Spark存储体系底层架构剖析-Spark商业环境实战

Spark商业环境实战及调优进阶系列

1. Spark存储体系组件关系解释

BlockInfoManger 主要提供读写锁控制，层级仅仅位于BlockManger之下，通常Spark读写操作都先调用BlockManger，然后咨询BlockInfoManger是否存在锁竞争，然后才会调用DiskStore和MemStore，进而调用DiskBlockManger来确定数据与位置映射，或者调用 MemoryManger来确定内存池的软边界和内存使用申请。

image

1.1 Driver 与 Executor 与 SparkEnv 与 BlockManger 组件关系：

Driver与 Executor 组件各自拥有任务执行的SparkEnv环境，而每一个SparkEnv 中都有一个BlockManger负责存储服务，作为高层抽象，BlockManger 之间需要通过 RPCEnv，ShuffleClient，及BlocakTransferService相互通讯。

1.1 BlockInfoManger 与 BlockInfo 共享锁和排它锁读写控制关系：

BlockInfo中具有读写锁的标志，通过标志可以判断是否进行写控制

  val NO_WRITER: Long = -1
  val NON_TASK_WRITER: Long = -1024
  
 * The task attempt id of the task which currently holds the write lock for this block, or
 * [[BlockInfo.NON_TASK_WRITER]] if the write lock is held by non-task code, or
 * [[BlockInfo.NO_WRITER]] if this block is not locked for writing.
 
 def writerTask: Long = _writerTask
 def writerTask_=(t: Long): Unit = {
 _writerTask = t
    checkInvariants()

BlockInfoManager具有BlockId与BlockInfo的映射关系以及任务id与BlockId的锁映射：

 private[this] val infos = new mutable.HashMap[BlockId, BlockInfo]  
 
 *Tracks the set of blocks that each task has locked for writing.
 private[this] val writeLocksByTask = new mutable.HashMap[TaskAttemptId, mutable.Set[BlockId]]
                                       with mutable.MultiMap[TaskAttemptId, BlockId]
 
 *Tracks the set of blocks that each task has locked for reading, along with the number of times
 *that a block has been locked (since our read locks are re-entrant).
 private[this] val readLocksByTask =
 new mutable.HashMap[TaskAttemptId, ConcurrentHashMultiset[BlockId]]

1.3 DiskBlockManager 与 DiskStore 组件关系：

可以看到DiskStore内部会调用DiskBlockManager来确定Block的读写位置：

以下是DiskStore的抽象写操作，需要传入FileOutputStream => Unit高阶函数：

  def put(blockId: BlockId)(writeFunc: FileOutputStream => Unit): Unit = {
  if (contains(blockId)) {
  throw new IllegalStateException(s"Block $blockId is already present in the disk store")
  }
  logDebug(s"Attempting to put block $blockId")
  val startTime = System.currentTimeMillis
  
  val file = diskManager.getFile(blockId)
  
  val fileOutputStream = new FileOutputStream(file)
  var threwException: Boolean = true
  try {
      writeFunc(fileOutputStream)
      threwException = false
  } finally {
   try {
      Closeables.close(fileOutputStream, threwException)
   } finally {
   if (threwException) {
    remove(blockId)
          }
      }
  }
  val finishTime = System.currentTimeMillis
  logDebug("Block %s stored as %s file on disk in %d ms".format(
  file.getName,
  Utils.bytesToString(file.length()),
  finishTime - startTime))
  }

以下是DiskStore的读操作，调用DiskBlockManager来获取数据位置：

  def getBytes(blockId: BlockId): ChunkedByteBuffer = {
  
  val file = diskManager.getFile(blockId.name)
 
  val channel = new RandomAccessFile(file, "r").getChannel
  Utils.tryWithSafeFinally {
* For small files, directly read rather than memory map
  if (file.length < minMemoryMapBytes) {
  val buf = ByteBuffer.allocate(file.length.toInt)
  channel.position(0)
  while (buf.remaining() != 0) {
    if (channel.read(buf) == -1) {
      throw new IOException("Reached EOF before filling buffer\n" +
        s"offset=0\nfile=${file.getAbsolutePath}\nbuf.remaining=${buf.remaining}")
    }
  }
  buf.flip()
  new ChunkedByteBuffer(buf)
  } else {
  new ChunkedByteBuffer(channel.map(MapMode.READ_ONLY, 0, file.length))
      }
  } {
  channel.close()
   }
  }

1.3 MemManager 与 MemStore 与 MemoryPool 组件关系：

在这里要强调的是：第一代大数据框架hadoop只将内存作为计算资源，而Spark不仅将内存作为计算资源外，还将内存的一部分纳入存储体系：

内存池模型：逻辑上分为堆内存和堆外内存，然后堆内存（或堆外内存）内部又分为StorageMemoryPool和ExecutionMemoryPool。
MemManager是抽象的，定义了内存管理器的接口规范，方便以后扩展，比如：老版的StaticMemoryManager和新版的UnifiedMemoryManager.
MemStore 依赖于UnifiedMemoryManager进行内存的申请和软边界变化或内存释放。
MemStore 内部同时负责存储真实的对象，比如内部成员变量：entries ，建立了内存中的BlockId与MemoryEntry(Block的内存的形式)之间的映射。
MemStore 内部的“占座”行为，如：内部变量offHeapUnrollMemoryMap 和onHeapUnrollMemoryMap。

1.4 BlockManagerMaster 与 BlockManager 组件关系：

BlockManagerMaster的作用就是对存在于Dirver或Executor上的BlockManger进行统一管理，这简直是代理行为，因为他持有BlockManagerMasterEndpointREf，进而和BlockManagerMasterEndpoint进行通讯。

2. Spark存储体系组件BlockTransferServic传输服务

未完待续

3. 总结

存储体系是Spark的基石，我争取把每一块细微的知识点进行剖析，和大部分博客不同的是，我会尽量采用最平实的语言，毕竟技术就是一层窗户纸。

秦凯新 20181031 凌晨