本文基于 Spark 2.1 进行解析
从 Spark 2.0 开始移除了Hash Based Shuffle,想要了解可参考Shuffle 过程,本文将讲解 Sort Based Shuffle。
ShuffleMapTask的结果(ShuffleMapStage中FinalRDD的数据)都将写入磁盘,以供后续Stage拉取,即整个Shuffle包括前Stage的Shuffle Write和后Stage的Shuffle Read,由于内容较多,本文先解析Shuffle Write。
概述:
执行一个ShuffleMapTask最终的执行逻辑是调用了ShuffleMapTask类的runTask()方法:
override def runTask(context: TaskContext): MapStatus = {
// Deserialize the RDD using the broadcast variable.
val deserializeStartTime = System.currentTimeMillis()
val ser = SparkEnv.get.closureSerializer.newInstance()
// 从广播变量中反序列化出finalRDD和dependency
val (rdd, dep) = ser.deserialize[(RDD[_], ShuffleDependency[_, _, _])](
ByteBuffer.wrap(taskBinary.value), Thread.currentThread.getContextClassLoader)
_executorDeserializeTime = System.currentTimeMillis() - deserializeStartTime
var writer: ShuffleWriter[Any, Any] = null
try {
// 获取shuffleManager
val manager = SparkEnv.get.shuffleManager
// 通过shuffleManager的getWriter()方法,获得shuffle的writer
writer = manager.getWriter[Any, Any](dep.shuffleHandle, partitionId, context)
// 通过rdd指定分区的迭代器iterator方法来遍历每一条数据,再之上再调用writer的write方法以写数据
writer.write(rdd.iterator(partition, context).asInstanceOf[Iterator[_ <: Product2[Any, Any]]])
writer.stop(success = true).get
} catch {
case e: Exception =>
try {
if (writer != null) {
writer.stop(success = false)
}
} catch {
case e: Exception =>
log.debug("Could not stop writer", e)
}
throw e
}
}
其中的finalRDD和dependency是在Driver端DAGScheluer中提交Stage的时候加入广播变量的。
接着通过SparkEnv获取shuffleManager,默认使用的是sort(对应的是org.apache.spark.shuffle.sort.SortShuffleManager),可通过spark.shuffle.manager设置。
然后调用了manager.getWriter方法,该方法中检测到满足Unsafe Shuffle条件会自动采用Unsafe Shuffle,否则采用Sort Shuffle。使用Unsafe Shuffle有几个限制,shuffle阶段不能有aggregate操作,分区数不能超过一定大小( 224 −1,这是可编码的最大parition id),所以像reduceByKey这类有aggregate操作的算子是不能使用Unsafe Shuffle。
这里暂时讨论Sort Shuffle的情况,即getWriter返回的是SortShuffleWriter,我们直接看writer.write发生了什么:
override def write(records: Iterator[Product2[K, V]]): Unit = {
sorter = if (dep.mapSideCombine) {
require(dep.aggregator.isDefined, "Map-side combine without Aggregator specified!")
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)
}
// 写内存缓冲区,超过阈值则溢写到磁盘文件
sorter.insertAll(records)
// 获取该task的最终输出文件
val output = shuffleBlockResolver.getDataFile(dep.shuffleId, mapId)
val tmp = Utils.tempFileWith(output)
try {
val blockId = ShuffleBlockId(dep.shuffleId, mapId, IndexShuffleBlockResolver.NOOP_REDUCE_ID)
// merge后写到data文件
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}")
}
}
}
先细看sorter.inster是怎么写到内存,并spill到磁盘文件的:
def insertAll(records: Iterator[Product2[K, V]]): Unit = {
// TODO: stop combining if we find that the reduction factor isn't high
val shouldCombine = aggregator.isDefined
// 若需要Combine
if (shouldCombine) {
// 获取对新value合并到聚合结果中的函数
val mergeValue = aggregator.get.mergeValue
// 获取创建初始聚合值的函数
val createCombiner = aggregator.get.createCombiner
var kv: Product2[K, V] = null
// 通过mergeValue 对已有的聚合结果的新value进行合并,通过createCombiner 对没有聚合结果的新value初始化聚合结果
val update = (hadValue: Boolean, oldValue: C) => {
if (hadValue) mergeValue(oldValue, kv._2) else createCombiner(kv._2)
}
// 遍历records
while (records.hasNext) {
addElementsRead()
kv = records.next()
// 使用update函数进行value的聚合
map.changeValue((getPartition(kv._1), kv._1), update)
// 是否需要spill到磁盘文件
maybeSpillCollection(usingMap = true)
}
// 不需要Combine
} else {
// Stick values into our buffer
while (records.hasNext) {
addElementsRead()
val kv = records.next()
buffer.insert(getPartition(kv._1), kv._1, kv._2.asInstanceOf[C])
maybeSpillCollection(usingMap = false)
}
}
}
需要聚合的情况,遍历records拿到record的KV,通过map的changeValue方法并根据update函数来对相同K的V进行聚合,这里的map是PartitionedAppendOnlyMap类型,只能添加数据不能删除数据,底层实现是一个数组,数组中存KV键值对的方式是[K1,V1,K2,V2…],每一次操作后都会判断是否要spill到磁盘。
不需要聚合的情况,直接将record放入buffer,然后判断是否要溢写到磁盘。
先看map.changeValue方法到底是怎么通过map实现对数据combine的:
override def changeValue(key: K, updateFunc: (Boolean, V) => V): V = {
// 通过聚合算法得到newValue
val newValue = super.changeValue(key, updateFunc)
// 跟新对map的大小采样
super.afterUpdate()
newValue
}
super.changeValue的实现:
def changeValue(key: K, updateFunc: (Boolean, V) => V): V = {
...
