spark core 2.0 Broadcast TorrentBroadcast

一个比特流类似的一个广播实现。

它的机制如下:

 Driver把序列化的对象分成小块,并且存在driver的BlockManager之中,并且汇报给BlockManagerMaster.

在每一个执行器中,执行器首先从他自己的BlockManager中读到对象。如果对象不存在,则从BlockManagerMaster获取对象的地址,这时是driver,然后从该地址中读出对象,

放到本地的BlockManager中,然后再告诉BlockManagerMaster 此BlockManager中有该对象的一个副本。这时其它的执行器从BlockManagerMaster获取该对象的地址时,地址就是增加之后的地址列表,这时随机找一个进行读。

通过重复上述步骤,BlockManagerMaster返回的地址越多,则从driver读取该对象的可能性就越小,所以可以减轻driver的压力。

/**
 * A BitTorrent-like implementation of [[org.apache.spark.broadcast.Broadcast]].
 *
 * The mechanism is as follows:
 *
 * The driver divides the serialized object into small chunks and
 * stores those chunks in the BlockManager of the driver.
 *
 * On each executor, the executor first attempts to fetch the object from its BlockManager. If
 * it does not exist, it then uses remote fetches to fetch the small chunks from the driver and/or
 * other executors if available. Once it gets the chunks, it puts the chunks in its own
 * BlockManager, ready for other executors to fetch from.
 *
 * This prevents the driver from being the bottleneck in sending out multiple copies of the
 * broadcast data (one per executor).
 *
 * When initialized, TorrentBroadcast objects read SparkEnv.get.conf.
 *
 * @param obj object to broadcast
 * @param id A unique identifier for the broadcast variable.
 */

private[spark] class TorrentBroadcast[T: ClassTag](obj: T, id: Long)
  extends Broadcast[T](id) with Logging with Serializable {

  /**
   * Value of the broadcast object on executors. This is reconstructed by [[readBroadcastBlock]],
   * which builds this value by reading blocks from the driver and/or other executors.
   *
   * On the driver, if the value is required, it is read lazily from the block manager.
   */
  @transient private lazy val _value: T = readBroadcastBlock()

  /** The compression codec to use, or None if compression is disabled */
  @transient private var compressionCodec: Option[CompressionCodec] = _
  /** Size of each block. Default value is 4MB.  This value is only read by the broadcaster. */
  @transient private var blockSize: Int = _

  private def setConf(conf: SparkConf) {
    compressionCodec = if (conf.getBoolean("spark.broadcast.compress", true)) {
      Some(CompressionCodec.createCodec(conf))
    } else {
      None
    }
    // Note: use getSizeAsKb (not bytes) to maintain compatibility if no units are provided
    blockSize = conf.getSizeAsKb("spark.broadcast.blockSize", "4m").toInt * 1024
  }
  setConf(SparkEnv.get.conf)

  private val broadcastId = BroadcastBlockId(id)

  /** Total number of blocks this broadcast variable contains. */
  private val numBlocks: Int = writeBlocks(obj)

  override protected def getValue() = {
    _value
  }

  /**
   * Divide the object into multiple blocks and put those blocks in the block manager.
   *
   * @param value the object to divide
   * @return number of blocks this broadcast variable is divided into
   */
  private def writeBlocks(value: T): Int = {
    import StorageLevel._
    // Store a copy of the broadcast variable in the driver so that tasks run on the driver
    // do not create a duplicate copy of the broadcast variable's value.
    val blockManager = SparkEnv.get.blockManager
    if (!blockManager.putSingle(broadcastId, value, MEMORY_AND_DISK, tellMaster = false)) {
      throw new SparkException(s"Failed to store $broadcastId in BlockManager")
    }
    val blocks =
      TorrentBroadcast.blockifyObject(value, blockSize, SparkEnv.get.serializer, compressionCodec)
    blocks.zipWithIndex.foreach { case (block, i) =>
      val pieceId = BroadcastBlockId(id, "piece" + i)
      val bytes = new ChunkedByteBuffer(block.duplicate())
      if (!blockManager.putBytes(pieceId, bytes, MEMORY_AND_DISK_SER, tellMaster = true)) {
        throw new SparkException(s"Failed to store $pieceId of $broadcastId in local BlockManager")
      }
    }
    blocks.length
  }

  /** Fetch torrent blocks from the driver and/or other executors. */
  private def readBlocks(): Array[ChunkedByteBuffer] = {
    // Fetch chunks of data. Note that all these chunks are stored in the BlockManager and reported
    // to the driver, so other executors can pull these chunks from this executor as well.
    val blocks = new Array[ChunkedByteBuffer](numBlocks)
    val bm = SparkEnv.get.blockManager

    for (pid <- Random.shuffle(Seq.range(0, numBlocks))) {
      val pieceId = BroadcastBlockId(id, "piece" + pid)
      logDebug(s"Reading piece $pieceId of $broadcastId")
      // First try getLocalBytes because there is a chance that previous attempts to fetch the
      // broadcast blocks have already fetched some of the blocks. In that case, some blocks
      // would be available locally (on this executor).
      bm.getLocalBytes(pieceId) match {
        case Some(block) =>
          blocks(pid) = block
          releaseLock(pieceId)
        case None =>
          bm.getRemoteBytes(pieceId) match {
            case Some(b) =>
              // We found the block from remote executors/driver's BlockManager, so put the block
              // in this executor's BlockManager.
              if (!bm.putBytes(pieceId, b, StorageLevel.MEMORY_AND_DISK_SER, tellMaster = true)) {
                throw new SparkException(
                  s"Failed to store $pieceId of $broadcastId in local BlockManager")
              }
              blocks(pid) = b
            case None =>
              throw new SparkException(s"Failed to get $pieceId of $broadcastId")
          }
      }
    }
    blocks
  }

