Spark源码阅读之存储体系--存储体系概述与shuffle服务

一、概述

根据《深入理解Spark：核心思想与源码分析》一书，结合最新的spark源代码master分支进行源码阅读，对新版本的代码加上自己的一些理解，如有错误，希望指出。

1.块管理器BlockManager的实现

块管理器是Spark存储体系的核心组件，Driver Application和Executor都会创建BlockManager，源代码位置在core/org.apache.spark.storage,部分代码如下。

private[spark] val externalShuffleServiceEnabled =
    conf.getBoolean("spark.shuffle.service.enabled", false)

  val diskBlockManager = {
    // Only perform cleanup if an external service is not serving our shuffle files.
    val deleteFilesOnStop =
      !externalShuffleServiceEnabled || executorId == SparkContext.DRIVER_IDENTIFIER
    new DiskBlockManager(conf, deleteFilesOnStop)
  }

  // Visible for testing
  private[storage] val blockInfoManager = new BlockInfoManager

  private val futureExecutionContext = ExecutionContext.fromExecutorService(
    ThreadUtils.newDaemonCachedThreadPool("block-manager-future", 128))

  // Actual storage of where blocks are kept
  private[spark] val memoryStore =
    new MemoryStore(conf, blockInfoManager, serializerManager, memoryManager, this)
  private[spark] val diskStore = new DiskStore(conf, diskBlockManager, securityManager)
  memoryManager.setMemoryStore(memoryStore)

  // Note: depending on the memory manager, `maxMemory` may actually vary over time.
  // However, since we use this only for reporting and logging, what we actually want here is
  // the absolute maximum value that `maxMemory` can ever possibly reach. We may need
  // to revisit whether reporting this value as the "max" is intuitive to the user.
  private val maxOnHeapMemory = memoryManager.maxOnHeapStorageMemory
  private val maxOffHeapMemory = memoryManager.maxOffHeapStorageMemory

  // Port used by the external shuffle service. In Yarn mode, this may be already be
  // set through the Hadoop configuration as the server is launched in the Yarn NM.
  private val externalShuffleServicePort = {
    val tmpPort = Utils.getSparkOrYarnConfig(conf, "spark.shuffle.service.port", "7337").toInt
    if (tmpPort == 0) {
      // for testing, we set "spark.shuffle.service.port" to 0 in the yarn config, so yarn finds
      // an open port.  But we still need to tell our spark apps the right port to use.  So
      // only if the yarn config has the port set to 0, we prefer the value in the spark config
      conf.get("spark.shuffle.service.port").toInt
    } else {
      tmpPort
    }
  }

  var blockManagerId: BlockManagerId = _

  // Address of the server that serves this executor's shuffle files. This is either an external
  // service, or just our own Executor's BlockManager.
  private[spark] var shuffleServerId: BlockManagerId = _

  // Client to read other executors' shuffle files. This is either an external service, or just the
  // standard BlockTransferService to directly connect to other Executors.
  private[spark] val shuffleClient = if (externalShuffleServiceEnabled) {
    val transConf = SparkTransportConf.fromSparkConf(conf, "shuffle", numUsableCores)
    new ExternalShuffleClient(transConf, securityManager,
      securityManager.isAuthenticationEnabled(), conf.get(config.SHUFFLE_REGISTRATION_TIMEOUT))
  } else {
    blockTransferService
  }

  // Max number of failures before this block manager refreshes the block locations from the driver
  private val maxFailuresBeforeLocationRefresh =
    conf.getInt("spark.block.failures.beforeLocationRefresh", 5)

  private val slaveEndpoint = rpcEnv.setupEndpoint(
    "BlockManagerEndpoint" + BlockManager.ID_GENERATOR.next,
    new BlockManagerSlaveEndpoint(rpcEnv, this, mapOutputTracker))

  // Pending re-registration action being executed asynchronously or null if none is pending.
  // Accesses should synchronize on asyncReregisterLock.
  private var asyncReregisterTask: Future[Unit] = null
  private val asyncReregisterLock = new Object

  // Field related to peer block managers that are necessary for block replication
  @volatile private var cachedPeers: Seq[BlockManagerId] = _
  private val peerFetchLock = new Object
  private var lastPeerFetchTime = 0L

  private var blockReplicationPolicy: BlockReplicationPolicy = _

上面代码中声明的BlockInfoManager用于管理BlockManager缓存BlockId及对应的BlockInfo，BlockInfoManager提供一些列的同步读写策略。BlockManager由以下部分组成。

