1996fanrui commented on code in PR #23635: URL: https://github.com/apache/flink/pull/23635#discussion_r1400131195
########## flink-runtime/src/main/java/org/apache/flink/runtime/scheduler/TaskBalancedPreferredSlotSharingStrategy.java: ########## @@ -0,0 +1,361 @@ +/* + * 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.api.java.tuple.Tuple2; +import org.apache.flink.runtime.jobgraph.JobVertexID; +import org.apache.flink.runtime.jobmanager.scheduler.CoLocationConstraint; +import org.apache.flink.runtime.jobmanager.scheduler.CoLocationGroup; +import org.apache.flink.runtime.jobmanager.scheduler.SlotSharingGroup; +import org.apache.flink.runtime.scheduler.strategy.ExecutionVertexID; +import org.apache.flink.runtime.scheduler.strategy.SchedulingExecutionVertex; +import org.apache.flink.runtime.scheduler.strategy.SchedulingTopology; + +import org.slf4j.Logger; +import org.slf4j.LoggerFactory; + +import javax.annotation.Nullable; + +import java.util.ArrayList; +import java.util.HashMap; +import java.util.LinkedHashMap; +import java.util.List; +import java.util.Map; +import java.util.Objects; +import java.util.Set; +import java.util.stream.Collectors; + +import static org.apache.flink.util.Preconditions.checkNotNull; + +/** + * This strategy tries to get a balanced tasks scheduling. Execution vertices, which are belong to + * the same SlotSharingGroup, tend to be put evenly in each ExecutionSlotSharingGroup. Co-location + * constraints will be respected. + */ +class TaskBalancedPreferredSlotSharingStrategy extends AbstractSlotSharingStrategy { + + public static final Logger LOG = + LoggerFactory.getLogger(TaskBalancedPreferredSlotSharingStrategy.class); + + TaskBalancedPreferredSlotSharingStrategy( + final SchedulingTopology topology, + final Set<SlotSharingGroup> slotSharingGroups, + final Set<CoLocationGroup> coLocationGroups) { + super(topology, slotSharingGroups, coLocationGroups); + } + + @Override + protected Map<ExecutionVertexID, ExecutionSlotSharingGroup> computeExecutionSlotSharingGroups( + SchedulingTopology schedulingTopology) { + return new TaskBalancedExecutionSlotSharingGroupBuilder( + schedulingTopology, this.logicalSlotSharingGroups, this.coLocationGroups) + .build(); + } + + static class Factory implements SlotSharingStrategy.Factory { + + public TaskBalancedPreferredSlotSharingStrategy create( + final SchedulingTopology topology, + final Set<SlotSharingGroup> slotSharingGroups, + final Set<CoLocationGroup> coLocationGroups) { + + return new TaskBalancedPreferredSlotSharingStrategy( + topology, slotSharingGroups, coLocationGroups); + } + } + + /** SlotSharingGroupBuilder class for balanced scheduling strategy. */ + private static class TaskBalancedExecutionSlotSharingGroupBuilder { + + private final SchedulingTopology topology; + + private final Map<JobVertexID, SlotSharingGroup> slotSharingGroupMap; + + /** Record the {@link ExecutionSlotSharingGroup}s for {@link SlotSharingGroup}s. */ + private final Map<SlotSharingGroup, List<ExecutionSlotSharingGroup>> + paralleledExecutionSlotSharingGroupsMap; + + /** + * Record the next round-robin {@link ExecutionSlotSharingGroup} index for {@link + * SlotSharingGroup}s. + */ + private final Map<SlotSharingGroup, Integer> slotSharingGroupIndexMap; + + private final Map<ExecutionVertexID, ExecutionSlotSharingGroup> + executionSlotSharingGroupMap; + + private final Map<JobVertexID, CoLocationGroup> coLocationGroupMap; + + private final Map<CoLocationConstraint, ExecutionSlotSharingGroup> + constraintToExecutionSlotSharingGroupMap; + + private TaskBalancedExecutionSlotSharingGroupBuilder( + final SchedulingTopology topology, + final Set<SlotSharingGroup> slotSharingGroups, + final Set<CoLocationGroup> coLocationGroups) { + this.topology = checkNotNull(topology); + + this.coLocationGroupMap = new HashMap<>(); + for (CoLocationGroup coLocationGroup : coLocationGroups) { + for (JobVertexID jobVertexId : coLocationGroup.getVertexIds()) { + coLocationGroupMap.put(jobVertexId, coLocationGroup); + } + } + + this.constraintToExecutionSlotSharingGroupMap = new HashMap<>(); + this.paralleledExecutionSlotSharingGroupsMap = new HashMap<>(slotSharingGroups.size()); + this.slotSharingGroupIndexMap = new HashMap<>(slotSharingGroups.size()); + this.slotSharingGroupMap = new HashMap<>(); + this.executionSlotSharingGroupMap = new HashMap<>(); + + for (SlotSharingGroup slotSharingGroup : slotSharingGroups) { + for (JobVertexID jobVertexId : slotSharingGroup.getJobVertexIds()) { + slotSharingGroupMap.put(jobVertexId, slotSharingGroup); + } + } + } + + private Map<ExecutionVertexID, ExecutionSlotSharingGroup> build() { + + final LinkedHashMap<JobVertexID, List<SchedulingExecutionVertex>> allVertices = + getExecutionVertices(topology); + + initParalleledExecutionSlotSharingGroupsMap(allVertices); + + // Loop on job vertices + for (Map.Entry<JobVertexID, List<SchedulingExecutionVertex>> executionVertexInfos : + allVertices.entrySet()) { + + JobVertexID jobVertexID = executionVertexInfos.getKey(); + List<SchedulingExecutionVertex> executionVertices = executionVertexInfos.getValue(); + final int jobVertexParallel = executionVertices.size(); + final SlotSharingGroup slotSharingGroup = slotSharingGroupMap.get(jobVertexID); + final int index = getSlotRoundRobinIndex(jobVertexParallel, slotSharingGroup); + + if (!coLocationGroupMap.containsKey(jobVertexID)) { + // For vertices without CoLocationConstraint. + allocateNonCoLocatedVertices(slotSharingGroup, executionVertices, index); + } else { + // For vertices with CoLocationConstraint. + allocateCoLocatedVertices(slotSharingGroup, executionVertices, index); + } + } + return executionSlotSharingGroupMap; + } + + private void initParalleledExecutionSlotSharingGroupsMap( + final LinkedHashMap<JobVertexID, List<SchedulingExecutionVertex>> allVertices) { + + allVertices.entrySet().stream() + .map( + jobVertexExecutionVertices -> + Tuple2.of( + slotSharingGroupMap.get( + jobVertexExecutionVertices.getKey()), + jobVertexExecutionVertices.getValue().size())) + .collect( + Collectors.groupingBy( + tuple -> tuple.f0, + Collectors.summarizingInt(tuple -> tuple.f1))) + .forEach( + (slotSharingGroup, statistics) -> { + int slotNum = statistics.getMax(); + paralleledExecutionSlotSharingGroupsMap.put( + slotSharingGroup, + createExecutionSlotSharingGroups( + slotSharingGroup, slotNum)); + }); + } + + private List<ExecutionSlotSharingGroup> createExecutionSlotSharingGroups( + SlotSharingGroup slotSharingGroup, int slotNum) { + final List<ExecutionSlotSharingGroup> executionSlotSharingGroups = + new ArrayList<>(slotNum); + for (int i = 0; i < slotNum; i++) { + final ExecutionSlotSharingGroup executionSlotSharingGroup = + new ExecutionSlotSharingGroup(slotSharingGroup); + executionSlotSharingGroups.add(i, executionSlotSharingGroup); + LOG.debug( + "Create {}th executionSlotSharingGroup {}.", i, executionSlotSharingGroup); + } + return executionSlotSharingGroups; + } + + private void allocateCoLocatedVertices( + SlotSharingGroup slotSharingGroup, + List<SchedulingExecutionVertex> executionVertices, + int index) { + final int slotNum = + paralleledExecutionSlotSharingGroupsMap.get(slotSharingGroup).size(); + final int jobVertexParallel = executionVertices.size(); + if (slotNum == jobVertexParallel) { + // For CoLocations with max parallelism of the slot sharing group. + allocateCoLocatedMaxParallelVertices(slotSharingGroup, executionVertices, index); + } else { + // For CoLocations with non-max parallelism of the slot sharing group. + allocateCoLocatedNonMaxParallelVertices(slotSharingGroup, executionVertices); + } + } + + private void allocateCoLocatedMaxParallelVertices( + SlotSharingGroup slotSharingGroup, + List<SchedulingExecutionVertex> executionVertices, + int index) { + final List<ExecutionSlotSharingGroup> executionSlotSharingGroups = + paralleledExecutionSlotSharingGroupsMap.get(slotSharingGroup); + for (SchedulingExecutionVertex executionVertex : executionVertices) { + final CoLocationConstraint coLocationConstraint = + getCoLocationConstraint(executionVertex); + ExecutionSlotSharingGroup executionSlotSharingGroup = + constraintToExecutionSlotSharingGroupMap.get(coLocationConstraint); + if (Objects.isNull(executionSlotSharingGroup)) { + executionSlotSharingGroup = executionSlotSharingGroups.get(index); + } + addVertexToExecutionSlotSharingGroup(executionSlotSharingGroup, executionVertex); + index = ++index % executionSlotSharingGroups.size(); + } + updateConstraintToExecutionSlotSharingGroupMap(executionVertices); + } + + private void allocateCoLocatedNonMaxParallelVertices( + SlotSharingGroup slotSharingGroup, + List<SchedulingExecutionVertex> executionVertices) { + List<ExecutionSlotSharingGroup> executionSlotSharingGroups = + paralleledExecutionSlotSharingGroupsMap.get(slotSharingGroup); + int index; + // For CoLocations with non-max parallelism of the slot sharing group. + for (SchedulingExecutionVertex executionVertex : executionVertices) { + final CoLocationConstraint coLocationConstraint = + getCoLocationConstraint(executionVertex); + ExecutionSlotSharingGroup executionSlotSharingGroup = + constraintToExecutionSlotSharingGroupMap.get(coLocationConstraint); + if (Objects.isNull(executionSlotSharingGroup)) { + index = getFittestSlotIndex(executionSlotSharingGroups, executionVertex); Review Comment: The time complexity here is O(n*n). I'm not sure whether it will affect the Flink job start duration. Anyway, it's better to do a benchmark for this FLIP in the end. ########## flink-runtime/src/test/java/org/apache/flink/runtime/scheduler/TaskBalancedPreferredSlotSharingStrategyTest.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.api.java.tuple.Tuple2; +import org.apache.flink.runtime.clusterframework.types.ResourceProfile; +import org.apache.flink.runtime.jobgraph.JobVertex; +import org.apache.flink.runtime.jobgraph.JobVertexID; +import org.apache.flink.runtime.jobmanager.scheduler.CoLocationGroup; +import org.apache.flink.runtime.jobmanager.scheduler.CoLocationGroupImpl; +import org.apache.flink.runtime.jobmanager.scheduler.SlotSharingGroup; +import org.apache.flink.runtime.scheduler.strategy.TestingSchedulingExecutionVertex; +import org.apache.flink.runtime.scheduler.strategy.TestingSchedulingTopology; +import org.apache.flink.util.Preconditions; + +import org.apache.flink.shaded.guava31.com.google.common.collect.Lists; +import org.apache.flink.shaded.guava31.com.google.common.collect.Sets; + +import org.assertj.core.data.Offset; +import org.junit.jupiter.api.Test; + +import javax.annotation.Nonnull; +import javax.annotation.Nullable; + +import java.util.ArrayList; +import java.util.Collections; +import java.util.Comparator; +import java.util.List; +import java.util.Optional; +import java.util.Set; +import java.util.stream.Collectors; + +import static org.assertj.core.api.Assertions.assertThat; + +/** Tests for {@link TaskBalancedPreferredSlotSharingStrategy}. */ +class TaskBalancedPreferredSlotSharingStrategyTest { Review Comment: One question for testing: is there any common testing for `TaskBalancedPreferredSlotSharingStrategyTest` and `LocalInputPreferredSlotSharingStrategyTest`? I see some test method names are same. Is it possible to extract an `AbstractSlotSharingStrategyTest`, moving some common tests to there and these 2 classes extend `AbstractSlotSharingStrategyTest`? ########## flink-runtime/src/main/java/org/apache/flink/runtime/scheduler/TaskBalancedPreferredSlotSharingStrategy.java: ########## @@ -0,0 +1,361 @@ +/* + * 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.api.java.tuple.Tuple2; +import org.apache.flink.runtime.jobgraph.JobVertexID; +import org.apache.flink.runtime.jobmanager.scheduler.CoLocationConstraint; +import org.apache.flink.runtime.jobmanager.scheduler.CoLocationGroup; +import org.apache.flink.runtime.jobmanager.scheduler.SlotSharingGroup; +import org.apache.flink.runtime.scheduler.strategy.ExecutionVertexID; +import org.apache.flink.runtime.scheduler.strategy.SchedulingExecutionVertex; +import org.apache.flink.runtime.scheduler.strategy.SchedulingTopology; + +import org.slf4j.Logger; +import org.slf4j.LoggerFactory; + +import javax.annotation.Nullable; + +import java.util.ArrayList; +import java.util.HashMap; +import java.util.LinkedHashMap; +import java.util.List; +import java.util.Map; +import java.util.Objects; +import java.util.Set; +import java.util.stream.Collectors; + +import static org.apache.flink.util.Preconditions.checkNotNull; + +/** + * This strategy tries to get a balanced tasks scheduling. Execution vertices, which are belong to + * the same SlotSharingGroup, tend to be put evenly in each ExecutionSlotSharingGroup. Co-location + * constraints will be respected. + */ +class TaskBalancedPreferredSlotSharingStrategy extends AbstractSlotSharingStrategy { + + public static final Logger LOG = + LoggerFactory.getLogger(TaskBalancedPreferredSlotSharingStrategy.class); + + TaskBalancedPreferredSlotSharingStrategy( + final SchedulingTopology topology, + final Set<SlotSharingGroup> slotSharingGroups, + final Set<CoLocationGroup> coLocationGroups) { + super(topology, slotSharingGroups, coLocationGroups); + } + + @Override + protected Map<ExecutionVertexID, ExecutionSlotSharingGroup> computeExecutionSlotSharingGroups( + SchedulingTopology schedulingTopology) { + return new TaskBalancedExecutionSlotSharingGroupBuilder( + schedulingTopology, this.logicalSlotSharingGroups, this.coLocationGroups) + .build(); + } + + static class Factory implements SlotSharingStrategy.Factory { + + public TaskBalancedPreferredSlotSharingStrategy create( + final SchedulingTopology topology, + final Set<SlotSharingGroup> slotSharingGroups, + final Set<CoLocationGroup> coLocationGroups) { + + return new TaskBalancedPreferredSlotSharingStrategy( + topology, slotSharingGroups, coLocationGroups); + } + } + + /** SlotSharingGroupBuilder class for balanced scheduling strategy. */ + private static class TaskBalancedExecutionSlotSharingGroupBuilder { + + private final SchedulingTopology topology; + + private final Map<JobVertexID, SlotSharingGroup> slotSharingGroupMap; + + /** Record the {@link ExecutionSlotSharingGroup}s for {@link SlotSharingGroup}s. */ + private final Map<SlotSharingGroup, List<ExecutionSlotSharingGroup>> + paralleledExecutionSlotSharingGroupsMap; + + /** + * Record the next round-robin {@link ExecutionSlotSharingGroup} index for {@link + * SlotSharingGroup}s. + */ + private final Map<SlotSharingGroup, Integer> slotSharingGroupIndexMap; + + private final Map<ExecutionVertexID, ExecutionSlotSharingGroup> + executionSlotSharingGroupMap; + + private final Map<JobVertexID, CoLocationGroup> coLocationGroupMap; + + private final Map<CoLocationConstraint, ExecutionSlotSharingGroup> + constraintToExecutionSlotSharingGroupMap; + + private TaskBalancedExecutionSlotSharingGroupBuilder( + final SchedulingTopology topology, + final Set<SlotSharingGroup> slotSharingGroups, + final Set<CoLocationGroup> coLocationGroups) { + this.topology = checkNotNull(topology); + + this.coLocationGroupMap = new HashMap<>(); + for (CoLocationGroup coLocationGroup : coLocationGroups) { + for (JobVertexID jobVertexId : coLocationGroup.getVertexIds()) { + coLocationGroupMap.put(jobVertexId, coLocationGroup); + } + } + + this.constraintToExecutionSlotSharingGroupMap = new HashMap<>(); + this.paralleledExecutionSlotSharingGroupsMap = new HashMap<>(slotSharingGroups.size()); + this.slotSharingGroupIndexMap = new HashMap<>(slotSharingGroups.size()); + this.slotSharingGroupMap = new HashMap<>(); + this.executionSlotSharingGroupMap = new HashMap<>(); + + for (SlotSharingGroup slotSharingGroup : slotSharingGroups) { + for (JobVertexID jobVertexId : slotSharingGroup.getJobVertexIds()) { + slotSharingGroupMap.put(jobVertexId, slotSharingGroup); + } + } + } + + private Map<ExecutionVertexID, ExecutionSlotSharingGroup> build() { + + final LinkedHashMap<JobVertexID, List<SchedulingExecutionVertex>> allVertices = + getExecutionVertices(topology); + + initParalleledExecutionSlotSharingGroupsMap(allVertices); + + // Loop on job vertices + for (Map.Entry<JobVertexID, List<SchedulingExecutionVertex>> executionVertexInfos : + allVertices.entrySet()) { + + JobVertexID jobVertexID = executionVertexInfos.getKey(); + List<SchedulingExecutionVertex> executionVertices = executionVertexInfos.getValue(); + final int jobVertexParallel = executionVertices.size(); + final SlotSharingGroup slotSharingGroup = slotSharingGroupMap.get(jobVertexID); + final int index = getSlotRoundRobinIndex(jobVertexParallel, slotSharingGroup); + + if (!coLocationGroupMap.containsKey(jobVertexID)) { + // For vertices without CoLocationConstraint. + allocateNonCoLocatedVertices(slotSharingGroup, executionVertices, index); + } else { + // For vertices with CoLocationConstraint. + allocateCoLocatedVertices(slotSharingGroup, executionVertices, index); + } + } + return executionSlotSharingGroupMap; + } + + private void initParalleledExecutionSlotSharingGroupsMap( + final LinkedHashMap<JobVertexID, List<SchedulingExecutionVertex>> allVertices) { + + allVertices.entrySet().stream() + .map( + jobVertexExecutionVertices -> + Tuple2.of( + slotSharingGroupMap.get( + jobVertexExecutionVertices.getKey()), + jobVertexExecutionVertices.getValue().size())) + .collect( + Collectors.groupingBy( + tuple -> tuple.f0, + Collectors.summarizingInt(tuple -> tuple.f1))) + .forEach( + (slotSharingGroup, statistics) -> { + int slotNum = statistics.getMax(); + paralleledExecutionSlotSharingGroupsMap.put( + slotSharingGroup, + createExecutionSlotSharingGroups( + slotSharingGroup, slotNum)); + }); + } + + private List<ExecutionSlotSharingGroup> createExecutionSlotSharingGroups( + SlotSharingGroup slotSharingGroup, int slotNum) { + final List<ExecutionSlotSharingGroup> executionSlotSharingGroups = + new ArrayList<>(slotNum); + for (int i = 0; i < slotNum; i++) { + final ExecutionSlotSharingGroup executionSlotSharingGroup = + new ExecutionSlotSharingGroup(slotSharingGroup); + executionSlotSharingGroups.add(i, executionSlotSharingGroup); + LOG.debug( + "Create {}th executionSlotSharingGroup {}.", i, executionSlotSharingGroup); + } + return executionSlotSharingGroups; + } + + private void allocateCoLocatedVertices( + SlotSharingGroup slotSharingGroup, + List<SchedulingExecutionVertex> executionVertices, + int index) { + final int slotNum = + paralleledExecutionSlotSharingGroupsMap.get(slotSharingGroup).size(); + final int jobVertexParallel = executionVertices.size(); + if (slotNum == jobVertexParallel) { + // For CoLocations with max parallelism of the slot sharing group. + allocateCoLocatedMaxParallelVertices(slotSharingGroup, executionVertices, index); + } else { + // For CoLocations with non-max parallelism of the slot sharing group. + allocateCoLocatedNonMaxParallelVertices(slotSharingGroup, executionVertices); + } + } + + private void allocateCoLocatedMaxParallelVertices( + SlotSharingGroup slotSharingGroup, + List<SchedulingExecutionVertex> executionVertices, + int index) { + final List<ExecutionSlotSharingGroup> executionSlotSharingGroups = + paralleledExecutionSlotSharingGroupsMap.get(slotSharingGroup); + for (SchedulingExecutionVertex executionVertex : executionVertices) { + final CoLocationConstraint coLocationConstraint = + getCoLocationConstraint(executionVertex); + ExecutionSlotSharingGroup executionSlotSharingGroup = + constraintToExecutionSlotSharingGroupMap.get(coLocationConstraint); + if (Objects.isNull(executionSlotSharingGroup)) { + executionSlotSharingGroup = executionSlotSharingGroups.get(index); Review Comment: Could we update `constraintToExecutionSlotSharingGroupMap` here directly? It's null, we can update directly. Currently, `updateConstraintToExecutionSlotSharingGroupMap` call a series of same logic with here. `allocateCoLocatedNonMaxParallelVertices` is same with here. ########## flink-runtime/src/main/java/org/apache/flink/runtime/scheduler/TaskBalancedPreferredSlotSharingStrategy.java: ########## @@ -0,0 +1,361 @@ +/* + * 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.api.java.tuple.Tuple2; +import org.apache.flink.runtime.jobgraph.JobVertexID; +import org.apache.flink.runtime.jobmanager.scheduler.CoLocationConstraint; +import org.apache.flink.runtime.jobmanager.scheduler.CoLocationGroup; +import org.apache.flink.runtime.jobmanager.scheduler.SlotSharingGroup; +import org.apache.flink.runtime.scheduler.strategy.ExecutionVertexID; +import org.apache.flink.runtime.scheduler.strategy.SchedulingExecutionVertex; +import org.apache.flink.runtime.scheduler.strategy.SchedulingTopology; + +import org.slf4j.Logger; +import org.slf4j.LoggerFactory; + +import javax.annotation.Nullable; + +import java.util.ArrayList; +import java.util.HashMap; +import java.util.LinkedHashMap; +import java.util.List; +import java.util.Map; +import java.util.Objects; +import java.util.Set; +import java.util.stream.Collectors; + +import static org.apache.flink.util.Preconditions.checkNotNull; + +/** + * This strategy tries to get a balanced tasks scheduling. Execution vertices, which are belong to + * the same SlotSharingGroup, tend to be put evenly in each ExecutionSlotSharingGroup. Co-location + * constraints will be respected. + */ +class TaskBalancedPreferredSlotSharingStrategy extends AbstractSlotSharingStrategy { + + public static final Logger LOG = + LoggerFactory.getLogger(TaskBalancedPreferredSlotSharingStrategy.class); + + TaskBalancedPreferredSlotSharingStrategy( + final SchedulingTopology topology, + final Set<SlotSharingGroup> slotSharingGroups, + final Set<CoLocationGroup> coLocationGroups) { + super(topology, slotSharingGroups, coLocationGroups); + } + + @Override + protected Map<ExecutionVertexID, ExecutionSlotSharingGroup> computeExecutionSlotSharingGroups( + SchedulingTopology schedulingTopology) { + return new TaskBalancedExecutionSlotSharingGroupBuilder( + schedulingTopology, this.logicalSlotSharingGroups, this.coLocationGroups) + .build(); + } + + static class Factory implements SlotSharingStrategy.Factory { + + public TaskBalancedPreferredSlotSharingStrategy create( + final SchedulingTopology topology, + final Set<SlotSharingGroup> slotSharingGroups, + final Set<CoLocationGroup> coLocationGroups) { + + return new TaskBalancedPreferredSlotSharingStrategy( + topology, slotSharingGroups, coLocationGroups); + } + } + + /** SlotSharingGroupBuilder class for balanced scheduling strategy. */ + private static class TaskBalancedExecutionSlotSharingGroupBuilder { + + private final SchedulingTopology topology; + + private final Map<JobVertexID, SlotSharingGroup> slotSharingGroupMap; + + /** Record the {@link ExecutionSlotSharingGroup}s for {@link SlotSharingGroup}s. */ + private final Map<SlotSharingGroup, List<ExecutionSlotSharingGroup>> + paralleledExecutionSlotSharingGroupsMap; + + /** + * Record the next round-robin {@link ExecutionSlotSharingGroup} index for {@link + * SlotSharingGroup}s. + */ + private final Map<SlotSharingGroup, Integer> slotSharingGroupIndexMap; + + private final Map<ExecutionVertexID, ExecutionSlotSharingGroup> + executionSlotSharingGroupMap; + + private final Map<JobVertexID, CoLocationGroup> coLocationGroupMap; + + private final Map<CoLocationConstraint, ExecutionSlotSharingGroup> + constraintToExecutionSlotSharingGroupMap; + + private TaskBalancedExecutionSlotSharingGroupBuilder( + final SchedulingTopology topology, + final Set<SlotSharingGroup> slotSharingGroups, + final Set<CoLocationGroup> coLocationGroups) { + this.topology = checkNotNull(topology); + + this.coLocationGroupMap = new HashMap<>(); + for (CoLocationGroup coLocationGroup : coLocationGroups) { + for (JobVertexID jobVertexId : coLocationGroup.getVertexIds()) { + coLocationGroupMap.put(jobVertexId, coLocationGroup); + } + } + + this.constraintToExecutionSlotSharingGroupMap = new HashMap<>(); + this.paralleledExecutionSlotSharingGroupsMap = new HashMap<>(slotSharingGroups.size()); + this.slotSharingGroupIndexMap = new HashMap<>(slotSharingGroups.size()); + this.slotSharingGroupMap = new HashMap<>(); + this.executionSlotSharingGroupMap = new HashMap<>(); + + for (SlotSharingGroup slotSharingGroup : slotSharingGroups) { + for (JobVertexID jobVertexId : slotSharingGroup.getJobVertexIds()) { + slotSharingGroupMap.put(jobVertexId, slotSharingGroup); + } + } + } + + private Map<ExecutionVertexID, ExecutionSlotSharingGroup> build() { + + final LinkedHashMap<JobVertexID, List<SchedulingExecutionVertex>> allVertices = + getExecutionVertices(topology); + + initParalleledExecutionSlotSharingGroupsMap(allVertices); + + // Loop on job vertices + for (Map.Entry<JobVertexID, List<SchedulingExecutionVertex>> executionVertexInfos : + allVertices.entrySet()) { + + JobVertexID jobVertexID = executionVertexInfos.getKey(); + List<SchedulingExecutionVertex> executionVertices = executionVertexInfos.getValue(); + final int jobVertexParallel = executionVertices.size(); + final SlotSharingGroup slotSharingGroup = slotSharingGroupMap.get(jobVertexID); + final int index = getSlotRoundRobinIndex(jobVertexParallel, slotSharingGroup); + + if (!coLocationGroupMap.containsKey(jobVertexID)) { + // For vertices without CoLocationConstraint. + allocateNonCoLocatedVertices(slotSharingGroup, executionVertices, index); + } else { + // For vertices with CoLocationConstraint. + allocateCoLocatedVertices(slotSharingGroup, executionVertices, index); + } + } + return executionSlotSharingGroupMap; + } + + private void initParalleledExecutionSlotSharingGroupsMap( + final LinkedHashMap<JobVertexID, List<SchedulingExecutionVertex>> allVertices) { + + allVertices.entrySet().stream() + .map( + jobVertexExecutionVertices -> + Tuple2.of( + slotSharingGroupMap.get( + jobVertexExecutionVertices.getKey()), + jobVertexExecutionVertices.getValue().size())) + .collect( + Collectors.groupingBy( + tuple -> tuple.f0, + Collectors.summarizingInt(tuple -> tuple.f1))) + .forEach( + (slotSharingGroup, statistics) -> { + int slotNum = statistics.getMax(); + paralleledExecutionSlotSharingGroupsMap.put( + slotSharingGroup, + createExecutionSlotSharingGroups( + slotSharingGroup, slotNum)); + }); + } + + private List<ExecutionSlotSharingGroup> createExecutionSlotSharingGroups( + SlotSharingGroup slotSharingGroup, int slotNum) { + final List<ExecutionSlotSharingGroup> executionSlotSharingGroups = + new ArrayList<>(slotNum); + for (int i = 0; i < slotNum; i++) { + final ExecutionSlotSharingGroup executionSlotSharingGroup = + new ExecutionSlotSharingGroup(slotSharingGroup); + executionSlotSharingGroups.add(i, executionSlotSharingGroup); + LOG.debug( + "Create {}th executionSlotSharingGroup {}.", i, executionSlotSharingGroup); + } + return executionSlotSharingGroups; + } + + private void allocateCoLocatedVertices( + SlotSharingGroup slotSharingGroup, + List<SchedulingExecutionVertex> executionVertices, + int index) { + final int slotNum = + paralleledExecutionSlotSharingGroupsMap.get(slotSharingGroup).size(); + final int jobVertexParallel = executionVertices.size(); + if (slotNum == jobVertexParallel) { + // For CoLocations with max parallelism of the slot sharing group. + allocateCoLocatedMaxParallelVertices(slotSharingGroup, executionVertices, index); + } else { + // For CoLocations with non-max parallelism of the slot sharing group. + allocateCoLocatedNonMaxParallelVertices(slotSharingGroup, executionVertices); + } + } + + private void allocateCoLocatedMaxParallelVertices( + SlotSharingGroup slotSharingGroup, + List<SchedulingExecutionVertex> executionVertices, + int index) { + final List<ExecutionSlotSharingGroup> executionSlotSharingGroups = + paralleledExecutionSlotSharingGroupsMap.get(slotSharingGroup); + for (SchedulingExecutionVertex executionVertex : executionVertices) { + final CoLocationConstraint coLocationConstraint = + getCoLocationConstraint(executionVertex); + ExecutionSlotSharingGroup executionSlotSharingGroup = + constraintToExecutionSlotSharingGroupMap.get(coLocationConstraint); + if (Objects.isNull(executionSlotSharingGroup)) { + executionSlotSharingGroup = executionSlotSharingGroups.get(index); + } + addVertexToExecutionSlotSharingGroup(executionSlotSharingGroup, executionVertex); + index = ++index % executionSlotSharingGroups.size(); + } + updateConstraintToExecutionSlotSharingGroupMap(executionVertices); + } + + private void allocateCoLocatedNonMaxParallelVertices( + SlotSharingGroup slotSharingGroup, + List<SchedulingExecutionVertex> executionVertices) { + List<ExecutionSlotSharingGroup> executionSlotSharingGroups = + paralleledExecutionSlotSharingGroupsMap.get(slotSharingGroup); + int index; + // For CoLocations with non-max parallelism of the slot sharing group. + for (SchedulingExecutionVertex executionVertex : executionVertices) { + final CoLocationConstraint coLocationConstraint = + getCoLocationConstraint(executionVertex); + ExecutionSlotSharingGroup executionSlotSharingGroup = + constraintToExecutionSlotSharingGroupMap.get(coLocationConstraint); + if (Objects.isNull(executionSlotSharingGroup)) { + index = getFittestSlotIndex(executionSlotSharingGroups, executionVertex); Review Comment: Most of code of `allocateCoLocatedNonMaxParallelVertices` and `allocateCoLocatedMaxParallelVertices` are same. IIUC, if we define a `getFittestSlotIndex` interface, these 2 methods can be merged to one method, right? - For max, `getFittestSlotIndex` return `executionVertex.getId().getSubtaskIndex()`. - For non-max, `getFittestSlotIndex` call currrent `getFittestSlotIndex` logic. ########## flink-runtime/src/main/java/org/apache/flink/runtime/scheduler/TaskBalancedPreferredSlotSharingStrategy.java: ########## @@ -0,0 +1,361 @@ +/* + * 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.api.java.tuple.Tuple2; +import org.apache.flink.runtime.jobgraph.JobVertexID; +import org.apache.flink.runtime.jobmanager.scheduler.CoLocationConstraint; +import org.apache.flink.runtime.jobmanager.scheduler.CoLocationGroup; +import org.apache.flink.runtime.jobmanager.scheduler.SlotSharingGroup; +import org.apache.flink.runtime.scheduler.strategy.ExecutionVertexID; +import org.apache.flink.runtime.scheduler.strategy.SchedulingExecutionVertex; +import org.apache.flink.runtime.scheduler.strategy.SchedulingTopology; + +import org.slf4j.Logger; +import org.slf4j.LoggerFactory; + +import javax.annotation.Nullable; + +import java.util.ArrayList; +import java.util.HashMap; +import java.util.LinkedHashMap; +import java.util.List; +import java.util.Map; +import java.util.Objects; +import java.util.Set; +import java.util.stream.Collectors; + +import static org.apache.flink.util.Preconditions.checkNotNull; + +/** + * This strategy tries to get a balanced tasks scheduling. Execution vertices, which are belong to + * the same SlotSharingGroup, tend to be put evenly in each ExecutionSlotSharingGroup. Co-location + * constraints will be respected. + */ +class TaskBalancedPreferredSlotSharingStrategy extends AbstractSlotSharingStrategy { + + public static final Logger LOG = + LoggerFactory.getLogger(TaskBalancedPreferredSlotSharingStrategy.class); + + TaskBalancedPreferredSlotSharingStrategy( + final SchedulingTopology topology, + final Set<SlotSharingGroup> slotSharingGroups, + final Set<CoLocationGroup> coLocationGroups) { + super(topology, slotSharingGroups, coLocationGroups); + } + + @Override + protected Map<ExecutionVertexID, ExecutionSlotSharingGroup> computeExecutionSlotSharingGroups( + SchedulingTopology schedulingTopology) { + return new