Spark2.x精通：Executor端BlockManager源码剖析

一、概述

    BlockManager是分布式块存储管理。核心机制是每个节点存储自己的内存空间和磁盘空间。BlockManagerMaster负责与其他节点的BlockManager通信并负责块在节点间的复制。BlockInfoManager负责管理块的元数据并提供读写锁的功能。当从本地的BlockManager获取不到块时，从远程节点Fetcher这个Block数据块。

    上篇文章我们已经讲了Driver端的BlockManagerMaster，这里我们结合源码剖析Executor端的BlockManager。

二、Executor端BlockManager源码剖析

    下面我们就来说下他的几个比较核心的方法：

1.Executor 在启动的时候，一定会实例化 BlockManager，我们这里先去看下initialize()函数，代码如下：

def initialize(appId: String): Unit = {   //初始化BlockTransferService,之前我们也说过，它主要是用于跨界点的数据传输    blockTransferService.init(this)   // 初始化ShuffleClient，读取其他executor上的shuffle文件的客户端    shuffleClient.init(appId)   //初始化 block复制分片策略 blockReplicationPolicy， 可以通过参数 spark.storage.replication.policy 来指定，   // 默认为 RandomBlockReplicationPolicy    blockReplicationPolicy = {      val priorityClass = conf.get(        "spark.storage.replication.policy", classOf[RandomBlockReplicationPolicy].getName)      val clazz = Utils.classForName(priorityClass)      val ret = clazz.newInstance.asInstanceOf[BlockReplicationPolicy]      logInfo(s"Using $priorityClass for block replication policy")      ret    }    //初始化BlockManagerId    val id =      BlockManagerId(executorId, blockTransferService.hostName, blockTransferService.port, None)    //向driver注册该BlockManager,消息会在BlockManagerMaster端进行处理    val idFromMaster = master.registerBlockManager(      id,      maxOnHeapMemory,      maxOffHeapMemory,      slaveEndpoint)
    blockManagerId = if (idFromMaster != null) idFromMaster else id     //如果启用了Shuffle辅助服务,用于Shuffle效率，这个由参数启用External shuffle Service服务控制，默认为false   //如果启用External shuffle Service服务，这里进行对应的初始化    shuffleServerId = if (externalShuffleServiceEnabled) {      logInfo(s"external shuffle service port = $externalShuffleServicePort")      BlockManagerId(executorId, blockTransferService.hostName, externalShuffleServicePort)    } else {      blockManagerId    }
    // Register Executors' configuration with the local shuffle service, if one should exist.    if (externalShuffleServiceEnabled && !blockManagerId.isDriver) {      registerWithExternalShuffleServer()    }
    logInfo(s"Initialized BlockManager: $blockManagerId")  }

2.向Driver端注册BlockManager，代码如下：

/**   * Re-register with the master and report all blocks to it. This will be called by the heart beat   * thread if our heartbeat to the block manager indicates that we were not registered.   *   * Note that this method must be called without any BlockInfo locks held.   */   //向Driver端注册BlockManager节点，  def reregister(): Unit = {    // TODO: We might need to rate limit re-registering.    logInfo(s"BlockManager $blockManagerId re-registering with master")    master.registerBlockManager(blockManagerId, maxOnHeapMemory, maxOffHeapMemory, slaveEndpoint)    //向Driver更新blockStatus信息    reportAllBlocks()  }
  /**   * Re-register with the master sometime soon.   */  private def asyncReregister(): Unit = {    asyncReregisterLock.synchronized {      if (asyncReregisterTask == null) {        asyncReregisterTask = Future[Unit] {          // This is a blocking action and should run in futureExecutionContext which is a cached          // thread pool          reregister()          asyncReregisterLock.synchronized {            asyncReregisterTask = null          }        }(futureExecutionContext)      }    }  }

3.读Block数据入口get()函数，代码如下：

def get[T: ClassTag](blockId: BlockId): Option[BlockResult] = {    //先从本地获取，调用getLocalValues方法    val local = getLocalValues(blockId)    if (local.isDefined) {      logInfo(s"Found block $blockId locally")      return local    }    //如果本地获取不到，调用getRemoteValues方法远程获取    val remote = getRemoteValues[T](blockId)    if (remote.isDefined) {      logInfo(s"Found block $blockId remotely")      return remote    }    None  }

