1.  /**
      *使用Hash的ShuffleManager具体实现子类，针对每个Mapper都会为各个Reduce分
      *区构建一个输出文件（也可能是多个任务复用文件）
2.    */
3.  private[spark]      class    HashShuffleManager(conf:         SparkConf)   extends
    ShuffleManager with Logging {
4.  ......

1.  private[spark] class HashShuffleWriter[K, V](
2.      shuffleBlockResolver: FileShuffleBlockResolver,
3.      handle: BaseShuffleHandle[K, V, _],
4.  mapId: Int,
5.      context: TaskContext)
6.    extends ShuffleWriter[K, V] with Logging {
7.
8.    //控制每个Writer输出时的切片个数，对应分区个数
9.    private val dep = handle.dependency
10.   private val numOutputSplits = dep.partitioner.numPartitions
11.
12.   ......
13.   //获取数据读写的块管理器
14.   private val blockManager = SparkEnv.get.blockManager
15.   private val ser = Serializer.getSerializer(dep.serializer.getOrElse(null))
16.
17. //从  FileShuffleBlockResolver      的 forMapTask方法中获取指定的    shuffleId  对应
    //的mapId
18. //对应分区个数构建的数据块写的ShuffleWriterGroup实例
19.   private val shuffle = shuffleBlockResolver.forMapTask(dep.shuffleId,
      mapId, numOutputSplits, ser, writeMetrics)
20.
21.
22.   /** Task输出时一组记录的写入 */
23.
24.   override def write(records: Iterator[Product2[K, V]]): Unit = {
25. //判断在写时是否需要先聚合，即定义了Map端Combine时，先对数据进行聚合再写入，否则
    //直接返回需要写入的一批记录
26.
27.   val iter = if (dep.aggregator.isDefined) {
28.       if (dep.mapSideCombine) {
29.         dep.aggregator.get.combineValuesByKey(records, context)
30.       } else {
31.         records
32.       }
33.     } else {
34.       require(!dep.mapSideCombine, "Map-side combine without Aggregator
          specified!")
35.       records
36.     }
37.
38.     //根据分区器，获取每条记录对应的bucketId（即所在Reduce序号），根据bucketId
        //从FileShuffleBlockResolver构建的ShuffleWriterGroup中，获取DiskBlock-
        //ObjectWriter实例，对应磁盘数据块的数据写入器
39.     for (elem<- iter) {
40.       val bucketId = dep.partitioner.getPartition(elem._1)
41.       shuffle.writers(bucketId).write(elem._1, elem._2)
42.     }
43.   }
44.      ......
45. }

1.  /**
2.    *针对给定的 Map Task，指定一个ShuffleWriterGroup实例，在数据块写入器成功
3.    *关闭时，会注册为完成状态
4.    */
5.
6.
7.   def forMapTask(shuffleId: Int, mapId: Int, numReduces: Int, serializer:
     Serializer,
8.  writeMetrics: ShuffleWriteMetrics): ShuffleWriterGroup = {
9.      new ShuffleWriterGroup {
10.       //在FileShuffleBlockResolver中维护着当前Map Task对应shuffleId标识的
          //Shuffle中，指定numReduces个数的Reduce的各个状态
11.
12. shuffleStates.putIfAbsent(shuffleId, new ShuffleState(numReduces))
13.       private val shuffleState = shuffleStates(shuffleId)
14.
15. ......
16.           //根据Reduce端的任务个数，构建元素类型为DiskBlockObjectWriter的数组，
              //DiskBlockObjectWriter负责具体数据的磁盘写入
17.    //原则上，Shuffle的输出可以存放在各种提供存储机制的系统上，但为了容错性等方面的
       //考虑，目前的Shuffle实行机制都会写入到磁盘中
18.
19.  val writers: Array[DiskBlockObjectWriter] = {
20.         //这里的逻辑Bucket的Id值即对应的Reduce的任务序号，或者说分区ID
21.         Array.tabulate[DiskBlockObjectWriter](numReduces) { bucketId =>
22.         //针对每个Map端分区的Id与Bucket的Id构建数据块的逻辑标识
23.           val blockId = ShuffleBlockId(shuffleId, mapId, bucketId)
24.           val blockFile = blockManager.diskBlockManager.getFile(blockId)
25.           val tmp = Utils.tempFileWith(blockFile)
26.blockManager.getDiskWriter(blockId, tmp, serializerInstance, bufferSize,
    writeMetrics)
27.         }
28.       }
29. ......
30.       //任务完成时回调的释放写入器方法
31.       override def releaseWriters(success: Boolean) {
32. shuffleState.completedMapTasks.add(mapId)
33.       }
34.     }
35.   }

1.  case class ShuffleBlockId(shuffleId: Int, mapId: Int, reduceId: Int)
    extends BlockId {
2.    override def name: String = "shuffle_" + shuffleId + "_" + mapId + "_"
      + reduceId
3.  }