  /**
   * Remove all persisted state associated with this Torrent broadcast on the executors.
   */
  override protected def doUnpersist(blocking: Boolean) {
    TorrentBroadcast.unpersist(id, removeFromDriver = false, blocking)
  }

  /**
   * Remove all persisted state associated with this Torrent broadcast on the executors
   * and driver.
   */
  override protected def doDestroy(blocking: Boolean) {
    TorrentBroadcast.unpersist(id, removeFromDriver = true, blocking)
  }

  /** Used by the JVM when serializing this object. */
  private def writeObject(out: ObjectOutputStream): Unit = Utils.tryOrIOException {
    assertValid()
    out.defaultWriteObject()
  }

  private def readBroadcastBlock(): T = Utils.tryOrIOException {
    TorrentBroadcast.synchronized {
      setConf(SparkEnv.get.conf)
      val blockManager = SparkEnv.get.blockManager
      blockManager.getLocalValues(broadcastId).map(_.data.next()) match {
        case Some(x) =>
          releaseLock(broadcastId)
          x.asInstanceOf[T]

        case None =>
          logInfo("Started reading broadcast variable " + id)
          val startTimeMs = System.currentTimeMillis()
          val blocks = readBlocks().flatMap(_.getChunks())
          logInfo("Reading broadcast variable " + id + " took" + Utils.getUsedTimeMs(startTimeMs))

          val obj = TorrentBroadcast.unBlockifyObject[T](
            blocks, SparkEnv.get.serializer, compressionCodec)
          // Store the merged copy in BlockManager so other tasks on this executor don't
          // need to re-fetch it.
          val storageLevel = StorageLevel.MEMORY_AND_DISK
          if (!blockManager.putSingle(broadcastId, obj, storageLevel, tellMaster = false)) {
            throw new SparkException(s"Failed to store $broadcastId in BlockManager")
          }
          obj
      }
    }
  }

  /**
   * If running in a task, register the given block's locks for release upon task completion.
   * Otherwise, if not running in a task then immediately release the lock.
   */
  private def releaseLock(blockId: BlockId): Unit = {
    val blockManager = SparkEnv.get.blockManager
    Option(TaskContext.get()) match {
      case Some(taskContext) =>
        taskContext.addTaskCompletionListener(_ => blockManager.releaseLock(blockId))
      case None =>
        // This should only happen on the driver, where broadcast variables may be accessed
        // outside of running tasks (e.g. when computing rdd.partitions()). In order to allow
        // broadcast variables to be garbage collected we need to free the reference here
        // which is slightly unsafe but is technically okay because broadcast variables aren't
        // stored off-heap.
        blockManager.releaseLock(blockId)
    }
  }

}


private object TorrentBroadcast extends Logging {

  def blockifyObject[T: ClassTag](
      obj: T,
      blockSize: Int,
      serializer: Serializer,
      compressionCodec: Option[CompressionCodec]): Array[ByteBuffer] = {
    val cbbos = new ChunkedByteBufferOutputStream(blockSize, ByteBuffer.allocate)
    val out = compressionCodec.map(c => c.compressedOutputStream(cbbos)).getOrElse(cbbos)
    val ser = serializer.newInstance()
    val serOut = ser.serializeStream(out)
    Utils.tryWithSafeFinally {
      serOut.writeObject[T](obj)
    } {
      serOut.close()
    }
    cbbos.toChunkedByteBuffer.getChunks()
  }

  def unBlockifyObject[T: ClassTag](
      blocks: Array[ByteBuffer],
      serializer: Serializer,
      compressionCodec: Option[CompressionCodec]): T = {
    require(blocks.nonEmpty, "Cannot unblockify an empty array of blocks")
    val is = new SequenceInputStream(
      blocks.iterator.map(new ByteBufferInputStream(_)).asJavaEnumeration)
    val in: InputStream = compressionCodec.map(c => c.compressedInputStream(is)).getOrElse(is)
    val ser = serializer.newInstance()
    val serIn = ser.deserializeStream(in)
    val obj = Utils.tryWithSafeFinally {
      serIn.readObject[T]()
    } {
      serIn.close()
    }
    obj
  }

  /**
   * Remove all persisted blocks associated with this torrent broadcast on the executors.
   * If removeFromDriver is true, also remove these persisted blocks on the driver.
   */
  def unpersist(id: Long, removeFromDriver: Boolean, blocking: Boolean): Unit = {
    logDebug(s"Unpersisting TorrentBroadcast $id")
    SparkEnv.get.blockManager.master.removeBroadcast(id, removeFromDriver, blocking)
  }
}





你可能感兴趣的:(spark)