3.1 这里先来看本地获取getLocalValues，代码如下：

def getLocalValues(blockId: BlockId): Option[BlockResult] = {    logDebug(s"Getting local block $blockId")    //从blockInfoManager中获取元数据信息    blockInfoManager.lockForReading(blockId) match {      case None =>        logDebug(s"Block $blockId was not found")        None      case Some(info) =>        //StorageLevel，按优先内存、其次磁盘的顺序考虑        val level = info.level        logDebug(s"Level for block $blockId is $level")        val taskAttemptId = Option(TaskContext.get()).map(_.taskAttemptId())         //如果数据是保存在内存中，则通过memoryStore直接从内存获取        if (level.useMemory && memoryStore.contains(blockId)) {          //这里判断是否有数据的序列化          val iter: Iterator[Any] = if (level.deserialized) {            memoryStore.getValues(blockId).get          } else {            serializerManager.dataDeserializeStream(              blockId, memoryStore.getBytes(blockId).get.toInputStream())(info.classTag)          }          // We need to capture the current taskId in case the iterator completion is triggered          // from a different thread which does not have TaskContext set; see SPARK-18406 for          // discussion.          val ci = CompletionIterator[Any, Iterator[Any]](iter, {            releaseLock(blockId, taskAttemptId)          })          Some(new BlockResult(ci, DataReadMethod.Memory, info.size))        //如果是保存在磁盘中，则通过diskStore从磁盘获取数据        } else if (level.useDisk && diskStore.contains(blockId)) {          val diskData = diskStore.getBytes(blockId)          val iterToReturn: Iterator[Any] = {          //跟上面的逻辑一致，判断是否有序列化            if (level.deserialized) {              val diskValues = serializerManager.dataDeserializeStream(                blockId,                diskData.toInputStream())(info.classTag)                //为加快读深度，会将磁盘数据，缓存到磁盘，maybeCacheDiskValuesInMemory函数就是干这个事情              maybeCacheDiskValuesInMemory(info, blockId, level, diskValues)            } else {              val stream = maybeCacheDiskBytesInMemory(info, blockId, level, diskData)                .map { _.toInputStream(dispose = false) }                .getOrElse { diskData.toInputStream() }              serializerManager.dataDeserializeStream(blockId, stream)(info.classTag)            }          }          //释放锁          val ci = CompletionIterator[Any, Iterator[Any]](iterToReturn, {            releaseLockAndDispose(blockId, diskData, taskAttemptId)          })          Some(new BlockResult(ci, DataReadMethod.Disk, info.size))        } else {          handleLocalReadFailure(blockId)        }    }  }

3.2 再去看看远程获取getRemoteValues，代码如下：

  private def getRemoteValues[T: ClassTag](blockId: BlockId): Option[BlockResult] = {    val ct = implicitly[ClassTag[T]]    //主要是这个方法    getRemoteBytes(blockId).map { data =>      val values =        serializerManager.dataDeserializeStream(blockId, data.toInputStream(dispose = true))(ct)      new BlockResult(values, DataReadMethod.Network, data.size)    }  }

    这里直接看getRemoteBytes()函数，代码如下：

def getRemoteBytes(blockId: BlockId): Option[ChunkedByteBuffer] = {    logDebug(s"Getting remote block $blockId")    require(blockId != null, "BlockId is null")    var runningFailureCount = 0    var totalFailureCount = 0
    // Because all the remote blocks are registered in driver, it is not necessary to ask    // all the slave executors to get block status.    //从BlockManagerMaster端获取block的状态和位置信息    //消息会在BlockManagerMasterEndpoint中进行处理返回结果    val locationsAndStatus = master.getLocationsAndStatus(blockId)    val blockSize = locationsAndStatus.map { b =>      b.status.diskSize.max(b.status.memSize)    }.getOrElse(0L)    //block所在BlockManager列表    val blockLocations = locationsAndStatus.map(_.locations).getOrElse(Seq.empty)
     //如果block块大于maxRemoteBlockToMem指定的阈值。    // 这里需要借助remoteBlockTempFileManager变量进行数据拉取     val tempFileManager = if (blockSize > maxRemoteBlockToMem) {      remoteBlockTempFileManager    } else {      null    }    //一个block可能存在多个blockmanager上，这里进行了排序    val locations = sortLocations(blockLocations)    val maxFetchFailures = locations.size    var locationIterator = locations.iterator    //循环尝试拉取数据，拉取成功跳出循环        while (locationIterator.hasNext) {      val loc = locationIterator.next()      logDebug(s"Getting remote block $blockId from $loc")      val data = try {        //这里blockTransferService服务进行远程数据的拉取       //更深层次的代码这里不跟了，最后实际是调用NettyBlockTransferService类中       //的 fetchBlocks()函数，从远程节点拉取数据到本地        blockTransferService.fetchBlockSync(          loc.host, loc.port, loc.executorId, blockId.toString, tempFileManager).nioByteBuffer()      } catch {        case NonFatal(e) =>          runningFailureCount += 1          totalFailureCount += 1
          if (totalFailureCount >= maxFetchFailures) {            // Give up trying anymore locations. Either we've tried all of the original locations,            // or we've refreshed the list of locations from the master, and have still            // hit failures after trying locations from the refreshed list.            logWarning(s"Failed to fetch block after $totalFailureCount fetch failures. " +              s"Most recent failure cause:", e)            return None          }
          logWarning(s"Failed to fetch remote block $blockId " +            s"from $loc (failed attempt $runningFailureCount)", e)
          // If there is a large number of executors then locations list can contain a          // large number of stale entries causing a large number of retries that may          // take a significant amount of time. To get rid of these stale entries          // we refresh the block locations after a certain number of fetch failures          if (runningFailureCount >= maxFailuresBeforeLocationRefresh) {            locationIterator = sortLocations(master.getLocations(blockId)).iterator            logDebug(s"Refreshed locations from the driver " +              s"after ${runningFailureCount} fetch failures.")            runningFailureCount = 0          }
          // This location failed, so we retry fetch from a different one by returning null here          null      }
      if (data != null) {        return Some(new ChunkedByteBuffer(data))      }      logDebug(s"The value of block $blockId is null")    }    logDebug(s"Block $blockId not found")    None  }

