Spark Core源码精读计划26 | 内存存储MemoryStore的具体实现

大数据真好玩

发布于 2019-08-21 15:32:52

7090

发布于 2019-08-21 15:32:52

前言
MemoryEntry
MemoryStore
- 构造与属性成员
- 直接写入字节
- 写入迭代器化的数据
- 读取字节与迭代器化的数据
- 淘汰缓存块
总结

前言

前面我们已经对内存池MemoryPool、内存管理器MemoryManager有了比较深入的了解，接下来要介绍的就是MemoryStore，它负责Spark内存存储的具体事项，将内存管理机制与存储块联系起来。本文先介绍与MemoryStore相关的MemoryEntry，然后详细分析MemoryStore的主要源码。

MemoryEntry

顾名思义，MemoryEntry就是内存中的一个“项”，或者说是块在内存中的抽象表示。它由一个特征定义。

代码#26.1 - o.a.s.memory.MemoryEntry特征

private sealed trait MemoryEntry[T] {
  def size: Long
  def memoryMode: MemoryMode
  def classTag: ClassTag[T]
}

其中，size表示该MemoryEntry代表的块大小，memoryMode表示块存储在堆内内存还是堆外内存，classTag则是该块所存储的对象的类型标记。MemoryEntry有序列化和反序列化的两种实现，如下所示。

代码#26.2 - o.a.s.memory.SerializedMemoryEntry/DeserializedMemoryEntry类

private case class DeserializedMemoryEntry[T](
    value: Array[T],
    size: Long,
    classTag: ClassTag[T]) extends MemoryEntry[T] {
  val memoryMode: MemoryMode = MemoryMode.ON_HEAP
}

private case class SerializedMemoryEntry[T](
    buffer: ChunkedByteBuffer,
    memoryMode: MemoryMode,
    classTag: ClassTag[T]) extends MemoryEntry[T] {
  def size: Long = buffer.size
}

可见，反序列化的DeserializedMemoryEntry只能用堆内内存存储，其数据是T类型的对象的数组。序列化的SerializedMemoryEntry能用堆内和堆外内存存储，数据用之前讲过的字节缓存ChunkedByteBuffer包装，并且其长度就是该SerializedMemoryEntry的大小。

MemoryStore

MemoryStore的内容比较多，仍然分块来看。

构造与属性成员

代码#26.3 - o.a.s.memory.MemoryStore类的构造与属性成员

private[spark] class MemoryStore(
    conf: SparkConf,
    blockInfoManager: BlockInfoManager,
    serializerManager: SerializerManager,
    memoryManager: MemoryManager,
    blockEvictionHandler: BlockEvictionHandler)
  extends Logging {

  private val entries = new LinkedHashMap[BlockId, MemoryEntry[_]](32, 0.75f, true)

  private val onHeapUnrollMemoryMap = mutable.HashMap[Long, Long]()
  private val offHeapUnrollMemoryMap = mutable.HashMap[Long, Long]()

  private val unrollMemoryThreshold: Long =
    conf.getLong("spark.storage.unrollMemoryThreshold", 1024 * 1024)

  private def maxMemory: Long = {
    memoryManager.maxOnHeapStorageMemory + memoryManager.maxOffHeapStorageMemory
  }

  if (maxMemory < unrollMemoryThreshold) {
    logWarning(s"Max memory ${Utils.bytesToString(maxMemory)} is less than the initial memory " +
      s"threshold ${Utils.bytesToString(unrollMemoryThreshold)} needed to store a block in " +
      s"memory. Please configure Spark with more memory.")
  }

  logInfo("MemoryStore started with capacity %s".format(Utils.bytesToString(maxMemory)))

  private def memoryUsed: Long = memoryManager.storageMemoryUsed

  def currentUnrollMemory: Long = memoryManager.synchronized {
    onHeapUnrollMemoryMap.values.sum + offHeapUnrollMemoryMap.values.sum
  }

  private def blocksMemoryUsed: Long = memoryManager.synchronized {
    memoryUsed - currentUnrollMemory
  }

可见，MemoryStore需要5个构造方法参数。前4个参数我们已经很熟悉了，不再多说。第5个参数是BlockEvictionHandler类型的，它实际上也是个特征，实现了该特征的类的作用就是将块从内存中淘汰掉。目前只有BlockManager实现了该特征，所以等到讲BlockManager时，再回头看它。