// 根据k 得到pos
var pos = rehash(k.hashCode) & mask
var i = 1
while (true) {
// 从data中获取该位置的原来的key
val curKey = data(2 * pos)
// 若原来的key和当前的key相等,则将两个值进行聚合
if (k.eq(curKey) || k.equals(curKey)) {
val newValue = updateFunc(true, data(2 * pos + 1).asInstanceOf[V])
data(2 * pos + 1) = newValue.asInstanceOf[AnyRef]
return newValue
// 若当前key对应的位置没有key,则将当前key作为该位置的key
// 并通过update方法初始化该位置的聚合结果
} else if (curKey.eq(null)) {
val newValue = updateFunc(false, null.asInstanceOf[V])
data(2 * pos) = k
data(2 * pos + 1) = newValue.asInstanceOf[AnyRef]
// 扩容
incrementSize()
return newValue
// 若对应位置有key但不和当前key相等,即hash冲突了,则继续向后遍历
} else {
val delta = i
pos = (pos + delta) & mask
i += 1
}
}
null.asInstanceOf[V] // Never reached but needed to keep compiler happy
}
根据K的hashCode再哈希与上掩码 得到 pos,2 * pos 为 k 应该所在的位置,2 * pos + 1 为 k 对应的 v 所在的位置,获取k应该所在位置的原来的key:
若原来的key不存在,则将当前k作为该位置的key,并通过update函数初始化该k对应的聚合结果,接着会通过incrementSize()方法进行扩容:
private def incrementSize() {
curSize += 1
if (curSize > growThreshold) {
growTable()
}
}
跟新curSize,若当前大小超过了阈值growThreshold(growThreshold是当前容量capacity的0.7倍),则通过growTable()来扩容:
protected def growTable() {
// 容量翻倍
val newCapacity = capacity * 2
require(newCapacity <= MAXIMUM_CAPACITY, s"Can't contain more than ${growThreshold} elements")
//生成新的数组来存数据
val newData = new Array[AnyRef](2 * newCapacity)
val newMask = newCapacity - 1
var oldPos = 0
while (oldPos < capacity) {
// 将旧数组中的数据重新计算位置放到新的数组中
if (!data(2 * oldPos).eq(null)) {
val key = data(2 * oldPos)
val value = data(2 * oldPos + 1)
var newPos = rehash(key.hashCode) & newMask
var i = 1
var keepGoing = true
while (keepGoing) {
val curKey = newData(2 * newPos)
if (curKey.eq(null)) {
newData(2 * newPos) = key
newData(2 * newPos + 1) = value
keepGoing = false
} else {
val delta = i
newPos = (newPos + delta) & newMask
i += 1
}
}
}
oldPos += 1
}
// 替换及跟新变量
data = newData
capacity = newCapacity
mask = newMask
growThreshold = (LOAD_FACTOR * newCapacity).toInt
}
这里重新创建了一个两倍capacity 的数组来存放数据,将原来数组中的数据通过重新计算位置放到新数组里,将data替换为新的数组,并跟新一些变量。
此时聚合已经完成,回到changeValue方面里面,接下来会执行super.afterUpdate()方法来对map的大小进行采样:
protected def afterUpdate(): Unit = {
numUpdates += 1
if (nextSampleNum == numUpdates) {
takeSample()
}
}
若每遍历跟新一条record,都来对map进行采样估计大小,假设采样一次需要1ms,100w次采样就会花上16.7分钟,性能大大降低。所以这里只有当update次数达到nextSampleNum 的时候才通过takeSample()采样一次:
private def takeSample(): Unit = {
samples.enqueue(Sample(SizeEstimator.estimate(this), numUpdates))
// Only use the last two samples to extrapolate
if (samples.size > 2) {
samples.dequeue()
}
// 估计每次跟新的变化量
val bytesDelta = samples.toList.reverse match {
case latest :: previous :: tail =>
(latest.size - previous.size).toDouble / (latest.numUpdates - previous.numUpdates)
// If fewer than 2 samples, assume no change
case _ => 0
}
// 跟新变化量
bytesPerUpdate = math.max(0, bytesDelta)
// 获取下次采样的次数
nextSampleNum = math.ceil(numUpdates * SAMPLE_GROWTH_RATE).toLong
}
这里估计每次跟新的变化量的逻辑是:(当前map大小-上次采样的时候的大小) / (当前update的次数 - 上次采样的时候的update次数)。
接着计算下次需要采样的update次数,该次数是指数级增长的,基数是1.1,第一次采样后,要1.1次进行第二次采样,第1.1*1.1次后进行第三次采样,以此类推,开始增长慢,后面增长跨度会非常大。
这里采样完成后回到insetAll方法,接着通过maybeSpillCollection方法判断是否需要spill:
private def maybeSpillCollection(usingMap: Boolean): Unit = {
var estimatedSize = 0L
if (usingMap) {
estimatedSize = map.estimateSize()
if (maybeSpill(map, estimatedSize)) {
map = new PartitionedAppendOnlyMap[K, C]
}
} else {
estimatedSize = buffer.estimateSize()
