spark core 2.0 SerializedShuffleHandle UnsafeShuffleWriter ShuffleExternalSorter

使用序列化Shuffle的三个条件：

 1. dependency.serializer.supportsRelocationOfSerializedObjects.

2.  !dependency.aggregator.isDefined,即不使用聚合函数。

3. !numPartitions > MAX_SHUFFLE_OUTPUT_PARTITIONS_FOR_SERIALIZED_MODE，即分区的数量小于1677216.

/**
   * Helper method for determining whether a shuffle should use an optimized serialized shuffle
   * path or whether it should fall back to the original path that operates on deserialized objects.
   */
  def canUseSerializedShuffle(dependency: ShuffleDependency[_, _, _]): Boolean = {
    val shufId = dependency.shuffleId
    val numPartitions = dependency.partitioner.numPartitions
    if (!dependency.serializer.supportsRelocationOfSerializedObjects) {
      log.debug(s"Can't use serialized shuffle for shuffle $shufId because the serializer, " +
        s"${dependency.serializer.getClass.getName}, does not support object relocation")
      false
    } else if (dependency.aggregator.isDefined) {
      log.debug(
        s"Can't use serialized shuffle for shuffle $shufId because an aggregator is defined")
      false
    } else if (numPartitions > MAX_SHUFFLE_OUTPUT_PARTITIONS_FOR_SERIALIZED_MODE) {
      log.debug(s"Can't use serialized shuffle for shuffle $shufId because it has more than " +
        s"$MAX_SHUFFLE_OUTPUT_PARTITIONS_FOR_SERIALIZED_MODE partitions")
      false
    } else {
      log.debug(s"Can use serialized shuffle for shuffle $shufId")
      true
    }
  }

如果handle是SerializedShuffleHandle ，则使用 UnsafeShuffleWriter来写数据。

 handle match {
      case unsafeShuffleHandle: SerializedShuffleHandle[K @unchecked, V @unchecked] =>
        new UnsafeShuffleWriter(
          env.blockManager,
          shuffleBlockResolver.asInstanceOf[IndexShuffleBlockResolver],
          context.taskMemoryManager(),
          unsafeShuffleHandle,
          mapId,
          context,
          env.conf)

UnsafeShuffleWriter的构造方法如下：

 public UnsafeShuffleWriter(
      BlockManager blockManager,
      IndexShuffleBlockResolver shuffleBlockResolver,
      TaskMemoryManager memoryManager,
      SerializedShuffleHandle<K, V> handle,
      int mapId,
      TaskContext taskContext,
      SparkConf sparkConf) throws IOException {
    final int numPartitions = handle.dependency().partitioner().numPartitions();
    if (numPartitions > SortShuffleManager.MAX_SHUFFLE_OUTPUT_PARTITIONS_FOR_SERIALIZED_MODE()) {
      throw new IllegalArgumentException(
        "UnsafeShuffleWriter can only be used for shuffles with at most " +
        SortShuffleManager.MAX_SHUFFLE_OUTPUT_PARTITIONS_FOR_SERIALIZED_MODE() +
        " reduce partitions");
    }
    this.blockManager = blockManager;
    this.shuffleBlockResolver = shuffleBlockResolver;
    this.memoryManager = memoryManager;
    this.mapId = mapId;
    final ShuffleDependency<K, V, V> dep = handle.dependency();
    this.shuffleId = dep.shuffleId();
    this.serializer = dep.serializer().newInstance();
    this.partitioner = dep.partitioner();
    this.writeMetrics = taskContext.taskMetrics().shuffleWriteMetrics();
    this.taskContext = taskContext;
    this.sparkConf = sparkConf;
    this.transferToEnabled = sparkConf.getBoolean("spark.file.transferTo", true);
    open();
  }

构造方法最后调用的  open方法如下：

  private void open() throws IOException {
    assert (sorter == null);
    sorter = new ShuffleExternalSorter(
      memoryManager,
      blockManager,
      taskContext,
      INITIAL_SORT_BUFFER_SIZE,
      partitioner.numPartitions(),
      sparkConf,
      writeMetrics);
    serBuffer = new MyByteArrayOutputStream(1024 * 1024);
    serOutputStream = serializer.serializeStream(serBuffer);
  }

