spark源码分析之shufflemanager

1、shufflemanager的实现类：sortshufflemanager

Spark 0.8及以前 Hash Based Shuffle

在Shuffle Write过程按照Hash的方式重组Partition的数据，不进行排序。每个map端的任务为每个reduce端的Task生成一个文件，通常会产生大量的文件（即对应为M*R个中间文件，其中M表示map端的Task个数，R表示reduce端的Task个数），伴随大量的随机磁盘IO操作与大量的内存开销。

Shuffle Read过程如果有combiner操作，那么它会把拉到的数据保存在一个Spark封装的哈希表（AppendOnlyMap）中进行合并。在代码结构上：

• org.apache.spark.storage.ShuffleBlockManager负责Shuffle Write
• org.apache.spark.BlockStoreShuffleFetcher负责Shuffle Read
• org.apache.spark.Aggregator负责combine，依赖于AppendOnlyMap

Spark 0.8.1 为Hash Based Shuffle引入File Consolidation机制

通过文件合并，中间文件的生成方式修改为每个执行单位（一个Executor中的执行单位等于Core的个数除以每个Task所需的Core数）为每个reduce端的任务生成一个文件。最终可以将文件个数从M*R修改为E*C/T*R，其中，E表示Executor的个数，C表示每个Executor中可用Core的个数，T表示Task所分配的Core的个数。是否采用Consolidate机制，需要配置spark.shuffle.consolidateFiles参数

Spark 0.9 引入ExternalAppendOnlyMap

在combine的时候，可以将数据spill到磁盘，然后通过堆排序merge

Spark 1.1 引入Sort Based Shuffle，但默认仍为Hash Based Shuffle

在Sort Based Shuffle的Shuffle Write阶段，map端的任务会按照Partition id以及key对记录进行排序。同时将全部结果写到一个数据文件中，同时生成一个索引文件，reduce端的Task可以通过该索引文件获取相关的数据。

在代码结构上：

从以前的ShuffleBlockManager中分离出ShuffleManager来专门管理Shuffle Writer和Shuffle Reader。两种Shuffle方式分别对应

org.apache.spark.shuffle.hash.HashShuffleManager和

org.apache.spark.shuffle.sort.SortShuffleManager，

可通过spark.shuffle.manager参数配置。两种Shuffle方式有各自的ShuffleWriter：org.apache.spark.shuffle.hash.HashShuffle和org.apache.spark.shuffle.sort.SortShuffleWriter；但共用一个ShuffleReader，即org.apache.spark.shuffle.hash.HashShuffleReader。

org.apache.spark.util.collection.ExternalSorter实现排序功能。可通过对spark.shuffle.spill参数配置，决定是否可以在排序时将临时数据Spill到磁盘。

Spark 1.2 默认的Shuffle方式改为Sort Based Shuffle

Spark 1.4 引入Tungsten-Sort Based Shuffle

将数据记录用序列化的二进制方式存储，把排序转化成指针数组的排序，引入堆外内存空间和新的内存管理模型，这些技术决定了使用Tungsten-Sort要符合一些严格的限制，比如Shuffle dependency不能带有aggregation、输出不能排序等。由于堆外内存的管理基于JDK Sun Unsafe API，故Tungsten-Sort Based Shuffle也被称为Unsafe Shuffle。

在代码层面：

• 新增org.apache.spark.shuffle.unsafe.UnsafeShuffleManager
• 新增org.apache.spark.shuffle.unsafe.UnsafeShuffleWriter(用java实现)
• ShuffleReader复用HashShuffleReader

Spark 1.6 Tungsten-sort并入Sort Based Shuffle

由SortShuffleManager自动判断选择最佳Shuffle方式，如果检测到满足Tungsten-sort条件会自动采用Tungsten-sort Based Shuffle，否则采用Sort Based Shuffle。

在代码方面：

• UnsafeShuffleManager合并到SortShuffleManager
• HashShuffleReader 重命名为BlockStoreShuffleReader，Sort Based Shuffle和Hash Based Shuffle仍共用ShuffleReader。

Spark 2.0 Hash Based Shuffle退出历史舞台，从此Spark只有Sort Based Shuffle，ShuffleManager的实现类就只有SortShufflemanager

2、sortshufflemanager.registerShuffle

3、sortshufflemanager.getReader

4、sortshufflemanager.getWriter

case unsafeShuffleHandle: SerializedShuffleHandle] =>
    new UnsafeShuffleWriter(.......)
  case bypassMergeSortHandle: BypassMergeSortShuffleHandle=>
    new BypassMergeSortShuffleWriter(......)
  case other: BaseShuffleHandle =>
    new SortShuffleWriter(shuffleBlockResolver, other, mapId, context)

