From 397d3816e1ee0351cd2814dc081f368fe45094c0 Mon Sep 17 00:00:00 2001 From: Reynold Xin Date: Wed, 19 Sep 2012 12:31:45 -0700 Subject: [PATCH] Separated ShuffledRDD into multiple classes: RepartitionShuffledRDD, ShuffledSortedRDD, and ShuffledAggregatedRDD. --- core/src/main/scala/spark/Aggregator.scala | 4 +- .../main/scala/spark/PairRDDFunctions.scala | 87 ++++++++++--------- core/src/main/scala/spark/ShuffledRDD.scala | 72 +++++++++++++-- .../spark/scheduler/ShuffleMapTask.scala | 9 +- core/src/test/scala/spark/ShuffleSuite.scala | 18 ++-- 5 files changed, 128 insertions(+), 62 deletions(-) diff --git a/core/src/main/scala/spark/Aggregator.scala b/core/src/main/scala/spark/Aggregator.scala index 6516bea157..b0daa70cfd 100644 --- a/core/src/main/scala/spark/Aggregator.scala +++ b/core/src/main/scala/spark/Aggregator.scala @@ -9,9 +9,9 @@ package spark * known as map-side aggregations. When set to false, * mergeCombiners function is not used. */ -class Aggregator[K, V, C] ( +case class Aggregator[K, V, C] ( val createCombiner: V => C, val mergeValue: (C, V) => C, val mergeCombiners: (C, C) => C, val mapSideCombine: Boolean = true) - extends Serializable + diff --git a/core/src/main/scala/spark/PairRDDFunctions.scala b/core/src/main/scala/spark/PairRDDFunctions.scala index 64018f8c6b..aa1d00c63c 100644 --- a/core/src/main/scala/spark/PairRDDFunctions.scala +++ b/core/src/main/scala/spark/PairRDDFunctions.scala @@ -1,11 +1,10 @@ package spark import java.io.EOFException -import java.net.URL import java.io.ObjectInputStream +import java.net.URL +import java.util.{Date, HashMap => JHashMap} import java.util.concurrent.atomic.AtomicLong -import java.util.{HashMap => JHashMap} -import java.util.Date import java.text.SimpleDateFormat import scala.collection.Map @@ -50,9 +49,18 @@ class PairRDDFunctions[K: ClassManifest, V: ClassManifest]( def combineByKey[C](createCombiner: V => C, mergeValue: (C, V) => C, mergeCombiners: (C, C) => C, - partitioner: Partitioner): RDD[(K, C)] = { - val aggregator = new Aggregator[K, V, C](createCombiner, mergeValue, mergeCombiners) - new ShuffledRDD(self, aggregator, partitioner) + partitioner: Partitioner, + mapSideCombine: Boolean = true): RDD[(K, C)] = { + val aggregator = + if (mapSideCombine) { + new Aggregator[K, V, C](createCombiner, mergeValue, mergeCombiners) + } else { + // Don't apply map-side combiner. + // A sanity check to make sure mergeCombiners is not defined. + assert(mergeCombiners == null) + new Aggregator[K, V, C](createCombiner, mergeValue, null, false) + } + new ShuffledAggregatedRDD(self, aggregator, partitioner) } def combineByKey[C](createCombiner: V => C, @@ -65,7 +73,7 @@ class PairRDDFunctions[K: ClassManifest, V: ClassManifest]( def reduceByKey(partitioner: Partitioner, func: (V, V) => V): RDD[(K, V)] = { combineByKey[V]((v: V) => v, func, func, partitioner) } - + def reduceByKeyLocally(func: (V, V) => V): Map[K, V] = { def reducePartition(iter: Iterator[(K, V)]): Iterator[JHashMap[K, V]] = { val map = new JHashMap[K, V] @@ -116,13 +124,24 @@ class PairRDDFunctions[K: ClassManifest, V: ClassManifest]( groupByKey(new HashPartitioner(numSplits)) } - def partitionBy(partitioner: Partitioner): RDD[(K, V)] = { - def