Initial work towards new RDD design

core/src/main/scala/spark/DaemonThreadFactory.scala

@@ -0,0 +1,14 @@
+package spark
+
+import java.util.concurrent.ThreadFactory
+
+/**
+ * A ThreadFactory that creates daemon threads
+ */
+private object DaemonThreadFactory extends ThreadFactory {
+  override def newThread(r: Runnable): Thread = {
+    val t = new Thread(r);
+    t.setDaemon(true)
+    return t
+  }
+}

core/src/main/scala/spark/LocalScheduler.scala

@@ -2,6 +2,8 @@ package spark
 
 import java.util.concurrent._
 
+import scala.collection.mutable.HashMap
+import scala.collection.mutable.HashSet
 import scala.collection.mutable.Map
 
 /**
@@ -14,28 +16,27 @@ private class LocalScheduler(threads: Int) extends Scheduler with Logging {
   override def start() {}
   
   override def waitForRegister() {}
-  
-  override def runTasks[T](tasks: Array[Task[T]])(implicit m: ClassManifest[T])
-      : Array[T] = {
-    val futures = new Array[Future[TaskResult[T]]](tasks.length)
-    
-    for (i <- 0 until tasks.length) {
-      futures(i) = threadPool.submit(new Callable[TaskResult[T]]() {
-        def call(): TaskResult[T] = {
+
+  val completionEvents = new LinkedBlockingQueue[CompletionEvent]
+
+  def submitTasks(tasks: Seq[Task[_]]) {
+    tasks.zipWithIndex.foreach { case (task, i) =>
+      threadPool.submit(new Runnable {
+        def run() {
           logInfo("Running task " + i)
           try {
             // Serialize and deserialize the task so that accumulators are
             // changed to thread-local ones; this adds a bit of unnecessary
-            // overhead but matches how the Nexus Executor works
+            // overhead but matches how the Mesos Executor works
             Accumulators.clear
             val bytes = Utils.serialize(tasks(i))
             logInfo("Size of task " + i + " is " + bytes.size + " bytes")
-            val task = Utils.deserialize[Task[T]](
+            val task = Utils.deserialize[Task[_]](
               bytes, currentThread.getContextClassLoader)
-            val value = task.run
+            val result: Any = task.run
             val accumUpdates = Accumulators.values
             logInfo("Finished task " + i)
-            new TaskResult[T](value, accumUpdates)
+            completionEvents.put(CompletionEvent(task, true, result, accumUpdates))
           } catch {
             case e: Exception => {
               // TODO: Do something nicer here
@@ -47,26 +48,233 @@ private class LocalScheduler(threads: Int) extends Scheduler with Logging {
         }
       })
     }
-    
-    val taskResults = futures.map(_.get)
-    for (result <- taskResults)
-      Accumulators.add(currentThread, result.accumUpdates)
-    return taskResults.map(_.value).toArray(m)
   }
   