1）shuffle客户端shuffleClient；

2）BlockManagerMaster，对存在于所有Executor上的BlockManager进行统一管理；

3）磁盘块管理器DiskBlockManager；

4）内存存储MemoryStore；

5）磁盘存储DiskStore；

BlockManager要生效必须要初始化，初始化代码如下，

def initialize(appId: String): Unit = {
    blockTransferService.init(this)
    shuffleClient.init(appId)

    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
    }

    val id =
      BlockManagerId(executorId, blockTransferService.hostName, blockTransferService.port, None)

    val idFromMaster = master.registerBlockManager(
      id,
      maxOnHeapMemory,
      maxOffHeapMemory,
      slaveEndpoint)

    blockManagerId = if (idFromMaster != null) idFromMaster else id

    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")
  }

1）BlockTransferService和shuffle客户端shuffleClient的初始化，ShuffleClien默认是BlockTransferService，当有外部的ShuffleService时，则调用外部的ExternalShuffleClient。

2）创建id为本地BlockManagerId，向BlockManagerMaster注册此id，获取从BlockManagerMaster的idFromMaster，如果idFromMaster为空则BlockManagerId为刚才创建的id，否则使用BlockManagerMaster注册到的idFromMaster。

3）ShuffleServerId的创建，当有外部的ShuffleService时，创建新的BlockManagerId作为ShuffleServerId。

4）当有外部的ShuffleService并且当前BlockMaId不是Driver端，则需要向ShuffleClient注册ShuffleServerId

2.Spark存储体系架构

1）1表示Executor的BlockManager与Driver的BlockManager进行消息通信，例如注册BlockManager、更新Block信息、获取Block所在的BlockManager、删除Executor等

2）2表示对BlockManager的读操作如get、doGetLocal等和写操作doPut、puSingle等

3）3表示当MemoryStore的内存不足时，写入DiskStore，而DiskStore实际依赖于DiskBlockManager

4）4表示通过访问远端节点的Executor的BlockManager中的TransportServer提供RPC服务下载或者上传Block

5）5表示远端节点的Executor的BlockManager访问本地Executor的BlockManager中的TransportServer提供的RPC服务下载或者上传Block。

二、shuffle服务与客户端

1.Block的RPC服务

当map任务与reduce任务处于不同的节点时，reduce任务需要从远端节点下载map任务的中间件输出，因此NettyBlockRpcServer提供打开，即下载Block文件的功能；一些情况下，为了容错，需要将Block的数据备份到其他节点上，所以NettyBlockRpcServer还提供了上传Block文件的RPC服务，实现见代码，代码位置：core/org.apache.spark.network.netty。

class NettyBlockRpcServer(
    appId: String,
    serializer: Serializer,
    blockManager: BlockDataManager)
  extends RpcHandler with Logging {

  private val streamManager = new OneForOneStreamManager()

  override def receive(
      client: TransportClient,
      rpcMessage: ByteBuffer,
      responseContext: RpcResponseCallback): Unit = {
    val message = BlockTransferMessage.Decoder.fromByteBuffer(rpcMessage)
    logTrace(s"Received request: $message")

    message match {
      case openBlocks: OpenBlocks =>
        val blocksNum = openBlocks.blockIds.length
        val blocks = for (i <- (0 until blocksNum).view)
          yield blockManager.getBlockData(BlockId.apply(openBlocks.blockIds(i)))
        val streamId = streamManager.registerStream(appId, blocks.iterator.asJava)
        logTrace(s"Registered streamId $streamId with $blocksNum buffers")
        responseContext.onSuccess(new StreamHandle(streamId, blocksNum).toByteBuffer)

      case uploadBlock: UploadBlock =>
        // StorageLevel and ClassTag are serialized as bytes using our JavaSerializer.
        val (level: StorageLevel, classTag: ClassTag[_]) = {
          serializer
            .newInstance()
            .deserialize(ByteBuffer.wrap(uploadBlock.metadata))
            .asInstanceOf[(StorageLevel, ClassTag[_])]
        }
        val data = new NioManagedBuffer(ByteBuffer.wrap(uploadBlock.blockData))
        val blockId = BlockId(uploadBlock.blockId)
        blockManager.putBlockData(blockId, data, level, classTag)
        responseContext.onSuccess(ByteBuffer.allocate(0))
    }
  }

  override def getStreamManager(): StreamManager = streamManager
}

 2.构造传输上下文TransportContext

代码位置，common/network-common/org.apache.spark.network

  public TransportContext(
      TransportConf conf,
      RpcHandler rpcHandler,
      boolean closeIdleConnections) {
    this.conf = conf;
    this.rpcHandler = rpcHandler;
    this.closeIdleConnections = closeIdleConnections;
  }

