shuai-xu commented on a change in pull request #8486: [FLINK-12372] [runtime] implement ExecutionSlotAllocator URL: https://github.com/apache/flink/pull/8486#discussion_r293211952
########## File path: flink-runtime/src/main/java/org/apache/flink/runtime/scheduler/DefaultExecutionSlotAllocator.java ########## @@ -0,0 +1,248 @@ +/* + * 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.runtime.scheduler; + +import org.apache.flink.annotation.VisibleForTesting; +import org.apache.flink.api.common.time.Time; +import org.apache.flink.runtime.clusterframework.types.AllocationID; +import org.apache.flink.runtime.clusterframework.types.SlotProfile; +import org.apache.flink.runtime.concurrent.FutureUtils; +import org.apache.flink.runtime.executiongraph.ExecutionVertex; +import org.apache.flink.runtime.instance.SlotSharingGroupId; +import org.apache.flink.runtime.jobmanager.scheduler.ScheduledUnit; +import org.apache.flink.runtime.jobmaster.LogicalSlot; +import org.apache.flink.runtime.jobmaster.SlotRequestId; +import org.apache.flink.runtime.jobmaster.slotpool.SlotProvider; +import org.apache.flink.runtime.scheduler.strategy.ExecutionVertexID; +import org.apache.flink.runtime.taskmanager.TaskManagerLocation; + +import org.slf4j.Logger; +import org.slf4j.LoggerFactory; + +import java.util.ArrayList; +import java.util.Arrays; +import java.util.Collection; +import java.util.Collections; +import java.util.HashMap; +import java.util.HashSet; +import java.util.List; +import java.util.Map; +import java.util.Optional; +import java.util.Set; +import java.util.concurrent.CompletableFuture; + +import static org.apache.flink.runtime.executiongraph.ExecutionVertex.MAX_DISTINCT_LOCATIONS_TO_CONSIDER; +import static org.apache.flink.util.Preconditions.checkNotNull; + +/** + * Default {@link ExecutionSlotAllocator} which will use {@link SlotProvider} to allocate slots and + * keep the unfulfilled requests for further cancellation. + */ +public class DefaultExecutionSlotAllocator implements ExecutionSlotAllocator { + + private static final Logger LOG = LoggerFactory.getLogger(DefaultExecutionSlotAllocator.class); + + /** + * Store the uncompleted slot assignments. + */ + private final Map<ExecutionVertexID, SlotExecutionVertexAssignment> pendingSlotAssignments; + + private final SlotProvider slotProvider; + + private final InputsLocationsRetriever inputsLocationsRetriever; + + private final Time allocationTimeout; + + public DefaultExecutionSlotAllocator( + SlotProvider slotProvider, + InputsLocationsRetriever inputsLocationsRetriever, + Time allocationTimeout) { + this.slotProvider = checkNotNull(slotProvider); + this.inputsLocationsRetriever = checkNotNull(inputsLocationsRetriever); + this.allocationTimeout = checkNotNull(allocationTimeout); + + pendingSlotAssignments = new HashMap<>(); + } + + @Override + public Collection<SlotExecutionVertexAssignment> allocateSlotsFor( + Collection<ExecutionVertexSchedulingRequirements> executionVertexSchedulingRequirements) { + + List<SlotExecutionVertexAssignment> slotExecutionVertexAssignments = + new ArrayList<>(executionVertexSchedulingRequirements.size()); + + Set<AllocationID> allPreviousAllocationIds = computeAllPriorAllocationIds(executionVertexSchedulingRequirements); + + for (ExecutionVertexSchedulingRequirements schedulingRequirements : executionVertexSchedulingRequirements) { + final ExecutionVertexID executionVertexId = schedulingRequirements.getExecutionVertexId(); + final SlotRequestId slotRequestId = new SlotRequestId(); + final SlotSharingGroupId slotSharingGroupId = schedulingRequirements.getSlotSharingGroupId(); + + LOG.debug("Allocate slot with id {} for execution {}", slotRequestId, executionVertexId); + + CompletableFuture<LogicalSlot> slotFuture = calculatePreferredLocations( + executionVertexId, + schedulingRequirements.getPreferredLocations(), + inputsLocationsRetriever).thenCompose( + (Collection<TaskManagerLocation> preferredLocations) -> + slotProvider.allocateSlot( + slotRequestId, + new ScheduledUnit( + executionVertexId.getJobVertexId(), + slotSharingGroupId, + schedulingRequirements.getCoLocationConstraint()), + new