5、BypassMergeSortShuffleWriter类似于hash shuffle，但是将output file合并成一个文件

1）、BypassMergeSortShuffleWriter.write

传参：partition的itearator

【如果record为空】

if (!records.hasNext()) {
  partitionLengths = new long[numPartitions];
  shuffleBlockResolver.writeIndexFileAndCommit(shuffleId, mapId, partitionLengths, null);
  mapStatus = MapStatus$.MODULE$.apply(blockManager.shuffleServerId(), partitionLengths);
  return;
}

【获取partition写入磁盘文件的writer】

partitionWriters = new DiskBlockObjectWriter[numPartitions];
for (int i = 0; i < numPartitions; i++) {
  final Tuple2<TempShuffleBlockId, File> tempShuffleBlockIdPlusFile =
    blockManager.diskBlockManager().createTempShuffleBlock();
  final File file = tempShuffleBlockIdPlusFile._2();
  final BlockId blockId = tempShuffleBlockIdPlusFile._1();
  partitionWriters[i] =
    blockManager.getDiskWriter(blockId, file, serInstance, fileBufferSize, writeMetrics);
}

【写文件】

while (records.hasNext()) {
  final Product2<K, V> record = records.next();
  final K key = record._1();
  partitionWriters[partitioner.getPartition(key)].write(key, record._2());
}

【获取每个ShuffleBlock，ShuffleBlock被称为FileSegment】

partitionWriterSegments = new FileSegment[numPartitions];
for (int i = 0; i < numPartitions; i++) {
  final DiskBlockObjectWriter writer = partitionWriters[i];
  partitionWriterSegments[i] = writer.commitAndGet();
  writer.close();
}

【合并文件以及写index文件】

File output = shuffleBlockResolver.getDataFile(shuffleId, mapId);
File tmp = Utils.tempFileWith(output);
try {
  partitionLengths = writePartitionedFile(tmp);
  shuffleBlockResolver.writeIndexFileAndCommit(shuffleId, mapId, partitionLengths, tmp);
} finally {
  if (tmp.exists() && !tmp.delete()) {
    logger.error("Error while deleting temp file {}", tmp.getAbsolutePath());
  }
}

2）、BypassMergeSortShuffleWriter.writePartitionedFile

传参：合并文件 File outputFile

final FileOutputStream out = new FileOutputStream(outputFile, true);
  for (int i = 0; i < numPartitions; i++) {
    final File file = partitionWriterSegments[i].file();
    if (file.exists()) {
      final FileInputStream in = new FileInputStream(file);
      boolean copyThrewException = true;
      try {
        lengths[i] = Utils.copyStream(in, out, false, transferToEnabled);
        copyThrewException = false;
      } finally {
        Closeables.close(in, copyThrewException);
      }
      if (!file.delete()) {
        logger.error("Unable to delete file for partition {}", i);
      }
    }
  }
  threwException = false;

返回文件的偏移量

6、SortShuffleWriter

1）、SortShuffleWriter.writer

【对rdd进行排序】

如果map端进行combine则，反之则不关心key在每个partition中是否被排序，既不传递aggregator也不传递ordering 
sorter = if (dep.mapSideCombine) {
  require(dep.aggregator.isDefined, "Map-side combine without Aggregator specified!")
//需要combine时，传递partitioner以及ordering
  new ExternalSorter[K, V, C](
    context, dep.aggregator, Some(dep.partitioner), dep.keyOrdering, dep.serializer)
} else {
  new ExternalSorter[K, V, V](
    context, aggregator = None, Some(dep.partitioner), ordering = None, dep.serializer)
}
//将数据存储在buffer或者map中，这是最关键的地方，根据需求（包括partition内的key排序，partitionID排序等等）排序，内存不够时数据会spill后写入spillfile
sorter.insertAll(records)

【写文件】

//blockManager.diskBlockManager.getFile(ShuffleDataBlockId(shuffleId, mapId, NOOP_REDUCE_ID))
val output = shuffleBlockResolver.getDataFile(dep.shuffleId, mapId)
//通过工具类创建临时文件
val tmp = Utils.tempFileWith(output)
try {
  val blockId = ShuffleBlockId(dep.shuffleId, mapId, IndexShuffleBlockResolver.NOOP_REDUCE_ID)
//将buffer或者map中的数据写入文件，各个partition
  val partitionLengths = sorter.writePartitionedFile(blockId, tmp)
//写index文件
  shuffleBlockResolver.writeIndexFileAndCommit(dep.shuffleId, mapId, partitionLengths, tmp)
  mapStatus = MapStatus(blockManager.shuffleServerId, partitionLengths)
} finally {
  if (tmp.exists() && !tmp.delete()) {
    logError(s"Error while deleting temp file ${tmp.getAbsolutePath}")
  }
}