ShuffleExternalSorter 为基于排序的shuflle的一个外部排序器。

/**
 * An external sorter that is specialized for sort-based shuffle.
 * <p>
 * Incoming records are appended to data pages. When all records have been inserted (or when the
 * current thread's shuffle memory limit is reached), the in-memory records are sorted according to
 * their partition ids (using a {@link ShuffleInMemorySorter}). The sorted records are then
 * written to a single output file (or multiple files, if we've spilled). The format of the output
 * files is the same as the format of the final output file written by
 * {@link org.apache.spark.shuffle.sort.SortShuffleWriter}: each output partition's records are
 * written as a single serialized, compressed stream that can be read with a new decompression and
 * deserialization stream.
 * <p>
 * Unlike {@link org.apache.spark.util.collection.ExternalSorter}, this sorter does not merge its
 * spill files. Instead, this merging is performed in {@link UnsafeShuffleWriter}, which uses a
 * specialized merge procedure that avoids extra serialization/deserialization.
 */
final class ShuffleExternalSorter extends MemoryConsumer {

The constructor method of ShuffleExternalSorter:

ShuffleExternalSorter(
      TaskMemoryManager memoryManager,
      BlockManager blockManager,
      TaskContext taskContext,
      int initialSize,
      int numPartitions,
      SparkConf conf,
      ShuffleWriteMetrics writeMetrics) {
    super(memoryManager,
      (int) Math.min(PackedRecordPointer.MAXIMUM_PAGE_SIZE_BYTES, memoryManager.pageSizeBytes()),
      memoryManager.getTungstenMemoryMode());
    this.taskMemoryManager = memoryManager;
    this.blockManager = blockManager;
    this.taskContext = taskContext;
    this.numPartitions = numPartitions;
    // Use getSizeAsKb (not bytes) to maintain backwards compatibility if no units are provided
    this.fileBufferSizeBytes = (int) conf.getSizeAsKb("spark.shuffle.file.buffer", "32k") * 1024;
    this.numElementsForSpillThreshold =
      conf.getLong("spark.shuffle.spill.numElementsForceSpillThreshold", 1024 * 1024 * 1024);
    this.writeMetrics = writeMetrics;
    this.inMemSorter = new ShuffleInMemorySorter(
      this, initialSize, conf.getBoolean("spark.shuffle.sort.useRadixSort", true));
    this.peakMemoryUsedBytes = getMemoryUsage();
  }

The main purpose of MyByteArrayOutputStream  is let buf can be accessed outside.

  /** Subclass of ByteArrayOutputStream that exposes `buf` directly. */
  private static final class MyByteArrayOutputStream extends ByteArrayOutputStream {
    MyByteArrayOutputStream(int size) { super(size); }
    public byte[] getBuf() { return buf; }
  }

 @Override
  public void write(scala.collection.Iterator<Product2<K, V>> records) throws IOException {
    // Keep track of success so we know if we encountered an exception
    // We do this rather than a standard try/catch/re-throw to handle
    // generic throwables.
    boolean success = false;
    try {
      while (records.hasNext()) {
        insertRecordIntoSorter(records.next());
      }
      closeAndWriteOutput();
      success = true;
    } finally {
      if (sorter != null) {
        try {
          sorter.cleanupResources();
        } catch (Exception e) {
          // Only throw this error if we won't be masking another
          // error.
          if (success) {
            throw e;
          } else {
            logger.error("In addition to a failure during writing, we failed during " +
                         "cleanup.", e);
          }
        }
      }
    }
  }

/**
   * Write a record to the shuffle sorter.
   */
  public void insertRecord(Object recordBase, long recordOffset, int length, int partitionId)
    throws IOException {

