Merge pull request #48 from ankurdave/bagel-new

Bagel: Large-scale graph processing on Spark

bagel/src/main/scala/spark/bagel/Bagel.scala

@@ -0,0 +1,159 @@
+package spark.bagel
+
+import spark._
+import spark.SparkContext._
+
+import scala.collection.mutable.ArrayBuffer
+
+object Bagel extends Logging {
+  def run[V <: Vertex : Manifest, M <: Message : Manifest, C : Manifest, A : Manifest](
+    sc: SparkContext,
+    verts: RDD[(String, V)],
+    msgs: RDD[(String, M)]
+  )(
+    combiner: Combiner[M, C] = new DefaultCombiner[M],
+    aggregator: Aggregator[V, A] = new NullAggregator[V],
+    superstep: Int = 0,
+    numSplits: Int = sc.numCores
+  )(
+    compute: (V, Option[C], A, Int) => (V, Iterable[M])
+  ): RDD[V] = {
+
+    logInfo("Starting superstep "+superstep+".")
+    val startTime = System.currentTimeMillis
+
+    val aggregated = agg(verts, aggregator)
+    val combinedMsgs = msgs.combineByKey(combiner.createCombiner, combiner.mergeMsg, combiner.mergeCombiners, numSplits)
+    val grouped = verts.groupWith(combinedMsgs)
+    val (processed, numMsgs, numActiveVerts) = comp[V, M, C](sc, grouped, compute(_, _, aggregated, superstep))
+
+    val timeTaken = System.currentTimeMillis - startTime
+    logInfo("Superstep %d took %d s".format(superstep, timeTaken / 1000))
+
+    // Check stopping condition and iterate
+    val noActivity = numMsgs == 0 && numActiveVerts == 0
+    if (noActivity) {
+      processed.map { case (id, (vert, msgs)) => vert }
+    } else {
+      val newVerts = processed.mapValues { case (vert, msgs) => vert }
+      val newMsgs = processed.flatMap {
+        case (id, (vert, msgs)) => msgs.map(m => (m.targetId, m))
+      }
+      run(sc, newVerts, newMsgs)(combiner, aggregator, superstep + 1, numSplits)(compute)
+    }
+  }
+
+  /**
+   * Aggregates the given vertices using the given aggregator, or does
+   * nothing if it is a NullAggregator.
+   */
+  def agg[V <: Vertex, A : Manifest](verts: RDD[(String, V)], aggregator: Aggregator[V, A]): A = aggregator match {
+    case _: NullAggregator[_] =>
+      None
+    case _ =>
+      verts.map {
+        case (id, vert) => aggregator.createAggregator(vert)
+      }.reduce(aggregator.mergeAggregators(_, _))
+  }
+
+  /**
+   * Processes the given vertex-message RDD using the compute
+   * function. Returns the processed RDD, the number of messages
+   * created, and the number of active vertices.
+   */
+  def comp[V <: Vertex, M <: Message, C](sc: SparkContext, grouped: RDD[(String, (Seq[V], Seq[C]))], compute: (V, Option[C]) => (V, Iterable[M])): (RDD[(String, (V, Iterable[M]))], Int, Int) = {
