[SPARK-24795][CORE] Implement barrier execution mode

core/src/main/scala/org/apache/spark/BarrierTaskContext.scala

@@ -0,0 +1,42 @@
+/*
+ * 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.spark
+
+import org.apache.spark.annotation.{Experimental, Since}
+
+/** A [[TaskContext]] with extra info and tooling for a barrier stage. */
+trait BarrierTaskContext extends TaskContext {
+
+  /**
+   * :: Experimental ::
+   * Sets a global barrier and waits until all tasks in this stage hit this barrier. Similar to
+   * MPI_Barrier function in MPI, the barrier() function call blocks until all tasks in the same
+   * stage have reached this routine.
+   */
+  @Experimental
+  @Since("2.4.0")
+  def barrier(): Unit
+
+  /**
+   * :: Experimental ::
+   * Returns the all task infos in this barrier stage, the task infos are ordered by partitionId.
+   */
+  @Experimental
+  @Since("2.4.0")
+  def getTaskInfos(): Array[BarrierTaskInfo]
+}

core/src/main/scala/org/apache/spark/BarrierTaskContextImpl.scala

@@ -0,0 +1,49 @@
+/*
+ * 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.spark
+
+import java.util.Properties
+
+import org.apache.spark.executor.TaskMetrics
+import org.apache.spark.memory.TaskMemoryManager
+import org.apache.spark.metrics.MetricsSystem
+
+/** A [[BarrierTaskContext]] implementation. */
+private[spark] class BarrierTaskContextImpl(
+    override val stageId: Int,
+    override val stageAttemptNumber: Int,
+    override val partitionId: Int,
+    override val taskAttemptId: Long,
+    override val attemptNumber: Int,
+    override val taskMemoryManager: TaskMemoryManager,
+    localProperties: Properties,
+    @transient private val metricsSystem: MetricsSystem,
+    // The default value is only used in tests.
+    override val taskMetrics: TaskMetrics = TaskMetrics.empty)
+  extends TaskContextImpl(stageId, stageAttemptNumber, partitionId, taskAttemptId, attemptNumber,
+      taskMemoryManager, localProperties, metricsSystem, taskMetrics)
+    with BarrierTaskContext {
+
+  // TODO SPARK-24817 implement global barrier.
+  override def barrier(): Unit = {}
+
+  override def getTaskInfos(): Array[BarrierTaskInfo] = {
+    val addressesStr = localProperties.getProperty("addresses", "")
+    addressesStr.split(",").map(_.trim()).map(new BarrierTaskInfo(_))
+  }
+}

core/src/main/scala/org/apache/spark/BarrierTaskInfo.scala

@@ -0,0 +1,31 @@
+/*
+ * 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.spark
+
+import org.apache.spark.annotation.{Experimental, Since}
+
+
+/**
+ * :: Experimental ::
+ * Carries all task infos of a barrier task.
+ *
+ * @param address the IPv4 address(host:port) of the executor that a barrier task is running on
+ */
+@Experimental
+@Since("2.4.0")
+class BarrierTaskInfo(val address: String)

core/src/main/scala/org/apache/spark/MapOutputTracker.scala

@@ -434,6 +434,18 @@ private[spark] class MapOutputTrackerMaster(
     }
   }
 
+  /** Unregister all map output information of the given shuffle. */
+  def unregisterAllMapOutput(shuffleId: Int) {
+    shuffleStatuses.get(shuffleId) match {
+      case Some(shuffleStatus) =>
+        shuffleStatus.removeOutputsByFilter(x => true)
+        incrementEpoch()
+      case None =>
+        throw new SparkException(
+          s"unregisterAllMapOutput called for nonexistent shuffle ID $shuffleId.")
+    }
+  }
+
   /** Unregister shuffle data */
   def unregisterShuffle(shuffleId: Int) {
     shuffleStatuses.remove(shuffleId).foreach { shuffleStatus =>

core/src/main/scala/org/apache/spark/rdd/MapPartitionsRDD.scala

@@ -27,7 +27,8 @@ import org.apache.spark.{Partition, TaskContext}
 private[spark] class MapPartitionsRDD[U: ClassTag, T: ClassTag](
     var prev: RDD[T],
     f: (TaskContext, Int, Iterator[T]) => Iterator[U],  // (TaskContext, partition index, iterator)
-    preservesPartitioning: Boolean = false)
+    preservesPartitioning: Boolean = false,
+    isFromBarrier: Boolean = false)
   extends RDD[U](prev) {
 
