/*
* Licensed to the Apache Software Foundation (ASF) under one
* or more contributor license agreements. See the NOTICE file
* distributed with this work for additional information
* regarding copyright ownership. The ASF licenses this file
* to you under the Apache License, Version 2.0 (the
* "License"); you may not use this file except in compliance
* with the License. You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
package org.apache.flink.optimizer.plantranslate;
import com.fasterxml.jackson.core.JsonFactory;
import org.apache.flink.api.common.JobID;
import org.apache.flink.api.common.aggregators.AggregatorRegistry;
import org.apache.flink.api.common.aggregators.AggregatorWithName;
import org.apache.flink.api.common.aggregators.ConvergenceCriterion;
import org.apache.flink.api.common.aggregators.LongSumAggregator;
import org.apache.flink.api.common.cache.DistributedCache;
import org.apache.flink.api.common.cache.DistributedCache.DistributedCacheEntry;
import org.apache.flink.api.common.distributions.DataDistribution;
import org.apache.flink.api.common.operators.util.UserCodeWrapper;
import org.apache.flink.api.common.typeutils.TypeSerializerFactory;
import org.apache.flink.configuration.GlobalConfiguration;
import org.apache.flink.optimizer.CompilerException;
import org.apache.flink.optimizer.dag.TempMode;
import org.apache.flink.optimizer.plan.BulkIterationPlanNode;
import org.apache.flink.optimizer.plan.BulkPartialSolutionPlanNode;
import org.apache.flink.optimizer.plan.Channel;
import org.apache.flink.optimizer.plan.DualInputPlanNode;
import org.apache.flink.optimizer.plan.IterationPlanNode;
import org.apache.flink.optimizer.plan.NAryUnionPlanNode;
import org.apache.flink.optimizer.plan.NamedChannel;
import org.apache.flink.optimizer.plan.OptimizedPlan;
import org.apache.flink.optimizer.plan.PlanNode;
import org.apache.flink.optimizer.plan.SingleInputPlanNode;
import org.apache.flink.optimizer.plan.SinkPlanNode;
import org.apache.flink.optimizer.plan.SolutionSetPlanNode;
import org.apache.flink.optimizer.plan.SourcePlanNode;
import org.apache.flink.optimizer.plan.WorksetIterationPlanNode;
import org.apache.flink.optimizer.plan.WorksetPlanNode;
import org.apache.flink.configuration.ConfigConstants;
import org.apache.flink.configuration.Configuration;
import org.apache.flink.optimizer.util.Utils;
import org.apache.flink.runtime.io.network.DataExchangeMode;
import org.apache.flink.runtime.io.network.partition.ResultPartitionType;
import org.apache.flink.runtime.iterative.convergence.WorksetEmptyConvergenceCriterion;
import org.apache.flink.runtime.iterative.task.IterationHeadTask;
import org.apache.flink.runtime.iterative.task.IterationIntermediateTask;
import org.apache.flink.runtime.iterative.task.IterationSynchronizationSinkTask;
import org.apache.flink.runtime.iterative.task.IterationTailTask;
import org.apache.flink.runtime.jobgraph.JobEdge;
import org.apache.flink.runtime.jobgraph.JobVertex;
import org.apache.flink.runtime.jobgraph.DistributionPattern;
import org.apache.flink.runtime.jobgraph.InputFormatVertex;
import org.apache.flink.runtime.jobgraph.JobGraph;
import org.apache.flink.runtime.jobgraph.OutputFormatVertex;
import org.apache.flink.runtime.jobmanager.scheduler.SlotSharingGroup;
import org.apache.flink.runtime.operators.BatchTask;
import org.apache.flink.runtime.operators.CoGroupDriver;
import org.apache.flink.runtime.operators.CoGroupWithSolutionSetFirstDriver;
import org.apache.flink.runtime.operators.CoGroupWithSolutionSetSecondDriver;
import org.apache.flink.runtime.operators.DataSinkTask;
import org.apache.flink.runtime.operators.DataSourceTask;
import org.apache.flink.runtime.operators.DriverStrategy;
import org.apache.flink.runtime.operators.JoinDriver;
import org.apache.flink.runtime.operators.JoinWithSolutionSetFirstDriver;
import org.apache.flink.runtime.operators.JoinWithSolutionSetSecondDriver;
import org.apache.flink.runtime.operators.NoOpDriver;
import org.apache.flink.runtime.operators.chaining.ChainedDriver;
import org.apache.flink.runtime.operators.shipping.ShipStrategyType;
import org.apache.flink.runtime.operators.util.LocalStrategy;
import org.apache.flink.runtime.operators.util.TaskConfig;
import org.apache.flink.util.StringUtils;
import org.apache.flink.util.Visitor;
import java.io.IOException;
import java.util.ArrayList;
import java.util.Collection;
import java.util.Collections;
import java.util.HashMap;
import java.util.Iterator;
import java.util.List;
import java.util.Map;
import java.util.Map.Entry;
/**
* This component translates the optimizer's resulting {@link org.apache.flink.optimizer.plan.OptimizedPlan}
* to a {@link org.apache.flink.runtime.jobgraph.JobGraph}. The translation is not strictly a one-to-one,
* because some nodes from the OptimizedPlan are collapsed into one job vertex.
*
* This translation does not make any decisions or assumptions. All degrees-of-freedom in the execution
* of the job are made by the Optimizer, so that this translation becomes a deterministic mapping.
*
* The basic method of operation is a top down traversal over the plan graph. On the way down, job vertices
* are created for the plan nodes, on the way back up, the nodes connect their predecessors.
*/
public class JobGraphGenerator implements Visitor<PlanNode> {
public static final String MERGE_ITERATION_AUX_TASKS_KEY = "compiler.merge-iteration-aux";
private static final boolean mergeIterationAuxTasks =
GlobalConfiguration.loadConfiguration().getBoolean(MERGE_ITERATION_AUX_TASKS_KEY, false);
private static final TaskInChain ALREADY_VISITED_PLACEHOLDER = new TaskInChain(null, null, null, null);
// ------------------------------------------------------------------------
private Map<PlanNode, JobVertex> vertices; // a map from optimizer nodes to job vertices
private Map<PlanNode, TaskInChain> chainedTasks; // a map from optimizer nodes to job vertices
private Map<IterationPlanNode, IterationDescriptor> iterations;
private List<TaskInChain> chainedTasksInSequence;
private List<JobVertex> auxVertices; // auxiliary vertices which are added during job graph generation
private final int defaultMaxFan;
private final float defaultSortSpillingThreshold;
private final boolean useLargeRecordHandler;
private int iterationIdEnumerator = 1;
private IterationPlanNode currentIteration; // the current the enclosing iteration
private List<IterationPlanNode> iterationStack; // stack of enclosing iterations
private SlotSharingGroup sharingGroup;
// ------------------------------------------------------------------------
/**
* Creates a new job graph generator that uses the default values for its resource configuration.
*/
public JobGraphGenerator() {
this.defaultMaxFan = ConfigConstants.DEFAULT_SPILLING_MAX_FAN;
this.defaultSortSpillingThreshold = ConfigConstants.DEFAULT_SORT_SPILLING_THRESHOLD;
this.useLargeRecordHandler = ConfigConstants.DEFAULT_USE_LARGE_RECORD_HANDLER;
}
public JobGraphGenerator(Configuration config) {
this.defaultMaxFan = config.getInteger(ConfigConstants.DEFAULT_SPILLING_MAX_FAN_KEY,
ConfigConstants.DEFAULT_SPILLING_MAX_FAN);
this.defaultSortSpillingThreshold = config.getFloat(ConfigConstants.DEFAULT_SORT_SPILLING_THRESHOLD_KEY,
ConfigConstants.DEFAULT_SORT_SPILLING_THRESHOLD);
this.useLargeRecordHandler = config.getBoolean(
ConfigConstants.USE_LARGE_RECORD_HANDLER_KEY,
ConfigConstants.DEFAULT_USE_LARGE_RECORD_HANDLER);
}
/**
* Translates a {@link org.apache.flink.optimizer.plan.OptimizedPlan} into a
* {@link org.apache.flink.runtime.jobgraph.JobGraph}.
*
* @param program Optimized plan that is translated into a JobGraph.
* @return JobGraph generated from the plan.