SlotProfile( + schedulingRequirements.getResourceProfile(), + preferredLocations, + Arrays.asList(schedulingRequirements.getPreviousAllocationId()), + allPreviousAllocationIds), + true, + allocationTimeout)); + + SlotExecutionVertexAssignment slotExecutionVertexAssignment = + new SlotExecutionVertexAssignment(executionVertexId, slotFuture); + // add to map first to avoid the future completed before added. + pendingSlotAssignments.put(executionVertexId, slotExecutionVertexAssignment); + + slotFuture.whenComplete( + (ignored, throwable) -> { + pendingSlotAssignments.remove(executionVertexId); + if (throwable != null) { + slotProvider.cancelSlotRequest(slotRequestId, slotSharingGroupId, throwable); + } + }); + + slotExecutionVertexAssignments.add(slotExecutionVertexAssignment); + } + + return slotExecutionVertexAssignments; + } + + @Override + public void cancel(ExecutionVertexID executionVertexId) { + SlotExecutionVertexAssignment slotExecutionVertexAssignment = pendingSlotAssignments.get(executionVertexId); + if (slotExecutionVertexAssignment != null) { + slotExecutionVertexAssignment.getLogicalSlotFuture().cancel(false); + } + } + + @Override + public CompletableFuture<Void> stop() { + List<ExecutionVertexID> executionVertexIds = new ArrayList<>(pendingSlotAssignments.keySet()); + executionVertexIds.stream().forEach(executionVertexID -> cancel(executionVertexID)); + + return CompletableFuture.completedFuture(null); + } + + /** + * Calculates the preferred locations for an execution. + * It will first try to use preferred locations based on state, + * if null, will use the preferred locations based on inputs. + */ + private static CompletableFuture<Collection<TaskManagerLocation>> calculatePreferredLocations( + ExecutionVertexID executionVertexId, + Collection<TaskManagerLocation> preferredLocationsBasedOnState, + InputsLocationsRetriever inputsLocationsRetriever) { + + if (!preferredLocationsBasedOnState.isEmpty()) { + return CompletableFuture.completedFuture(preferredLocationsBasedOnState); + } + + return getPreferredLocationsBasedOnInputs(executionVertexId, inputsLocationsRetriever); + } + + /** + * Gets the location preferences of the execution, as determined by the locations + * of the predecessors from which it receives input data. + * If there are more than {@link ExecutionVertex#MAX_DISTINCT_LOCATIONS_TO_CONSIDER} different locations of source data, + * or neither the sources have not been started nor will be started with the execution together, + * this method returns an empty collection to indicate no location preference. + * + * @return The preferred locations based in input streams, or an empty iterable, + * if there is no input-based preference. + */ + @VisibleForTesting + static CompletableFuture<Collection<TaskManagerLocation>> getPreferredLocationsBasedOnInputs( + ExecutionVertexID executionVertexId, + InputsLocationsRetriever inputsLocationsRetriever) { + CompletableFuture<Collection<TaskManagerLocation>> preferredLocations = + CompletableFuture.completedFuture(Collections.emptyList()); + + Collection<CompletableFuture<TaskManagerLocation>> locationsFutures = new ArrayList<>(); + + Collection<Collection<ExecutionVertexID>> allProducers = + inputsLocationsRetriever.getConsumedResultPartitionsProducers(executionVertexId); + for (Collection<ExecutionVertexID> producers : allProducers) { + + for (ExecutionVertexID producer : producers) { + Optional<CompletableFuture<TaskManagerLocation>> optionalLocationFuture = + inputsLocationsRetriever.getTaskManagerLocation(producer); + optionalLocationFuture.ifPresent( + (locationFuture) -> locationsFutures.add(locationFuture)); + // If the parallelism is large, wait for all futures coming back may cost a long time. Review comment: I think this is the main reason. As user will change the parallelism according to the total size of data to keep the size of each partition stable. We didn't see the case waiting for a long time since the logic is same in legacy code and for almost all jobs in alibaba, the parallelism is much larger than 8. ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