   override def stop() {}
 
   override def numCores() = threads
-}
 
+  var nextStageId = 0
 
-/**
- * A ThreadFactory that creates daemon threads
- */
-private object DaemonThreadFactory extends ThreadFactory {
-  override def newThread(r: Runnable): Thread = {
-    val t = new Thread(r);
-    t.setDaemon(true)
-    return t
+  def newStageId() = {
+    var res = nextStageId
+    nextStageId += 1
+    res
+  }
+
+  val idToStage = new HashMap[Int, Stage]
+
+  val shuffleToMapStage = new HashMap[ShuffleDependency[_,_,_], Stage]
+
+  val cacheLocs = new HashMap[RDD[_], Array[List[String]]]
+
+  def getCacheLocs(rdd: RDD[_]): Array[List[String]] = {
+    cacheLocs.getOrElseUpdate(rdd, Array.fill[List[String]](rdd.splits.size)(Nil))
+  }
+
+  def addCacheLoc(rdd: RDD[_], partition: Int, host: String) {
+    val locs = getCacheLocs(rdd)
+    locs(partition) = host :: locs(partition)
+  }
+
+  def removeCacheLoc(rdd: RDD[_], partition: Int, host: String) {
+    val locs = getCacheLocs(rdd)
+    locs(partition) -= host
+  }
+
+  def getShuffleMapStage(shuf: ShuffleDependency[_,_,_]): Stage = {
+    shuffleToMapStage.get(shuf) match {
+      case Some(stage) => stage
+      case None =>
+        val stage = newStage(
+          true, shuf.rdd, shuf.spec.partitioner.numPartitions)
+        shuffleToMapStage(shuf) = stage
+        stage
+    }
+  }
+
+  def newStage(isShuffleMap: Boolean, rdd: RDD[_], numPartitions: Int): Stage = {
+    val id = newStageId()
+    val parents = getParentStages(rdd)
+    val stage = new Stage(id, isShuffleMap, rdd, parents, numPartitions)
+    idToStage(id) = stage
+    stage
+  }
+
+  def getParentStages(rdd: RDD[_]): List[Stage] = {
+    val parents = new HashSet[Stage]
+    val visited = new HashSet[RDD[_]]
+    def visit(r: RDD[_]) {
+      if (!visited(r)) {
+        visited += r
+        for (dep <- r.dependencies) {
+          dep match {
+            case shufDep: ShuffleDependency[_,_,_] =>
+              parents += getShuffleMapStage(shufDep)
+            case _ =>
+              visit(dep.rdd)
+          }
+        }
+      }
+    }
+    visit(rdd)
+    parents.toList
+  }
+
+  def getMissingParentStages(stage: Stage): List[Stage] = {
+    val missing = new HashSet[Stage]
+    val visited = new HashSet[RDD[_]]
+    def visit(rdd: RDD[_]) {
+      if (!visited(rdd)) {
+        visited += rdd
+        val locs = getCacheLocs(rdd)
+        for (p <- 0 until rdd.splits.size) {
+          if (locs(p) == Nil) {
+            for (dep <- rdd.dependencies) {
+              dep match {
+                case shufDep: ShuffleDependency[_,_,_] =>
+                  val stage = getShuffleMapStage(shufDep)
+                  if (!stage.isAvailable)
+                    missing += stage
+                case narrowDep: NarrowDependency[_] =>
+                  visit(narrowDep.rdd)
+              }
+            }
+          }
+        }
+      }
+    }
+    visit(stage.rdd)
+    missing.toList
+  }
+
+  override def runJob[T, U](rdd: RDD[T], func: Iterator[T] => U)(implicit m: ClassManifest[U])
+      : Array[U] = {
+    val numOutputParts: Int = rdd.splits.size
+    val finalStage = newStage(false, rdd, numOutputParts)
+    val results = new Array[U](numOutputParts)
+    val finished = new Array[Boolean](numOutputParts)
+    var numFinished = 0
+
+    val waiting = new HashSet[Stage]
+    val running = new HashSet[Stage]
+    val pendingTasks = new HashMap[Stage, HashSet[Task[_]]]
+
+    def submitStage(stage: Stage) {
+      if (!waiting(stage) && !running(stage)) {