以下是MemoryStore类的属性成员：

entries：块ID与对应的MemoryEntry的映射关系，用LinkedHashMap结构存储，初始容量为32，负载因子0.75。
onHeapUnrollMemoryMap/offHeapUnrollMemoryMap：分别存储TaskAttempId与该Task在堆内、堆外内存占用的展开内存大小映射关系。
unrollMemoryThreshold：在展开块之前申请的初始展开内存大小，由spark.storage.unrollMemoryThreshold配置项来控制，默认1MB。

除此之外，还有四个Getter方法，它们负责返回对应内存的量：

maxMemory：堆内与堆外存储内存之和。如果内存管理器为StaticMemoryManager，该值为定值；如果内存管理器为UnifiedMemoryManager，该值会浮动。
memoryUsed：已经使用了的堆内与堆外存储内存之和。
currentUnrollMemory：当前展开内存占用的大小，由上面的onHeapUnrollMemoryMap/offHeapUnrollMemoryMap统计而来。
blocksMemoryUsed：当前除展开内存之外的存储内存（即真正存储块的内存）大小，即memoryUsed与currentUnrollMemory之差。

下面我们分别来看向MemoryStore写入以及从MemoryStore读取数据的方法。

直接写入字节

该方法名为putBytes()，代码如下。

代码#26.4 - o.a.s.memory.MemoryStore.putBytes()方法

  def putBytes[T: ClassTag](
      blockId: BlockId,
      size: Long,
      memoryMode: MemoryMode,
      _bytes: () => ChunkedByteBuffer): Boolean = {
    require(!contains(blockId), s"Block $blockId is already present in the MemoryStore")
    if (memoryManager.acquireStorageMemory(blockId, size, memoryMode)) {
      
      val bytes = _bytes()
      assert(bytes.size == size)
      val entry = new SerializedMemoryEntry[T](bytes, memoryMode, implicitly[ClassTag[T]])
      entries.synchronized {
        entries.put(blockId, entry)
      }
      logInfo("Block %s stored as bytes in memory (estimated size %s, free %s)".format(
        blockId, Utils.bytesToString(size), Utils.bytesToString(maxMemory - blocksMemoryUsed)))
      true
    } else {
      false
    }
  }

该方法的实现比较简单：首先调用MemoryManager.acquireStorageMemory()方法申请所需的内存，然后调用参数中传入的偏函数_bytes，获取已经转化为ChunkedByteBuffer的数据。再创建出对应的SerializedMemoryEntry，并将该MemoryEntry放入entries映射。注意LinkedHashMap本身不是线程安全的，因此对其并发访问都要加锁。

写入迭代器化的数据

所谓迭代器化的数据，就是指用Iterator[T]形式表示的块数据。之所以会这样表示，是因为有时单个块对应的数据可能过大，不能一次性存入内存。为了避免造成OOM，就可以一边遍历迭代器，一边周期性地写内存，并检查内存是否够用，就像翻书一样。“展开”（Unroll）这个词形象地说明了该过程，其对应的方法是putIteratorAsValues()与putIteratorAsBytes()，分别产生DeserializedMemoryEntry与SerializedMemoryEntry。由于两个方法的逻辑类似，因此我们只以putIteratorAsValues()来讲解。

代码很长，但我还是不“Unroll”了，全部放在下面。

代码#26.5 - o.a.s.memory.MemoryStore.putIteratorAsValues()

  private[storage] def putIteratorAsValues[T](
      blockId: BlockId,
      values: Iterator[T],
      classTag: ClassTag[T]): Either[PartiallyUnrolledIterator[T], Long] = {
    require(!contains(blockId), s"Block $blockId is already present in the MemoryStore")

    var elementsUnrolled = 0
    var keepUnrolling = true
    val initialMemoryThreshold = unrollMemoryThreshold
    val memoryCheckPeriod = conf.get(UNROLL_MEMORY_CHECK_PERIOD)
    var memoryThreshold = initialMemoryThreshold
    val memoryGrowthFactor = conf.get(UNROLL_MEMORY_GROWTH_FACTOR)
    var unrollMemoryUsedByThisBlock = 0L
    var vector = new SizeTrackingVector[T]()(classTag)