*/
public JobGraph compileJobGraph(OptimizedPlan program) {
return compileJobGraph(program, null);
}
public JobGraph compileJobGraph(OptimizedPlan program, JobID jobId) {
if (program == null) {
throw new NullPointerException("Program is null, did you called " +
"ExecutionEnvironment.execute()");
}
if (jobId == null) {
jobId = JobID.generate();
}
this.vertices = new HashMap<PlanNode, JobVertex>();
this.chainedTasks = new HashMap<PlanNode, TaskInChain>();
this.chainedTasksInSequence = new ArrayList<TaskInChain>();
this.auxVertices = new ArrayList<JobVertex>();
this.iterations = new HashMap<IterationPlanNode, IterationDescriptor>();
this.iterationStack = new ArrayList<IterationPlanNode>();
this.sharingGroup = new SlotSharingGroup();
// this starts the traversal that generates the job graph
program.accept(this);
// sanity check that we are not somehow in an iteration at the end
if (this.currentIteration != null) {
throw new CompilerException("The graph translation ended prematurely, leaving an unclosed iteration.");
}
// finalize the iterations
for (IterationDescriptor iteration : this.iterations.values()) {
if (iteration.getIterationNode() instanceof BulkIterationPlanNode) {
finalizeBulkIteration(iteration);
} else if (iteration.getIterationNode() instanceof WorksetIterationPlanNode) {
finalizeWorksetIteration(iteration);
} else {
throw new CompilerException();
}
}
// now that the traversal is done, we have the chained tasks write their configs into their
// parents' configurations
for (TaskInChain tic : this.chainedTasksInSequence) {
TaskConfig t = new TaskConfig(tic.getContainingVertex().getConfiguration());
t.addChainedTask(tic.getChainedTask(), tic.getTaskConfig(), tic.getTaskName());
}
// ----- attach the additional info to the job vertices, for display in the runtime monitor
attachOperatorNamesAndDescriptions();
// ----------- finalize the job graph -----------
// create the job graph object
JobGraph graph = new JobGraph(jobId, program.getJobName());
try {
graph.setExecutionConfig(program.getOriginalPlan().getExecutionConfig());
}
catch (IOException e) {
throw new CompilerException("Could not serialize the ExecutionConfig." +
"This indicates that non-serializable types (like custom serializers) were registered");
}
graph.setAllowQueuedScheduling(false);
graph.setSessionTimeout(program.getOriginalPlan().getSessionTimeout());
// add vertices to the graph
for (JobVertex vertex : this.vertices.values()) {
graph.addVertex(vertex);
}
for (JobVertex vertex : this.auxVertices) {
graph.addVertex(vertex);
vertex.setSlotSharingGroup(sharingGroup);
}
// add registered cache file into job configuration
for (Entry<String, DistributedCacheEntry> e : program.getOriginalPlan().getCachedFiles()) {
DistributedCache.writeFileInfoToConfig(e.getKey(), e.getValue(), graph.getJobConfiguration());
}
// release all references again
this.vertices = null;
this.chainedTasks = null;
this.chainedTasksInSequence = null;
this.auxVertices = null;
this.iterations = null;
this.iterationStack = null;
// return job graph
return graph;
}
/**
* This methods implements the pre-visiting during a depth-first traversal. It create the job vertex and
* sets local strategy.
*
* @param node
* The node that is currently processed.
* @return True, if the visitor should descend to the node's children, false if not.
* @see org.apache.flink.util.Visitor#preVisit(org.apache.flink.util.Visitable)
*/
@Override
public boolean preVisit(PlanNode node) {
// check if we have visited this node before. in non-tree graphs, this happens
if (this.vertices.containsKey(node) || this.chainedTasks.containsKey(node) || this.iterations.containsKey(node)) {
// return false to prevent further descend
return false;
}
// the vertex to be created for the current node
final JobVertex vertex;
try {
if (node instanceof SinkPlanNode) {
vertex = createDataSinkVertex((SinkPlanNode) node);
}
else if (node instanceof SourcePlanNode) {
vertex = createDataSourceVertex((SourcePlanNode) node);
}
else if (node instanceof BulkIterationPlanNode) {
BulkIterationPlanNode iterationNode = (BulkIterationPlanNode) node;
// for the bulk iteration, we skip creating anything for now. we create the graph
// for the step function in the post visit.
// check that the root of the step function has the same parallelism as the iteration.
// because the tail must have the same parallelism as the head, we can only merge the last
// operator with the tail, if they have the same parallelism. not merging is currently not
// implemented
PlanNode root = iterationNode.getRootOfStepFunction();
if (root.getParallelism() != node.getParallelism())
{
throw new CompilerException("Error: The final operator of the step " +
"function has a different parallelism than the iteration operator itself.");
}
IterationDescriptor descr = new IterationDescriptor(iterationNode, this.iterationIdEnumerator++);
this.iterations.put(iterationNode, descr);
vertex = null;
}
else if (node instanceof WorksetIterationPlanNode) {
WorksetIterationPlanNode iterationNode = (WorksetIterationPlanNode) node;
// we have the same constraints as for the bulk iteration
PlanNode nextWorkSet = iterationNode.getNextWorkSetPlanNode();
PlanNode solutionSetDelta = iterationNode.getSolutionSetDeltaPlanNode();
if (nextWorkSet.getParallelism() != node.getParallelism())
{
throw new CompilerException("It is currently not supported that the final operator of the step " +
"function has a different parallelism than the iteration operator itself.");
}
if (solutionSetDelta.getParallelism() != node.getParallelism())
{
throw new CompilerException("It is currently not supported that the final operator of the step " +
"function has a different parallelism than the iteration operator itself.");
}
IterationDescriptor descr = new IterationDescriptor(iterationNode, this.iterationIdEnumerator++);
this.iterations.put(iterationNode, descr);
vertex = null;
}
else if (node instanceof SingleInputPlanNode) {
vertex = createSingleInputVertex((SingleInputPlanNode) node);
}
else if (node instanceof DualInputPlanNode) {
vertex = createDualInputVertex((DualInputPlanNode) node);
}
else if (node instanceof NAryUnionPlanNode) {
// skip the union for now
vertex = null;
}
else if (node instanceof BulkPartialSolutionPlanNode) {
// create a head node (or not, if it is merged into its successor)
vertex = createBulkIterationHead((BulkPartialSolutionPlanNode) node);
}
else if (node instanceof SolutionSetPlanNode) {
// this represents an access into the solution set index.
// we do not create a vertex for the solution set here (we create the head at the workset place holder)
// we adjust the joins / cogroups that go into the solution set here
for (Channel c : node.getOutgoingChannels()) {
DualInputPlanNode target = (DualInputPlanNode) c.getTarget();
JobVertex accessingVertex = this.vertices.get(target);
TaskConfig conf = new TaskConfig(accessingVertex.getConfiguration());
int inputNum = c == target.getInput1() ? 0 : c == target.getInput2() ? 1 : -1;
// sanity checks
if (inputNum == -1) {
throw new CompilerException();
}
// adjust the driver
if (conf.getDriver().equals(JoinDriver.class)) {
conf.setDriver(inputNum == 0 ? JoinWithSolutionSetFirstDriver.class : JoinWithSolutionSetSecondDriver.class);
}
else if (conf.getDriver().equals(CoGroupDriver.class)) {
conf.setDriver(inputNum == 0 ? CoGroupWithSolutionSetFirstDriver.class : CoGroupWithSolutionSetSecondDriver.class);
}
else {
throw new CompilerException("Found join with solution set using incompatible operator (only Join/CoGroup are valid).");
}
}
// make sure we do not visit this node again. for that, we add a 'already seen' entry into one of the sets
this.chainedTasks.put(node, ALREADY_VISITED_PLACEHOLDER);
vertex = null;
}
else if (node instanceof WorksetPlanNode) {
// create the iteration head here
vertex = createWorksetIterationHead((WorksetPlanNode) node);
}
else {
throw new CompilerException("Unrecognized node type: " + node.getClass().getName());
}
}
catch (Exception e) {
throw new CompilerException("Error translating node '" + node + "': " + e.getMessage(), e);
}
// check if a vertex was created, or if it was chained or skipped
if (vertex != null) {
// set parallelism
int pd = node.getParallelism();
vertex.setParallelism(pd);
vertex.setMaxParallelism(pd);
vertex.setSlotSharingGroup(sharingGroup);
// check whether this vertex is part of an iteration step function
if (this.currentIteration != null) {
// check that the task has the same parallelism as the iteration as such
PlanNode iterationNode = (PlanNode) this.currentIteration;
if (iterationNode.getParallelism() < pd) {
throw new CompilerException("Error: All functions that are part of an iteration must have the same, or a lower, parallelism than the iteration operator.");
}
// store the id of the iterations the step functions participate in
IterationDescriptor descr = this.iterations.get(this.currentIteration);
new TaskConfig(vertex.getConfiguration()).setIterationId(descr.getId());
}
// store in the map
this.vertices.put(node, vertex);
}
// returning true causes deeper descend
return true;
}
/**
* This method implements the post-visit during the depth-first traversal. When the post visit happens,
* all of the descendants have been processed, so this method connects all of the current node's
* predecessors to the current node.
*
* @param node
* The node currently processed during the post-visit.