createCombiner(v: V) = ArrayBuffer(v) - def mergeValue(buf: ArrayBuffer[V], v: V) = buf += v - def mergeCombiners(b1: ArrayBuffer[V], b2: ArrayBuffer[V]) = b1 ++= b2 - val bufs = combineByKey[ArrayBuffer[V]]( - createCombiner _, mergeValue _, mergeCombiners _, partitioner) - bufs.flatMapValues(buf => buf) + /** + * Repartition the RDD using the specified partitioner. If mapSideCombine is + * true, Spark will group values of the same key together on the map side + * before the repartitioning. If a large number of duplicated keys are + * expected, and the size of the keys are large, mapSideCombine should be set + * to true. + */ + def partitionBy(partitioner: Partitioner, mapSideCombine: Boolean = false): RDD[(K, V)] = { + if (mapSideCombine) { + def createCombiner(v: V) = ArrayBuffer(v) + def mergeValue(buf: ArrayBuffer[V], v: V) = buf += v + def mergeCombiners(b1: ArrayBuffer[V], b2: ArrayBuffer[V]) = b1 ++= b2 + val bufs = combineByKey[ArrayBuffer[V]]( + createCombiner _, mergeValue _, mergeCombiners _, partitioner) + bufs.flatMapValues(buf => buf) + } else { + new RepartitionShuffledRDD(self, partitioner) + } } def join[W](other: RDD[(K, W)], partitioner: Partitioner): RDD[(K, (V, W))] = { @@ -194,17 +213,17 @@ class PairRDDFunctions[K: ClassManifest, V: ClassManifest]( } def collectAsMap(): Map[K, V] = HashMap(self.collect(): _*) - + def mapValues[U](f: V => U): RDD[(K, U)] = { val cleanF = self.context.clean(f) new MappedValuesRDD(self, cleanF) } - + def flatMapValues[U](f: V => TraversableOnce[U]): RDD[(K, U)] = { val cleanF = self.context.clean(f) new FlatMappedValuesRDD(self, cleanF) } - + def cogroup[W](other: RDD[(K, W)], partitioner: Partitioner): RDD[(K, (Seq[V], Seq[W]))] = { val cg = new CoGroupedRDD[K]( Seq(self.asInstanceOf[RDD[(_, _)]], other.asInstanceOf[RDD[(_, _)]]), @@ -215,12 +234,12 @@ class PairRDDFunctions[K: ClassManifest, V: ClassManifest]( (vs.asInstanceOf[Seq[V]], ws.asInstanceOf[Seq[W]]) } } - + def cogroup[W1, W2](other1: RDD[(K, W1)], other2: RDD[(K, W2)], partitioner: Partitioner) : RDD[(K, (Seq[V], Seq[W1], Seq[W2]))] = { val cg = new CoGroupedRDD[K]( Seq(self.asInstanceOf[RDD[(_, _)]], - other1.asInstanceOf[RDD[(_, _)]], + other1.asInstanceOf[RDD[(_, _)]], other2.asInstanceOf[RDD[(_, _)]]), partitioner) val prfs = new PairRDDFunctions[K, Seq[Seq[_]]](cg)(classManifest[K], Manifests.seqSeqManifest) @@ -289,7 +308,7 @@ class PairRDDFunctions[K: ClassManifest, V: ClassManifest]( def saveAsHadoopFile[F <: OutputFormat[K, V]](path: String)(implicit fm: ClassManifest[F]) { saveAsHadoopFile(path, getKeyClass, getValueClass, fm.erasure.asInstanceOf[Class[F]]) } - + def saveAsNewAPIHadoopFile[F <: NewOutputFormat[K, V]](path: String)(implicit fm: ClassManifest[F]) { saveAsNewAPIHadoopFile(path, getKeyClass, getValueClass, fm.erasure.asInstanceOf[Class[F]]) } @@ -363,7 +382,7 @@ class PairRDDFunctions[K: ClassManifest, V: ClassManifest]( FileOutputFormat.setOutputPath(conf, HadoopWriter.createPathFromString(path, conf)) saveAsHadoopDataset(conf) } - + def saveAsHadoopDataset(conf: JobConf) { val outputFormatClass = conf.getOutputFormat val keyClass = conf.getOutputKeyClass @@ -377,7 +396,7 @@ class PairRDDFunctions[K: ClassManifest, V: ClassManifest]( if (valueClass == null) { throw new SparkException("Output value class not set") } - + logInfo("Saving as hadoop file of type (" + keyClass.getSimpleName+ ", " + valueClass.getSimpleName+ ")") val writer = new HadoopWriter(conf) @@ -390,14 +409,14 @@ class PairRDDFunctions[K: ClassManifest, V: ClassManifest]( writer.setup(context.stageId, context.splitId, attemptNumber) writer.open() - + var count = 0 while(iter.hasNext) { val record = iter.next count += 1 writer.write(record._1.asInstanceOf[AnyRef], record._2.asInstanceOf[AnyRef]) } - + writer.close() writer.commit() } @@ -413,28 +432,14 @@ class PairRDDFunctions[K: ClassManifest, V: ClassManifest]( class OrderedRDDFunctions[K <% Ordered[K]: ClassManifest, V: ClassManifest]( self: RDD[(K, V)]) - extends Logging + extends Logging with Serializable { def sortByKey(ascending: Boolean = true): RDD[(K,V)] = { - val rangePartitionedRDD = self.partitionBy(new RangePartitioner(self.splits.size, self, ascending)) - new SortedRDD(rangePartitionedRDD, ascending) + new ShuffledSortedRDD(self, ascending) } } -class SortedRDD[K <% Ordered[K], V](prev: RDD[(K, V)], ascending: Boolean) - extends RDD[(K, V)](prev.context) { - - override def splits = prev.splits - override val partitioner = prev.partitioner - override val dependencies = List(new OneToOneDependency(prev)) - - override def compute(split: Split) = { - prev.iterator(split).toArray - .sortWith((x, y) => if (ascending) x._1 < y._1 else x._1 > y._1).iterator - } -} - class MappedValuesRDD[K, V, U](prev: RDD[(K, V)], f: V => U) extends RDD[(K, U)](prev.context) { override def splits = prev.splits override val dependencies = List(new OneToOneDependency(prev)) @@ -444,7 +449,7 @@ class MappedValuesRDD[K, V, U](prev: RDD[(K, V)], f: V => U) extends RDD[(K, U)] class FlatMappedValuesRDD[K, V, U](prev: RDD[(K, V)], f: V => TraversableOnce[U]) extends RDD[(K, U)](prev.context) { - + override def splits = prev.splits override val dependencies = List(new OneToOneDependency(prev)) override val partitioner = prev.partitioner diff --git a/core/src/main/scala/spark/ShuffledRDD.scala b/core/src/main/scala/spark/ShuffledRDD.scala index 3616d8e47e..a7346060b3 100644 --- a/core/src/main/scala/spark/ShuffledRDD.scala +++ b/core/src/main/scala/spark/ShuffledRDD.scala @@ -1,29 +1,89 @@ package spark +import scala.collection.mutable.ArrayBuffer import java.util.{HashMap => JHashMap} + class ShuffledRDDSplit(val idx: Int) extends Split { override val index = idx override def hashCode(): Int = idx } -class ShuffledRDD[K, V, C]( + +/** + * The resulting RDD from a shuffle (e.g. repartitioning of data). + */ +abstract class ShuffledRDD[K, V, C]( @transient parent: RDD[(K, V)], aggregator: Aggregator[K, V, C], - part : Partitioner) + part : Partitioner) extends RDD[(K, C)](parent.context) { - //override val partitioner = Some(part) + override val partitioner = Some(part) - + @transient val splits_ = Array.tabulate[Split](part.numPartitions)(i => new ShuffledRDDSplit(i)) override def splits = splits_ - + override def preferredLocations(split: Split) = Nil - + val dep = new ShuffleDependency(context.newShuffleId, parent, aggregator, part) override val dependencies = List(dep) +} + + +/** + * Repartition a