    keepUnrolling =
      reserveUnrollMemoryForThisTask(blockId, initialMemoryThreshold, MemoryMode.ON_HEAP)

    if (!keepUnrolling) {
      logWarning(s"Failed to reserve initial memory threshold of " +
        s"${Utils.bytesToString(initialMemoryThreshold)} for computing block $blockId in memory.")
    } else {
      unrollMemoryUsedByThisBlock += initialMemoryThreshold
    }

    while (values.hasNext && keepUnrolling) {
      vector += values.next()
      if (elementsUnrolled % memoryCheckPeriod == 0) {
        val currentSize = vector.estimateSize()
        if (currentSize >= memoryThreshold) {
          val amountToRequest = (currentSize * memoryGrowthFactor - memoryThreshold).toLong
          keepUnrolling =
            reserveUnrollMemoryForThisTask(blockId, amountToRequest, MemoryMode.ON_HEAP)
          if (keepUnrolling) {
            unrollMemoryUsedByThisBlock += amountToRequest
          }
          memoryThreshold += amountToRequest
        }
      }
      elementsUnrolled += 1
    }

    if (keepUnrolling) {
      val arrayValues = vector.toArray
      vector = null
      val entry =
        new DeserializedMemoryEntry[T](arrayValues, SizeEstimator.estimate(arrayValues), classTag)
      val size = entry.size

      def transferUnrollToStorage(amount: Long): Unit = {
        memoryManager.synchronized {
          releaseUnrollMemoryForThisTask(MemoryMode.ON_HEAP, amount)
          val success = memoryManager.acquireStorageMemory(blockId, amount, MemoryMode.ON_HEAP)
          assert(success, "transferring unroll memory to storage memory failed")
        }
      }

      val enoughStorageMemory = {
        if (unrollMemoryUsedByThisBlock <= size) {
          val acquiredExtra =
            memoryManager.acquireStorageMemory(
              blockId, size - unrollMemoryUsedByThisBlock, MemoryMode.ON_HEAP)
          if (acquiredExtra) {
            transferUnrollToStorage(unrollMemoryUsedByThisBlock)
          }
          acquiredExtra
        } else {
          val excessUnrollMemory = unrollMemoryUsedByThisBlock - size
          releaseUnrollMemoryForThisTask(MemoryMode.ON_HEAP, excessUnrollMemory)
          transferUnrollToStorage(size)
          true
        }
      }

      if (enoughStorageMemory) {
        entries.synchronized {
          entries.put(blockId, entry)
        }
        logInfo("Block %s stored as values in memory (estimated size %s, free %s)".format(
          blockId, Utils.bytesToString(size), Utils.bytesToString(maxMemory - blocksMemoryUsed)))
        Right(size)
      } else {
        assert(currentUnrollMemoryForThisTask >= unrollMemoryUsedByThisBlock,
          "released too much unroll memory")
        Left(new PartiallyUnrolledIterator(
          this,
          MemoryMode.ON_HEAP,
          unrollMemoryUsedByThisBlock,
          unrolled = arrayValues.toIterator,
          rest = Iterator.empty))
      }
    } else {
      logUnrollFailureMessage(blockId, vector.estimateSize())
      Left(new PartiallyUnrolledIterator(
        this,
        MemoryMode.ON_HEAP,
        unrollMemoryUsedByThisBlock,
        unrolled = vector.iterator,
        rest = values))
    }
  }

在具体看逻辑之前，先弄明白两个配置项：

UNROLL_MEMORY_CHECK_PERIOD，对应参数spark.storage.unrollMemoryCheckPeriod，表示在迭代过程中检查内存是否够用的周期，默认值16，即每16个元素检查一次。
UNROLL_MEMORY_GROWTH_FACTOR，对应参数spark.storage.unrollMemoryGrowthFactor，表示申请新的展开内存时扩展的倍数，默认值1.5。