TaskBalancedExecutionSlotSharingGroupBuilder( + schedulingTopology, this.logicalSlotSharingGroups, this.coLocationGroups) + .build(); + } + + static class Factory implements SlotSharingStrategy.Factory { + + public TaskBalancedPreferredSlotSharingStrategy create( + final SchedulingTopology topology, + final Set<SlotSharingGroup> slotSharingGroups, + final Set<CoLocationGroup> coLocationGroups) { + + return new TaskBalancedPreferredSlotSharingStrategy( + topology, slotSharingGroups, coLocationGroups); + } + } + + /** SlotSharingGroupBuilder class for balanced scheduling strategy. */ + private static class TaskBalancedExecutionSlotSharingGroupBuilder { + + private final SchedulingTopology topology; + + private final Map<JobVertexID, SlotSharingGroup> slotSharingGroupMap; + + /** Record the {@link ExecutionSlotSharingGroup}s for {@link SlotSharingGroup}s. */ + private final Map<SlotSharingGroup, List<ExecutionSlotSharingGroup>> + paralleledExecutionSlotSharingGroupsMap; + + /** + * Record the next round-robin {@link ExecutionSlotSharingGroup} index for {@link + * SlotSharingGroup}s. + */ + private final Map<SlotSharingGroup, Integer> slotSharingGroupIndexMap; + + private final Map<ExecutionVertexID, ExecutionSlotSharingGroup> + executionSlotSharingGroupMap; + + private final Map<JobVertexID, CoLocationGroup> coLocationGroupMap; + + private final Map<CoLocationConstraint, ExecutionSlotSharingGroup> + constraintToExecutionSlotSharingGroupMap; + + private TaskBalancedExecutionSlotSharingGroupBuilder( + final SchedulingTopology topology, + final Set<SlotSharingGroup> slotSharingGroups, + final Set<CoLocationGroup> coLocationGroups) { + this.topology = checkNotNull(topology); + + this.coLocationGroupMap = new HashMap<>(); + for (CoLocationGroup coLocationGroup : coLocationGroups) { + for (JobVertexID jobVertexId : coLocationGroup.getVertexIds()) { + coLocationGroupMap.put(jobVertexId, coLocationGroup); + } + } + + this.constraintToExecutionSlotSharingGroupMap = new HashMap<>(); + this.paralleledExecutionSlotSharingGroupsMap = new HashMap<>(slotSharingGroups.size()); + this.slotSharingGroupIndexMap = new HashMap<>(slotSharingGroups.size()); + this.slotSharingGroupMap = new HashMap<>(); + this.executionSlotSharingGroupMap = new HashMap<>(); + + for (SlotSharingGroup slotSharingGroup : slotSharingGroups) { + for (JobVertexID jobVertexId : slotSharingGroup.getJobVertexIds()) { + slotSharingGroupMap.put(jobVertexId, slotSharingGroup); + } + } + } + + private Map<ExecutionVertexID, ExecutionSlotSharingGroup> build() { + + final LinkedHashMap<JobVertexID, List<SchedulingExecutionVertex>> allVertices = + getExecutionVertices(topology); + + initParalleledExecutionSlotSharingGroupsMap(allVertices); + + // Loop on job vertices + for (Map.Entry<JobVertexID, List<SchedulingExecutionVertex>> executionVertexInfos : + allVertices.entrySet()) { + + JobVertexID jobVertexID = executionVertexInfos.getKey(); + List<SchedulingExecutionVertex> executionVertices = executionVertexInfos.getValue(); + final int jobVertexParallel = executionVertices.size(); + final SlotSharingGroup slotSharingGroup = slotSharingGroupMap.get(jobVertexID); + final int index = getSlotRoundRobinIndex(jobVertexParallel, slotSharingGroup); + + if (!coLocationGroupMap.containsKey(jobVertexID)) { + // For vertices without CoLocationConstraint. + allocateNonCoLocatedVertices(slotSharingGroup, executionVertices, index); + } else { + // For vertices with CoLocationConstraint. + allocateCoLocatedVertices(slotSharingGroup, executionVertices, index); + } + } + return executionSlotSharingGroupMap; + } + + private void initParalleledExecutionSlotSharingGroupsMap( + final LinkedHashMap<JobVertexID, List<SchedulingExecutionVertex>> allVertices) { + + allVertices.entrySet().stream() + .map( + jobVertexExecutionVertices -> + Tuple2.of( + slotSharingGroupMap.get( + jobVertexExecutionVertices.getKey()), + jobVertexExecutionVertices.getValue().size())) + .collect( + Collectors.groupingBy( + tuple -> tuple.f0, + Collectors.summarizingInt(tuple -> tuple.f1))) + .forEach( + (slotSharingGroup, statistics) -> { + int slotNum = statistics.getMax(); + paralleledExecutionSlotSharingGroupsMap.put( + slotSharingGroup, + createExecutionSlotSharingGroups( + slotSharingGroup, slotNum)); + }); + } + + private List<ExecutionSlotSharingGroup> createExecutionSlotSharingGroups( + SlotSharingGroup slotSharingGroup, int slotNum) { + final List<ExecutionSlotSharingGroup> executionSlotSharingGroups = + new ArrayList<>(slotNum); + for (int i = 0; i < slotNum; i++) { + final ExecutionSlotSharingGroup executionSlotSharingGroup = + new ExecutionSlotSharingGroup(slotSharingGroup); + executionSlotSharingGroups.add(i, executionSlotSharingGroup); + LOG.debug( + "Create {}th executionSlotSharingGroup {}.", i, executionSlotSharingGroup); + } + return executionSlotSharingGroups; + } + + private void allocateCoLocatedVertices( + SlotSharingGroup slotSharingGroup, + List<SchedulingExecutionVertex> executionVertices, + int index) { + final int slotNum = + paralleledExecutionSlotSharingGroupsMap.get(slotSharingGroup).size(); + final int jobVertexParallel = executionVertices.size(); + if (slotNum == jobVertexParallel) { + // For CoLocations with max parallelism of the slot sharing group. + allocateCoLocatedMaxParallelVertices(slotSharingGroup, executionVertices, index); + } else { + // For CoLocations with non-max parallelism of the slot sharing group. + allocateCoLocatedNonMaxParallelVertices(slotSharingGroup, executionVertices); + } + } + + private void allocateCoLocatedMaxParallelVertices( + SlotSharingGroup slotSharingGroup, + List<SchedulingExecutionVertex> executionVertices, + int index) { + final List<ExecutionSlotSharingGroup> executionSlotSharingGroups = + paralleledExecutionSlotSharingGroupsMap.get(slotSharingGroup); + for (SchedulingExecutionVertex executionVertex : executionVertices) { + final CoLocationConstraint coLocationConstraint = + getCoLocationConstraint(executionVertex); + ExecutionSlotSharingGroup executionSlotSharingGroup = + constraintToExecutionSlotSharingGroupMap.get(coLocationConstraint); + if (Objects.isNull(executionSlotSharingGroup)) { + executionSlotSharingGroup = executionSlotSharingGroups.get(index); + } + addVertexToExecutionSlotSharingGroup(executionSlotSharingGroup, executionVertex); + index = ++index % executionSlotSharingGroups.size(); + } + updateConstraintToExecutionSlotSharingGroupMap(executionVertices); + } + + private void allocateCoLocatedNonMaxParallelVertices( + SlotSharingGroup slotSharingGroup, + List<SchedulingExecutionVertex> executionVertices) { + List<ExecutionSlotSharingGroup> executionSlotSharingGroups = + paralleledExecutionSlotSharingGroupsMap.get(slotSharingGroup); + int index; + // For CoLocations with non-max parallelism of the slot sharing group. + for (SchedulingExecutionVertex executionVertex : executionVertices) { + final CoLocationConstraint coLocationConstraint = + getCoLocationConstraint(executionVertex); + ExecutionSlotSharingGroup executionSlotSharingGroup = + constraintToExecutionSlotSharingGroupMap.get(coLocationConstraint); + if (Objects.isNull(executionSlotSharingGroup)) { + index = getFittestSlotIndex(executionSlotSharingGroups, executionVertex); + executionSlotSharingGroup = executionSlotSharingGroups.get(index); + } + addVertexToExecutionSlotSharingGroup(executionSlotSharingGroup, executionVertex); + } + updateConstraintToExecutionSlotSharingGroupMap(executionVertices); + index = getFittestSlotIndex(executionSlotSharingGroups, null); + final int jobVertexParallel = executionVertices.size(); + updateSlotRoundRobinIndexIfNeeded(jobVertexParallel, slotSharingGroup, index); + } + + private void allocateNonCoLocatedVertices( + SlotSharingGroup slotSharingGroup, + List<SchedulingExecutionVertex> executionVertices, + int index) { + final List<ExecutionSlotSharingGroup> executionSlotSharingGroups = + paralleledExecutionSlotSharingGroupsMap.get(slotSharingGroup); + for (SchedulingExecutionVertex executionVertex : executionVertices) { + addVertexToExecutionSlotSharingGroup( + executionSlotSharingGroups.get(index), executionVertex); + index = ++index % executionSlotSharingGroups.size(); + } + updateSlotRoundRobinIndexIfNeeded(executionVertices.size(), slotSharingGroup, index); + } + + private void addVertexToExecutionSlotSharingGroup( + ExecutionSlotSharingGroup executionSlotSharingGroup, + SchedulingExecutionVertex executionVertex) { + final ExecutionVertexID executionVertexId = executionVertex.getId(); + executionSlotSharingGroup.addVertex(executionVertexId); + executionSlotSharingGroupMap.put(executionVertexId, executionSlotSharingGroup); + } + + private void updateConstraintToExecutionSlotSharingGroupMap( + final List<SchedulingExecutionVertex> executionVertices) { + + for (SchedulingExecutionVertex executionVertex : executionVertices) { + final ExecutionVertexID executionVertexId = executionVertex.getId(); + final CoLocationGroup coLocationGroup = + coLocationGroupMap.get(executionVertexId.getJobVertexId()); + if (Objects.nonNull(coLocationGroup)) { + final CoLocationConstraint constraint = + getCoLocationConstraint(executionVertex); + + constraintToExecutionSlotSharingGroupMap.put( + constraint, executionSlotSharingGroupMap.get(executionVertexId)); + } + } + } + + private CoLocationConstraint getCoLocationConstraint(SchedulingExecutionVertex sev) { + final JobVertexID jobVertexID = sev.getId().getJobVertexId(); + final int subtaskIndex = sev.getId().getSubtaskIndex(); + return coLocationGroupMap.get(jobVertexID).getLocationConstraint(subtaskIndex); + } + + private int getSlotRoundRobinIndex( + final int jobVertexParallelism, SlotSharingGroup slotSharingGroup) { + final List<ExecutionSlotSharingGroup> executionSlotSharingGroups = + paralleledExecutionSlotSharingGroupsMap.get(slotSharingGroup); + final boolean maxParallel = jobVertexParallelism == executionSlotSharingGroups.size(); + return maxParallel ? 0 : slotSharingGroupIndexMap.getOrDefault(slotSharingGroup, 0); + } + + private int getFittestSlotIndex( + final List<ExecutionSlotSharingGroup> executionSlotSharingGroups, + final @Nullable SchedulingExecutionVertex executionVertex) { + int indexWithLeastExecutionVertices = 0; + int leastExecutionVertices = Integer.MAX_VALUE; + for (int index = 0; index < executionSlotSharingGroups.size(); index++) { + final ExecutionSlotSharingGroup executionSlotSharingGroup = + executionSlotSharingGroups.get(index); + final int executionVertices = + executionSlotSharingGroup.getExecutionVertexIds().size(); + if (leastExecutionVertices > executionVertices + && allocatable(executionSlotSharingGroup, executionVertex)) { + indexWithLeastExecutionVertices = index; + leastExecutionVertices = executionVertices; + } + } + return indexWithLeastExecutionVertices; + } + + private void updateSlotRoundRobinIndexIfNeeded( Review Comment: The two methods `updateSlotRoundRobinIndexIfNeeded` and `getSlotRoundRobinIndex` are best placed next to each other. And it's better to extract an method to check whether current paralleslism is the maxParallel of current ssg. These 2 methods will use it. Currently, you write it twice. ########## flink-runtime/src/main/java/org/apache/flink/runtime/scheduler/TaskBalancedPreferredSlotSharingStrategy.java: ########## @@ -0,0 +1,361 @@ +/* + * 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.api.java.tuple.Tuple2; +import org.apache.flink.runtime.jobgraph.JobVertexID; +import org.apache.flink.runtime.jobmanager.scheduler.CoLocationConstraint; +import org.apache.flink.runtime.jobmanager.scheduler.CoLocationGroup; +import org.apache.flink.runtime.jobmanager.scheduler.SlotSharingGroup; +import org.apache.flink.runtime.scheduler.strategy.ExecutionVertexID; +import org.apache.flink.runtime.scheduler.strategy.SchedulingExecutionVertex; +import org.apache.flink.runtime.scheduler.strategy.SchedulingTopology; + +import org.slf4j.Logger; +import org.slf4j.LoggerFactory; + +import javax.annotation.Nullable; + +import java.util.ArrayList; +import java.util.HashMap; +import java.util.LinkedHashMap; +import java.util.List; +import java.util.Map; +import java.util.Objects; +import java.util.Set; +import java.util.stream.Collectors; + +import static org.apache.flink.util.Preconditions.checkNotNull; + +/** + * This strategy tries to get a balanced tasks scheduling. Execution vertices, which are belong to + * the same SlotSharingGroup, tend to be put evenly in each ExecutionSlotSharingGroup. Co-location + * constraints will be respected. + */ +class TaskBalancedPreferredSlotSharingStrategy extends AbstractSlotSharingStrategy { + + public static final Logger LOG = + LoggerFactory.getLogger(TaskBalancedPreferredSlotSharingStrategy.class); + + TaskBalancedPreferredSlotSharingStrategy( + final SchedulingTopology topology, + final Set<SlotSharingGroup> slotSharingGroups, + final Set<CoLocationGroup> coLocationGroups) { + super(topology, slotSharingGroups, coLocationGroups); + } + + @Override + protected Map<ExecutionVertexID, ExecutionSlotSharingGroup> computeExecutionSlotSharingGroups( + SchedulingTopology schedulingTopology) { + return new TaskBalancedExecutionSlotSharingGroupBuilder( + schedulingTopology, this.logicalSlotSharingGroups, this.coLocationGroups) + .build(); + } + + static class Factory implements SlotSharingStrategy.Factory { + + public TaskBalancedPreferredSlotSharingStrategy create( + final SchedulingTopology topology, + final Set<SlotSharingGroup> slotSharingGroups, + final Set<CoLocationGroup> coLocationGroups) { + + return new TaskBalancedPreferredSlotSharingStrategy( + topology, slotSharingGroups, coLocationGroups); + } + } + + /** SlotSharingGroupBuilder class for balanced scheduling strategy. */ + private static class TaskBalancedExecutionSlotSharingGroupBuilder { + + private final SchedulingTopology topology; + + private final Map<JobVertexID, SlotSharingGroup> slotSharingGroupMap; + + /** Record the {@link ExecutionSlotSharingGroup}s for {@link SlotSharingGroup}s. */ + private final Map<SlotSharingGroup, List<ExecutionSlotSharingGroup>> + paralleledExecutionSlotSharingGroupsMap; + + /** + * Record the next round-robin {@link ExecutionSlotSharingGroup} index for {@link + * SlotSharingGroup}s. + */ + private final Map<SlotSharingGroup, Integer> slotSharingGroupIndexMap; + + private final Map<ExecutionVertexID, ExecutionSlotSharingGroup> + executionSlotSharingGroupMap; + + private final Map<JobVertexID, CoLocationGroup> coLocationGroupMap; + + private final Map<CoLocationConstraint, ExecutionSlotSharingGroup> + constraintToExecutionSlotSharingGroupMap; + + private TaskBalancedExecutionSlotSharingGroupBuilder( + final SchedulingTopology topology, + final Set<SlotSharingGroup> slotSharingGroups, + final Set<CoLocationGroup> coLocationGroups) { + this.topology = checkNotNull(topology); + + this.coLocationGroupMap = new HashMap<>(); + for (CoLocationGroup coLocationGroup : coLocationGroups) { + for (JobVertexID jobVertexId : coLocationGroup.getVertexIds()) { + coLocationGroupMap.put(jobVertexId, coLocationGroup); + } + } + + this.constraintToExecutionSlotSharingGroupMap = new HashMap<>(); + this.paralleledExecutionSlotSharingGroupsMap = new HashMap<>(slotSharingGroups.size()); + this.slotSharingGroupIndexMap = new HashMap<>(slotSharingGroups.size()); + this.slotSharingGroupMap = new HashMap<>(); + this.executionSlotSharingGroupMap = new HashMap<>(); + + for (SlotSharingGroup slotSharingGroup : slotSharingGroups) { + for (JobVertexID jobVertexId : slotSharingGroup.getJobVertexIds()) { + slotSharingGroupMap.put(jobVertexId, slotSharingGroup); + } + } + } + + private Map<ExecutionVertexID, ExecutionSlotSharingGroup> build() { + + final LinkedHashMap<JobVertexID, List<SchedulingExecutionVertex>> allVertices = + getExecutionVertices(topology); + + initParalleledExecutionSlotSharingGroupsMap(allVertices); + + // Loop on job vertices + for (Map.Entry<JobVertexID, List<SchedulingExecutionVertex>> executionVertexInfos : + allVertices.entrySet()) { + + JobVertexID jobVertexID = executionVertexInfos.getKey(); + List<SchedulingExecutionVertex> executionVertices = executionVertexInfos.getValue(); + final int jobVertexParallel = executionVertices.size(); + final SlotSharingGroup slotSharingGroup = slotSharingGroupMap.get(jobVertexID); + final int index = getSlotRoundRobinIndex(jobVertexParallel, slotSharingGroup); + + if (!coLocationGroupMap.containsKey(jobVertexID)) { + // For vertices without CoLocationConstraint. + allocateNonCoLocatedVertices(slotSharingGroup, executionVertices, index); + } else { + // For vertices with CoLocationConstraint. + allocateCoLocatedVertices(slotSharingGroup, executionVertices, index); + } + } + return executionSlotSharingGroupMap; + } + + private void initParalleledExecutionSlotSharingGroupsMap( + final LinkedHashMap<JobVertexID, List<SchedulingExecutionVertex>> allVertices) { + + allVertices.entrySet().stream() + .map( + jobVertexExecutionVertices -> + Tuple2.of( + slotSharingGroupMap.get( + jobVertexExecutionVertices.getKey()), + jobVertexExecutionVertices.getValue().size())) + .collect( + Collectors.groupingBy( + tuple -> tuple.f0, + Collectors.summarizingInt(tuple -> tuple.f1))) + .forEach( + (slotSharingGroup, statistics) -> { + int slotNum = statistics.getMax(); + paralleledExecutionSlotSharingGroupsMap.put( + slotSharingGroup, + createExecutionSlotSharingGroups( + slotSharingGroup, slotNum)); + }); + } + + private List<ExecutionSlotSharingGroup> createExecutionSlotSharingGroups( + SlotSharingGroup slotSharingGroup, int slotNum) { + final List<ExecutionSlotSharingGroup> executionSlotSharingGroups = + new ArrayList<>(slotNum); + for (int i = 0; i < slotNum; i++) { + final ExecutionSlotSharingGroup executionSlotSharingGroup = + new ExecutionSlotSharingGroup(slotSharingGroup); + executionSlotSharingGroups.add(i, executionSlotSharingGroup); + LOG.debug( + "Create {}th executionSlotSharingGroup {}.", i, executionSlotSharingGroup); + } + return executionSlotSharingGroups; + } + + private void allocateCoLocatedVertices( + SlotSharingGroup slotSharingGroup, + List<SchedulingExecutionVertex> executionVertices, + int index) { Review Comment: This index may be not needed, right? For max parallelism, using the `executionVertex.getId().getSubtaskIndex()` as the index is enough. For non-max, I see the index isn't used. ########## flink-runtime/src/main/java/org/apache/flink/runtime/scheduler/TaskBalancedPreferredSlotSharingStrategy.java: ########## @@ -0,0 +1,361 @@ +/* + * 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.api.java.tuple.Tuple2; +import org.apache.flink.runtime.jobgraph.JobVertexID; +import org.apache.flink.runtime.jobmanager.scheduler.CoLocationConstraint; +import org.apache.flink.runtime.jobmanager.scheduler.CoLocationGroup; +import org.apache.flink.runtime.jobmanager.scheduler.SlotSharingGroup; +import org.apache.flink.runtime.scheduler.strategy.ExecutionVertexID; +import org.apache.flink.runtime.scheduler.strategy.SchedulingExecutionVertex; +import org.apache.flink.runtime.scheduler.strategy.SchedulingTopology; + +import org.slf4j.Logger; +import org.slf4j.LoggerFactory; + +import javax.annotation.Nullable; + +import java.util.ArrayList; +import java.util.HashMap; +import java.util.LinkedHashMap; +import java.util.List; +import java.util.Map; +import java.util.Objects; +import java.util.Set; +import java.util.stream.Collectors; + +import static org.apache.flink.util.Preconditions.checkNotNull; + +/** + * This strategy tries to get a balanced tasks scheduling. Execution vertices, which are belong to + * the same SlotSharingGroup, tend to be put evenly in each ExecutionSlotSharingGroup. Co-location + * constraints will be respected. + */ +class TaskBalancedPreferredSlotSharingStrategy extends AbstractSlotSharingStrategy { + + public static final Logger LOG = + LoggerFactory.getLogger(TaskBalancedPreferredSlotSharingStrategy.class); + + TaskBalancedPreferredSlotSharingStrategy( + final SchedulingTopology topology, + final Set<SlotSharingGroup> slotSharingGroups, + final Set<CoLocationGroup> coLocationGroups) { + super(topology, slotSharingGroups, coLocationGroups); + } + + @Override + protected Map<ExecutionVertexID, ExecutionSlotSharingGroup> computeExecutionSlotSharingGroups( + SchedulingTopology schedulingTopology) { + return new TaskBalancedExecutionSlotSharingGroupBuilder( + schedulingTopology, this.logicalSlotSharingGroups, this.coLocationGroups) + .build(); + } + + static class Factory implements SlotSharingStrategy.Factory { + + public TaskBalancedPreferredSlotSharingStrategy create( + final SchedulingTopology topology, + final Set<SlotSharingGroup> slotSharingGroups, + final Set<CoLocationGroup> coLocationGroups) { + + return new TaskBalancedPreferredSlotSharingStrategy( + topology, slotSharingGroups, coLocationGroups); + } + } + + /** SlotSharingGroupBuilder class for balanced scheduling strategy. */ + private static class TaskBalancedExecutionSlotSharingGroupBuilder { + + private final SchedulingTopology topology; + + private final Map<JobVertexID, SlotSharingGroup> slotSharingGroupMap; + + /** Record the {@link ExecutionSlotSharingGroup}s for {@link SlotSharingGroup}s. */ + private final Map<SlotSharingGroup, List<ExecutionSlotSharingGroup>> + paralleledExecutionSlotSharingGroupsMap; + + /** + * Record the next round-robin {@link ExecutionSlotSharingGroup} index for {@link + * SlotSharingGroup}s. + */ + private final Map<SlotSharingGroup, Integer> slotSharingGroupIndexMap; + + private final Map<ExecutionVertexID, ExecutionSlotSharingGroup> + executionSlotSharingGroupMap; + + private final Map<JobVertexID, CoLocationGroup> coLocationGroupMap; + + private final Map<CoLocationConstraint, ExecutionSlotSharingGroup> + constraintToExecutionSlotSharingGroupMap; + + private TaskBalancedExecutionSlotSharingGroupBuilder( + final SchedulingTopology topology, + final Set<SlotSharingGroup> slotSharingGroups, + final Set<CoLocationGroup> coLocationGroups) { + this.topology = checkNotNull(topology); + + this.coLocationGroupMap = new HashMap<>(); + for (CoLocationGroup coLocationGroup : coLocationGroups) { + for (JobVertexID jobVertexId : coLocationGroup.getVertexIds()) { + coLocationGroupMap.put(jobVertexId, coLocationGroup); + } + } + + this.constraintToExecutionSlotSharingGroupMap = new HashMap<>(); + this.paralleledExecutionSlotSharingGroupsMap = new HashMap<>(slotSharingGroups.size()); + this.slotSharingGroupIndexMap = new HashMap<>(slotSharingGroups.size()); + this.slotSharingGroupMap = new HashMap<>(); + this.executionSlotSharingGroupMap = new HashMap<>(); + + for (SlotSharingGroup slotSharingGroup : slotSharingGroups) { + for (JobVertexID jobVertexId : slotSharingGroup.getJobVertexIds()) { + slotSharingGroupMap.put(jobVertexId, slotSharingGroup); + } + } + } + + private Map<ExecutionVertexID, ExecutionSlotSharingGroup> build() { + + final LinkedHashMap<JobVertexID, List<SchedulingExecutionVertex>> allVertices = + getExecutionVertices(topology); + + initParalleledExecutionSlotSharingGroupsMap(allVertices); + + // Loop on job vertices + for (Map.Entry<JobVertexID, List<SchedulingExecutionVertex>> executionVertexInfos : + allVertices.entrySet()) { + + JobVertexID jobVertexID = executionVertexInfos.getKey(); + List<SchedulingExecutionVertex> executionVertices = executionVertexInfos.getValue(); + final int jobVertexParallel = executionVertices.size(); + final SlotSharingGroup slotSharingGroup = slotSharingGroupMap.get(jobVertexID); + final int index = getSlotRoundRobinIndex(jobVertexParallel, slotSharingGroup); + + if (!coLocationGroupMap.containsKey(jobVertexID)) { + // For vertices without