spark 筆記 16： BlockManager

先看一下原理性的文章：http://jerryshao.me/architecture/2013/10/08/spark-storage-module-analysis/?，http://jerryshao.me/architecture/2013/10/08/spark-storage-module-analysis/? , 另外，spark的存儲使用了Segment File的概念（http://en.wikipedia.org/wiki/Segmented_file_transfer?），概括的說，它是把文件劃分成多個段，分別存儲在不同的服務器上；在讀取的時候，同時從這些服務器上讀取。（這也是BT的基礎）。

之前分析shuffle的調用關系的時候，其實已經包含了很多的BlockManager的流程，但還是有必要系統的看一遍它的代碼。

getLocalFromDisk這個函數，是前面看shuffleManager的終點，但卻是BlockManager的起點。即使是到遠端獲取block的操作，也是發送一個消息到遠端服務器上執行getLocalFromDisk，然后再把結果發送回來。

->diskStore.getValues(blockId, serializer)

============================BlockManager============================

-> BlockManager::getLocalFromDisk

->diskStore.getValues(blockId, serializer)

->getBytes(blockId).map(bytes => blockManager.dataDeserialize(blockId, bytes, serializer))

->val segment = diskManager.getBlockLocation(blockId) --DiskBlockManager的方法，獲取block在一個文件中的一個塊位置

->if ?blockId.isShuffle and env.shuffleManager.isInstanceOf[SortShuffleManager] --如果是hash類型shuffle，

->sortShuffleManager.getBlockLocation(blockId.asInstanceOf[ShuffleBlockId], this) --For sort-based shuffle, let it figure out its blocks

->else if blockId.isShuffle and shuffleBlockManager.consolidateShuffleFiles --聯合文件模式

->shuffleBlockManager.getBlockLocation(blockId.asInstanceOf[ShuffleBlockId]) --For hash-based shuffle with consolidated files

->val shuffleState = shuffleStates(id.shuffleId) --

->for (fileGroup <- shuffleState.allFileGroups)

->val segment = fileGroup.getFileSegmentFor(id.mapId, id.reduceId) --次函數單獨分析

->if (segment.isDefined) { return segment.get }

->else

->val file = getFile(blockId.name)--getFile(filename: String): File

->val hash = Utils.nonNegativeHash(filename)

->val dirId = hash % localDirs.length

->val subDirId = (hash / localDirs.length) % subDirsPerLocalDir

->var subDir = subDirs(dirId)(subDirId)

->new File(subDir, filename)

->new FileSegment(file, 0, file.length())

->val channel = new RandomAccessFile(segment.file, "r").getChannel

->if (segment.length < minMemoryMapBytes)

->channel.position(segment.offset)

->channel.read(buf)

->return buf

->else

->return Some(channel.map(MapMode.READ_ONLY, segment.offset, segment.length))

ShuffleFileGroup：如何通過mapId和reduceId在ShuffleBlockManager 中獲取數據：getFileSegmentFor函數

->根據reduceId從ShuffleFileGroup的屬性val files: Array[File]里面找到reduce的文件句柄fd

? ? ->根據mapId從mapIdToIndex找到index，

? ?? ???->根據reduce找到blockOffset向量和blockLen向量，

? ??? ??? ??->再通過index從向量里面找到offset和len，

? ??? ??? ?? ???->最后通過offset和len從fd里面讀取到需要的數據

從遠本地取數據

->BlockManager::doGetLocal

->val info = blockInfo.get(blockId).orNull

->val level = info.level

->if (level.useMemory) --Look for the block in memory

->val result = if (asBlockResult)

->memoryStore.getValues(blockId).map(new BlockResult(_, DataReadMethod.Memory, info.size))

->esle

->memoryStore.getBytes(blockId)

->if (level.useOffHeap) -- Look for the block in Tachyon

->tachyonStore.getBytes(blockId)

->if (level.useDisk)

->val bytes: ByteBuffer = diskStore.getBytes(blockId)

->if (!level.useMemory) // If the block shouldn't be stored in memory, we can just return it

->if (asBlockResult)

->return Some(new BlockResult(dataDeserialize(blockId, bytes), DataReadMethod.Disk, info.size))

->else

->return Some(bytes)

->else --memory// Otherwise, we also have to store something in the memory store

->if (!level.deserialized || !asBlockResult) 不序列化或者不block"memory serialized", or if it should be cached as objects in memory