然后就可以具体探究该方法的执行流程了：

调用reserveUnrollMemoryForThisTask()，申请初始的展开内存，并随时记录该块使用了多少展开内存。
循环迭代块的数据，将其放入一个SizeTrackingVector中。该数据结构可以动态估算其中存储的元素的大小，后面会详细分析。
每当到了检查的时机，如果已经展开的数据大小超过了当前的展开内存阈值，就再次调用reserveUnrollMemoryForThisTask()方法，试图申请新的展开内存（注意上面的扩展倍数的用法）。申请到之后，同时更新阈值。
所有数据都展开之后，标志keepUnrolling为真，表示展开成功。将SizeTrackingVector中的数据封装为DeserializedMemoryEntry。
检查申请到的展开内存是否比实际大小还大。如果是，就调用嵌套定义的transferUnrollToStorage()方法（实际又调用了releaseUnrollMemoryForThisTask()方法），释放掉多余的展开内存，并将它们返还给存储内存。
一切成功，将块ID与DeserializedMemoryEntry的映射放入entries，并返回Right。注意这个方法返回值的类型是Either类型，它在Scala中表示不相交的两个结果的集合，即可能返回错误的结果（Left），或者正确的结果（Right）。
如果没有足够的展开内存，或者展开所有数据后keepUnrolling标志为假，都表示这次写入不成功，返回Left，其中又包含PartiallyUnrolledIterator，表示一个没有完全展开的迭代器。

对于这种又臭又长的单个方法，多读几遍自然就能通顺。下面贴出它调用的申请与释放展开内存的方法，与上面的一大坨相比已经是毛毛雨了，不再赘述了。

代码#26.6 - o.a.s.memory.MemoryStore.reserveUnrollMemoryForThisTask()/releaseUnrollMemoryForThisTask()方法

  def reserveUnrollMemoryForThisTask(
      blockId: BlockId,
      memory: Long,
      memoryMode: MemoryMode): Boolean = {
    memoryManager.synchronized {
      val success = memoryManager.acquireUnrollMemory(blockId, memory, memoryMode)
      if (success) {
        val taskAttemptId = currentTaskAttemptId()
        val unrollMemoryMap = memoryMode match {
          case MemoryMode.ON_HEAP => onHeapUnrollMemoryMap
          case MemoryMode.OFF_HEAP => offHeapUnrollMemoryMap
        }
        unrollMemoryMap(taskAttemptId) = unrollMemoryMap.getOrElse(taskAttemptId, 0L) + memory
      }
      success
    }
  }

  def releaseUnrollMemoryForThisTask(memoryMode: MemoryMode, memory: Long = Long.MaxValue): Unit = {
    val taskAttemptId = currentTaskAttemptId()
    memoryManager.synchronized {
      val unrollMemoryMap = memoryMode match {
        case MemoryMode.ON_HEAP => onHeapUnrollMemoryMap
        case MemoryMode.OFF_HEAP => offHeapUnrollMemoryMap
      }
      if (unrollMemoryMap.contains(taskAttemptId)) {
        val memoryToRelease = math.min(memory, unrollMemoryMap(taskAttemptId))
        if (memoryToRelease > 0) {
          unrollMemoryMap(taskAttemptId) -= memoryToRelease
          memoryManager.releaseUnrollMemory(memoryToRelease, memoryMode)
        }
        if (unrollMemoryMap(taskAttemptId) == 0) {
          unrollMemoryMap.remove(taskAttemptId)
        }
      }
    }
  }

读取字节与迭代器化的数据

前者对应的是SerializedMemoryEntry，由getBytes()方法实现。后者对应的是DeserializedMemoryEntry，由getValues()方法实现。

代码#26.7 - o.a.s.memory.MemoryStore.getBytes()/getValues()方法

  def getBytes(blockId: BlockId): Option[ChunkedByteBuffer] = {
    val entry = entries.synchronized { entries.get(blockId) }
    entry match {
      case null => None
      case e: DeserializedMemoryEntry[_] =>
        throw new IllegalArgumentException("should only call getBytes on serialized blocks")
      case SerializedMemoryEntry(bytes, _, _) => Some(bytes)
    }
  }

  def getValues(blockId: BlockId): Option[Iterator[_]] = {
    val entry = entries.synchronized { entries.get(blockId) }
    entry match {
      case null => None
      case e: SerializedMemoryEntry[_] =>
        throw new IllegalArgumentException("should only call getValues on deserialized blocks")
      case DeserializedMemoryEntry(values, _, _) =>
        val x = Some(values)
        x.map(_.iterator)
    }
  }