+        val missing = getMissingParentStages(stage)
+        if (missing == Nil) {
+          logInfo("Submitting " + stage + ", which has no missing parents")
+          submitMissingTasks(stage)
+          running += stage
+        } else {
+          for (parent <- missing)
+            submitStage(parent)
+          waiting += stage
+        }
+      }
+    }
+
+    def submitMissingTasks(stage: Stage) {
+      var tasks: List[Task[_]] = Nil
+      if (stage == finalStage) {
+        for (p <- 0 until numOutputParts if (!finished(p))) {
+          val locs = getPreferredLocs(rdd, p)
+          tasks = new ResultTask(rdd, func, p, locs) :: tasks
+        }
+      }
+      submitTasks(tasks)
+    }
+
+    submitStage(finalStage)
+
+    while (numFinished != numOutputParts) {
+      val evt = completionEvents.take()
+      if (evt.successful) {
+        evt.task match {
+          case rt: ResultTask[_, _] =>
+            results(rt.partition) = evt.result.asInstanceOf[U]
+            finished(rt.partition) = true
+            numFinished += 1
+          // case smt: ShuffleMapTask
+        }
+      } else {
+        throw new SparkException("Task failed: " + evt.task)
+        // TODO: Kill the running job
+      }
+    }
+
+    return results
+  }
+
+  def getPreferredLocs(rdd: RDD[_], partition: Int): List[String] = {
+    // If the partition is cached, return the cache locations
+    val cached = getCacheLocs(rdd)(partition)
+    if (cached != Nil) {
+      return cached
+    }
+    // If the RDD has some placement preferences (as is the case for input RDDs), get those
+    val rddPrefs = rdd.preferredLocations(rdd.splits(partition)).toList
+    if (rddPrefs != Nil) {
+      return rddPrefs
+    }
+    // If the RDD has narrow dependencies, pick the first partition of the first narrow dep
+    // that has any placement preferences. Ideally we would choose based on transfer sizes,
+    // but this will do for now.
+    rdd.dependencies.foreach(_ match {
+      case n: NarrowDependency[_] =>
+        for (inPart <- n.getParents(partition)) {
+          val locs = getPreferredLocs(n.rdd, inPart)
+          if (locs != Nil)
+            return locs;
+        }
+    })
+    return Nil
   }
 }
+
+case class CompletionEvent(task: Task[_], successful: Boolean, result: Any, accumUpdates: Map[Long, Any])
+
+class ResultTask[T, U](rdd: RDD[T], func: Iterator[T] => U, val partition: Int, locs: Seq[String])
+extends Task[U] {
+  val split = rdd.splits(partition)
+
+  override def run: U = {
+    func(rdd.iterator(split))
+  }
+
+  override def preferredLocations: Seq[String] = locs
+
+  override def toString = "ResultTask " + partition
+}
+
+class Stage(val id: Int, val isShuffleMap: Boolean, val rdd: RDD[_], val parents: List[Stage], val numPartitions: Int) {
+  val outputLocs = Array.fill[List[String]](numPartitions)(Nil)
+  var numAvailableOutputs = 0
+
+  def isAvailable: Boolean = {
+    if (parents.size == 0 && !isShuffleMap)
+      true
+    else
+      numAvailableOutputs == numPartitions
+  }
+
+  def addOutputLoc(partition: Int, host: String) {
+    val prevList = outputLocs(partition)
+    outputLocs(partition) = host :: prevList
+    if (prevList == Nil)
+      numAvailableOutputs += 1
+  }
+
+  def removeOutputLoc(partition: Int, host: String) {
+    val prevList = outputLocs(partition)
+    val newList = prevList - host
+    outputLocs(partition) = newList
+    if (prevList != Nil && newList == Nil)
+      numAvailableOutputs -= 1
+  }
+
+  override def toString = "Stage " + id
+
+  override def hashCode(): Int = id
+}