* @see org.apache.flink.util.Visitor#postVisit(org.apache.flink.util.Visitable) t
*/
@Override
public void postVisit(PlanNode node) {
try {
// --------- check special cases for which we handle post visit differently ----------
// skip data source node (they have no inputs)
// also, do nothing for union nodes, we connect them later when gathering the inputs for a task
// solution sets have no input. the initial solution set input is connected when the iteration node is in its postVisit
if (node instanceof SourcePlanNode || node instanceof NAryUnionPlanNode || node instanceof SolutionSetPlanNode) {
return;
}
// check if we have an iteration. in that case, translate the step function now
if (node instanceof IterationPlanNode) {
// prevent nested iterations
if (node.isOnDynamicPath()) {
throw new CompilerException("Nested Iterations are not possible at the moment!");
}
// if we recursively go into an iteration (because the constant path of one iteration contains
// another one), we push the current one onto the stack
if (this.currentIteration != null) {
this.iterationStack.add(this.currentIteration);
}
this.currentIteration = (IterationPlanNode) node;
this.currentIteration.acceptForStepFunction(this);
// pop the current iteration from the stack
if (this.iterationStack.isEmpty()) {
this.currentIteration = null;
} else {
this.currentIteration = this.iterationStack.remove(this.iterationStack.size() - 1);
}
// inputs for initial bulk partial solution or initial workset are already connected to the iteration head in the head's post visit.
// connect the initial solution set now.
if (node instanceof WorksetIterationPlanNode) {
// connect the initial solution set
WorksetIterationPlanNode wsNode = (WorksetIterationPlanNode) node;
JobVertex headVertex = this.iterations.get(wsNode).getHeadTask();
TaskConfig headConfig = new TaskConfig(headVertex.getConfiguration());
int inputIndex = headConfig.getDriverStrategy().getNumInputs();
headConfig.setIterationHeadSolutionSetInputIndex(inputIndex);
translateChannel(wsNode.getInitialSolutionSetInput(), inputIndex, headVertex, headConfig, false);
}
return;
}
final JobVertex targetVertex = this.vertices.get(node);
// --------- Main Path: Translation of channels ----------
//
// There are two paths of translation: One for chained tasks (or merged tasks in general),
// which do not have their own task vertex. The other for tasks that have their own vertex,
// or are the primary task in a vertex (to which the others are chained).
// check whether this node has its own task, or is merged with another one
if (targetVertex == null) {
// node's task is merged with another task. it is either chained, of a merged head vertex
// from an iteration
final TaskInChain chainedTask;
if ((chainedTask = this.chainedTasks.get(node)) != null) {
// Chained Task. Sanity check first...
final Iterator<Channel> inConns = node.getInputs().iterator();
if (!inConns.hasNext()) {
throw new CompilerException("Bug: Found chained task with no input.");
}
final Channel inConn = inConns.next();
if (inConns.hasNext()) {
throw new CompilerException("Bug: Found a chained task with more than one input!");
}
if (inConn.getLocalStrategy() != null && inConn.getLocalStrategy() != LocalStrategy.NONE) {
throw new CompilerException("Bug: Found a chained task with an input local strategy.");
}
if (inConn.getShipStrategy() != null && inConn.getShipStrategy() != ShipStrategyType.FORWARD) {
throw new CompilerException("Bug: Found a chained task with an input ship strategy other than FORWARD.");
}
JobVertex container = chainedTask.getContainingVertex();
if (container == null) {
final PlanNode sourceNode = inConn.getSource();
container = this.vertices.get(sourceNode);
if (container == null) {
// predecessor is itself chained
container = this.chainedTasks.get(sourceNode).getContainingVertex();
if (container == null) {
throw new IllegalStateException("Bug: Chained task predecessor has not been assigned its containing vertex.");
}
} else {
// predecessor is a proper task job vertex and this is the first chained task. add a forward connection entry.
new TaskConfig(container.getConfiguration()).addOutputShipStrategy(ShipStrategyType.FORWARD);
}
chainedTask.setContainingVertex(container);
}
// add info about the input serializer type
chainedTask.getTaskConfig().setInputSerializer(inConn.getSerializer(), 0);
// update name of container task
String containerTaskName = container.getName();
if (containerTaskName.startsWith("CHAIN ")) {
container.setName(containerTaskName + " -> " + chainedTask.getTaskName());
} else {
container.setName("CHAIN " + containerTaskName + " -> " + chainedTask.getTaskName());
}
//update resource of container task
container.setResources(container.getMinResources().merge(node.getMinResources()),
container.getPreferredResources().merge(node.getPreferredResources()));
this.chainedTasksInSequence.add(chainedTask);
return;
}
else if (node instanceof BulkPartialSolutionPlanNode ||
node instanceof WorksetPlanNode)
{
// merged iteration head task. the task that the head is merged with will take care of it
return;
} else {
throw new CompilerException("Bug: Unrecognized merged task vertex.");
}
}
// -------- Here, we translate non-chained tasks -------------
if (this.currentIteration != null) {
JobVertex head = this.iterations.get(this.currentIteration).getHeadTask();
// Exclude static code paths from the co-location constraint, because otherwise
// their execution determines the deployment slots of the co-location group
if (node.isOnDynamicPath()) {
targetVertex.setStrictlyCoLocatedWith(head);
}
}
// create the config that will contain all the description of the inputs
final TaskConfig targetVertexConfig = new TaskConfig(targetVertex.getConfiguration());
// get the inputs. if this node is the head of an iteration, we obtain the inputs from the
// enclosing iteration node, because the inputs are the initial inputs to the iteration.
final Iterator<Channel> inConns;
if (node instanceof BulkPartialSolutionPlanNode) {
inConns = ((BulkPartialSolutionPlanNode) node).getContainingIterationNode().getInputs().iterator();
// because the partial solution has its own vertex, is has only one (logical) input.
// note this in the task configuration
targetVertexConfig.setIterationHeadPartialSolutionOrWorksetInputIndex(0);
} else if (node instanceof WorksetPlanNode) {
WorksetPlanNode wspn = (WorksetPlanNode) node;
// input that is the initial workset
inConns = Collections.singleton(wspn.getContainingIterationNode().getInput2()).iterator();
// because we have a stand-alone (non-merged) workset iteration head, the initial workset will
// be input 0 and the solution set will be input 1
targetVertexConfig.setIterationHeadPartialSolutionOrWorksetInputIndex(0);
targetVertexConfig.setIterationHeadSolutionSetInputIndex(1);
} else {
inConns = node.getInputs().iterator();
}
if (!inConns.hasNext()) {
throw new CompilerException("Bug: Found a non-source task with no input.");
}
int inputIndex = 0;
while (inConns.hasNext()) {
Channel input = inConns.next();
inputIndex += translateChannel(input, inputIndex, targetVertex, targetVertexConfig, false);
}
// broadcast variables
int broadcastInputIndex = 0;
for (NamedChannel broadcastInput: node.getBroadcastInputs()) {
int broadcastInputIndexDelta = translateChannel(broadcastInput, broadcastInputIndex, targetVertex, targetVertexConfig, true);
targetVertexConfig.setBroadcastInputName(broadcastInput.getName(), broadcastInputIndex);
targetVertexConfig.setBroadcastInputSerializer(broadcastInput.getSerializer(), broadcastInputIndex);
broadcastInputIndex += broadcastInputIndexDelta;
}
} catch (Exception e) {
throw new CompilerException(
"An error occurred while translating the optimized plan to a JobGraph: " + e.getMessage(), e);
}
}
private int translateChannel(Channel input, int inputIndex, JobVertex targetVertex,
TaskConfig targetVertexConfig, boolean isBroadcast) throws Exception
{
final PlanNode inputPlanNode = input.getSource();
final Iterator<Channel> allInChannels;
if (inputPlanNode instanceof NAryUnionPlanNode) {
allInChannels = ((NAryUnionPlanNode) inputPlanNode).getListOfInputs().iterator();
// If the union node has a batch data exchange, we have to adopt the exchange mode of
// the inputs of the union as well, because the optimizer has a separate union
// node, which does not exist in the JobGraph. Otherwise, this can result in
// deadlocks when closing a branching flow at runtime.
