[SPARK-4550] In sort-based shuffle, store map outputs in serialized form

Refer to the JIRA for the design doc and some perf results.

I wanted to call out some of the more possibly controversial changes up front:
* Map outputs are only stored in serialized form when Kryo is in use.  I'm still unsure whether Java-serialized objects can be relocated.  At the very least, Java serialization writes out a stream header which causes problems with the current approach, so I decided to leave investigating this to future work.
* The shuffle now explicitly operates on key-value pairs instead of any object.  Data is written to shuffle files in alternating keys and values instead of key-value tuples.  `BlockObjectWriter.write` now accepts a key argument and a value argument instead of any object.
* The map output buffer can hold a max of Integer.MAX_VALUE bytes.  Though this wouldn't be terribly difficult to change.
* When spilling occurs, the objects that still in memory at merge time end up serialized and deserialized an extra time.

Author: Sandy Ryza <[email protected]>

Closes apache#4450 from sryza/sandy-spark-4550 and squashes the following commits:

8c70dd9 [Sandy Ryza] Fix serialization
9c16fe6 [Sandy Ryza] Fix a couple tests and move getAutoReset to KryoSerializerInstance
6c54e06 [Sandy Ryza] Fix scalastyle
d8462d8 [Sandy Ryza] SPARK-4550

Author: sryza

core/src/main/scala/org/apache/spark/serializer/KryoSerializer.scala

@@ -200,6 +200,16 @@ private[spark] class KryoSerializerInstance(ks: KryoSerializer) extends Serializ
   override def deserializeStream(s: InputStream): DeserializationStream = {
     new KryoDeserializationStream(kryo, s)
   }
+
+  /**
+   * Returns true if auto-reset is on. The only reason this would be false is if the user-supplied
+   * registrator explicitly turns auto-reset off.
+   */
+  def getAutoReset(): Boolean = {
+    val field = classOf[Kryo].getDeclaredField("autoReset")
+    field.setAccessible(true)
+    field.get(kryo).asInstanceOf[Boolean]
+  }
 }
 
 /**

core/src/main/scala/org/apache/spark/serializer/Serializer.scala

@@ -101,7 +101,12 @@ abstract class SerializerInstance {
  */
 @DeveloperApi
 abstract class SerializationStream {
+  /** The most general-purpose method to write an object. */
   def writeObject[T: ClassTag](t: T): SerializationStream
+  /** Writes the object representing the key of a key-value pair. */
+  def writeKey[T: ClassTag](key: T): SerializationStream = writeObject(key)
+  /** Writes the object representing the value of a key-value pair. */
+  def writeValue[T: ClassTag](value: T): SerializationStream = writeObject(value)
   def flush(): Unit
   def close(): Unit
 
@@ -120,7 +125,12 @@ abstract class SerializationStream {
  */
 @DeveloperApi
 abstract class DeserializationStream {
+  /** The most general-purpose method to read an object. */
   def readObject[T: ClassTag](): T
+  /** Reads the object representing the key of a key-value pair. */
+  def readKey[T: ClassTag](): T = readObject[T]()
+  /** Reads the object representing the value of a key-value pair. */
+  def readValue[T: ClassTag](): T = readObject[T]()
   def close(): Unit
 
   /**
@@ -141,4 +151,25 @@ abstract class DeserializationStream {
       DeserializationStream.this.close()
     }
   }
+
+  /**
+   * Read the elements of this stream through an iterator over key-value pairs. This can only be
+   * called once, as reading each element will consume data from the input source.
+   */
+  def asKeyValueIterator: Iterator[(Any, Any)] = new NextIterator[(Any, Any)] {
+    override protected def getNext() = {
+      try {
+        (readKey[Any](), readValue[Any]())
+      } catch {
+        case eof: EOFException => {
+          finished = true
+          null
+        }
+      }
+    }
+
+    override protected def close() {
+      DeserializationStream.this.close()
+    }
+  }
 }

core/src/main/scala/org/apache/spark/shuffle/hash/HashShuffleWriter.scala

@@ -63,7 +63,7 @@ private[spark] class HashShuffleWriter[K, V](
 
     for (elem <- iter) {
       val bucketId = dep.partitioner.getPartition(elem._1)
-      shuffle.writers(bucketId).write(elem)
+      shuffle.writers(bucketId).write(elem._1, elem._2)
     }
   }
 

core/src/main/scala/org/apache/spark/storage/BlockObjectWriter.scala

@@ -33,7 +33,7 @@ import org.apache.spark.util.Utils
  * This interface does not support concurrent writes. Also, once the writer has
  * been opened, it cannot be reopened again.
  */
-private[spark] abstract class BlockObjectWriter(val blockId: BlockId) {
+private[spark] abstract class BlockObjectWriter(val blockId: BlockId) extends OutputStream {
 
   def open(): BlockObjectWriter
 
@@ -54,9 +54,14 @@ private[spark] abstract class BlockObjectWriter(val blockId: BlockId) {
   def revertPartialWritesAndClose()
 