key-value pair RDD. + */ +class RepartitionShuffledRDD[K, V]( + @transient parent: RDD[(K, V)], + part : Partitioner) + extends ShuffledRDD[K, V, V]( + parent, + Aggregator[K, V, V](null, null, null, false), + part) { + + override def compute(split: Split): Iterator[(K, V)] = { + val buf = new ArrayBuffer[(K, V)] + val fetcher = SparkEnv.get.shuffleFetcher + def addTupleToBuffer(k: K, v: V) = { buf += Tuple(k, v) } + fetcher.fetch[K, V](dep.shuffleId, split.index, addTupleToBuffer) + buf.iterator + } +} + + +/** + * A sort-based shuffle (that doesn't apply aggregation). It does so by first + * repartitioning the RDD by range, and then sort within each range. + */ +class ShuffledSortedRDD[K <% Ordered[K]: ClassManifest, V]( + @transient parent: RDD[(K, V)], + ascending: Boolean) + extends RepartitionShuffledRDD[K, V]( + parent, + new RangePartitioner(parent.splits.size, parent, ascending)) { + + override def compute(split: Split): Iterator[(K, V)] = { + // By separating this from RepartitionShuffledRDD, we avoided a + // buf.iterator.toArray call, thus avoiding building up the buffer twice. + val buf = new ArrayBuffer[(K, V)] + def addTupleToBuffer(k: K, v: V) = { buf += Tuple(k, v) } + SparkEnv.get.shuffleFetcher.fetch[K, V](dep.shuffleId, split.index, addTupleToBuffer) + buf.sortWith((x, y) => if (ascending) x._1 < y._1 else x._1 > y._1).iterator + } +} + + +/** + * The resulting RDD from shuffle and running (hash-based) aggregation. + */ +class ShuffledAggregatedRDD[K, V, C]( + @transient parent: RDD[(K, V)], + aggregator: Aggregator[K, V, C], + part : Partitioner) + extends ShuffledRDD[K, V, C](parent, aggregator, part) { override def compute(split: Split): Iterator[(K, C)] = { val combiners = new JHashMap[K, C] diff --git a/core/src/main/scala/spark/scheduler/ShuffleMapTask.scala b/core/src/main/scala/spark/scheduler/ShuffleMapTask.scala index 3bcc588015..745aa0c939 100644 --- a/core/src/main/scala/spark/scheduler/ShuffleMapTask.scala +++ b/core/src/main/scala/spark/scheduler/ShuffleMapTask.scala @@ -44,7 +44,8 @@ object ShuffleMapTask { } // Since both the JarSet and FileSet have the same format this is used for both. - def serializeFileSet(set : HashMap[String, Long], stageId: Int, cache : JHashMap[Int, Array[Byte]]) : Array[Byte] = { + def serializeFileSet( + set : HashMap[String, Long], stageId: Int, cache : JHashMap[Int, Array[Byte]]) : Array[Byte] = { val old = cache.get(stageId) if (old != null) { return old @@ -59,7 +60,6 @@ object ShuffleMapTask { } } - def deserializeInfo(stageId: Int, bytes: Array[Byte]): (RDD[_], ShuffleDependency[_,_,_]) = { synchronized { val loader = Thread.currentThread.getContextClassLoader @@ -113,7 +113,8 @@ class ShuffleMapTask( out.writeInt(bytes.length) out.write(bytes) - val fileSetBytes = ShuffleMapTask.serializeFileSet(fileSet, stageId, ShuffleMapTask.fileSetCache) + val fileSetBytes = ShuffleMapTask.serializeFileSet( + fileSet, stageId, ShuffleMapTask.fileSetCache) out.writeInt(fileSetBytes.length) out.write(fileSetBytes) val jarSetBytes = ShuffleMapTask.serializeFileSet(jarSet, stageId, ShuffleMapTask.jarSetCache) @@ -172,7 +173,7 @@ class ShuffleMapTask( buckets.map(_.iterator) } else { // No combiners (no map-side aggregation). Simply