for (Channel in : inputPlanNode.getInputs()) {
if (input.getDataExchangeMode().equals(DataExchangeMode.BATCH)) {
in.setDataExchangeMode(DataExchangeMode.BATCH);
}
if (isBroadcast) {
in.setShipStrategy(ShipStrategyType.BROADCAST, in.getDataExchangeMode());
}
}
}
else if (inputPlanNode instanceof BulkPartialSolutionPlanNode) {
if (this.vertices.get(inputPlanNode) == null) {
// merged iteration head
final BulkPartialSolutionPlanNode pspn = (BulkPartialSolutionPlanNode) inputPlanNode;
final BulkIterationPlanNode iterationNode = pspn.getContainingIterationNode();
// check if the iteration's input is a union
if (iterationNode.getInput().getSource() instanceof NAryUnionPlanNode) {
allInChannels = (iterationNode.getInput().getSource()).getInputs().iterator();
} else {
allInChannels = Collections.singletonList(iterationNode.getInput()).iterator();
}
// also, set the index of the gate with the partial solution
targetVertexConfig.setIterationHeadPartialSolutionOrWorksetInputIndex(inputIndex);
} else {
// standalone iteration head
allInChannels = Collections.singletonList(input).iterator();
}
} else if (inputPlanNode instanceof WorksetPlanNode) {
if (this.vertices.get(inputPlanNode) == null) {
// merged iteration head
final WorksetPlanNode wspn = (WorksetPlanNode) inputPlanNode;
final WorksetIterationPlanNode iterationNode = wspn.getContainingIterationNode();
// check if the iteration's input is a union
if (iterationNode.getInput2().getSource() instanceof NAryUnionPlanNode) {
allInChannels = (iterationNode.getInput2().getSource()).getInputs().iterator();
} else {
allInChannels = Collections.singletonList(iterationNode.getInput2()).iterator();
}
// also, set the index of the gate with the partial solution
targetVertexConfig.setIterationHeadPartialSolutionOrWorksetInputIndex(inputIndex);
} else {
// standalone iteration head
allInChannels = Collections.singletonList(input).iterator();
}
} else if (inputPlanNode instanceof SolutionSetPlanNode) {
// for now, skip connections with the solution set node, as this is a local index access (later to be parameterized here)
// rather than a vertex connection
return 0;
} else {
allInChannels = Collections.singletonList(input).iterator();
}
// check that the type serializer is consistent
TypeSerializerFactory<?> typeSerFact = null;
// accounting for channels on the dynamic path
int numChannelsTotal = 0;
int numChannelsDynamicPath = 0;
int numDynamicSenderTasksTotal = 0;
// expand the channel to all the union channels, in case there is a union operator at its source
while (allInChannels.hasNext()) {
final Channel inConn = allInChannels.next();
// sanity check the common serializer
if (typeSerFact == null) {
typeSerFact = inConn.getSerializer();
} else if (!typeSerFact.equals(inConn.getSerializer())) {
throw new CompilerException("Conflicting types in union operator.");
}
final PlanNode sourceNode = inConn.getSource();
JobVertex sourceVertex = this.vertices.get(sourceNode);
TaskConfig sourceVertexConfig;
if (sourceVertex == null) {
// this predecessor is chained to another task or an iteration
final TaskInChain chainedTask;
final IterationDescriptor iteration;
if ((chainedTask = this.chainedTasks.get(sourceNode)) != null) {
// push chained task
if (chainedTask.getContainingVertex() == null) {
throw new IllegalStateException("Bug: Chained task has not been assigned its containing vertex when connecting.");
}
sourceVertex = chainedTask.getContainingVertex();
sourceVertexConfig = chainedTask.getTaskConfig();
} else if ((iteration = this.iterations.get(sourceNode)) != null) {
// predecessor is an iteration
sourceVertex = iteration.getHeadTask();
sourceVertexConfig = iteration.getHeadFinalResultConfig();
} else {
throw new CompilerException("Bug: Could not resolve source node for a channel.");
}
} else {
// predecessor is its own vertex
sourceVertexConfig = new TaskConfig(sourceVertex.getConfiguration());
}
DistributionPattern pattern = connectJobVertices(
inConn, inputIndex, sourceVertex, sourceVertexConfig, targetVertex, targetVertexConfig, isBroadcast);
// accounting on channels and senders
numChannelsTotal++;
if (inConn.isOnDynamicPath()) {
numChannelsDynamicPath++;
numDynamicSenderTasksTotal += getNumberOfSendersPerReceiver(pattern,
sourceVertex.getParallelism(), targetVertex.getParallelism());
}
}
// for the iterations, check that the number of dynamic channels is the same as the number
// of channels for this logical input. this condition is violated at the moment, if there
// is a union between nodes on the static and nodes on the dynamic path
if (numChannelsDynamicPath > 0 && numChannelsTotal != numChannelsDynamicPath) {
throw new CompilerException("Error: It is currently not supported to union between dynamic and static path in an iteration.");
}
if (numDynamicSenderTasksTotal > 0) {
if (isBroadcast) {
targetVertexConfig.setBroadcastGateIterativeWithNumberOfEventsUntilInterrupt(inputIndex, numDynamicSenderTasksTotal);
} else {
targetVertexConfig.setGateIterativeWithNumberOfEventsUntilInterrupt(inputIndex, numDynamicSenderTasksTotal);
}
}
// the local strategy is added only once. in non-union case that is the actual edge,
// in the union case, it is the edge between union and the target node
addLocalInfoFromChannelToConfig(input, targetVertexConfig, inputIndex, isBroadcast);
return 1;
}
private int getNumberOfSendersPerReceiver(DistributionPattern pattern, int numSenders, int numReceivers) {
if (pattern == DistributionPattern.ALL_TO_ALL) {
return numSenders;
} else if (pattern == DistributionPattern.POINTWISE) {
if (numSenders != numReceivers) {
if (numReceivers == 1) {
return numSenders;
}
else if (numSenders == 1) {
return 1;
}
else {
throw new CompilerException("Error: A changing parallelism is currently " +
"not supported between tasks within an iteration.");
}
} else {
return 1;
}
} else {
throw new CompilerException("Unknown distribution pattern for channels: " + pattern);
}
}
// ------------------------------------------------------------------------
// Methods for creating individual vertices
// ------------------------------------------------------------------------
private JobVertex createSingleInputVertex(SingleInputPlanNode node) throws CompilerException {
final String taskName = node.getNodeName();
final DriverStrategy ds = node.getDriverStrategy();
// check, whether chaining is possible
boolean chaining;
{
Channel inConn = node.getInput();
PlanNode pred = inConn.getSource();
chaining = ds.getPushChainDriverClass() != null &&
!(pred instanceof NAryUnionPlanNode) && // first op after union is stand-alone, because union is merged
!(pred instanceof BulkPartialSolutionPlanNode) && // partial solution merges anyways
!(pred instanceof WorksetPlanNode) && // workset merges anyways
!(pred instanceof IterationPlanNode) && // cannot chain with iteration heads currently
inConn.getShipStrategy() == ShipStrategyType.FORWARD &&
inConn.getLocalStrategy() == LocalStrategy.NONE &&
pred.getOutgoingChannels().size() == 1 &&
node.getParallelism() == pred.getParallelism() &&
node.getBroadcastInputs().isEmpty();
// cannot chain the nodes that produce the next workset or the next solution set, if they are not the
// in a tail
if (this.currentIteration != null && this.currentIteration instanceof WorksetIterationPlanNode &&
node.getOutgoingChannels().size() > 0)
{
WorksetIterationPlanNode wspn = (WorksetIterationPlanNode) this.currentIteration;
if (wspn.getSolutionSetDeltaPlanNode() == pred || wspn.getNextWorkSetPlanNode() == pred) {
chaining = false;
}
}
// cannot chain the nodes that produce the next workset in a bulk iteration if a termination criterion follows
if (this.currentIteration != null && this.currentIteration instanceof BulkIterationPlanNode)
{
BulkIterationPlanNode wspn = (BulkIterationPlanNode) this.currentIteration;
if (node == wspn.getRootOfTerminationCriterion() && wspn.getRootOfStepFunction() == pred){
chaining = false;
}else if(node.getOutgoingChannels().size() > 0 &&(wspn.getRootOfStepFunction() == pred ||
wspn.getRootOfTerminationCriterion() == pred)) {
chaining = false;
}
}
}
final JobVertex vertex;
final TaskConfig config;
if (chaining) {
vertex = null;
config = new TaskConfig(new Configuration());
this.chainedTasks.put(node, new TaskInChain(node, ds.getPushChainDriverClass(), config, taskName));
} else {
// create task vertex
vertex = new JobVertex(taskName);
vertex.setResources(node.getMinResources(), node.getPreferredResources());
vertex.setInvokableClass((this.currentIteration != null && node.isOnDynamicPath()) ? IterationIntermediateTask.class : BatchTask.class);
config = new TaskConfig(vertex.getConfiguration());
config.setDriver(ds.getDriverClass());
}
// set user code
config.setStubWrapper(node.getProgramOperator().getUserCodeWrapper());
config.setStubParameters(node.getProgramOperator().getParameters());
// set the driver strategy
config.setDriverStrategy(ds);
for (int i = 0; i < ds.getNumRequiredComparators(); i++) {
config.setDriverComparator(node.getComparator(i), i);
}
// assign memory, file-handles, etc.