4.写Block数据入口putBytes()函数，代码如下：

  def putBytes[T: ClassTag](      blockId: BlockId,      bytes: ChunkedByteBuffer,      level: StorageLevel,      tellMaster: Boolean = true): Boolean = {    require(bytes != null, "Bytes is null")    doPutBytes(blockId, bytes, level, implicitly[ClassTag[T]], tellMaster)  }

 4.1 直接看doPutBytes()函数，代码如下:

private def doPutBytes[T](      blockId: BlockId,      bytes: ChunkedByteBuffer,      level: StorageLevel,      classTag: ClassTag[T],      tellMaster: Boolean = true,      keepReadLock: Boolean = false): Boolean = {    doPut(blockId, level, classTag, tellMaster = tellMaster, keepReadLock = keepReadLock) { info =>      val startTimeMs = System.currentTimeMillis      // Since we're storing bytes, initiate the replication before storing them locally.      // This is faster as data is already serialized and ready to send.      //如果block块的复制策略大于1个副本 这里就需要向其远程BlockManager写数据      val replicationFuture = if (level.replication > 1) {        Future {          // This is a blocking action and should run in futureExecutionContext which is a cached          // thread pool. The ByteBufferBlockData wrapper is not disposed of to avoid releasing          // buffers that are owned by the caller.          replicate(blockId, new ByteBufferBlockData(bytes, false), level, classTag)        }(futureExecutionContext)      } else {        null      }
      val size = bytes.size      //如果持久化级别为内存 ，通过memoryStore写数据      if (level.useMemory) {        // Put it in memory first, even if it also has useDisk set to true;        // We will drop it to disk later if the memory store can't hold it.        //判断是否需要序列化         val putSucceeded = if (level.deserialized) {          val values =            serializerManager.dataDeserializeStream(blockId, bytes.toInputStream())(classTag)          memoryStore.putIteratorAsValues(blockId, values, classTag) match {            case Right(_) => true            case Left(iter) =>              // If putting deserialized values in memory failed, we will put the bytes directly to              // disk, so we don't need this iterator and can close it to free resources earlier.              iter.close()              false          }        } else {          val memoryMode = level.memoryMode          memoryStore.putBytes(blockId, size, memoryMode, () => {            if (memoryMode == MemoryMode.OFF_HEAP &&                bytes.chunks.exists(buffer => !buffer.isDirect)) {              bytes.copy(Platform.allocateDirectBuffer)            } else {              bytes            }          })        }        if (!putSucceeded && level.useDisk) {          logWarning(s"Persisting block $blockId to disk instead.")          diskStore.putBytes(blockId, bytes)        }    //如果持久化级别为磁盘 ，通过diskStore写数据      } else if (level.useDisk) {        diskStore.putBytes(blockId, bytes)      }      //获取最新的blockStatus，并向BlockManagerMaster发送消息，通知更新      val putBlockStatus = getCurrentBlockStatus(blockId, info)      val blockWasSuccessfullyStored = putBlockStatus.storageLevel.isValid      if (blockWasSuccessfullyStored) {        // Now that the block is in either the memory or disk store,        // tell the master about it.        info.size = size        if (tellMaster && info.tellMaster) {          reportBlockStatus(blockId, putBlockStatus)        }        addUpdatedBlockStatusToTaskMetrics(blockId, putBlockStatus)      }      logDebug("Put block %s locally took %s".format(blockId, Utils.getUsedTimeMs(startTimeMs)))      //如果副本数大于1  等待远程节点返回处理结果      if (level.replication > 1) {        // Wait for asynchronous replication to finish        try {          ThreadUtils.awaitReady(replicationFuture, Duration.Inf)        } catch {          case NonFatal(t) =>            throw new Exception("Error occurred while waiting for replication to finish", t)        }      }      if (blockWasSuccessfullyStored) {        None      } else {        Some(bytes)      }    }.isEmpty  }