core/src/main/scala/spark/MesosScheduler.scala

@@ -105,7 +105,7 @@ extends MScheduler with spark.Scheduler with Logging
    * The primary means to submit a job to the scheduler. Given a list of tasks,
    * runs them and returns an array of the results.
    */
-  override def runTasks[T: ClassManifest](tasks: Array[Task[T]]): Array[T] = {
+  def runTasks[T: ClassManifest](tasks: Array[Task[T]]): Array[T] = {
     waitForRegister()
     val jobId = newJobId()
     val myJob = new SimpleJob(this, tasks, jobId)
@@ -291,4 +291,9 @@ extends MScheduler with spark.Scheduler with Logging
     // Serialize the map as an array of (String, String) pairs
     return Utils.serialize(props.toArray)
   }
+
+  override def runJob[T, U](rdd: RDD[T], func: Iterator[T] => U)(implicit m: ClassManifest[U])
+      : Array[U] = {
+    new Array[U](0)
+  }
 }

core/src/main/scala/spark/RDD.scala

@@ -12,25 +12,51 @@ import SparkContext._
 
 import mesos._
 
+abstract class Dependency[T](val rdd: RDD[T], val isShuffle: Boolean)
+
+abstract class NarrowDependency[T](rdd: RDD[T])
+extends Dependency(rdd, false) {
+  def getParents(outputPartition: Int): Seq[Int]
+}
+
+class ShuffleDependency[K, V, C](
+  rdd: RDD[(K, V)],
+  val spec: ShuffleSpec[K, V, C]
+) extends Dependency(rdd, true)
+
+class ShuffleSpec[K, V, C] (
+  val createCombiner: V => C,
+  val mergeValue: (C, V) => C,
+  val mergeCombiners: (C, C) => C,
+  val partitioner: Partitioner[K]
+)
+
+abstract class Partitioner[K] {
+  def numPartitions: Int
+  def getPartition(key: K): Int
+}
 
 @serializable
 abstract class RDD[T: ClassManifest](@transient sc: SparkContext) {
   def splits: Array[Split]
   def iterator(split: Split): Iterator[T]
   def preferredLocations(split: Split): Seq[String]
+  
+  def dependencies: List[Dependency[_]] = Nil
+  def partitioner: Option[Partitioner[_]] = None
 
   def taskStarted(split: Split, slot: SlaveOffer) {}
 
   def sparkContext = sc
 
-  def map[U: ClassManifest](f: T => U) = new MappedRDD(this, sc.clean(f))
-  def filter(f: T => Boolean) = new FilteredRDD(this, sc.clean(f))
+  def map[U: ClassManifest](f: T => U): RDD[U] = new MappedRDD(this, sc.clean(f))
+  def filter(f: T => Boolean): RDD[T] = new FilteredRDD(this, sc.clean(f))
   def cache() = new CachedRDD(this)
 
-  def sample(withReplacement: Boolean, frac: Double, seed: Int) =
+  def sample(withReplacement: Boolean, frac: Double, seed: Int): RDD[T] =
     new SampledRDD(this, withReplacement, frac, seed)
 
-  def flatMap[U: ClassManifest](f: T => Traversable[U]) =
+  def flatMap[U: ClassManifest](f: T => Traversable[U]): RDD[U] =
     new FlatMappedRDD(this, sc.clean(f))
 
   def foreach(f: T => Unit) {
@@ -40,8 +66,7 @@ abstract class RDD[T: ClassManifest](@transient sc: SparkContext) {
   }
 
   def collect(): Array[T] = {
-    val tasks = splits.map(s => new CollectTask(this, s))
-    val results = sc.runTaskObjects(tasks)
+    val results = sc.scheduler.runJob(this, (iter: Iterator[T]) => iter.toArray)
     Array.concat(results: _*)
   }
 
@@ -49,9 +74,15 @@ abstract class RDD[T: ClassManifest](@transient sc: SparkContext) {
 
   def reduce(f: (T, T) => T): T = {
     val cleanF = sc.clean(f)
-    val tasks = splits.map(s => new ReduceTask(this, s, f))
+    val reducePartition: Iterator[T] => Option[T] = iter => {
+      if (iter.hasNext)
+        Some(iter.reduceLeft(f))
+      else
+        None
+    }
+    val options = sc.scheduler.runJob(this, reducePartition)
     val results = new ArrayBuffer[T]
-    for (option <- sc.runTaskObjects(tasks); elem <- option)
+    for (opt <- options; elem <- opt)
       results += elem
     if (results.size == 0)
       throw new UnsupportedOperationException("empty collection")
@@ -77,20 +108,20 @@ abstract class RDD[T: ClassManifest](@transient sc: SparkContext) {
   }
 
   def count(): Long = {
-    try { 
+    try {
       map(x => 1L).reduce(_+_) 
     } catch { 
       case e: UnsupportedOperationException => 0L // No elements in RDD
     }
   }
 
-  def union(other: RDD[T]) = new UnionRDD(sc, Array(this, other))
+  def union(other: RDD[T]): RDD[T] = new UnionRDD(sc, Array(this, other))
 