+    var numMsgs = sc.accumulator(0)
+    var numActiveVerts = sc.accumulator(0)
+    val processed = grouped.flatMapValues {
+      case (Seq(), _) => None
+      case (Seq(v), c) =>
+          val (newVert, newMsgs) =
+            compute(v, c match {
+              case Seq(comb) => Some(comb)
+              case Seq() => None
+            })
+
+          numMsgs += newMsgs.size
+          if (newVert.active)
+            numActiveVerts += 1
+
+          Some((newVert, newMsgs))
+    }.cache
+
+    // Force evaluation of processed RDD for accurate performance measurements
+    processed.foreach(x => {})
+
+    (processed, numMsgs.value, numActiveVerts.value)
+  }
+
+  /**
+   * Converts a compute function that doesn't take an aggregator to
+   * one that does, so it can be passed to Bagel.run.
+   */
+  implicit def addAggregatorArg[
+    V <: Vertex : Manifest, M <: Message : Manifest, C
+  ](
+    compute: (V, Option[C], Int) => (V, Iterable[M])
+  ): (V, Option[C], Option[Nothing], Int) => (V, Iterable[M]) = {
+    (vert: V, messages: Option[C], aggregator: Option[Nothing], superstep: Int) => compute(vert, messages, superstep)
+  }
+}
+
+// TODO: Simplify Combiner interface and make it more OO.
+trait Combiner[M, C] {
+  def createCombiner(msg: M): C
+  def mergeMsg(combiner: C, msg: M): C
+  def mergeCombiners(a: C, b: C): C
+}
+
+trait Aggregator[V, A] {
+  def createAggregator(vert: V): A
+  def mergeAggregators(a: A, b: A): A
+}
+
+@serializable
+class DefaultCombiner[M] extends Combiner[M, ArrayBuffer[M]] {
+  def createCombiner(msg: M): ArrayBuffer[M] =
+    ArrayBuffer(msg)
+  def mergeMsg(combiner: ArrayBuffer[M], msg: M): ArrayBuffer[M] =
+    combiner += msg
+  def mergeCombiners(a: ArrayBuffer[M], b: ArrayBuffer[M]): ArrayBuffer[M] =
+    a ++= b
+}
+
+@serializable
+class NullAggregator[V] extends Aggregator[V, Option[Nothing]] {
+  def createAggregator(vert: V): Option[Nothing] = None
+  def mergeAggregators(a: Option[Nothing], b: Option[Nothing]): Option[Nothing] = None
+}
+
+/**
+ * Represents a Bagel vertex.
+ *
+ * Subclasses may store state along with each vertex and must be
+ * annotated with @serializable.
+ */
+trait Vertex {
+  def id: String
+  def active: Boolean
+}
+
+/**
+ * Represents a Bagel message to a target vertex.
+ *
+ * Subclasses may contain a payload to deliver to the target vertex
+ * and must be annotated with @serializable.
+ */
+trait Message {
+  def targetId: String
+}
+
+/**
+ * Represents a directed edge between two vertices.
+ *
+ * Subclasses may store state along each edge and must be annotated
+ * with @serializable.
+ */
+trait Edge {
+  def targetId: String
+}