   override val partitioner = if (preservesPartitioning) firstParent[T].partitioner else None
@@ -41,4 +42,7 @@ private[spark] class MapPartitionsRDD[U: ClassTag, T: ClassTag](
     super.clearDependencies()
     prev = null
   }
+
+  @transient protected lazy override val isBarrier_ : Boolean =
+    isFromBarrier || dependencies.exists(_.rdd.isBarrier())
 }

core/src/main/scala/org/apache/spark/rdd/RDD.scala

@@ -33,7 +33,7 @@ import org.apache.hadoop.mapred.TextOutputFormat
 
 import org.apache.spark._
 import org.apache.spark.Partitioner._
-import org.apache.spark.annotation.{DeveloperApi, Since}
+import org.apache.spark.annotation.{DeveloperApi, Experimental, Since}
 import org.apache.spark.api.java.JavaRDD
 import org.apache.spark.internal.Logging
 import org.apache.spark.partial.BoundedDouble
@@ -1647,6 +1647,14 @@ abstract class RDD[T: ClassTag](
     }
   }
 
+  /**
+   * :: Experimental ::
+   * Indicates that Spark must launch the tasks together for the current stage.
+   */
+  @Experimental
+  @Since("2.4.0")
+  def barrier(): RDDBarrier[T] = withScope(new RDDBarrier[T](this))
+
   // =======================================================================
   // Other internal methods and fields
   // =======================================================================
@@ -1839,6 +1847,20 @@ abstract class RDD[T: ClassTag](
   def toJavaRDD() : JavaRDD[T] = {
     new JavaRDD(this)(elementClassTag)
   }
+
+  /**
+   * Whether the RDD is in a barrier stage. Spark must launch all the tasks at the same time for a
+   * barrier stage.
+   *
+   * An RDD is in a barrier stage, if at least one of its parent RDD(s), or itself, are mapped from
+   * a RDDBarrier. This function always returns false for a [[ShuffledRDD]], since a
+   * [[ShuffledRDD]] indicates start of a new stage.
+   */
+  private[spark] def isBarrier(): Boolean = isBarrier_
+
+  // From performance concern, cache the value to avoid repeatedly compute `isBarrier()` on a long
+  // RDD chain.
+  @transient protected lazy val isBarrier_ : Boolean = dependencies.exists(_.rdd.isBarrier())
 }
 
 

core/src/main/scala/org/apache/spark/rdd/RDDBarrier.scala

@@ -0,0 +1,52 @@
+/*
+ * 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.spark.rdd
+
+import scala.reflect.ClassTag
+
+import org.apache.spark.BarrierTaskContext
+import org.apache.spark.TaskContext
+import org.apache.spark.annotation.{Experimental, Since}
+
+/** Represents an RDD barrier, which forces Spark to launch tasks of this stage together. */
+class RDDBarrier[T: ClassTag](rdd: RDD[T]) {
+
+  /**
+   * :: Experimental ::
+   * Maps partitions together with a provided BarrierTaskContext.
+   *
+   * `preservesPartitioning` indicates whether the input function preserves the partitioner, which
+   * should be `false` unless `rdd` is a pair RDD and the input function doesn't modify the keys.
+   */
+  @Experimental
+  @Since("2.4.0")
+  def mapPartitions[S: ClassTag](
+      f: (Iterator[T], BarrierTaskContext) => Iterator[S],
+      preservesPartitioning: Boolean = false): RDD[S] = rdd.withScope {
+    val cleanedF = rdd.sparkContext.clean(f)
+    new MapPartitionsRDD(
+      rdd,
+      (context: TaskContext, index: Int, iter: Iterator[T]) =>
+        cleanedF(iter, context.asInstanceOf[BarrierTaskContext]),
+      preservesPartitioning,
+      isFromBarrier = true
+    )
+  }
+
+  /** TODO extra conf(e.g. timeout) */
+}

core/src/main/scala/org/apache/spark/rdd/ShuffledRDD.scala

@@ -110,4 +110,6 @@ class ShuffledRDD[K: ClassTag, V: ClassTag, C: ClassTag](
     super.clearDependencies()
     prev = null
   }
+
+  private[spark] override def isBarrier(): Boolean = false
 }

core/src/main/scala/org/apache/spark/scheduler/ActiveJob.scala

@@ -60,4 +60,10 @@ private[spark] class ActiveJob(
   val finished = Array.fill[Boolean](numPartitions)(false)
 
   var numFinished = 0
+
+  // Mark all the partitions of the stage to be not finished.
+  def resetAllPartitions(): Unit = {
+    (0 until numPartitions).foreach(finished.update(_, false))
+    numFinished = 0
+  }
 }