    // for tests
    assert(inMemSorter != null);
    if (inMemSorter.numRecords() >= numElementsForSpillThreshold) {
      logger.info("Spilling data because number of spilledRecords crossed the threshold " +
        numElementsForSpillThreshold);
      spill();
    }

    growPointerArrayIfNecessary();
    // Need 4 bytes to store the record length.
    final int required = length + 4;
    acquireNewPageIfNecessary(required);

    assert(currentPage != null);
    final Object base = currentPage.getBaseObject();
    final long recordAddress = taskMemoryManager.encodePageNumberAndOffset(currentPage, pageCursor);
    Platform.putInt(base, pageCursor, length);
    pageCursor += 4;
    Platform.copyMemory(recordBase, recordOffset, base, pageCursor, length);
    pageCursor += length;
    inMemSorter.insertRecord(recordAddress, partitionId);
  }

/**
   * Sort and spill the current records in response to memory pressure.
   */
  @Override
  public long spill(long size, MemoryConsumer trigger) throws IOException {
    if (trigger != this || inMemSorter == null || inMemSorter.numRecords() == 0) {
      return 0L;
    }

    logger.info("Thread {} spilling sort data of {} to disk ({} {} so far)",
      Thread.currentThread().getId(),
      Utils.bytesToString(getMemoryUsage()),
      spills.size(),
      spills.size() > 1 ? " times" : " time");

    writeSortedFile(false);
    final long spillSize = freeMemory();
    inMemSorter.reset();
    // Reset the in-memory sorter's pointer array only after freeing up the memory pages holding the
    // records. Otherwise, if the task is over allocated memory, then without freeing the memory
    // pages, we might not be able to get memory for the pointer array.
    taskContext.taskMetrics().incMemoryBytesSpilled(spillSize);
    return spillSize;
  }

 /**
   * Sorts the in-memory records and writes the sorted records to an on-disk file.
   * This method does not free the sort data structures.
   *
   * @param isLastFile if true, this indicates that we're writing the final output file and that the
   *                   bytes written should be counted towards shuffle spill metrics rather than
   *                   shuffle write metrics.
   */
  private void writeSortedFile(boolean isLastFile) throws IOException {

    final ShuffleWriteMetrics writeMetricsToUse;

    if (isLastFile) {
      // We're writing the final non-spill file, so we _do_ want to count this as shuffle bytes.
      writeMetricsToUse = writeMetrics;
    } else {
      // We're spilling, so bytes written should be counted towards spill rather than write.
      // Create a dummy WriteMetrics object to absorb these metrics, since we don't want to count
      // them towards shuffle bytes written.
      writeMetricsToUse = new ShuffleWriteMetrics();
    }

    // This call performs the actual sort.
    final ShuffleInMemorySorter.ShuffleSorterIterator sortedRecords =
      inMemSorter.getSortedIterator();

    // Currently, we need to open a new DiskBlockObjectWriter for each partition; we can avoid this
    // after SPARK-5581 is fixed.
    DiskBlockObjectWriter writer;

    // Small writes to DiskBlockObjectWriter will be fairly inefficient. Since there doesn't seem to
    // be an API to directly transfer bytes from managed memory to the disk writer, we buffer
    // data through a byte array. This array does not need to be large enough to hold a single
    // record;
    final byte[] writeBuffer = new byte[DISK_WRITE_BUFFER_SIZE];

    // Because this output will be read during shuffle, its compression codec must be controlled by
    // spark.shuffle.compress instead of spark.shuffle.spill.compress, so we need to use
    // createTempShuffleBlock here; see SPARK-3426 for more details.
    final Tuple2<TempShuffleBlockId, File> spilledFileInfo =
      blockManager.diskBlockManager().createTempShuffleBlock();
    final File file = spilledFileInfo._2();
    final TempShuffleBlockId blockId = spilledFileInfo._1();
    final SpillInfo spillInfo = new SpillInfo(numPartitions, file, blockId);