4.2 上面的第18行代码中，如果block副本大于1，需要调用replicate()函数，复制到远程的BlockManager中，代码如下：

 private def replicate(      blockId: BlockId,      data: BlockData,      level: StorageLevel,      classTag: ClassTag[_],      existingReplicas: Set[BlockManagerId] = Set.empty): Unit = {    //获取复制最大失败次数，默认是1，参数spark.storage.maxReplicationFailures控制    val maxReplicationFailures = conf.getInt("spark.storage.maxReplicationFailures", 1)    val tLevel = StorageLevel(      useDisk = level.useDisk,      useMemory = level.useMemory,      useOffHeap = level.useOffHeap,      deserialized = level.deserialized,      replication = 1)   //需要复制的副本数    val numPeersToReplicateTo = level.replication - 1    val startTime = System.nanoTime
    val peersReplicatedTo = mutable.HashSet.empty ++ existingReplicas    val peersFailedToReplicateTo = mutable.HashSet.empty[BlockManagerId]    var numFailures = 0    //获取集群中的所有blockmanager节点，过滤掉本地blockmanager    val initialPeers = getPeers(false).filterNot(existingReplicas.contains)
    var peersForReplication = blockReplicationPolicy.prioritize(      blockManagerId,      initialPeers,      peersReplicatedTo,      blockId,      numPeersToReplicateTo)    //下面就是循环向远程节点传递数据了    while(numFailures <= maxReplicationFailures &&      !peersForReplication.isEmpty &&      peersReplicatedTo.size < numPeersToReplicateTo) {      val peer = peersForReplication.head      try {        val onePeerStartTime = System.nanoTime        logTrace(s"Trying to replicate $blockId of ${data.size} bytes to $peer")        //这里还是通过blockTransferService来上传数据到其他blockManager节点        blockTransferService.uploadBlockSync(          peer.host,          peer.port,          peer.executorId,          blockId,          new BlockManagerManagedBuffer(blockInfoManager, blockId, data, false),          tLevel,          classTag)        logTrace(s"Replicated $blockId of ${data.size} bytes to $peer" +          s" in ${(System.nanoTime - onePeerStartTime).toDouble / 1e6} ms")        peersForReplication = peersForReplication.tail        peersReplicatedTo += peer      } catch {        case NonFatal(e) =>          logWarning(s"Failed to replicate $blockId to $peer, failure #$numFailures", e)          peersFailedToReplicateTo += peer          // we have a failed replication, so we get the list of peers again          // we don't want peers we have already replicated to and the ones that          // have failed previously          val filteredPeers = getPeers(true).filter { p =>            !peersFailedToReplicateTo.contains(p) && !peersReplicatedTo.contains(p)          }
          numFailures += 1          peersForReplication = blockReplicationPolicy.prioritize(            blockManagerId,            filteredPeers,            peersReplicatedTo,            blockId,            numPeersToReplicateTo - peersReplicatedTo.size)      }    }    logDebug(s"Replicating $blockId of ${data.size} bytes to " +      s"${peersReplicatedTo.size} peer(s) took ${(System.nanoTime - startTime) / 1e6} ms")    if (peersReplicatedTo.size < numPeersToReplicateTo) {      logWarning(s"Block $blockId replicated to only " +        s"${peersReplicatedTo.size} peer(s) instead of $numPeersToReplicateTo peers")    }
    logDebug(s"block $blockId replicated to ${peersReplicatedTo.mkString(", ")}")  }

    这里只是讲了BlockManager中的核心函数，很简单无非就是数据的的本地、远程读写，更细的东西自己去看一下源码，感谢关注！！！

标签：val,level,Spark2,blockId,bytes,源码,Executor,block,size
来源： https://blog.51cto.com/15080019/2653895