   /**
-   * Writes an object.
+   * Writes a key-value pair.
    */
-  def write(value: Any)
+  def write(key: Any, value: Any)
+
+  /**
+   * Notify the writer that a record worth of bytes has been written with writeBytes.
+   */
+  def recordWritten()
 
   /**
    * Returns the file segment of committed data that this Writer has written.
@@ -203,12 +208,32 @@ private[spark] class DiskBlockObjectWriter(
     }
   }
 
-  override def write(value: Any) {
+  override def write(key: Any, value: Any) {
+    if (!initialized) {
+      open()
+    }
+
+    objOut.writeKey(key)
+    objOut.writeValue(value)
+    numRecordsWritten += 1
+    writeMetrics.incShuffleRecordsWritten(1)
+
+    if (numRecordsWritten % 32 == 0) {
+      updateBytesWritten()
+    }
+  }
+
+  override def write(b: Int): Unit = throw new UnsupportedOperationException()
+
+  override def write(kvBytes: Array[Byte], offs: Int, len: Int): Unit = {
     if (!initialized) {
       open()
     }
 
-    objOut.writeObject(value)
+    bs.write(kvBytes, offs, len)
+  }
+
+  override def recordWritten(): Unit = {
     numRecordsWritten += 1
     writeMetrics.incShuffleRecordsWritten(1)
 
@@ -238,7 +263,7 @@ private[spark] class DiskBlockObjectWriter(
   }
 
   // For testing
-  private[spark] def flush() {
+  private[spark] override def flush() {
     objOut.flush()
     bs.flush()
   }

core/src/main/scala/org/apache/spark/storage/ShuffleBlockFetcherIterator.scala

@@ -17,14 +17,12 @@
 
 package org.apache.spark.storage
 
-import java.io.{InputStream, IOException}
 import java.util.concurrent.LinkedBlockingQueue
 
 import scala.collection.mutable.{ArrayBuffer, HashSet, Queue}
-import scala.util.{Failure, Success, Try}
+import scala.util.{Failure, Try}
 
 import org.apache.spark.{Logging, TaskContext}
-import org.apache.spark.network.BlockTransferService
 import org.apache.spark.network.shuffle.{BlockFetchingListener, ShuffleClient}
 import org.apache.spark.network.buffer.ManagedBuffer
 import org.apache.spark.serializer.{SerializerInstance, Serializer}
@@ -301,7 +299,7 @@ final class ShuffleBlockFetcherIterator(
         // the scheduler gets a FetchFailedException.
         Try(buf.createInputStream()).map { is0 =>
           val is = blockManager.wrapForCompression(blockId, is0)
-          val iter = serializerInstance.deserializeStream(is).asIterator
+          val iter = serializerInstance.deserializeStream(is).asKeyValueIterator
           CompletionIterator[Any, Iterator[Any]](iter, {
             // Once the iterator is exhausted, release the buffer and set currentResult to null
             // so we don't release it again in cleanup.