bagel/src/main/scala/spark/bagel/examples/ShortestPath.scala

@@ -0,0 +1,96 @@
+package spark.bagel.examples
+
+import spark._
+import spark.SparkContext._
+
+import scala.math.min
+
+import spark.bagel._
+import spark.bagel.Bagel._
+
+object ShortestPath {
+  def main(args: Array[String]) {
+    if (args.length < 4) {
+      System.err.println("Usage: ShortestPath <graphFile> <startVertex> " +
+                         "<numSplits> <host>")
+      System.exit(-1)
+    }
+
+    val graphFile = args(0)
+    val startVertex = args(1)
+    val numSplits = args(2).toInt
+    val host = args(3)
+    val sc = new SparkContext(host, "ShortestPath")
+
+    // Parse the graph data from a file into two RDDs, vertices and messages
+    val lines =
+      (sc.textFile(graphFile)
+       .filter(!_.matches("^\\s*#.*"))
+       .map(line => line.split("\t")))
+
+    val vertices: RDD[(String, SPVertex)] =
+      (lines.groupBy(line => line(0))
+       .map {
+         case (vertexId, lines) => {
+           val outEdges = lines.collect {
+             case Array(_, targetId, edgeValue) =>
+               new SPEdge(targetId, edgeValue.toInt)
+           }
+           
+           (vertexId, new SPVertex(vertexId, Int.MaxValue, outEdges, true))
+         }
+       })
+
+    val messages: RDD[(String, SPMessage)] =
+      (lines.filter(_.length == 2)
+       .map {
+         case Array(vertexId, messageValue) =>
+           (vertexId, new SPMessage(vertexId, messageValue.toInt))
+       })
+    
+    System.err.println("Read "+vertices.count()+" vertices and "+
+                       messages.count()+" messages.")
+
+    // Do the computation
+    val compute = addAggregatorArg {
+      (self: SPVertex, messageMinValue: Option[Int], superstep: Int) =>
+        val newValue = messageMinValue match {
+          case Some(minVal) => min(self.value, minVal)
+          case None => self.value
+        }
+
+        val outbox =
+          if (newValue != self.value)
+            self.outEdges.map(edge =>
+              new SPMessage(edge.targetId, newValue + edge.value))
+          else
+            List()
+
+        (new SPVertex(self.id, newValue, self.outEdges, false), outbox)
+    }
+    val result = Bagel.run(sc, vertices, messages)(combiner = MinCombiner, numSplits = numSplits)(compute)
+
+    // Print the result
+    System.err.println("Shortest path from "+startVertex+" to all vertices:")
+    val shortest = result.map(vertex =>
+      "%s\t%s\n".format(vertex.id, vertex.value match {
+        case x if x == Int.MaxValue => "inf"
+        case x => x
+      })).collect.mkString
+    println(shortest)
+  }
+}
+
+@serializable
+object MinCombiner extends Combiner[SPMessage, Int] {
+  def createCombiner(msg: SPMessage): Int =
+    msg.value
+  def mergeMsg(combiner: Int, msg: SPMessage): Int =
+    min(combiner, msg.value)
+  def mergeCombiners(a: Int, b: Int): Int =
+    min(a, b)
+}
+
+@serializable class SPVertex(val id: String, val value: Int, val outEdges: Seq[SPEdge], val active: Boolean) extends Vertex
+@serializable class SPEdge(val targetId: String, val value: Int) extends Edge
+@serializable class SPMessage(val targetId: String, val value: Int) extends Message