->val copyForMemory = ByteBuffer.allocate(bytes.limit)

->copyForMemory.put(bytes)

->memoryStore.putBytes(blockId, copyForMemory, level)

->if (!asBlockResult)

->return Some(bytes)

->else --需要序列化再寫內存

->val values = dataDeserialize(blockId, bytes)

->if (level.deserialized) // Cache the values before returning them

->val putResult = memoryStore.putIterator(blockId, values, level, returnValues = true, allowPersistToDisk = false)

->putResult.data match case Left(it) return Some(new BlockResult(it, DataReadMethod.Disk, info.size))

->else

->return Some(new BlockResult(values, DataReadMethod.Disk, info.size))

->val values = dataDeserialize(blockId, bytes)

從遠端獲取數據

->BlockManager::doGetRemote

->val locations = Random.shuffle(master.getLocations(blockId)) --隨機打散

->for (loc <- locations) --遍歷所有地址

->val data = BlockManagerWorker.syncGetBlock(GetBlock(blockId), ConnectionManagerId(loc.host, loc.port))

->val blockMessage = BlockMessage.fromGetBlock(msg)

->val newBlockMessage = new BlockMessage()

->newBlockMessage.set(getBlock)

->typ = BlockMessage.TYPE_GET_BLOCK

->id = getBlock.id

->val blockMessageArray = new BlockMessageArray(blockMessage)

-> val responseMessage = Try(Await.result(connectionManager.sendMessageReliably(toConnManagerId, blockMessageArray.toBufferMessage), Duration.Inf))

->responseMessage match {case Success(message) => ?val bufferMessage = message.asInstanceOf[BufferMessage]

->logDebug("Response message received " + bufferMessage)

->BlockMessageArray.fromBufferMessage(bufferMessage).foreach(blockMessage =>?

->logDebug("Found " + blockMessage)

->return blockMessage.getData

->return Some(data)

===========================end=================================

再次引用這個圖：多個map可以對應一個文件，其中每個map對應文件中的某些段。這樣做是為了減少文件數量。

（圖片來源：http://jerryshao.me/architecture/2014/01/04/spark-shuffle-detail-investigation/?）

獲取block數據返回的數據結構

/* Class for returning a fetched block and associated metrics. */
private[spark] class BlockResult(
val data: Iterator[Any],
    readMethod: DataReadMethod.Value,
    bytes: Long) {
val inputMetrics = new InputMetrics(readMethod)
inputMetrics.bytesRead = bytes
}

private[spark] class BlockManager(
    executorId: String,
    actorSystem: ActorSystem,
    val master: BlockManagerMaster,
    defaultSerializer: Serializer,
    maxMemory: Long,
    val conf: SparkConf,
    securityManager: SecurityManager,
    mapOutputTracker: MapOutputTracker,
    shuffleManager: ShuffleManager)
extends BlockDataProvider with Logging {

shuffle狀態，主要包含了unusedFileGroups、allFileGroups兩個屬性，記錄當前已經使用和未使用的ShuffleFileGroup

/**
 * Contains all the state related to a particular shuffle. This includes a pool of unused
 * ShuffleFileGroups, as well as all ShuffleFileGroups that have been created for the shuffle.
 */
private class ShuffleState(val numBuckets: Int) {
val nextFileId = new AtomicInteger(0)
val unusedFileGroups = new ConcurrentLinkedQueue[ShuffleFileGroup]()
val allFileGroups = new ConcurrentLinkedQueue[ShuffleFileGroup]()

/**
   * The mapIds of all map tasks completed on this Executor for this shuffle.
   * NB: This is only populated if consolidateShuffleFiles is FALSE. We don't need it otherwise.
   */
  val completedMapTasks = new ConcurrentLinkedQueue[Int]()
}

shuffleStates 是一個基于時間戳的hash table?

private val shuffleStates = new TimeStampedHashMap[ShuffleId, ShuffleState]

private val metadataCleaner =
new MetadataCleaner(MetadataCleanerType.SHUFFLE_BLOCK_MANAGER, this.cleanup, conf)

Used by sort-based shuffle： shuffle結束時將結果注冊到shuffleStates

/**
 * Register a completed map without getting a ShuffleWriterGroup. Used by sort-based shuffle
 * because it just writes a single file by itself.
 */
def addCompletedMap(shuffleId: Int, mapId: Int, numBuckets: Int): Unit = {
shuffleStates.putIfAbsent(shuffleId, new ShuffleState(numBuckets))
val shuffleState = shuffleStates(shuffleId)
  shuffleState.completedMapTasks.add(mapId)
}

將自己注冊給master?