core/src/main/scala/org/apache/spark/util/collection/ChainedBuffer.scala

@@ -0,0 +1,144 @@
+/*
+ * 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.util.collection
+
+import java.io.OutputStream
+
+import scala.collection.mutable.ArrayBuffer
+
+/**
+ * A logical byte buffer that wraps a list of byte arrays. All the byte arrays have equal size. The
+ * advantage of this over a standard ArrayBuffer is that it can grow without claiming large amounts
+ * of memory and needing to copy the full contents. The disadvantage is that the contents don't
+ * occupy a contiguous segment of memory.
+ */
+private[spark] class ChainedBuffer(chunkSize: Int) {
+  private val chunkSizeLog2 = (math.log(chunkSize) / math.log(2)).toInt
+  assert(math.pow(2, chunkSizeLog2).toInt == chunkSize,
+    s"ChainedBuffer chunk size $chunkSize must be a power of two")
+  private val chunks: ArrayBuffer[Array[Byte]] = new ArrayBuffer[Array[Byte]]()
+  private var _size: Int = _
+
+  /**
+   * Feed bytes from this buffer into a BlockObjectWriter.
+   *
+   * @param pos Offset in the buffer to read from.
+   * @param os OutputStream to read into.
+   * @param len Number of bytes to read.
+   */
+  def read(pos: Int, os: OutputStream, len: Int): Unit = {
+    if (pos + len > _size) {
+      throw new IndexOutOfBoundsException(
+        s"Read of $len bytes at position $pos would go past size ${_size} of buffer")
+    }
+    var chunkIndex = pos >> chunkSizeLog2
+    var posInChunk = pos - (chunkIndex << chunkSizeLog2)
+    var written = 0
+    while (written < len) {
+      val toRead = math.min(len - written, chunkSize - posInChunk)
+      os.write(chunks(chunkIndex), posInChunk, toRead)
+      written += toRead
+      chunkIndex += 1
+      posInChunk = 0
+    }
+  }
+
+  /**
+   * Read bytes from this buffer into a byte array.
+   *
+   * @param pos Offset in the buffer to read from.
+   * @param bytes Byte array to read into.
+   * @param offs Offset in the byte array to read to.
+   * @param len Number of bytes to read.
+   */
+  def read(pos: Int, bytes: Array[Byte], offs: Int, len: Int): Unit = {
+    if (pos + len > _size) {
+      throw new IndexOutOfBoundsException(
+        s"Read of $len bytes at position $pos would go past size of buffer")
+    }
+    var chunkIndex = pos >> chunkSizeLog2
+    var posInChunk = pos - (chunkIndex << chunkSizeLog2)
+    var written = 0
+    while (written < len) {
+      val toRead = math.min(len - written, chunkSize - posInChunk)
+      System.arraycopy(chunks(chunkIndex), posInChunk, bytes, offs + written, toRead)
+      written += toRead
+      chunkIndex += 1
+      posInChunk = 0
+    }
+  }
+
+  /**
+   * Write bytes from a byte array into this buffer.
+   *
+   * @param pos Offset in the buffer to write to.
+   * @param bytes Byte array to write from.
+   * @param offs Offset in the byte array to write from.
+   * @param len Number of bytes to write.
+   */
+  def write(pos: Int, bytes: Array[Byte], offs: Int, len: Int): Unit = {
+    if (pos > _size) {
+      throw new IndexOutOfBoundsException(
+        s"Write at position $pos starts after end of buffer ${_size}")
+    }
+    // Grow if needed
+    val endChunkIndex = (pos + len - 1) >> chunkSizeLog2
+    while (endChunkIndex >= chunks.length) {
+      chunks += new Array[Byte](chunkSize)
+    }
+
+    var chunkIndex = pos >> chunkSizeLog2
+    var posInChunk = pos - (chunkIndex << chunkSizeLog2)
+    var written = 0
+    while (written < len) {
+      val toWrite = math.min(len - written, chunkSize - posInChunk)
+      System.arraycopy(bytes, offs + written, chunks(chunkIndex), posInChunk, toWrite)
+      written += toWrite
+      chunkIndex += 1
+      posInChunk = 0
+    }
+
+    _size = math.max(_size, pos + len)
+  }
+
+  /**
+   * Total size of buffer that can be written to without allocating additional memory.
+   */
+  def capacity: Int = chunks.size * chunkSize
+
+  /**
+   * Size of the logical buffer.
+   */
+  def size: Int = _size
+}
+
+/**
+ * Output stream that writes to a ChainedBuffer.
+ */
+private[spark] class ChainedBufferOutputStream(chainedBuffer: ChainedBuffer) extends OutputStream {
+  private var pos = 0
+
+  override def write(b: Int): Unit = {
+    throw new UnsupportedOperationException()
+  }
+
+  override def write(bytes: Array[Byte], offs: Int, len: Int): Unit = {
+    chainedBuffer.write(pos, bytes, offs, len)
+    pos += len
+  }
+}

core/src/main/scala/org/apache/spark/util/collection/ExternalAppendOnlyMap.scala

@@ -174,7 +174,7 @@ class ExternalAppendOnlyMap[K, V, C](
       val it = currentMap.destructiveSortedIterator(keyComparator)
       while (it.hasNext) {
         val kv = it.next()
-        writer.write(kv)
+        writer.write(kv._1, kv._2)
         objectsWritten += 1
 
         if (objectsWritten == serializerBatchSize) {
@@ -435,7 +435,9 @@ class ExternalAppendOnlyMap[K, V, C](
      */
     private def readNextItem(): (K, C) = {
       try {
-        val item = deserializeStream.readObject().asInstanceOf[(K, C)]
+        val k = deserializeStream.readKey().asInstanceOf[K]
+        val c = deserializeStream.readValue().asInstanceOf[C]
+        val item = (k, c)
         objectsRead += 1
         if (objectsRead == serializerBatchSize) {
           objectsRead = 0