partition the map output. - val buckets = Array.tabulate(numOutputSplits)(_ => new ArrayBuffer[(Any, Any)]) + val buckets = Array.fill(numOutputSplits)(new ArrayBuffer[(Any, Any)]) for (elem <- rdd.iterator(split)) { val pair = elem.asInstanceOf[(Any, Any)] val bucketId = partitioner.getPartition(pair._1) diff --git a/core/src/test/scala/spark/ShuffleSuite.scala b/core/src/test/scala/spark/ShuffleSuite.scala index f622c413f7..9d7e2591f1 100644 --- a/core/src/test/scala/spark/ShuffleSuite.scala +++ b/core/src/test/scala/spark/ShuffleSuite.scala @@ -15,16 +15,16 @@ import scala.collection.mutable.ArrayBuffer import SparkContext._ class ShuffleSuite extends FunSuite with ShouldMatchers with BeforeAndAfter { - + var sc: SparkContext = _ - + after { if (sc != null) { sc.stop() sc = null } } - + test("groupByKey") { sc = new SparkContext("local", "test") val pairs = sc.parallelize(Array((1, 1), (1, 2), (1, 3), (2, 1))) @@ -57,7 +57,7 @@ class ShuffleSuite extends FunSuite with ShouldMatchers with BeforeAndAfter { val valuesFor2 = groups.find(_._1 == 2).get._2 assert(valuesFor2.toList.sorted === List(1)) } - + test("groupByKey with many output partitions") { sc = new SparkContext("local", "test") val pairs = sc.parallelize(Array((1, 1), (1, 2), (1, 3), (2, 1))) @@ -187,7 +187,7 @@ class ShuffleSuite extends FunSuite with ShouldMatchers with BeforeAndAfter { (4, (ArrayBuffer(), ArrayBuffer('w'))) )) } - + test("zero-partition RDD") { sc = new SparkContext("local", "test") val emptyDir = Files.createTempDir() @@ -195,7 +195,7 @@ class ShuffleSuite extends FunSuite with ShouldMatchers with BeforeAndAfter { assert(file.splits.size == 0) assert(file.collect().toList === Nil) // Test that a shuffle on the file works, because this used to be a bug - assert(file.map(line => (line, 1)).reduceByKey(_ + _).collect().toList === Nil) + assert(file.map(line => (line, 1)).reduceByKey(_ + _).collect().toList === Nil) } test("map-side combine") { @@ -212,7 +212,7 @@ class ShuffleSuite extends FunSuite with ShouldMatchers with BeforeAndAfter { _+_, _+_, false) - val shuffledRdd = new ShuffledRDD( + val shuffledRdd = new ShuffledAggregatedRDD( pairs, aggregator, new HashPartitioner(2)) assert(shuffledRdd.collect().toSet === Set((1, 8), (2, 1))) @@ -220,7 +220,7 @@ class ShuffleSuite extends FunSuite with ShouldMatchers with BeforeAndAfter { // not see an exception because mergeCombine should not have been called. val aggregatorWithException = new Aggregator[Int, Int, Int]( (v: Int) => v, _+_, ShuffleSuite.mergeCombineException, false) - val shuffledRdd1 = new ShuffledRDD( + val shuffledRdd1 = new ShuffledAggregatedRDD( pairs, aggregatorWithException, new HashPartitioner(2)) assert(shuffledRdd1.collect().toSet === Set((1, 8), (2, 1))) @@ -228,7 +228,7 @@ class ShuffleSuite extends FunSuite with ShouldMatchers with BeforeAndAfter { // expect to see an exception thrown. val aggregatorWithException1 = new Aggregator[Int, Int, Int]( (v: Int) => v, _+_, ShuffleSuite.mergeCombineException) - val shuffledRdd2 = new ShuffledRDD( + val shuffledRdd2 = new ShuffledAggregatedRDD( pairs, aggregatorWithException1, new HashPartitioner(2)) evaluating { shuffledRdd2.collect() } should produce [SparkException] }