/**
 * Initialize the BlockManager. Register to the BlockManagerMaster, and start the
 * BlockManagerWorker actor.
 */
private def initialize(): Unit = {
  master.registerBlockManager(blockManagerId, maxMemory, slaveActor)
  BlockManagerWorker.startBlockManagerWorker(this)
}

從本地磁盤獲取一個block數據。為了方便使用

/**
 * A short-circuited method to get blocks directly from disk. This is used for getting
 * shuffle blocks. It is safe to do so without a lock on block info since disk store
 * never deletes (recent) items.
 */
def getLocalFromDisk(blockId: BlockId, serializer: Serializer): Option[Iterator[Any]] = {
diskStore.getValues(blockId, serializer).orElse {
throw new BlockException(blockId, s"Block $blockId not found on disk, though it should be")
  }
}

ShuffleWriterGroup：每個shuffleMapTask都有一組shuffleWriter，它給每個reducer分配了一個writer。當前只有HashShufflle使用了，唯一一個實例化是在forMapTask返回的，給HashShuffleWriter的shuffle屬性使用：

/** A group of writers for a ShuffleMapTask, one writer per reducer. */
private[spark] trait ShuffleWriterGroup {
val writers: Array[BlockObjectWriter]

/** @param success Indicates all writes were successful. If false, no blocks will be recorded. */
  def releaseWriters(success: Boolean)
}

/**
 * Manages assigning disk-based block writers to shuffle tasks. Each shuffle task gets one file
 * per reducer (this set of files is called a ShuffleFileGroup).
 *
 * As an optimization to reduce the number of physical shuffle files produced, multiple shuffle
 * blocks are aggregated into the same file. There is one "combined shuffle file" per reducer
 * per concurrently executing shuffle task. As soon as a task finishes writing to its shuffle
 * files, it releases them for another task.
 * Regarding the implementation of this feature, shuffle files are identified by a 3-tuple:
 *   - shuffleId: The unique id given to the entire shuffle stage.
 *   - bucketId: The id of the output partition (i.e., reducer id)
 *   - fileId: The unique id identifying a group of "combined shuffle files." Only one task at a
 *       time owns a particular fileId, and this id is returned to a pool when the task finishes.
 * Each shuffle file is then mapped to a FileSegment, which is a 3-tuple (file, offset, length)
 * that specifies where in a given file the actual block data is located.
 *
 * Shuffle file metadata is stored in a space-efficient manner. Rather than simply mapping
 * ShuffleBlockIds directly to FileSegments, each ShuffleFileGroup maintains a list of offsets for
 * each block stored in each file. In order to find the location of a shuffle block, we search the
 * files within a ShuffleFileGroups associated with the block's reducer.
 */
// TODO: Factor this into a separate class for each ShuffleManager implementation
private[spark]
class ShuffleBlockManager(blockManager: BlockManager,
                          shuffleManager: ShuffleManager) extends Logging {

ShuffleFileGroup是一組文件，每個reducer對應一個。每個map將會對應一個這個文件（但多個map可以對應一個文件）。多個map對應一個文件時，它們寫入是分段寫入的（mapId，ReduceId）通過getFileSegmentFor函數獲取到這個塊的內容

privateobject /**
   * .
   * .
   */
  private class val Int, val Int, val private var numBlocksInt 0

    /**
     *  For instance,
     * if mapId 5 is the first block in each file, mapIdToIndex(5) = 0.
     */
    private val mapIdToIndex new Int, Int/**
     * Stores consecutive offsets and lengths of blocks into each reducer file, ordered by
     * position in the file.
     * Note:      * .
     */
    private val blockOffsetsByReducer fillLongnew Longprivate val blockLengthsByReducer fillLongnew Longdef applyIntdef recordMapOutputInt, Long, LongassertmapIdToIndexnumBlocks
      numBlocks 1
      for 0 blockOffsetsByReducerblockLengthsByReducer/** Returns the FileSegment associated with the given map task, or None if no entry exists. */
    def getFileSegmentForInt, Intval val blockOffsetsByReducerval blockLengthsByReducerval mapIdToIndex, 1if 0val val Somenew , , else 

來自為知筆記(Wiz)