CoLocationConstraint. + allocateNonCoLocatedVertices(slotSharingGroup, executionVertices, index); + } else { + // For vertices with CoLocationConstraint. + allocateCoLocatedVertices(slotSharingGroup, executionVertices, index); + } + } + return executionSlotSharingGroupMap; + } + + private void initParalleledExecutionSlotSharingGroupsMap( + final LinkedHashMap<JobVertexID, List<SchedulingExecutionVertex>> allVertices) { + + allVertices.entrySet().stream() + .map( + jobVertexExecutionVertices -> + Tuple2.of( + slotSharingGroupMap.get( + jobVertexExecutionVertices.getKey()), + jobVertexExecutionVertices.getValue().size())) + .collect( + Collectors.groupingBy( + tuple -> tuple.f0, + Collectors.summarizingInt(tuple -> tuple.f1))) + .forEach( + (slotSharingGroup, statistics) -> { + int slotNum = statistics.getMax(); + paralleledExecutionSlotSharingGroupsMap.put( + slotSharingGroup, + createExecutionSlotSharingGroups( + slotSharingGroup, slotNum)); + }); + } + + private List<ExecutionSlotSharingGroup> createExecutionSlotSharingGroups( + SlotSharingGroup slotSharingGroup, int slotNum) { + final List<ExecutionSlotSharingGroup> executionSlotSharingGroups = + new ArrayList<>(slotNum); + for (int i = 0; i < slotNum; i++) { + final ExecutionSlotSharingGroup executionSlotSharingGroup = + new ExecutionSlotSharingGroup(slotSharingGroup); + executionSlotSharingGroups.add(i, executionSlotSharingGroup); + LOG.debug( + "Create {}th executionSlotSharingGroup {}.", i, executionSlotSharingGroup); + } + return executionSlotSharingGroups; + } + + private void allocateCoLocatedVertices( + SlotSharingGroup slotSharingGroup, + List<SchedulingExecutionVertex> executionVertices, + int index) { + final int slotNum = + paralleledExecutionSlotSharingGroupsMap.get(slotSharingGroup).size(); + final int jobVertexParallel = executionVertices.size(); + if (slotNum == jobVertexParallel) { + // For CoLocations with max parallelism of the slot sharing group. + allocateCoLocatedMaxParallelVertices(slotSharingGroup, executionVertices, index); + } else { + // For CoLocations with non-max parallelism of the slot sharing group. + allocateCoLocatedNonMaxParallelVertices(slotSharingGroup, executionVertices); + } + } + + private void allocateCoLocatedMaxParallelVertices( + SlotSharingGroup slotSharingGroup, + List<SchedulingExecutionVertex> executionVertices, + int index) { + final List<ExecutionSlotSharingGroup> executionSlotSharingGroups = + paralleledExecutionSlotSharingGroupsMap.get(slotSharingGroup); + for (SchedulingExecutionVertex executionVertex : executionVertices) { + final CoLocationConstraint coLocationConstraint = + getCoLocationConstraint(executionVertex); + ExecutionSlotSharingGroup executionSlotSharingGroup = + constraintToExecutionSlotSharingGroupMap.get(coLocationConstraint); + if (Objects.isNull(executionSlotSharingGroup)) { + executionSlotSharingGroup = executionSlotSharingGroups.get(index); + } + addVertexToExecutionSlotSharingGroup(executionSlotSharingGroup, executionVertex); + index = ++index % executionSlotSharingGroups.size(); + } + updateConstraintToExecutionSlotSharingGroupMap(executionVertices); + } + + private void allocateCoLocatedNonMaxParallelVertices( + SlotSharingGroup slotSharingGroup, + List<SchedulingExecutionVertex> executionVertices) { + List<ExecutionSlotSharingGroup> executionSlotSharingGroups = + paralleledExecutionSlotSharingGroupsMap.get(slotSharingGroup); + int index; + // For CoLocations with non-max parallelism of the slot sharing group. + for (SchedulingExecutionVertex executionVertex : executionVertices) { + final CoLocationConstraint coLocationConstraint = + getCoLocationConstraint(executionVertex); + ExecutionSlotSharingGroup executionSlotSharingGroup = + constraintToExecutionSlotSharingGroupMap.get(coLocationConstraint); + if (Objects.isNull(executionSlotSharingGroup)) { + index = getFittestSlotIndex(executionSlotSharingGroups, executionVertex); + executionSlotSharingGroup = executionSlotSharingGroups.get(index); + } + addVertexToExecutionSlotSharingGroup(executionSlotSharingGroup, executionVertex); + } + updateConstraintToExecutionSlotSharingGroupMap(executionVertices); + index = getFittestSlotIndex(executionSlotSharingGroups, null); + final int jobVertexParallel = executionVertices.size(); + updateSlotRoundRobinIndexIfNeeded(jobVertexParallel, slotSharingGroup, index); + } + + private void allocateNonCoLocatedVertices( + SlotSharingGroup slotSharingGroup, + List<SchedulingExecutionVertex> executionVertices, + int index) { + final List<ExecutionSlotSharingGroup> executionSlotSharingGroups = + paralleledExecutionSlotSharingGroupsMap.get(slotSharingGroup); + for (SchedulingExecutionVertex executionVertex : executionVertices) { + addVertexToExecutionSlotSharingGroup( + executionSlotSharingGroups.get(index), executionVertex); + index = ++index % executionSlotSharingGroups.size(); + } + updateSlotRoundRobinIndexIfNeeded(executionVertices.size(), slotSharingGroup, index); + } + + private void addVertexToExecutionSlotSharingGroup( + ExecutionSlotSharingGroup executionSlotSharingGroup, + SchedulingExecutionVertex executionVertex) { + final ExecutionVertexID executionVertexId = executionVertex.getId(); + executionSlotSharingGroup.addVertex(executionVertexId); + executionSlotSharingGroupMap.put(executionVertexId, executionSlotSharingGroup); + } + + private void updateConstraintToExecutionSlotSharingGroupMap( + final List<SchedulingExecutionVertex> executionVertices) { + + for (SchedulingExecutionVertex executionVertex : executionVertices) { + final ExecutionVertexID executionVertexId = executionVertex.getId(); + final CoLocationGroup coLocationGroup = + coLocationGroupMap.get(executionVertexId.getJobVertexId()); + if (Objects.nonNull(coLocationGroup)) { + final CoLocationConstraint constraint = + getCoLocationConstraint(executionVertex); + + constraintToExecutionSlotSharingGroupMap.put( + constraint, executionSlotSharingGroupMap.get(executionVertexId)); + } + } + } + + private CoLocationConstraint getCoLocationConstraint(SchedulingExecutionVertex sev) { + final JobVertexID jobVertexID = sev.getId().getJobVertexId(); + final int subtaskIndex = sev.getId().getSubtaskIndex(); + return coLocationGroupMap.get(jobVertexID).getLocationConstraint(subtaskIndex); + } + + private int getSlotRoundRobinIndex( + final int jobVertexParallelism, SlotSharingGroup slotSharingGroup) { + final List<ExecutionSlotSharingGroup> executionSlotSharingGroups = + paralleledExecutionSlotSharingGroupsMap.get(slotSharingGroup); + final boolean maxParallel = jobVertexParallelism == executionSlotSharingGroups.size(); + return maxParallel ? 0 : slotSharingGroupIndexMap.getOrDefault(slotSharingGroup, 0); + } + + private int getFittestSlotIndex( + final List<ExecutionSlotSharingGroup> executionSlotSharingGroups, + final @Nullable SchedulingExecutionVertex executionVertex) { + int indexWithLeastExecutionVertices = 0; + int leastExecutionVertices = Integer.MAX_VALUE; + for (int index = 0; index < executionSlotSharingGroups.size(); index++) { + final ExecutionSlotSharingGroup executionSlotSharingGroup = + executionSlotSharingGroups.get(index); + final int executionVertices = + executionSlotSharingGroup.getExecutionVertexIds().size(); + if (leastExecutionVertices > executionVertices + && allocatable(executionSlotSharingGroup, executionVertex)) { + indexWithLeastExecutionVertices = index; + leastExecutionVertices = executionVertices; + } + } + return indexWithLeastExecutionVertices; + } + + private void updateSlotRoundRobinIndexIfNeeded( Review Comment: Also, I found `allocateCoLocatedVertices` also check whether current paralleslism is the maxParallel of current ssg. So 3 callers can call the common method. -- 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. To unsubscribe, e-mail: issues-unsubscr...@flink.apache.org For queries about this service, please contact Infrastructure at: us...@infra.apache.org