1.BlockManagerMaster创建
BlockManagerMaster要负责整个应用程序在运行期间block元数据的管理和维护，以及向从节点发送指令执行命令，它是在构造SparkEnv的时候创建的，Driver端是创建SparkContext的时候创建SparkEnv，SparkEnv中对应的初始化代码如下：
val blockManagerMaster = new BlockManagerMaster(registerOrLookupEndpoint( BlockManagerMaster.DRIVER_ENDPOINT_NAME, new BlockManagerMasterEndpoint(rpcEnv, isLocal, conf, listenerBus)), conf, isDriver)
这里可以看到在构造blockManagerMaster时，会创建一个BlockManagerMasterEndpoint实例并注册到了rpcEnv中，Executor中的blockManager通过Driver端BlockManagerMasterEndpoint的引用BlockManagerMasterRef与blockManagerMaster进行通信。
2.BlockManagerMaster成员函数：
1).removeExecutor()函数，代码如下：
//向BlockManagerMasterEndpoint发送RemoveExecutor消息，移除挂掉的Exeutor //这个函数只会在driver端调用 def removeExecutor(execId: String) { tell(RemoveExecutor(execId)) logInfo("Removed " + execId + " successfully in removeExecutor") }
2).removeExecutorAsync()函数，代码如下：
// 跟1)作用差不多，移除挂掉的Executor，这里是非阻塞的异步方法 def removeExecutorAsync(execId: String) { driverEndpoint.ask[Boolean](RemoveExecutor(execId)) logInfo("Removal of executor " + execId + " requested") }
3).registerBlockManager()函数，代码如下：
//Executor端的BlockManager启动会，会向BlockManagerMaster进行注册// BlockManagerMaster会保存在master的blockManagerInfo中 def registerBlockManager( blockManagerId: BlockManagerId, maxOnHeapMemSize: Long, maxOffHeapMemSize: Long, slaveEndpoint: RpcEndpointRef): BlockManagerId = { logInfo(s"Registering BlockManager $blockManagerId") val updatedId = driverEndpoint.askSync[BlockManagerId]( RegisterBlockManager(blockManagerId, maxOnHeapMemSize, maxOffHeapMemSize, slaveEndpoint)) logInfo(s"Registered BlockManager $updatedId") updatedId }
3).updateBlockInfo()函数，代码如下：
//更新block数据块信息 def updateBlockInfo( blockManagerId: BlockManagerId, blockId: BlockId, storageLevel: StorageLevel, memSize: Long, diskSize: Long): Boolean = { //向BlockManagerMasterEndpoint发送UpdateBlockInfo消息，并且返回结果 val res = driverEndpoint.askSync[Boolean]( UpdateBlockInfo(blockManagerId, blockId, storageLevel, memSize, diskSize)) logDebug(s"Updated info of block $blockId") res }
4).getLocations()函数，代码如下：
//获取block所在的BockManager节点信息，这里返回的是Seq集合，//如果block的Replication>1 一个block块，可能会在多个blockmanager//节点上存在def getLocations(blockId: BlockId): Seq[BlockManagerId] = {//向BlockManagerMasterEndpoint发送GetLocations消息driverEndpoint.askSync[Seq[BlockManagerId]](GetLocations(blockId))}
5).getPeers()函数，代码如下：
//获取参数blockManagerId之外的其他BlockManagerId， //上面说了一个block，可能会在多个blockmanager节点上存在 def getPeers(blockManagerId: BlockManagerId): Seq[BlockManagerId] = { //向BlockManagerMasterEndpoint发送GetPeers消息 driverEndpoint.askSync[Seq[BlockManagerId]](GetPeers(blockManagerId)) }
6).getExecutorEndpointRef()函数，代码如下：
//这里就是获取BlockManagerMasterEndpoint的引用，与其进行通信 private def getExecutorEndpointRef(executorId: String): Option[RpcEndpointRef] = { for ( blockManagerId <- blockManagerIdByExecutor.get(executorId); info <- blockManagerInfo.get(blockManagerId) ) yield { info.slaveEndpoint } }
7).getBlockStatus()函数，代码如下：
//获取一个Block的状态信息，位置，占用内存和磁盘大小def getBlockStatus( blockId: BlockId, askSlaves: Boolean = true): Map[BlockManagerId, BlockStatus] = { val msg = GetBlockStatus(blockId, askSlaves) val response = driverEndpoint. askSync[Map[BlockManagerId, Future[Option[BlockStatus]]]](msg) val (blockManagerIds, futures) = response.unzip implicit val sameThread = ThreadUtils.sameThread val cbf = implicitly[ CanBuildFrom[Iterable[Future[Option[BlockStatus]]], Option[BlockStatus], Iterable[Option[BlockStatus]]]] val blockStatus = timeout.awaitResult( Future.sequence[Option[BlockStatus], Iterable](futures)(cbf, ThreadUtils.sameThread)) if (blockStatus == null) { throw new SparkException("BlockManager returned null for BlockStatus query: " + blockId) } blockManagerIds.zip(blockStatus).flatMap { case (blockManagerId, status) => status.map { s => (blockManagerId, s) } }.toMap }
总结一下BlockManagerMaster里面的各种函数处理其实都在 BlockManagerMasterEndpoint实例中，后面我们会详细剖析BlockManagerMasterEndpoint类的各个消息的具体处理流程。