assignDriverResources(node, config);
return vertex;
}
private JobVertex createDualInputVertex(DualInputPlanNode node) throws CompilerException {
final String taskName = node.getNodeName();
final DriverStrategy ds = node.getDriverStrategy();
final JobVertex vertex = new JobVertex(taskName);
final TaskConfig config = new TaskConfig(vertex.getConfiguration());
vertex.setResources(node.getMinResources(), node.getPreferredResources());
vertex.setInvokableClass( (this.currentIteration != null && node.isOnDynamicPath()) ? IterationIntermediateTask.class : BatchTask.class);
// set user code
config.setStubWrapper(node.getProgramOperator().getUserCodeWrapper());
config.setStubParameters(node.getProgramOperator().getParameters());
// set the driver strategy
config.setDriver(ds.getDriverClass());
config.setDriverStrategy(ds);
if (node.getComparator1() != null) {
config.setDriverComparator(node.getComparator1(), 0);
}
if (node.getComparator2() != null) {
config.setDriverComparator(node.getComparator2(), 1);
}
if (node.getPairComparator() != null) {
config.setDriverPairComparator(node.getPairComparator());
}
// assign memory, file-handles, etc.
assignDriverResources(node, config);
return vertex;
}
private InputFormatVertex createDataSourceVertex(SourcePlanNode node) throws CompilerException {
final InputFormatVertex vertex = new InputFormatVertex(node.getNodeName());
final TaskConfig config = new TaskConfig(vertex.getConfiguration());
vertex.setResources(node.getMinResources(), node.getPreferredResources());
vertex.setInvokableClass(DataSourceTask.class);
vertex.setFormatDescription(getDescriptionForUserCode(node.getProgramOperator().getUserCodeWrapper()));
// set user code
config.setStubWrapper(node.getProgramOperator().getUserCodeWrapper());
config.setStubParameters(node.getProgramOperator().getParameters());
config.setOutputSerializer(node.getSerializer());
return vertex;
}
private JobVertex createDataSinkVertex(SinkPlanNode node) throws CompilerException {
final OutputFormatVertex vertex = new OutputFormatVertex(node.getNodeName());
final TaskConfig config = new TaskConfig(vertex.getConfiguration());
vertex.setResources(node.getMinResources(), node.getPreferredResources());
vertex.setInvokableClass(DataSinkTask.class);
vertex.setFormatDescription(getDescriptionForUserCode(node.getProgramOperator().getUserCodeWrapper()));
// set user code
config.setStubWrapper(node.getProgramOperator().getUserCodeWrapper());
config.setStubParameters(node.getProgramOperator().getParameters());
return vertex;
}
private JobVertex createBulkIterationHead(BulkPartialSolutionPlanNode pspn) {
// get the bulk iteration that corresponds to this partial solution node
final BulkIterationPlanNode iteration = pspn.getContainingIterationNode();
// check whether we need an individual vertex for the partial solution, or whether we
// attach ourselves to the vertex of the parent node. We can combine the head with a node of
// the step function, if
// 1) There is one parent that the partial solution connects to via a forward pattern and no
// local strategy
// 2) parallelism and the number of subtasks per instance does not change
// 3) That successor is not a union
// 4) That successor is not itself the last node of the step function
// 5) There is no local strategy on the edge for the initial partial solution, as
// this translates to a local strategy that would only be executed in the first iteration
final boolean merge;
if (mergeIterationAuxTasks && pspn.getOutgoingChannels().size() == 1) {
final Channel c = pspn.getOutgoingChannels().get(0);
final PlanNode successor = c.getTarget();
merge = c.getShipStrategy() == ShipStrategyType.FORWARD &&
c.getLocalStrategy() == LocalStrategy.NONE &&
c.getTempMode() == TempMode.NONE &&
successor.getParallelism() == pspn.getParallelism() &&
!(successor instanceof NAryUnionPlanNode) &&
successor != iteration.getRootOfStepFunction() &&
iteration.getInput().getLocalStrategy() == LocalStrategy.NONE;
} else {
merge = false;
}
// create or adopt the head vertex
final JobVertex toReturn;
final JobVertex headVertex;
final TaskConfig headConfig;
if (merge) {
final PlanNode successor = pspn.getOutgoingChannels().get(0).getTarget();
headVertex = this.vertices.get(successor);
if (headVertex == null) {
throw new CompilerException(
"Bug: Trying to merge solution set with its successor, but successor has not been created.");
}
// reset the vertex type to iteration head
headVertex.setInvokableClass(IterationHeadTask.class);
headConfig = new TaskConfig(headVertex.getConfiguration());
toReturn = null;
} else {
// instantiate the head vertex and give it a no-op driver as the driver strategy.
// everything else happens in the post visit, after the input (the initial partial solution)
// is connected.
headVertex = new JobVertex("PartialSolution ("+iteration.getNodeName()+")");
headVertex.setResources(iteration.getMinResources(), iteration.getPreferredResources());
headVertex.setInvokableClass(IterationHeadTask.class);
headConfig = new TaskConfig(headVertex.getConfiguration());
headConfig.setDriver(NoOpDriver.class);
toReturn = headVertex;
}
// create the iteration descriptor and the iteration to it
IterationDescriptor descr = this.iterations.get(iteration);
if (descr == null) {
throw new CompilerException("Bug: Iteration descriptor was not created at when translating the iteration node.");
}
descr.setHeadTask(headVertex, headConfig);
return toReturn;
}
private JobVertex createWorksetIterationHead(WorksetPlanNode wspn) {
// get the bulk iteration that corresponds to this partial solution node
final WorksetIterationPlanNode iteration = wspn.getContainingIterationNode();
// check whether we need an individual vertex for the partial solution, or whether we
// attach ourselves to the vertex of the parent node. We can combine the head with a node of
// the step function, if
// 1) There is one parent that the partial solution connects to via a forward pattern and no
// local strategy
// 2) parallelism and the number of subtasks per instance does not change
// 3) That successor is not a union
// 4) That successor is not itself the last node of the step function
// 5) There is no local strategy on the edge for the initial workset, as
// this translates to a local strategy that would only be executed in the first superstep
final boolean merge;
if (mergeIterationAuxTasks && wspn.getOutgoingChannels().size() == 1) {
final Channel c = wspn.getOutgoingChannels().get(0);
final PlanNode successor = c.getTarget();
merge = c.getShipStrategy() == ShipStrategyType.FORWARD &&
c.getLocalStrategy() == LocalStrategy.NONE &&
c.getTempMode() == TempMode.NONE &&
successor.getParallelism() == wspn.getParallelism() &&
!(successor instanceof NAryUnionPlanNode) &&
successor != iteration.getNextWorkSetPlanNode() &&
iteration.getInitialWorksetInput().getLocalStrategy() == LocalStrategy.NONE;
} else {
merge = false;
}
// create or adopt the head vertex
final JobVertex toReturn;
final JobVertex headVertex;
final TaskConfig headConfig;
if (merge) {
final PlanNode successor = wspn.getOutgoingChannels().get(0).getTarget();
headVertex = this.vertices.get(successor);
if (headVertex == null) {
throw new CompilerException(
"Bug: Trying to merge solution set with its sucessor, but successor has not been created.");
}
// reset the vertex type to iteration head
headVertex.setInvokableClass(IterationHeadTask.class);
headConfig = new TaskConfig(headVertex.getConfiguration());
toReturn = null;
} else {
// instantiate the head vertex and give it a no-op driver as the driver strategy.
// everything else happens in the post visit, after the input (the initial partial solution)
// is connected.
headVertex = new JobVertex("IterationHead("+iteration.getNodeName()+")");
headVertex.setResources(iteration.getMinResources(), iteration.getPreferredResources());
headVertex.setInvokableClass(IterationHeadTask.class);
headConfig = new TaskConfig(headVertex.getConfiguration());
headConfig.setDriver(NoOpDriver.class);
toReturn = headVertex;
}
headConfig.setSolutionSetUnmanaged(iteration.getIterationNode().getIterationContract().isSolutionSetUnManaged());
// create the iteration descriptor and the iteration to it
IterationDescriptor descr = this.iterations.get(iteration);
if (descr == null) {
throw new CompilerException("Bug: Iteration descriptor was not created at when translating the iteration node.");
}
descr.setHeadTask(headVertex, headConfig);
return toReturn;
}
private void assignDriverResources(PlanNode node, TaskConfig config) {
final double relativeMem = node.getRelativeMemoryPerSubTask();
if (relativeMem > 0) {
config.setRelativeMemoryDriver(relativeMem);
config.setFilehandlesDriver(this.defaultMaxFan);
config.setSpillingThresholdDriver(this.defaultSortSpillingThreshold);
}
}
private void assignLocalStrategyResources(Channel c, TaskConfig config, int inputNum) {
if (c.getRelativeMemoryLocalStrategy() > 0) {
config.setRelativeMemoryInput(inputNum, c.getRelativeMemoryLocalStrategy());
config.setFilehandlesInput(inputNum, this.defaultMaxFan);
config.setSpillingThresholdInput(inputNum, this.defaultSortSpillingThreshold);
config.setUseLargeRecordHandler(this.useLargeRecordHandler);
}
}
// ------------------------------------------------------------------------
// Connecting Vertices
// ------------------------------------------------------------------------
/**
* NOTE: The channel for global and local strategies are different if we connect a union. The global strategy
* channel is then the channel into the union node, the local strategy channel the one from the union to the
* actual target operator.