if (maybeSpill(buffer, estimatedSize)) {
buffer = new PartitionedPairBuffer[K, C]
}
}
if (estimatedSize > _peakMemoryUsedBytes) {
_peakMemoryUsedBytes = estimatedSize
}
}
通过集合的estimateSize方法估计map的大小,若需要spill则将集合中的数据spill到磁盘文件,并且为集合创建一个新的对象放数据。先看看估计大小的方法estimateSize:
def estimateSize(): Long = {
assert(samples.nonEmpty)
val extrapolatedDelta = bytesPerUpdate * (numUpdates - samples.last.numUpdates)
(samples.last.size + extrapolatedDelta).toLong
}
以上次采样完更新的bytePerUpdate 作为最近平均每次跟新的大小,估计当前占用内存:(当前update次数-上次采样时的update次数) * 每次跟新大小 + 上次采样记录的大小。
获取到当前集合的大小后调用maybeSpill判断是否需要spill:
protected def maybeSpill(collection: C, currentMemory: Long): Boolean = {
var shouldSpill = false
if (elementsRead % 32 == 0 && currentMemory >= myMemoryThreshold) {
// Claim up to double our current memory from the shuffle memory pool
val amountToRequest = 2 * currentMemory - myMemoryThreshold
val granted = acquireMemory(amountToRequest)
// 跟新申请到的内存
myMemoryThreshold += granted
// 集合大小还是比申请到的内存大?spill : no spill
shouldSpill = currentMemory >= myMemoryThreshold
}
shouldSpill = shouldSpill || _elementsRead > numElementsForceSpillThreshold
// Actually spill
if (shouldSpill) {
_spillCount += 1
logSpillage(currentMemory)
spill(collection)
_elementsRead = 0
_memoryBytesSpilled += currentMemory
releaseMemory()
}
shouldSpill
}
这里有两种情况都可导致spill:
若需要spill,则跟新spill次数,调用spill(collection)方法进行溢写磁盘,并释放内存。
跟进spill方法看看其具体实现:
override protected[this] def spill(collection: WritablePartitionedPairCollection[K, C]): Unit = {
// 传入comparator将集合中的数据先根据partition排序再通过key排序后返回一个迭代器
val inMemoryIterator = collection.destructiveSortedWritablePartitionedIterator(comparator)
// 写到磁盘文件,并返回一个对该文件的描述对象SpilledFile
val spillFile = spillMemoryIteratorToDisk(inMemoryIterator)
// 添加到spill文件数组
spills.append(spillFile)
}
继续跟进看看spillMemoryIteratorToDisk的实现:
private[this] def spillMemoryIteratorToDisk(inMemoryIterator: WritablePartitionedIterator)
: SpilledFile = {
// 生成临时文件和blockId
val (blockId, file) = diskBlockManager.createTempShuffleBlock()
// 这些值在每次flush后会被重置
var objectsWritten: Long = 0
var spillMetrics: ShuffleWriteMetrics = null
var writer: DiskBlockObjectWriter = null
def openWriter(): Unit = {
assert (writer == null && spillMetrics == null)
spillMetrics = new ShuffleWriteMetrics
writer = blockManager.getDiskWriter(blockId, file, serInstance, fileBufferSize, spillMetrics)
}
openWriter()
// 按写入磁盘的顺序记录分支的大小
val batchSizes = new ArrayBuffer[Long]
// 记录每个分区有多少元素
val elementsPerPartition = new Array[Long](numPartitions)
// Flush writer 内容到磁盘,并更新相关变量
def flush(): Unit = {
val w = writer
writer = null
w.commitAndClose()
_diskBytesSpilled += spillMetrics.bytesWritten
batchSizes.append(spillMetrics.bytesWritten)
spillMetrics = null
objectsWritten = 0
}
var success = false
try {
// 遍历迭代器
while (inMemoryIterator.hasNext) {
val partitionId = inMemoryIterator.nextPartition()
require(partitionId >= 0 && partitionId < numPartitions,
s"partition Id: ${partitionId} should be in the range [0, ${numPartitions})")
inMemoryIterator.writeNext(writer)
elementsPerPartition(partitionId) += 1
objectsWritten += 1
// 元素个数达到批量序列化大小则flush到磁盘
if (objectsWritten == serializerBatchSize) {
flush()
openWriter()
}
}
// 将剩余的数据flush
if (objectsWritten > 0) {
flush()
} else if (writer != null) {
val w = writer
writer = null
w.revertPartialWritesAndClose()
}
success = true
} finally {
...