淘汰缓存块

该方法名为evictBlocksToFreeSpace()，用途为淘汰现有的一些块，以为新的块腾出空间。它在StorageMemoryPool.acquireMemory()方法（代码#23.4）中调用，如果忘记了的话，可以返回去看看。这个方法的代码也比较长，但稍微容易理解一些。

代码#26.8 - o.a.s.memory.MemoryStore.evictBlocksToFreeSpace()方法

  private[spark] def evictBlocksToFreeSpace(
      blockId: Option[BlockId],
      space: Long,
      memoryMode: MemoryMode): Long = {
    assert(space > 0)
    memoryManager.synchronized {
      var freedMemory = 0L
      val rddToAdd = blockId.flatMap(getRddId)
      val selectedBlocks = new ArrayBuffer[BlockId]

      def blockIsEvictable(blockId: BlockId, entry: MemoryEntry[_]): Boolean = {
        entry.memoryMode == memoryMode && (rddToAdd.isEmpty || rddToAdd != getRddId(blockId))
      }

      entries.synchronized {
        val iterator = entries.entrySet().iterator()
        while (freedMemory < space && iterator.hasNext) {
          val pair = iterator.next()
          val blockId = pair.getKey
          val entry = pair.getValue
          if (blockIsEvictable(blockId, entry)) {
            if (blockInfoManager.lockForWriting(blockId, blocking = false).isDefined) {
              selectedBlocks += blockId
              freedMemory += pair.getValue.size
            }
          }
        }
      }

      def dropBlock[T](blockId: BlockId, entry: MemoryEntry[T]): Unit = {
        val data = entry match {
          case DeserializedMemoryEntry(values, _, _) => Left(values)
          case SerializedMemoryEntry(buffer, _, _) => Right(buffer)
        }
        val newEffectiveStorageLevel =
          blockEvictionHandler.dropFromMemory(blockId, () => data)(entry.classTag)
        if (newEffectiveStorageLevel.isValid) {
          blockInfoManager.unlock(blockId)
        } else {
          blockInfoManager.removeBlock(blockId)
        }
      }

      if (freedMemory >= space) {
        var lastSuccessfulBlock = -1
        try {
          logInfo(s"${selectedBlocks.size} blocks selected for dropping " +
            s"(${Utils.bytesToString(freedMemory)} bytes)")
          (0 until selectedBlocks.size).foreach { idx =>
            val blockId = selectedBlocks(idx)
            val entry = entries.synchronized {
              entries.get(blockId)
            }

            if (entry != null) {
              dropBlock(blockId, entry)
              afterDropAction(blockId)
            }
            lastSuccessfulBlock = idx
          }
          logInfo(s"After dropping ${selectedBlocks.size} blocks, " +
            s"free memory is ${Utils.bytesToString(maxMemory - blocksMemoryUsed)}")
          freedMemory
        } finally {
          if (lastSuccessfulBlock != selectedBlocks.size - 1) {
            (lastSuccessfulBlock + 1 until selectedBlocks.size).foreach { idx =>
              val blockId = selectedBlocks(idx)
              blockInfoManager.unlock(blockId)
            }
          }
        }
      } else {
        blockId.foreach { id =>
          logInfo(s"Will not store $id")
        }
        selectedBlocks.foreach { id =>
          blockInfoManager.unlock(id)
        }
        0L
      }
    }
  }

方法参数中的space就表示需要腾出多大的空间。其执行流程如下：

循环遍历entries映射中的块，找出其中能够被淘汰的块。所谓能够被淘汰，是指MemoryMode相同（即堆内对堆内，堆外对堆外，不能交叉），并且块ID对应的块数据不属于RDD。
为这些块加写锁，保证当前正在被读取的块不会被淘汰掉。记录将要被淘汰的块ID。
如果腾出的空间已经达到了目标值，就调用嵌套定义的dropBlock()方法真正地移除这些块，最终仍然调用了BlockManager.dropFromMemory()方法。该方法会产生两种结果：一是块仍然存在，只是StorageLevel发生变化（比如转存到了磁盘），就只需解开它的写锁；二是块被彻底地移除，就得调用BlockInfoManager.remove()方法删掉它。最后将剩余未处理的块解锁。
如果腾出的空间最终仍然不能达到目标值，就不会执行淘汰动作，新的块也不会被存入。