*
* @throws CompilerException
*/
private DistributionPattern connectJobVertices(Channel channel, int inputNumber,
final JobVertex sourceVertex, final TaskConfig sourceConfig,
final JobVertex targetVertex, final TaskConfig targetConfig, boolean isBroadcast)
throws CompilerException
{
// ------------ connect the vertices to the job graph --------------
final DistributionPattern distributionPattern;
switch (channel.getShipStrategy()) {
case FORWARD:
distributionPattern = DistributionPattern.POINTWISE;
break;
case PARTITION_RANDOM:
case BROADCAST:
case PARTITION_HASH:
case PARTITION_CUSTOM:
case PARTITION_RANGE:
case PARTITION_FORCED_REBALANCE:
distributionPattern = DistributionPattern.ALL_TO_ALL;
break;
default:
throw new RuntimeException("Unknown runtime ship strategy: " + channel.getShipStrategy());
}
final ResultPartitionType resultType;
switch (channel.getDataExchangeMode()) {
case PIPELINED:
resultType = ResultPartitionType.PIPELINED;
break;
case BATCH:
// BLOCKING results are currently not supported in closed loop iterations
//
// See https://issues.apache.org/jira/browse/FLINK-1713 for details
resultType = channel.getSource().isOnDynamicPath()
? ResultPartitionType.PIPELINED
: ResultPartitionType.BLOCKING;
break;
case PIPELINE_WITH_BATCH_FALLBACK:
throw new UnsupportedOperationException("Data exchange mode " +
channel.getDataExchangeMode() + " currently not supported.");
default:
throw new UnsupportedOperationException("Unknown data exchange mode.");
}
JobEdge edge = targetVertex.connectNewDataSetAsInput(sourceVertex, distributionPattern, resultType);
// -------------- configure the source task's ship strategy strategies in task config --------------
final int outputIndex = sourceConfig.getNumOutputs();
sourceConfig.addOutputShipStrategy(channel.getShipStrategy());
if (outputIndex == 0) {
sourceConfig.setOutputSerializer(channel.getSerializer());
}
if (channel.getShipStrategyComparator() != null) {
sourceConfig.setOutputComparator(channel.getShipStrategyComparator(), outputIndex);
}
if (channel.getShipStrategy() == ShipStrategyType.PARTITION_RANGE) {
final DataDistribution dataDistribution = channel.getDataDistribution();
if (dataDistribution != null) {
sourceConfig.setOutputDataDistribution(dataDistribution, outputIndex);
} else {
throw new RuntimeException("Range partitioning requires data distribution.");
}
}
if (channel.getShipStrategy() == ShipStrategyType.PARTITION_CUSTOM) {
if (channel.getPartitioner() != null) {
sourceConfig.setOutputPartitioner(channel.getPartitioner(), outputIndex);
} else {
throw new CompilerException("The ship strategy was set to custom partitioning, but no partitioner was set.");
}
}
// ---------------- configure the receiver -------------------
if (isBroadcast) {
targetConfig.addBroadcastInputToGroup(inputNumber);
} else {
targetConfig.addInputToGroup(inputNumber);
}
// ---------------- attach the additional infos to the job edge -------------------
String shipStrategy = JsonMapper.getShipStrategyString(channel.getShipStrategy());
if (channel.getShipStrategyKeys() != null && channel.getShipStrategyKeys().size() > 0) {
shipStrategy += " on " + (channel.getShipStrategySortOrder() == null ?
channel.getShipStrategyKeys().toString() :
Utils.createOrdering(channel.getShipStrategyKeys(), channel.getShipStrategySortOrder()).toString());
}
String localStrategy;
if (channel.getLocalStrategy() == null || channel.getLocalStrategy() == LocalStrategy.NONE) {
localStrategy = null;
}
else {
localStrategy = JsonMapper.getLocalStrategyString(channel.getLocalStrategy());
if (localStrategy != null && channel.getLocalStrategyKeys() != null && channel.getLocalStrategyKeys().size() > 0) {
localStrategy += " on " + (channel.getLocalStrategySortOrder() == null ?
channel.getLocalStrategyKeys().toString() :
Utils.createOrdering(channel.getLocalStrategyKeys(), channel.getLocalStrategySortOrder()).toString());
}
}
String caching = channel.getTempMode() == TempMode.NONE ? null : channel.getTempMode().toString();
edge.setShipStrategyName(shipStrategy);
edge.setPreProcessingOperationName(localStrategy);
edge.setOperatorLevelCachingDescription(caching);
return distributionPattern;
}
private void addLocalInfoFromChannelToConfig(Channel channel, TaskConfig config, int inputNum, boolean isBroadcastChannel) {
// serializer
if (isBroadcastChannel) {
config.setBroadcastInputSerializer(channel.getSerializer(), inputNum);
if (channel.getLocalStrategy() != LocalStrategy.NONE || (channel.getTempMode() != null && channel.getTempMode() != TempMode.NONE)) {
throw new CompilerException("Found local strategy or temp mode on a broadcast variable channel.");
} else {
return;
}
} else {
config.setInputSerializer(channel.getSerializer(), inputNum);
}
// local strategy
if (channel.getLocalStrategy() != LocalStrategy.NONE) {
config.setInputLocalStrategy(inputNum, channel.getLocalStrategy());
if (channel.getLocalStrategyComparator() != null) {
config.setInputComparator(channel.getLocalStrategyComparator(), inputNum);
}
}
assignLocalStrategyResources(channel, config, inputNum);
// materialization / caching
if (channel.getTempMode() != null) {
final TempMode tm = channel.getTempMode();
boolean needsMemory = false;
// Don't add a pipeline breaker if the data exchange is already blocking, EXCEPT the channel is within an iteration.
if (tm.breaksPipeline() &&
(channel.isOnDynamicPath() || channel.getDataExchangeMode() != DataExchangeMode.BATCH) ) {
config.setInputAsynchronouslyMaterialized(inputNum, true);
needsMemory = true;
}
if (tm.isCached()) {
config.setInputCached(inputNum, true);
needsMemory = true;
}
if (needsMemory) {
// sanity check
if (tm == TempMode.NONE || channel.getRelativeTempMemory() <= 0) {
throw new CompilerException("Bug in compiler: Inconsistent description of input materialization.");
}
config.setRelativeInputMaterializationMemory(inputNum, channel.getRelativeTempMemory());
}
}
}
private void finalizeBulkIteration(IterationDescriptor descr) {
final BulkIterationPlanNode bulkNode = (BulkIterationPlanNode) descr.getIterationNode();
final JobVertex headVertex = descr.getHeadTask();
final TaskConfig headConfig = new TaskConfig(headVertex.getConfiguration());
final TaskConfig headFinalOutputConfig = descr.getHeadFinalResultConfig();
// ------------ finalize the head config with the final outputs and the sync gate ------------
final int numStepFunctionOuts = headConfig.getNumOutputs();
final int numFinalOuts = headFinalOutputConfig.getNumOutputs();
if (numStepFunctionOuts == 0) {
throw new CompilerException("The iteration has no operation inside the step function.");
}
headConfig.setIterationHeadFinalOutputConfig(headFinalOutputConfig);
headConfig.setIterationHeadIndexOfSyncOutput(numStepFunctionOuts + numFinalOuts);
final double relativeMemForBackChannel = bulkNode.getRelativeMemoryPerSubTask();
if (relativeMemForBackChannel <= 0) {
throw new CompilerException("Bug: No memory has been assigned to the iteration back channel.");
}
headConfig.setRelativeBackChannelMemory(relativeMemForBackChannel);
// --------------------------- create the sync task ---------------------------
final JobVertex sync = new JobVertex("Sync(" + bulkNode.getNodeName() + ")");
sync.setResources(bulkNode.getMinResources(), bulkNode.getPreferredResources());
sync.setInvokableClass(IterationSynchronizationSinkTask.class);
sync.setParallelism(1);
sync.setMaxParallelism(1);
this.auxVertices.add(sync);
final TaskConfig syncConfig = new TaskConfig(sync.getConfiguration());
syncConfig.setGateIterativeWithNumberOfEventsUntilInterrupt(0, headVertex.getParallelism());
// set the number of iteration / convergence criterion for the sync
final int maxNumIterations = bulkNode.getIterationNode().getIterationContract().getMaximumNumberOfIterations();
if (maxNumIterations < 1) {
throw new CompilerException("Cannot create bulk iteration with unspecified maximum number of iterations.");
}
syncConfig.setNumberOfIterations(maxNumIterations);
// connect the sync task
sync.connectNewDataSetAsInput(headVertex, DistributionPattern.POINTWISE, ResultPartitionType.PIPELINED);
// ----------------------------- create the iteration tail ------------------------------
final PlanNode rootOfTerminationCriterion = bulkNode.getRootOfTerminationCriterion();
final PlanNode rootOfStepFunction = bulkNode.getRootOfStepFunction();
final TaskConfig tailConfig;
JobVertex rootOfStepFunctionVertex = this.vertices.get(rootOfStepFunction);
if (rootOfStepFunctionVertex == null) {
// last op is chained
final TaskInChain taskInChain = this.chainedTasks.get(rootOfStepFunction);
if (taskInChain == null) {
throw new CompilerException("Bug: Tail of step function not found as vertex or chained task.");
}
rootOfStepFunctionVertex = taskInChain.getContainingVertex();
// the fake channel is statically typed to pact record. no data is sent over this channel anyways.