TransportContext既可以创建Netty服务，也可以创建Netty访问客户端，组成部分如下。

1）TransportConf：主要控制Netty框架提供的shuffle的I/O交互的客户端和服务端线程数量等

2）RpcHandler：负责shuffle的I/O服务端在接收到客户端的RPC请求后，提供打开Block或者上传Block的RPC处理，此处实现为NettyBlockRpcServer

3）是否关闭闲置连接

3.RPC客户端工厂TransportClientFactory

 public TransportClientFactory createClientFactory(List<TransportClientBootstrap> bootstraps) {
    return new TransportClientFactory(this, bootstraps);
  }

以下分析TransportClientFactory代码。

public TransportClientFactory(
      TransportContext context,
      List<TransportClientBootstrap> clientBootstraps) {
    this.context = Preconditions.checkNotNull(context);
    this.conf = context.getConf();
    this.clientBootstraps = Lists.newArrayList(Preconditions.checkNotNull(clientBootstraps));
    this.connectionPool = new ConcurrentHashMap<>();
    this.numConnectionsPerPeer = conf.numConnectionsPerPeer();
    this.rand = new Random();

    IOMode ioMode = IOMode.valueOf(conf.ioMode());
    this.socketChannelClass = NettyUtils.getClientChannelClass(ioMode);
    this.workerGroup = NettyUtils.createEventLoop(
        ioMode,
        conf.clientThreads(),
        conf.getModuleName() + "-client");
    this.pooledAllocator = NettyUtils.createPooledByteBufAllocator(
      conf.preferDirectBufs(), false /* allowCache */, conf.clientThreads());
  }

1）clientBootstraps：用于缓存客户端列表

2）connectionPool：用户缓存客户端连接

3）numConnectionsPerPeer：节点之间取数据的连接数，可以使用属性spark.shuffle.io.numConnectionsPerPeer来配置，默认为1

4）SocketChannelClass：客户端channel被创建时使用的类，可以使用属性spark.shuffle.io.mode来配置

5）workerGroup：根据Netty的规范，客户端只有worker组，所以此处创建workerGroup，实际是NioEventLoopGroup

6）pooledAllocator：汇集ByteBuf但对本地线程缓存禁用的分配器。

4.Netty服务器TransportServer

TransportServer提供了Netty实现的服务器端，用于提供RPC服务，如上传、下载等，代码如下。

 public TransportServer createServer(
      String host, int port, List<TransportServerBootstrap> bootstraps) {
    return new TransportServer(this, host, port, rpcHandler, bootstraps);
  }

TransportServer构造器如下

public TransportServer(
      TransportContext context,
      String hostToBind,
      int portToBind,
      RpcHandler appRpcHandler,
      List<TransportServerBootstrap> bootstraps) {
    this.context = context;
    this.conf = context.getConf();
    this.appRpcHandler = appRpcHandler;
    this.bootstraps = Lists.newArrayList(Preconditions.checkNotNull(bootstraps));

    try {
      init(hostToBind, portToBind);
    } catch (RuntimeException e) {
      JavaUtils.closeQuietly(this);
      throw e;
    }
  }

init方法对TransportServer进行初始化，通过使用Netty框架的EventLoopGroup、ServerBootstrap等API创建shuffle的I/O交互的服务端，主要代码见清单。

  private void init(String hostToBind, int portToBind) {

    IOMode ioMode = IOMode.valueOf(conf.ioMode());
    EventLoopGroup bossGroup =
      NettyUtils.createEventLoop(ioMode, conf.serverThreads(), conf.getModuleName() + "-server");
    EventLoopGroup workerGroup = bossGroup;

    PooledByteBufAllocator allocator = NettyUtils.createPooledByteBufAllocator(
      conf.preferDirectBufs(), true /* allowCache */, conf.serverThreads());

    bootstrap = new ServerBootstrap()
      .group(bossGroup, workerGroup)
      .channel(NettyUtils.getServerChannelClass(ioMode))
      .option(ChannelOption.ALLOCATOR, allocator)
      .childOption(ChannelOption.ALLOCATOR, allocator);

    if (conf.backLog() > 0) {
      bootstrap.option(ChannelOption.SO_BACKLOG, conf.backLog());
    }

    if (conf.receiveBuf() > 0) {
      bootstrap.childOption(ChannelOption.SO_RCVBUF, conf.receiveBuf());
    }

    if (conf.sendBuf() > 0) {
      bootstrap.childOption(ChannelOption.SO_SNDBUF, conf.sendBuf());
    }

    bootstrap.childHandler(new ChannelInitializer<SocketChannel>() {
      @Override
      protected void initChannel(SocketChannel ch) throws Exception {
        RpcHandler rpcHandler = appRpcHandler;
        for (TransportServerBootstrap bootstrap : bootstraps) {
          rpcHandler = bootstrap.doBootstrap(ch, rpcHandler);
        }
        context.initializePipeline(ch, rpcHandler);
      }
    });