bagel/src/main/scala/spark/bagel/examples/WikipediaPageRank.scala

@@ -0,0 +1,158 @@
+package spark.bagel.examples
+
+import spark._
+import spark.SparkContext._
+
+import spark.bagel._
+import spark.bagel.Bagel._
+
+import scala.collection.mutable.ArrayBuffer
+import scala.xml.{XML,NodeSeq}
+
+import java.io.{Externalizable,ObjectInput,ObjectOutput,DataOutputStream,DataInputStream}
+
+import com.esotericsoftware.kryo._
+
+object WikipediaPageRank {
+  def main(args: Array[String]) {
+    if (args.length < 4) {
+      System.err.println("Usage: WikipediaPageRank <inputFile> <threshold> <numSplits> <host> [<noCombiner>]")
+      System.exit(-1)
+    }
+
+    System.setProperty("spark.serialization", "spark.KryoSerialization")
+    System.setProperty("spark.kryo.registrator", classOf[PRKryoRegistrator].getName)
+
+    val inputFile = args(0)
+    val threshold = args(1).toDouble
+    val numSplits = args(2).toInt
+    val host = args(3)
+    val noCombiner = args.length > 4 && args(4).nonEmpty
+    val sc = new SparkContext(host, "WikipediaPageRank")
+
+    // Parse the Wikipedia page data into a graph
+    val input = sc.textFile(inputFile)
+
+    println("Counting vertices...")
+    val numVertices = input.count()
+    println("Done counting vertices.")
+
+    println("Parsing input file...")
+    val vertices: RDD[(String, PRVertex)] = input.map(line => {
+      val fields = line.split("\t")
+      val (title, body) = (fields(1), fields(3).replace("\\n", "\n"))
+      val links =
+        if (body == "\\N")
+          NodeSeq.Empty
+        else
+          try {
+            XML.loadString(body) \\ "link" \ "target"
+          } catch {
+            case e: org.xml.sax.SAXParseException =>
+              System.err.println("Article \""+title+"\" has malformed XML in body:\n"+body)
+            NodeSeq.Empty
+          }
+      val outEdges = ArrayBuffer(links.map(link => new PREdge(new String(link.text))): _*)
+      val id = new String(title)
+      (id, new PRVertex(id, 1.0 / numVertices, outEdges, true))
+    }).cache
+    println("Done parsing input file.")
+
+    // Do the computation
+    val epsilon = 0.01 / numVertices
+    val messages = sc.parallelize(List[(String, PRMessage)]())
+    val result =
+      if (noCombiner) {
+        Bagel.run(sc, vertices, messages)(numSplits = numSplits)(PRNoCombiner.compute(numVertices, epsilon))
+      } else {
+        Bagel.run(sc, vertices, messages)(combiner = PRCombiner, numSplits = numSplits)(PRCombiner.compute(numVertices, epsilon))
+      }
+
+    // Print the result
+    System.err.println("Articles with PageRank >= "+threshold+":")
+    val top = result.filter(_.value >= threshold).map(vertex =>
+      "%s\t%s\n".format(vertex.id, vertex.value)).collect.mkString
+    println(top)
+  }
+}
+
+@serializable
+object PRCombiner extends Combiner[PRMessage, Double] {
+  def createCombiner(msg: PRMessage): Double =
+    msg.value
+  def mergeMsg(combiner: Double, msg: PRMessage): Double =
+    combiner + msg.value
+  def mergeCombiners(a: Double, b: Double): Double =
+    a + b
+
+  def compute(numVertices: Long, epsilon: Double)(self: PRVertex, messageSum: Option[Double], superstep: Int): (PRVertex, Iterable[PRMessage]) = {
+    val newValue = messageSum match {
+      case Some(msgSum) if msgSum != 0 =>
+        0.15 / numVertices + 0.85 * msgSum
+      case _ => self.value
+    }
+
+    val terminate = (superstep >= 10 && (newValue - self.value).abs < epsilon) || superstep >= 30
+
+    val outbox =
+      if (!terminate)
+        self.outEdges.map(edge =>
+          new PRMessage(edge.targetId, newValue / self.outEdges.size))
+      else
+        ArrayBuffer[PRMessage]()
+
+    (new PRVertex(self.id, newValue, self.outEdges, !terminate), outbox)
+  }
+}
+
+@serializable
+object PRNoCombiner extends DefaultCombiner[PRMessage] {
+  def compute(numVertices: Long, epsilon: Double)(self: PRVertex, messages: Option[ArrayBuffer[PRMessage]], superstep: Int): (PRVertex, Iterable[PRMessage]) =
+    PRCombiner.compute(numVertices, epsilon)(self, messages match {
+      case Some(msgs) => Some(msgs.map(_.value).sum)
+      case None => None
+    }, superstep)
+}
+
+@serializable class PRVertex() extends Vertex {
+  var id: String = _
+  var value: Double = _
+  var outEdges: ArrayBuffer[PREdge] = _
+  var active: Boolean = true
+
+  def this(id: String, value: Double, outEdges: ArrayBuffer[PREdge], active: Boolean) {
+    this()
+    this.id = id
+    this.value = value
+    this.outEdges = outEdges
+    this.active = active
+  }
+}
+
+@serializable class PRMessage() extends Message {
+  var targetId: String = _
+  var value: Double = _
+
+  def this(targetId: String, value: Double) {
+    this()
+    this.targetId = targetId
+    this.value = value
+  }
+}
+
+@serializable class PREdge() extends Edge {
+  var targetId: String = _
+
+  def this(targetId: String) {
+    this()
+    this.targetId = targetId
+   }
+}
+
+class PRKryoRegistrator extends KryoRegistrator {
+  def registerClasses(kryo: Kryo) {
+    kryo.register(classOf[PRVertex])
+    kryo.register(classOf[PRMessage])
+    kryo.register(classOf[PREdge])
+  }
+}

