Spark Core（十八）Spark的BlockManager原理与源码分析

BlockManager定义
1. BlockManager是Spark的分布式存储系统，与我们平常说的分布式存储系统是有区别的，区别就是这个分布式存储系统只会管理Block块数据，它运行在所有节点上。BlockManager的结构是Maser-Slave架构，Master就是Driver上的BlockManagerMaster，Slave就是每个Executor上的BlockManager。BlockManagerMaster负责接受Executor上的BlockManager的注册以及管理BlockManager的元数据信息
BlockManager原理
1. 从上边的定义我们已经得知，BlockManager是分布式的，运行在各个节点上的。从BlockManager的创建过程来看，其实Block是运行在Driver和每个Executor的。因为在创建SparkContext的时候，会调用SparkEnv.blockManager.initialize方法实例化BlockManager对象，在创建Executor对象的时候也会创建BlockManager。
2. 在初始化BlockManager的时候，第一步会初始化BlockTransferService的init方法（子类NettyBlockTransferService实现了init方法），这个方法的作用就是初始化Netty服务，为拉取block数据提供服务。第二步是调用shuffleClient的init方法，shuffleClient这个引用有可能是BlockTransferService有可能是ExternalShuffleClient，取决于我们的配置文件是否配置了externalShuffleServiceEnabled未开启状态，其实无论是哪种，都是为了对外提供服务，能够使block数据再节点之间流动起来。
3. BlockManagerMaster调用registerBlockManager方法，向BlockManagerMaster（其实BlockManagerMasterEndpoint）发送BlockManager的注册请求。
4. BlockManagerMaster（其实BlockManagerMasterEndpoint）接受到BlockManager的注册请求后。会调用register方法，开始注册Executor上的BlockManager，注册完成以后将BlockManagerId返回给对应Executor上的BlockManager。

BlockManager源码

当我们的程序启动的时候，首先会创建SparkContext对象，在创建SparkContext对象的时候就会调用_env.blockManager.initialize(_applicationId)创建BlockManager对象，这个BlockManager就是Driver上的BlockManager，它负责管理集群中Executor上的BlockManager

SparkContext里创建BlockManager代码片段

//为Driver创建BlockManager
_env.blockManager.initialize(_applicationId)

创建Executor的时候，Executor内部会调用_env.blockManager.initialize(conf.getAppId)方法创建BlockManager

if (!isLocal) {
    env.metricsSystem.registerSource(executorSource)
    env.blockManager.initialize(conf.getAppId)
  }

BlockManager类里的initialize方法，该方法作用是创建BlockManager，并且向BlockManagerMaster进行注册

def initialize(appId: String): Unit = {
    //初始化BlockTransferService，其实是它的子类NettyBlockTransferService是下了init方法，
    //该方法的作用就是初始化传输服务，通过传输服务可以从不同的节点上拉取Block数据
    blockTransferService.init(this)
    shuffleClient.init(appId)

    //设置block的复制分片策略，由spark.storage.replication.policy指定
    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
    }

    //根据给定参数为对对应的Executor封装一个BlockManagerId对象（块存储的唯一标识）
    //executorID：executor的Id，blockTransferService.hostName：传输Block数据的服务的主机名
    //blockTransferService.port：传输Block数据的服务的主机名
    val id = BlockManagerId(executorId, blockTransferService.hostName, blockTransferService.port, None)

    //调用BlockManagerMaster的registerBlockManager方法向Driver上的BlockManagerMaster注册
    val idFromMaster = master.registerBlockManager(
      id,
      maxMemory,
      slaveEndpoint)
    //更新BlockManagerId
    blockManagerId = if (idFromMaster != null) idFromMaster else id

    //判断是否开了外部shuffle服务
    shuffleServerId = if (externalShuffleServiceEnabled) {
      logInfo(s"external shuffle service port = $externalShuffleServicePort")
      BlockManagerId(executorId, blockTransferService.hostName, externalShuffleServicePort)
    } else {
      blockManagerId
    }

    // 如果开启了外部shuffle服务，并且该节点是Driver的话就调用registerWithExternalShuffleServer方法
    //将BlockManager注册在本地
    if (externalShuffleServiceEnabled && !blockManagerId.isDriver) {
      registerWithExternalShuffleServer()
    }

    logInfo(s"Initialized BlockManager: $blockManagerId")
  }

BlockManagerMaster类里的registerBlockManager方法，向Driver发送RegisterBlockManager消息进行注册

def registerBlockManager(blockManagerId: BlockManagerId,maxMemSize: Long,slaveEndpoint: RpcEndpointRef): BlockManagerId = {
    logInfo(s"Registering BlockManager $blockManagerId")
    //向Driver发送注册BlockManager请求
    //blockManagerId：块存储的唯一标识，里边封装了该BlockManager所在的executorId，提供Netty服务的主机名和端口
    //maxMemSize最大的内存
    val updatedId = driverEndpoint.askWithRetry[BlockManagerId](
      RegisterBlockManager(blockManagerId, maxMemSize, slaveEndpoint))
    logInfo(s"Registered BlockManager $updatedId")
    updatedId
  }