    // Unfortunately, we need a serializer instance in order to construct a DiskBlockObjectWriter.
    // Our write path doesn't actually use this serializer (since we end up calling the `write()`
    // OutputStream methods), but DiskBlockObjectWriter still calls some methods on it. To work
    // around this, we pass a dummy no-op serializer.
    final SerializerInstance ser = DummySerializerInstance.INSTANCE;

    writer = blockManager.getDiskWriter(blockId, file, ser, fileBufferSizeBytes, writeMetricsToUse);

    int currentPartition = -1;
    while (sortedRecords.hasNext()) {
      sortedRecords.loadNext();
      final int partition = sortedRecords.packedRecordPointer.getPartitionId();
      assert (partition >= currentPartition);
      if (partition != currentPartition) {
        // Switch to the new partition
        if (currentPartition != -1) {
          writer.commitAndClose();
          spillInfo.partitionLengths[currentPartition] = writer.fileSegment().length();
        }
        currentPartition = partition;
        writer =
          blockManager.getDiskWriter(blockId, file, ser, fileBufferSizeBytes, writeMetricsToUse);
      }

      final long recordPointer = sortedRecords.packedRecordPointer.getRecordPointer();
      final Object recordPage = taskMemoryManager.getPage(recordPointer);
      final long recordOffsetInPage = taskMemoryManager.getOffsetInPage(recordPointer);
      int dataRemaining = Platform.getInt(recordPage, recordOffsetInPage);
      long recordReadPosition = recordOffsetInPage + 4; // skip over record length
      while (dataRemaining > 0) {
        final int toTransfer = Math.min(DISK_WRITE_BUFFER_SIZE, dataRemaining);
        Platform.copyMemory(
          recordPage, recordReadPosition, writeBuffer, Platform.BYTE_ARRAY_OFFSET, toTransfer);
        writer.write(writeBuffer, 0, toTransfer);
        recordReadPosition += toTransfer;
        dataRemaining -= toTransfer;
      }
      writer.recordWritten();
    }

    if (writer != null) {
      writer.commitAndClose();
      // If `writeSortedFile()` was called from `closeAndGetSpills()` and no records were inserted,
      // then the file might be empty. Note that it might be better to avoid calling
      // writeSortedFile() in that case.
      if (currentPartition != -1) {
        spillInfo.partitionLengths[currentPartition] = writer.fileSegment().length();
        spills.add(spillInfo);
      }
    }

    if (!isLastFile) {  // i.e. this is a spill file
      // The current semantics of `shuffleRecordsWritten` seem to be that it's updated when records
      // are written to disk, not when they enter the shuffle sorting code. DiskBlockObjectWriter
      // relies on its `recordWritten()` method being called in order to trigger periodic updates to
      // `shuffleBytesWritten`. If we were to remove the `recordWritten()` call and increment that
      // counter at a higher-level, then the in-progress metrics for records written and bytes
      // written would get out of sync.
      //
      // When writing the last file, we pass `writeMetrics` directly to the DiskBlockObjectWriter;
      // in all other cases, we pass in a dummy write metrics to capture metrics, then copy those
      // metrics to the true write metrics here. The reason for performing this copying is so that
      // we can avoid reporting spilled bytes as shuffle write bytes.
      //
      // Note that we intentionally ignore the value of `writeMetricsToUse.shuffleWriteTime()`.
      // Consistent with ExternalSorter, we do not count this IO towards shuffle write time.
      // This means that this IO time is not accounted for anywhere; SPARK-3577 will fix this.
      writeMetrics.incRecordsWritten(writeMetricsToUse.recordsWritten());
      taskContext.taskMetrics().incDiskBytesSpilled(writeMetricsToUse.bytesWritten());
    }
  }

/**
 * Metadata for a block of data written by {@link ShuffleExternalSorter}.
 */
final class SpillInfo {
  final long[] partitionLengths;
  final File file;
  final TempShuffleBlockId blockId;

  SpillInfo(int numPartitions, File file, TempShuffleBlockId blockId) {
    this.partitionLengths = new long[numPartitions];
    this.file = file;
    this.blockId = blockId;
  }
}