bagel/src/test/scala/bagel/BagelSuite.scala

@@ -0,0 +1,53 @@
+package spark.bagel
+
+import org.scalatest.{FunSuite, Assertions}
+import org.scalatest.prop.Checkers
+import org.scalacheck.Arbitrary._
+import org.scalacheck.Gen
+import org.scalacheck.Prop._
+
+import scala.collection.mutable.ArrayBuffer
+
+import spark._
+
+import spark.bagel.Bagel._
+
+@serializable class TestVertex(val id: String, val active: Boolean, val age: Int) extends Vertex
+@serializable class TestMessage(val targetId: String) extends Message
+
+class BagelSuite extends FunSuite with Assertions {
+  test("halting by voting") {
+    val sc = new SparkContext("local", "test")
+    val verts = sc.parallelize(Array("a", "b", "c", "d").map(id => (id, new TestVertex(id, true, 0))))
+    val msgs = sc.parallelize(Array[(String, TestMessage)]())
+    val numSupersteps = 5
+    val result =
+      Bagel.run(sc, verts, msgs)()(addAggregatorArg {
+      (self: TestVertex, msgs: Option[ArrayBuffer[TestMessage]], superstep: Int) =>
+        (new TestVertex(self.id, superstep < numSupersteps - 1, self.age + 1), Array[TestMessage]())
+    })
+    for (vert <- result.collect)
+      assert(vert.age === numSupersteps)
+  }
+
+  test("halting by message silence") {
+    val sc = new SparkContext("local", "test")
+    val verts = sc.parallelize(Array("a", "b", "c", "d").map(id => (id, new TestVertex(id, false, 0))))
+    val msgs = sc.parallelize(Array("a" -> new TestMessage("a")))
+    val numSupersteps = 5
+    val result =
+      Bagel.run(sc, verts, msgs)()(addAggregatorArg {
+      (self: TestVertex, msgs: Option[ArrayBuffer[TestMessage]], superstep: Int) =>
+        val msgsOut =
+          msgs match {
+            case Some(ms) if (superstep < numSupersteps - 1) =>
+              ms
+            case _ =>
+              new ArrayBuffer[TestMessage]()
+          }
+        (new TestVertex(self.id, self.active, self.age + 1), msgsOut)
+    })
+    for (vert <- result.collect)
+      assert(vert.age === numSupersteps)
+  }
+}

project/build/SparkProject.scala

@@ -14,6 +14,8 @@ extends ParentProject(info) with IdeaProject
   lazy val examples =
     project("examples", "Spark Examples", new ExamplesProject(_), core)
 
+  lazy val bagel = project("bagel", "Bagel", new BagelProject(_), core)
+
   class CoreProject(info: ProjectInfo)
   extends DefaultProject(info) with Eclipsify with IdeaProject with DepJar with XmlTestReport
   {}
@@ -21,6 +23,10 @@ extends ParentProject(info) with IdeaProject
   class ExamplesProject(info: ProjectInfo)
   extends DefaultProject(info) with Eclipsify with IdeaProject
   {}
+
+  class BagelProject(info: ProjectInfo)
+  extends DefaultProject(info) with DepJar with XmlTestReport
+  {}
 }
 
 

run

@@ -35,6 +35,7 @@ export JAVA_OPTS
 
 CORE_DIR=$FWDIR/core
 EXAMPLES_DIR=$FWDIR/examples
+BAGEL_DIR=$FWDIR/bagel
 
 # Build up classpath
 CLASSPATH="$SPARK_CLASSPATH:$CORE_DIR/target/scala_2.8.1/classes:$MESOS_CLASSPATH"
@@ -60,6 +61,7 @@ CLASSPATH+=:$EXAMPLES_DIR/target/scala_2.8.1/classes
 for jar in $CORE_DIR/lib/hadoop-0.20.2/lib/*.jar; do
   CLASSPATH+=:$jar
 done
+CLASSPATH+=:$BAGEL_DIR/target/scala_2.8.1/classes
 export CLASSPATH # Needed for spark-shell
 
 if [ -n "$SCALA_HOME" ]; then