this.jobClassLoaderService = ((JetServiceBackend) nodeEngine.getService(JetServiceBackend.SERVICE_NAME)).getJobClassLoaderService();
this.executionId = executionId;
initProcSuppliers(jobId, tempDirectories, jobSerializationService);
initDag(jobSerializationService);
this.ptionArrgmt = new PartitionArrangement(partitionAssignment, nodeEngine.getThisAddress());
Set<Integer> higherPriorityVertices = VertexDef.getHigherPriorityVertices(vertices);
for (Address destAddr : remoteMembers.get()) {
Connection conn = getMemberConnection(nodeEngine, destAddr);
if (conn == null) {
throw new TopologyChangedException("no connection to job participant: " + destAddr);
}
memberConnections.put(destAddr, conn);
}
for (VertexDef vertex : vertices) {
ClassLoader processorClassLoader = isLightJob ? null : jobClassLoaderService.getProcessorClassLoader(jobId, vertex.name());
Collection<? extends Processor> processors = doWithClassLoader(processorClassLoader, () -> createProcessors(vertex, vertex.localParallelism()));
String jobPrefix = prefix(jobConfig.getName(), jobId, vertex.name());
// create StoreSnapshotTasklet and the queues to it
ConcurrentConveyor<Object> ssConveyor = null;
if (!isLightJob) {
// Note that we create the snapshot queues for all non-light jobs, even if they don't have
// processing guarantee enabled, because in EE one can request a snapshot also for
// non-snapshotted jobs.
@SuppressWarnings("unchecked") QueuedPipe<Object>[] snapshotQueues = new QueuedPipe[vertex.localParallelism()];
Arrays.setAll(snapshotQueues, i -> new OneToOneConcurrentArrayQueue<>(SNAPSHOT_QUEUE_SIZE));
ssConveyor = ConcurrentConveyor.concurrentConveyor(null, snapshotQueues);
ILogger storeSnapshotLogger = prefixedLogger(nodeEngine.getLogger(StoreSnapshotTasklet.class), jobPrefix);
StoreSnapshotTasklet ssTasklet = new StoreSnapshotTasklet(snapshotContext, ConcurrentInboundEdgeStream.create(ssConveyor, 0, 0, true, jobPrefix + "/ssFrom", null), new AsyncSnapshotWriterImpl(nodeEngine, snapshotContext, vertex.name(), memberIndex, memberCount, jobSerializationService), storeSnapshotLogger, vertex.name(), higherPriorityVertices.contains(vertex.vertexId()));
tasklets.add(ssTasklet);
}
int localProcessorIdx = 0;
for (Processor processor : processors) {
int globalProcessorIndex = memberIndex * vertex.localParallelism() + localProcessorIdx;
String processorPrefix = prefix(jobConfig.getName(), jobId, vertex.name(), globalProcessorIndex);
ILogger logger = prefixedLogger(nodeEngine.getLogger(processor.getClass()), processorPrefix);
ProcCtx context = new ProcCtx(nodeEngine, jobId, executionId, getJobConfig(), logger, vertex.name(), localProcessorIdx, globalProcessorIndex, isLightJob, partitionAssignment, vertex.localParallelism(), memberIndex, memberCount, tempDirectories, jobSerializationService, subject, processorClassLoader);
// createOutboundEdgeStreams() populates localConveyorMap and edgeSenderConveyorMap.
// Also populates instance fields: senderMap, receiverMap, tasklets.
List<OutboundEdgeStream> outboundStreams = createOutboundEdgeStreams(vertex, localProcessorIdx, jobPrefix, jobSerializationService);
List<InboundEdgeStream> inboundStreams = createInboundEdgeStreams(vertex, localProcessorIdx, jobPrefix, globalProcessorIndex);
OutboundCollector snapshotCollector = ssConveyor == null ? null : new ConveyorCollector(ssConveyor, localProcessorIdx, null);
// vertices which are only used for snapshot restore will not be marked as "source=true" in metrics
// also do not consider snapshot restore edges for determining source tag
boolean isSource = vertex.inboundEdges().stream().allMatch(EdgeDef::isSnapshotRestoreEdge) && !vertex.isSnapshotVertex();
ProcessorTasklet processorTasklet = new ProcessorTasklet(context, nodeEngine.getExecutionService().getExecutor(TASKLET_INIT_CLOSE_EXECUTOR_NAME), jobSerializationService, processor, inboundStreams, outboundStreams, snapshotContext, snapshotCollector, isSource);
tasklets.add(processorTasklet);
this.processors.add(processor);
localProcessorIdx++;
}
}
List<ReceiverTasklet> allReceivers = receiverMap.values().stream().flatMap(o -> o.values().stream()).flatMap(a -> a.values().stream()).collect(toList());
tasklets.addAll(allReceivers);
}
use of com.hazelcast.jet.impl.execution.ConveyorCollector in project hazelcast by hazelcast.