1.  /**
2.    *获取一个针对特定Map Task的ShuffleWriterGroup
3.    */
4.
5.
6.    def forMapTask(shuffleId: Int, mapId: Int, numBuckets: Int, serializer:
      Serializer,
7.  writeMetrics: ShuffleWriteMetrics): ShuffleWriterGroup = {
8.      new ShuffleWriterGroup {
9.  ......
10.       val writers: Array[DiskBlockObjectWriter] = if
         (consolidateShuffleFiles) {
11. //获取未使用的文件组
12. fileGroup = getUnusedFileGroup()
13.         Array.tabulate[DiskBlockObjectWriter](numBuckets) { bucketId =>
14. val blockId = ShuffleBlockId(shuffleId, mapId, bucketId)
15. //注意获取磁盘写入器时，传入的第二个参数与未使用文件合并机制时的差异
16. //fileGroup(bucketId)：构造器方式调用，对应apply的方法调用
17.blockManager.getDiskWriter(blockId, fileGroup(bucketId), serializerInstance,
    bufferSize,
18. writeMetrics)
19.         }
20.       } else {
21.         Array.tabulate[DiskBlockObjectWriter](numBuckets) { bucketId =>
22. val blockId = ShuffleBlockId(shuffleId, mapId, bucketId)
23. //根据ShuffleBlockId信息获取文件名
24.           val blockFile = blockManager.diskBlockManager.getFile(blockId)
25.           val tmp = Utils.tempFileWith(blockFile)
26.blockManager.getDiskWriter(blockId, tmp, serializerInstance, bufferSize,
    writeMetrics)
27.         }
28.       }
29. ......
30. writeMetrics.incShuffleWriteTime(System.nanoTime - openStartTime)
31.       override def releaseWriters(success: Boolean) {
32. //带文件合并机制时，写入器在释放后的处理
33. //3个关键信息mapId、offsets、lengths
34.         if (consolidateShuffleFiles) {
35.           if (success) {
36.             val offsets = writers.map(_.fileSegment().offset)
37.             val lengths = writers.map(_.fileSegment().length)
38. fileGroup.recordMapOutput(mapId, offsets, lengths)
39.           }
40. //回收文件组，便于后续复用
41. recycleFileGroup(fileGroup)
42.         } else {
43. shuffleState.completedMapTasks.add(mapId)
44.         }
45.       }

1.  private def getUnusedFileGroup(): ShuffleFileGroup = {
2.  //获取已经构建但未使用的文件组，如果获取失败，则重新构建一个文件组
3.          val fileGroup = shuffleState.unusedFileGroups.poll()
4.          if (fileGroup != null) fileGroup else newFileGroup()
5.        }
6.  //重新构建一个文件组的源码
7.        private def newFileGroup(): ShuffleFileGroup = {
8.  //构建后会对文件编号进行递增，该文件编号最终用在生成的文件名中
9.          val fileId = shuffleState.nextFileId.getAndIncrement()
10.         val files = Array.tabulate[File](numBuckets) { bucketId =>
11. //最终的文件名，可以通过文件名的组成及取值细节，加深对实现细节在文件个数上的差异的理解
12.
13.           val filename = physicalFileName(shuffleId, bucketId, fileId)
14. blockManager.diskBlockManager.getFile(filename)
15.         }
16. //构建并添加到shuffleState中，便于后续复用
17.         val fileGroup = new ShuffleFileGroup(shuffleId, fileId, files)
18. shuffleState.allFileGroups.add(fileGroup)
19. fileGroup
20.       }

1.  private def physicalFileName(shuffleId: Int, bucketId: Int, fileId: Int) = {
2.      "merged_shuffle_%d_%d_%d".format(shuffleId, bucketId, fileId)
3.    }

1.  private def resourceOfferSingleTaskSet(
2.  taskSet: TaskSetManager,
3.  maxLocality: TaskLocality,
4.  shuffledOffers: Seq[WorkerOffer],
5.  availableCpus: Array[Int],
6.      tasks: Seq[ArrayBuffer[TaskDescription]]) : Boolean = {
7.  var launchedTask = false
8.    for (i <- 0 until shuffledOffers.size) {
9.  val execId = shuffledOffers(i).executorId
10. val host = shuffledOffers(i).host
11. //判断当前Executor工作点上可用的内核个数是否满足Task所需的内核个数
12. //CPUS_PER_TASK：表示设置的每个Task所需的内核个数
13. if (availableCpus(i) >= CPUS_PER_TASK) {
14. try {
15. for (task <- taskSet.resourceOffer(execId, host, maxLocality)) {
16. ......
17. launchedTask = true
18. }
19.      } catch {
20. ......
21.       }
22.     }
23.   }
24. return launchedTask
25. }

1.  //每个任务请求的CPU个数
2.  val CPUS_PER_TASK = conf.getInt("spark.task.cpus", 1)