tailConfig = taskInChain.getTaskConfig();
} else {
tailConfig = new TaskConfig(rootOfStepFunctionVertex.getConfiguration());
}
tailConfig.setIsWorksetUpdate();
// No following termination criterion
if (rootOfStepFunction.getOutgoingChannels().isEmpty()) {
rootOfStepFunctionVertex.setInvokableClass(IterationTailTask.class);
tailConfig.setOutputSerializer(bulkNode.getSerializerForIterationChannel());
}
// create the fake output task for termination criterion, if needed
final TaskConfig tailConfigOfTerminationCriterion;
// If we have a termination criterion and it is not an intermediate node
if(rootOfTerminationCriterion != null && rootOfTerminationCriterion.getOutgoingChannels().isEmpty()) {
JobVertex rootOfTerminationCriterionVertex = this.vertices.get(rootOfTerminationCriterion);
if (rootOfTerminationCriterionVertex == null) {
// last op is chained
final TaskInChain taskInChain = this.chainedTasks.get(rootOfTerminationCriterion);
if (taskInChain == null) {
throw new CompilerException("Bug: Tail of termination criterion not found as vertex or chained task.");
}
rootOfTerminationCriterionVertex = taskInChain.getContainingVertex();
// the fake channel is statically typed to pact record. no data is sent over this channel anyways.
tailConfigOfTerminationCriterion = taskInChain.getTaskConfig();
} else {
tailConfigOfTerminationCriterion = new TaskConfig(rootOfTerminationCriterionVertex.getConfiguration());
}
rootOfTerminationCriterionVertex.setInvokableClass(IterationTailTask.class);
// Hack
tailConfigOfTerminationCriterion.setIsSolutionSetUpdate();
tailConfigOfTerminationCriterion.setOutputSerializer(bulkNode.getSerializerForIterationChannel());
// tell the head that it needs to wait for the solution set updates
headConfig.setWaitForSolutionSetUpdate();
}
// ------------------- register the aggregators -------------------
AggregatorRegistry aggs = bulkNode.getIterationNode().getIterationContract().getAggregators();
Collection<AggregatorWithName<?>> allAggregators = aggs.getAllRegisteredAggregators();
headConfig.addIterationAggregators(allAggregators);
syncConfig.addIterationAggregators(allAggregators);
String convAggName = aggs.getConvergenceCriterionAggregatorName();
ConvergenceCriterion<?> convCriterion = aggs.getConvergenceCriterion();
if (convCriterion != null || convAggName != null) {
if (convCriterion == null) {
throw new CompilerException("Error: Convergence criterion aggregator set, but criterion is null.");
}
if (convAggName == null) {
throw new CompilerException("Error: Aggregator convergence criterion set, but aggregator is null.");
}
syncConfig.setConvergenceCriterion(convAggName, convCriterion);
}
}
private void finalizeWorksetIteration(IterationDescriptor descr) {
final WorksetIterationPlanNode iterNode = (WorksetIterationPlanNode) descr.getIterationNode();
final JobVertex headVertex = descr.getHeadTask();
final TaskConfig headConfig = new TaskConfig(headVertex.getConfiguration());
final TaskConfig headFinalOutputConfig = descr.getHeadFinalResultConfig();
// ------------ finalize the head config with the final outputs and the sync gate ------------
{
final int numStepFunctionOuts = headConfig.getNumOutputs();
final int numFinalOuts = headFinalOutputConfig.getNumOutputs();
if (numStepFunctionOuts == 0) {
throw new CompilerException("The workset iteration has no operation on the workset inside the step function.");
}
headConfig.setIterationHeadFinalOutputConfig(headFinalOutputConfig);
headConfig.setIterationHeadIndexOfSyncOutput(numStepFunctionOuts + numFinalOuts);
final double relativeMemory = iterNode.getRelativeMemoryPerSubTask();
if (relativeMemory <= 0) {
throw new CompilerException("Bug: No memory has been assigned to the workset iteration.");
}
headConfig.setIsWorksetIteration();
headConfig.setRelativeBackChannelMemory(relativeMemory / 2);
headConfig.setRelativeSolutionSetMemory(relativeMemory / 2);
// set the solution set serializer and comparator
headConfig.setSolutionSetSerializer(iterNode.getSolutionSetSerializer());
headConfig.setSolutionSetComparator(iterNode.getSolutionSetComparator());
}
// --------------------------- create the sync task ---------------------------
final TaskConfig syncConfig;
{
final JobVertex sync = new JobVertex("Sync (" + iterNode.getNodeName() + ")");
sync.setResources(iterNode.getMinResources(), iterNode.getPreferredResources());
sync.setInvokableClass(IterationSynchronizationSinkTask.class);
sync.setParallelism(1);
sync.setMaxParallelism(1);
this.auxVertices.add(sync);
syncConfig = new TaskConfig(sync.getConfiguration());
syncConfig.setGateIterativeWithNumberOfEventsUntilInterrupt(0, headVertex.getParallelism());
// set the number of iteration / convergence criterion for the sync
final int maxNumIterations = iterNode.getIterationNode().getIterationContract().getMaximumNumberOfIterations();
if (maxNumIterations < 1) {
throw new CompilerException("Cannot create workset iteration with unspecified maximum number of iterations.");
}
syncConfig.setNumberOfIterations(maxNumIterations);
// connect the sync task
sync.connectNewDataSetAsInput(headVertex, DistributionPattern.POINTWISE, ResultPartitionType.PIPELINED);
}
// ----------------------------- create the iteration tails -----------------------------
// ----------------------- for next workset and solution set delta-----------------------
{
// we have three possible cases:
// 1) Two tails, one for workset update, one for solution set update
// 2) One tail for workset update, solution set update happens in an intermediate task
// 3) One tail for solution set update, workset update happens in an intermediate task
final PlanNode nextWorksetNode = iterNode.getNextWorkSetPlanNode();
final PlanNode solutionDeltaNode = iterNode.getSolutionSetDeltaPlanNode();
final boolean hasWorksetTail = nextWorksetNode.getOutgoingChannels().isEmpty();
final boolean hasSolutionSetTail = (!iterNode.isImmediateSolutionSetUpdate()) || (!hasWorksetTail);
{
// get the vertex for the workset update
final TaskConfig worksetTailConfig;
JobVertex nextWorksetVertex = this.vertices.get(nextWorksetNode);
if (nextWorksetVertex == null) {
// nextWorksetVertex is chained
TaskInChain taskInChain = this.chainedTasks.get(nextWorksetNode);
if (taskInChain == null) {
throw new CompilerException("Bug: Next workset node not found as vertex or chained task.");
}
nextWorksetVertex = taskInChain.getContainingVertex();
worksetTailConfig = taskInChain.getTaskConfig();
} else {
worksetTailConfig = new TaskConfig(nextWorksetVertex.getConfiguration());
}
// mark the node to perform workset updates
worksetTailConfig.setIsWorksetIteration();
worksetTailConfig.setIsWorksetUpdate();
if (hasWorksetTail) {
nextWorksetVertex.setInvokableClass(IterationTailTask.class);
worksetTailConfig.setOutputSerializer(iterNode.getWorksetSerializer());
}
}
{
final TaskConfig solutionDeltaConfig;
JobVertex solutionDeltaVertex = this.vertices.get(solutionDeltaNode);
if (solutionDeltaVertex == null) {
// last op is chained
TaskInChain taskInChain = this.chainedTasks.get(solutionDeltaNode);
if (taskInChain == null) {
throw new CompilerException("Bug: Solution Set Delta not found as vertex or chained task.");
}
solutionDeltaVertex = taskInChain.getContainingVertex();
solutionDeltaConfig = taskInChain.getTaskConfig();
} else {
solutionDeltaConfig = new TaskConfig(solutionDeltaVertex.getConfiguration());
}
solutionDeltaConfig.setIsWorksetIteration();
solutionDeltaConfig.setIsSolutionSetUpdate();
if (hasSolutionSetTail) {
solutionDeltaVertex.setInvokableClass(IterationTailTask.class);
solutionDeltaConfig.setOutputSerializer(iterNode.getSolutionSetSerializer());