    InetSocketAddress address = hostToBind == null ?
        new InetSocketAddress(portToBind): new InetSocketAddress(hostToBind, portToBind);
    channelFuture = bootstrap.bind(address);
    channelFuture.syncUninterruptibly();

    port = ((InetSocketAddress) channelFuture.channel().localAddress()).getPort();
    logger.debug("Shuffle server started on port: {}", port);
  }

1）ioMode：NIO或者EPOLL

2） ChannelOption.ALLOCATOR：在Netty 4中实现了一个新的ByteBuf内存池，它是一个纯Java版本的 jemalloc （Facebook也在用）。现在，Netty不会再因为用零填充缓冲区而浪费内存带宽了。不过，由于它不依赖于GC，开发人员需要小心内存泄漏。如果忘记在处理程序中释放缓冲区，那么内存使用率会无限地增长。Netty默认不使用内存池，需要在创建客户端或者服务端的时候进行指定，使用内存池之后，内存的申请和释放必须成对出现，即retain()和release()要成对出现，否则会导致内存泄露。

3）RpcHandler处理接收到的数据逻辑

5.获取远程shuffle文件

NettyBlockTransferService的fetchBlocks方法用于获取远程的shuffle文件，实际是使用NettyBlockTransferService中创建的Netty服务。

override def fetchBlocks(
      host: String,
      port: Int,
      execId: String,
      blockIds: Array[String],
      listener: BlockFetchingListener,
      tempShuffleFileManager: TempShuffleFileManager): Unit = {
    logTrace(s"Fetch blocks from $host:$port (executor id $execId)")
    try {
      val blockFetchStarter = new RetryingBlockFetcher.BlockFetchStarter {
        override def createAndStart(blockIds: Array[String], listener: BlockFetchingListener) {
          val client = clientFactory.createClient(host, port)
          new OneForOneBlockFetcher(client, appId, execId, blockIds, listener,
            transportConf, tempShuffleFileManager).start()
        }
      }

      val maxRetries = transportConf.maxIORetries()
      if (maxRetries > 0) {
        // Note this Fetcher will correctly handle maxRetries == 0; we avoid it just in case there's
        // a bug in this code. We should remove the if statement once we're sure of the stability.
        new RetryingBlockFetcher(transportConf, blockFetchStarter, blockIds, listener).start()
      } else {
        blockFetchStarter.createAndStart(blockIds, listener)
      }
    } catch {
      case e: Exception =>
        logError("Exception while beginning fetchBlocks", e)
        blockIds.foreach(listener.onBlockFetchFailure(_, e))
    }
  }

6.上传shuffle文件

NettyBlockTransferService的uploadBlock方法用于上传shuffle文件到远程的Executor，实际也是用NettyBlockTransferService中创建的Netty服务，步骤如下。

override def uploadBlock(
      hostname: String,
      port: Int,
      execId: String,
      blockId: BlockId,
      blockData: ManagedBuffer,
      level: StorageLevel,
      classTag: ClassTag[_]): Future[Unit] = {
    val result = Promise[Unit]()
    val client = clientFactory.createClient(hostname, port)

    // StorageLevel and ClassTag are serialized as bytes using our JavaSerializer.
    // Everything else is encoded using our binary protocol.
    val metadata = JavaUtils.bufferToArray(serializer.newInstance().serialize((level, classTag)))

    // Convert or copy nio buffer into array in order to serialize it.
    val array = JavaUtils.bufferToArray(blockData.nioByteBuffer())

    client.sendRpc(new UploadBlock(appId, execId, blockId.toString, metadata, array).toByteBuffer,
      new RpcResponseCallback {
        override def onSuccess(response: ByteBuffer): Unit = {
          logTrace(s"Successfully uploaded block $blockId")
          result.success((): Unit)
        }
        override def onFailure(e: Throwable): Unit = {
          logError(s"Error while uploading block $blockId", e)
          result.failure(e)
        }
      })

    result.future
  }

1）创建Netty服务的客户端，客户端连接的hostname和port正是BlockManager的hostname和port

2）将Block的存储级别StorageLevel和类标签序列化

3）将Block的ByteBuffer转化为数据，便于序列化

4）将appId、execId、blockId、metadata、转化为数组的Block封装为UploadBlock，并将其序列化为字节数组

5）最终调用Netty客户端的sendRpc方法将字节数组上传，回掉函数RpcResponseCallback根据RPC的结果更改上传状态。