the class ExecutionPlan method createLocalOutboundCollector.
private OutboundCollector createLocalOutboundCollector(EdgeDef edge, int processorIndex, int totalPartitionCount, int[][] partitionsPerProcessor) {
int upstreamParallelism = edge.sourceVertex().localParallelism();
int downstreamParallelism = edge.destVertex().localParallelism();
int queueSize = edge.getConfig().getQueueSize();
int numRemoteMembers = ptionArrgmt.getRemotePartitionAssignment().size();
if (edge.routingPolicy() == RoutingPolicy.ISOLATED) {
ConcurrentConveyor<Object>[] localConveyors = localConveyorMap.computeIfAbsent(edge.edgeId(), edgeId -> {
int queueCount = upstreamParallelism / downstreamParallelism;
int remainder = upstreamParallelism % downstreamParallelism;
return Stream.concat(Arrays.stream(createConveyorArray(remainder, queueCount + 1, queueSize)), Arrays.stream(createConveyorArray(downstreamParallelism - remainder, Math.max(1, queueCount), queueSize))).toArray((IntFunction<ConcurrentConveyor<Object>[]>) ConcurrentConveyor[]::new);
});
OutboundCollector[] localCollectors = IntStream.range(0, downstreamParallelism).filter(i -> i % upstreamParallelism == processorIndex % downstreamParallelism).mapToObj(i -> new ConveyorCollector(localConveyors[i], processorIndex / downstreamParallelism, null)).toArray(OutboundCollector[]::new);
return compositeCollector(localCollectors, edge, totalPartitionCount, true);
} else {
ConcurrentConveyor<Object>[] localConveyors = localConveyorMap.computeIfAbsent(edge.edgeId(), edgeId -> {
int queueCount = upstreamParallelism + (!edge.isLocal() ? numRemoteMembers : 0);
return createConveyorArray(downstreamParallelism, queueCount, queueSize);
});
OutboundCollector[] localCollectors = new OutboundCollector[downstreamParallelism];
Arrays.setAll(localCollectors, n -> new ConveyorCollector(localConveyors[n], processorIndex, partitionsPerProcessor[n]));
return compositeCollector(localCollectors, edge, totalPartitionCount, true);
}
}
use of com.hazelcast.jet.impl.execution.ConveyorCollector in project hazelcast by hazelcast.
the class ExecutionPlan method createIfAbsentReceiverTasklet.
private void createIfAbsentReceiverTasklet(EdgeDef edge, String jobPrefix, int[][] ptionsPerProcessor, int totalPtionCount, InternalSerializationService jobSerializationService) {
final ConcurrentConveyor<Object>[] localConveyors = localConveyorMap.get(edge.edgeId());
receiverMap.computeIfAbsent(edge.destVertex().vertexId(), x -> new HashMap<>()).computeIfAbsent(edge.destOrdinal(), x -> {
Map<Address, ReceiverTasklet> addrToTasklet = new HashMap<>();
// create a receiver per address
int offset = 0;
for (Address addr : ptionArrgmt.getRemotePartitionAssignment().keySet()) {
final OutboundCollector[] collectors = new OutboundCollector[ptionsPerProcessor.length];
// assign the queues starting from end
final int queueOffset = --offset;
Arrays.setAll(collectors, n -> new ConveyorCollector(localConveyors[n], localConveyors[n].queueCount() + queueOffset, ptionsPerProcessor[n]));
final OutboundCollector collector = compositeCollector(collectors, edge, totalPtionCount, true);
ReceiverTasklet receiverTasklet = new ReceiverTasklet(collector, jobSerializationService, edge.getConfig().getReceiveWindowMultiplier(), getJetConfig().getFlowControlPeriodMs(), nodeEngine.getLoggingService(), addr, edge.destOrdinal(), edge.destVertex().name(), memberConnections.get(addr), jobPrefix);
addrToTasklet.put(addr, receiverTasklet);
}
return addrToTasklet;
});
}
Aggregations