// tell the head that it needs to wait for the solution set updates
headConfig.setWaitForSolutionSetUpdate();
}
else {
// no tail, intermediate update. must be immediate update
if (!iterNode.isImmediateSolutionSetUpdate()) {
throw new CompilerException("A solution set update without dedicated tail is not set to perform immediate updates.");
}
solutionDeltaConfig.setIsSolutionSetUpdateWithoutReprobe();
}
}
}
// ------------------- register the aggregators -------------------
AggregatorRegistry aggs = iterNode.getIterationNode().getIterationContract().getAggregators();
Collection<AggregatorWithName<?>> allAggregators = aggs.getAllRegisteredAggregators();
for (AggregatorWithName<?> agg : allAggregators) {
if (agg.getName().equals(WorksetEmptyConvergenceCriterion.AGGREGATOR_NAME)) {
throw new CompilerException("User defined aggregator used the same name as built-in workset " +
"termination check aggregator: " + WorksetEmptyConvergenceCriterion.AGGREGATOR_NAME);
}
}
headConfig.addIterationAggregators(allAggregators);
syncConfig.addIterationAggregators(allAggregators);
String convAggName = aggs.getConvergenceCriterionAggregatorName();
ConvergenceCriterion<?> convCriterion = aggs.getConvergenceCriterion();
if (convCriterion != null || convAggName != null) {
if (convCriterion == null) {
throw new CompilerException("Error: Convergence criterion aggregator set, but criterion is null.");
}
if (convAggName == null) {
throw new CompilerException("Error: Aggregator convergence criterion set, but aggregator is null.");
}
syncConfig.setConvergenceCriterion(convAggName, convCriterion);
}
headConfig.addIterationAggregator(WorksetEmptyConvergenceCriterion.AGGREGATOR_NAME, new LongSumAggregator());
syncConfig.addIterationAggregator(WorksetEmptyConvergenceCriterion.AGGREGATOR_NAME, new LongSumAggregator());
syncConfig.setImplicitConvergenceCriterion(WorksetEmptyConvergenceCriterion.AGGREGATOR_NAME, new WorksetEmptyConvergenceCriterion());
}
private String getDescriptionForUserCode(UserCodeWrapper<?> wrapper) {
try {
if (wrapper.hasObject()) {
try {
return wrapper.getUserCodeObject().toString();
}
catch (Throwable t) {
return wrapper.getUserCodeClass().getName();
}
}
else {
return wrapper.getUserCodeClass().getName();
}
}
catch (Throwable t) {
return null;
}
}
private void attachOperatorNamesAndDescriptions() {
JsonFactory jsonFactory = new JsonFactory();
// we go back to front
// start with the in chains
for (int i = chainedTasksInSequence.size() - 1; i >= 0; i--) {
TaskInChain next = chainedTasksInSequence.get(i);
PlanNode planNode = next.getPlanNode();
JobVertex vertex = next.getContainingVertex();
// operator
String opName = planNode.getOptimizerNode().getOperatorName();
if (vertex.getOperatorName() == null) {
vertex.setOperatorName(opName);
}
else {
vertex.setOperatorName(opName + " -> " + vertex.getOperatorName());
}
// operator description
String opDescription = JsonMapper.getOperatorStrategyString(planNode.getDriverStrategy());
if (vertex.getOperatorDescription() == null) {
vertex.setOperatorDescription(opDescription);
}
else {
vertex.setOperatorDescription(opDescription + "\n -> " + vertex.getOperatorDescription());
}
// pretty name
String prettyName = StringUtils.showControlCharacters(planNode.getNodeName());
if (vertex.getOperatorPrettyName() == null) {
vertex.setOperatorPrettyName(prettyName);
}
else {
vertex.setOperatorPrettyName(prettyName + "\n -> " + vertex.getOperatorPrettyName());
}
// optimizer output properties
if (vertex.getResultOptimizerProperties() == null) {
// since we go backwards, this must be the last in its chain
String outputProps =
JsonMapper.getOptimizerPropertiesJson(jsonFactory, planNode);
vertex.setResultOptimizerProperties(outputProps);
}
}
// finish back-to-front traversal by going over the head vertices
for (Map.Entry<PlanNode, JobVertex> entry : vertices.entrySet()) {
PlanNode node = entry.getKey();
JobVertex vertex = entry.getValue();
// get the predecessors
String input1name = null;
String input2name = null;
int num = 0;
for (Channel c : node.getInputs()) {
if (num == 0) {
input1name = c.getSource().getNodeName();
}
else if (num == 1) {
input2name = c.getSource().getNodeName();
}
num++;
}
// operator
String opName = node.getOptimizerNode().getOperatorName();
if (vertex.getOperatorName() == null) {
vertex.setOperatorName(opName);
}
else {
vertex.setOperatorName(opName + " -> " + vertex.getOperatorName());
}
// operator description
String opStrategy = JsonMapper.getOperatorStrategyString(
node.getDriverStrategy(),
input1name != null ? input1name : "(unnamed)",
input2name != null ? input2name : "(unnamed)");
if (vertex.getOperatorDescription() == null) {
vertex.setOperatorDescription(opStrategy);
}
else {
vertex.setOperatorDescription(opStrategy + "\n -> " + vertex.getOperatorDescription());
}
// pretty name
String prettyName = StringUtils.showControlCharacters(node.getNodeName());
if (vertex.getOperatorPrettyName() == null) {
vertex.setOperatorPrettyName(prettyName);
}
else {
vertex.setOperatorPrettyName(prettyName + "\n -> " + vertex.getOperatorPrettyName());
}
// if there is not yet an output from a chained task, we set this output
if (vertex.getResultOptimizerProperties() == null) {
vertex.setResultOptimizerProperties(
JsonMapper.getOptimizerPropertiesJson(jsonFactory, node));
}
}
}
// -------------------------------------------------------------------------------------
// Descriptors for tasks / configurations that are chained or merged with other tasks
// -------------------------------------------------------------------------------------
/**
* Utility class that describes a task in a sequence of chained tasks. Chained tasks are tasks that run
* together in one thread.
*/
private static final class TaskInChain {
private final Class<? extends ChainedDriver<?, ?>> chainedTask;
private final TaskConfig taskConfig;
private final String taskName;
private final PlanNode planNode;
private JobVertex containingVertex;
TaskInChain(PlanNode planNode, Class<? extends ChainedDriver<?, ?>> chainedTask,
TaskConfig taskConfig, String taskName) {
this.planNode = planNode;
this.chainedTask = chainedTask;
this.taskConfig = taskConfig;
this.taskName = taskName;
}
public PlanNode getPlanNode() {
return planNode;
}
public Class<? extends ChainedDriver<?, ?>> getChainedTask() {
return this.chainedTask;
}
public TaskConfig getTaskConfig() {
return this.taskConfig;
}
public String getTaskName() {
return this.taskName;
}
public JobVertex getContainingVertex() {
return this.containingVertex;
}
public void setContainingVertex(JobVertex containingVertex) {
this.containingVertex = containingVertex;
}
}
private static final class IterationDescriptor {
private final IterationPlanNode iterationNode;
private JobVertex headTask;
private TaskConfig headConfig;
private TaskConfig headFinalResultConfig;
private final int id;
public IterationDescriptor(IterationPlanNode iterationNode, int id) {
this.iterationNode = iterationNode;
this.id = id;
}
public IterationPlanNode getIterationNode() {
return iterationNode;
}
public void setHeadTask(JobVertex headTask, TaskConfig headConfig) {
this.headTask = headTask;
this.headFinalResultConfig = new TaskConfig(new Configuration());
// check if we already had a configuration, for example if the solution set was
if (this.headConfig != null) {
headConfig.getConfiguration().addAll(this.headConfig.getConfiguration());
}
this.headConfig = headConfig;
}
public JobVertex getHeadTask() {
return headTask;
}
public TaskConfig getHeadFinalResultConfig() {
return headFinalResultConfig;
}
public int getId() {
return this.id;
}
}
}