BlockManagerMasterEndpoint类里的receiveAndReply方法，这个方法就是接受请求的消息，然后并处理请求

def receiveAndReply(context: RpcCallContext): PartialFunction[Any, Unit] = {
    //BlocManagerMasterEndPoint接收到来自Executor上的BlockManager注册请求的时候，
    //调用register方法开始注册BlockManager，
    case RegisterBlockManager(blockManagerId, maxMemSize, slaveEndpoint) =>
      context.reply(register(blockManagerId, maxMemSize, slaveEndpoint))
    //....其余代码省略
  }

BlockManagerMasterEndpoint类里的register方法，该方法的作用就是开始注册executor上的BlockManager


// BlockManager到BlockManaInfo的映射
private val blockManagerInfo = new mutable.HashMap[BlockManagerId, BlockManagerInfo]

//executorId到BlockManagerId的映射
private val blockManagerIdByExecutor = new mutable.HashMap[String, BlockManagerId]

    private def register(
      idWithoutTopologyInfo: BlockManagerId,
      maxMemSize: Long,
      slaveEndpoint: RpcEndpointRef): BlockManagerId = {
    //利用从Executor上传过来的BlockManagerId信息重新封装BlockManagerId，并且
    //之前传过来没有拓扑信息，这次直接将拓扑信息也封装进去，得到一个更完整的BlockManagerId
    val id = BlockManagerId(
      idWithoutTopologyInfo.executorId,
      idWithoutTopologyInfo.host,
      idWithoutTopologyInfo.port,
      topologyMapper.getTopologyForHost(idWithoutTopologyInfo.host))

    val time = System.currentTimeMillis()
    //判断当前这个BlockManagerId是否注册过，注册结构为：HashMap[BlockManagerId, BlockManagerInfo]
    //如果没注册过就向下执行开始注册
    if (!blockManagerInfo.contains(id)) {
      //首先会根据executorId查找内存缓存结构中是否有对应的BlockManagerId,如果为存在那么就将调用removeExecutor方法，
      //将executor从BlockManagerMaster中移除，首先会移除executorId对应的BlockManagerId,然后在移除该旧的BlockManager
      //其实就是移除以前的注册过的旧数据
      blockManagerIdByExecutor.get(id.executorId) match {
        case Some(oldId) =>
          // A block manager of the same executor already exists, so remove it (assumed dead)
          logError("Got two different block manager registrations on same executor - "
              + s" will replace old one $oldId with new one $id")
          removeExecutor(id.executorId)
        case None =>
      }
      logInfo("Registering block manager %s with %s RAM, %s".format(
        id.hostPort, Utils.bytesToString(maxMemSize), id))

      //将executorId与BlockManagerId映射起来，放到内存缓存中
      blockManagerIdByExecutor(id.executorId) = id

      //将BlockManagerId与BlockManagerInfo映射起来，放入内存缓存中
      //BlockManagerInfo封住了BlockMangerId，当前注册的事件，最大的内存
      blockManagerInfo(id) = new BlockManagerInfo(
        id, System.currentTimeMillis(), maxMemSize, slaveEndpoint)
    }
    listenerBus.post(SparkListenerBlockManagerAdded(time, id, maxMemSize))
    id
  }

BlockManagerMasterEndpoint类里的removeExecutor方法，该方法的作用就是移除掉之前注册过的旧数据

private def removeExecutor(execId: String) {
    logInfo("Trying to remove executor " + execId + " from BlockManagerMaster.")
    //根据当前的executorId获取对应的BlockManager，这个BlockManager已经是旧数据了，然后调用removeBlockManager方法继续移除其他的旧数据
    blockManagerIdByExecutor.get(execId).foreach(removeBlockManager)
  }

BlockManagerMasterEndpoint类里的removeBlockManager方法，该方法的作用就是开始移除之前注册过的并且冲突的旧数据

 private def removeBlockManager(blockManagerId: BlockManagerId) {

    //根据就得BlockManager找到对应的BlockManagerInfo
    val info = blockManagerInfo(blockManagerId)

    //将旧的BlockManager里的executorId对应的BlockManagerId移除掉
    blockManagerIdByExecutor -= blockManagerId.executorId

    // 然后移除掉BlockManagerId对应的BlockManagerInfo
    blockManagerInfo.remove(blockManagerId)
    val iterator = info.blocks.keySet.iterator
    while (iterator.hasNext) {
      val blockId = iterator.next
      val locations = blockLocations.get(blockId)
      locations -= blockManagerId
      if (locations.size == 0) {
        blockLocations.remove(blockId)
      }
    }
    listenerBus.post(SparkListenerBlockManagerRemoved(System.currentTimeMillis(), blockManagerId))
    logInfo(s"Removing block manager $blockManagerId")
  }

Spark Core（十八）Spark的BlockManager原理与源码分析

猜你喜欢