-  def ++(other: RDD[T]) = this.union(other)
+  def ++(other: RDD[T]): RDD[T] = this.union(other)
 
-  def splitRdd() = new SplitRDD(this)
+  def splitRdd(): RDD[Array[T]] = new SplitRDD(this)
 
-  def cartesian[U: ClassManifest](other: RDD[U]) =
+  def cartesian[U: ClassManifest](other: RDD[U]): RDD[(T, U)] =
     new CartesianRDD(sc, this, other)
 
   def groupBy[K](func: T => K, numSplits: Int): RDD[(K, Seq[T])] =

core/src/main/scala/spark/Scheduler.scala

@@ -4,7 +4,8 @@ package spark
 private trait Scheduler {
   def start()
   def waitForRegister()
-  def runTasks[T](tasks: Array[Task[T]])(implicit m: ClassManifest[T]): Array[T]
+  //def runTasks[T](tasks: Array[Task[T]])(implicit m: ClassManifest[T]): Array[T]
+  def runJob[T, U](rdd: RDD[T], func: Iterator[T] => U)(implicit m: ClassManifest[U]): Array[U]
   def stop()
   def numCores(): Int
 }

core/src/main/scala/spark/SparkContext.scala

@@ -14,7 +14,7 @@ class SparkContext(
   val sparkHome: String = null,
   val jars: Seq[String] = Nil)
 extends Logging {
-  private var scheduler: Scheduler = {
+  private[spark] var scheduler: Scheduler = {
     // Regular expression used for local[N] master format
     val LOCAL_N_REGEX = """local\[([0-9]+)\]""".r
     master match {
@@ -126,19 +126,17 @@ extends Logging {
       None
   }
 
-  // Submit an array of tasks (passed as functions) to the scheduler
-  def runTasks[T: ClassManifest](tasks: Array[() => T]): Array[T] = {
-    runTaskObjects(tasks.map(f => new FunctionTask(f)))
-  }
-
   // Run an array of spark.Task objects
   private[spark] def runTaskObjects[T: ClassManifest](tasks: Seq[Task[T]])
       : Array[T] = {
+    return null;
+    /*
     logInfo("Running " + tasks.length + " tasks in parallel")
     val start = System.nanoTime
     val result = scheduler.runTasks(tasks.toArray)
     logInfo("Tasks finished in " + (System.nanoTime - start) / 1e9 + " s")
     return result
+    */
   }
 
   // Clean a closure to make it ready to serialized and send to tasks

examples/src/main/scala/spark/examples/CpuHog.scala

@@ -1,26 +0,0 @@
-package spark.examples
-
-import spark._
-
-object CpuHog {
-  def main(args: Array[String]) {
-    if (args.length != 3) {
-      System.err.println("Usage: CpuHog <master> <tasks> <threads_per_task>");
-      System.exit(1)
-    }
-    val sc = new SparkContext(args(0), "CPU hog")
-    val tasks = args(1).toInt
-    val threads = args(2).toInt
-    def task {
-      for (i <- 0 until threads-1) {
-        new Thread() {
-          override def run {
-            while(true) {}
-          }
-        }.start()
-      }
-      while(true) {}
-    }
-    sc.runTasks(Array.make(tasks, () => task))
-  }
-}

examples/src/main/scala/spark/examples/SleepJob.scala

@@ -1,21 +0,0 @@
-package spark.examples
-
-import spark._
-
-object SleepJob {
-  def main(args: Array[String]) {
-    if (args.length != 3) {
-      System.err.println("Usage: SleepJob <master> <tasks> <task_duration>");
-      System.exit(1)
-    }
-    val sc = new SparkContext(args(0), "Sleep job")
-    val tasks = args(1).toInt
-    val duration = args(2).toInt
-    def task {
-      val start = System.currentTimeMillis
-      while (System.currentTimeMillis - start < duration * 1000L)
-        Thread.sleep(200)
-    }
-    sc.runTasks(Array.make(tasks, () => task))
-  }
-}