}
// 返回SpilledFile
SpilledFile(file, blockId, batchSizes.toArray, elementsPerPartition)
}
通过diskBlockManager创建临时文件和blockID,临时文件名格式为是 “temp_shuffle_” + id,遍历内存数据迭代器,并调用Writer(DiskBlockObjectWriter)的write方法,当写的次数达到序列化大小则flush到磁盘文件,并重新打开writer,及跟新batchSizes等信息。
最后返回一个SpilledFile对象,该对象包含了溢写的临时文件File,blockId,每次flush的到磁盘的大小,每个partition对应的数据条数。
spill完成,并且insertAll方法也执行完成,回到开始的SortShuffleWriter的write方法:
override def write(records: Iterator[Product2[K, V]]): Unit = {
...
// 写内存缓冲区,超过阈值则溢写到磁盘文件
sorter.insertAll(records)
// 获取该task的最终输出文件
val output = shuffleBlockResolver.getDataFile(dep.shuffleId, mapId)
val tmp = Utils.tempFileWith(output)
try {
val blockId = ShuffleBlockId(dep.shuffleId, mapId, IndexShuffleBlockResolver.NOOP_REDUCE_ID)
// merge后写到data文件
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}")
}
}
}
获取最后的输出文件名及blockId,文件格式:
"shuffle_" + shuffleId + "_" + mapId + "_" + reduceId + ".data"
接着通过sorter.writePartitionedFile方法来写文件,其中包括内存及所有spill文件的merge操作,看看起具体实现:
def writePartitionedFile(
blockId: BlockId,
outputFile: File): Array[Long] = {
val writeMetrics = context.taskMetrics().shuffleWriteMetrics
// 跟踪每个分区在文件中的range
val lengths = new Array[Long](numPartitions)
// 数据只存在内存中
if (spills.isEmpty) {
val collection = if (aggregator.isDefined) map else buffer
// 将内存中的数据先通过partitionId再通过k排序后返回一个迭代器
val it = collection.destructiveSortedWritablePartitionedIterator(comparator)
// 遍历数据写入磁盘
while (it.hasNext) {
val writer = blockManager.getDiskWriter(
blockId, outputFile, serInstance, fileBufferSize, writeMetrics)
val partitionId = it.nextPartition()
//等待一个partition的数据写完后刷新到磁盘文件
while (it.hasNext && it.nextPartition() == partitionId) {
it.writeNext(writer)
}
writer.commitAndClose()
val segment = writer.fileSegment()
// 记录每个partition数据长度
lengths(partitionId) = segment.length
}
} else {
// 有数据spill到磁盘,先merge
for ((id, elements) <- this.partitionedIterator) {
if (elements.hasNext) {
val writer = blockManager.getDiskWriter(
blockId, outputFile, serInstance, fileBufferSize, writeMetrics)
for (elem <- elements) {
writer.write(elem._1, elem._2)
}
writer.commitAndClose()
val segment = writer.fileSegment()
lengths(id) = segment.length
}
}
}
context.taskMetrics().incMemoryBytesSpilled(memoryBytesSpilled)
context.taskMetrics().incDiskBytesSpilled(diskBytesSpilled)
context.taskMetrics().incPeakExecutionMemory(peakMemoryUsedBytes)
lengths
}
接下来看看通过this.partitionedIterator方法是怎么将内存及spill文件的数据进行merge-sort的:
def partitionedIterator: Iterator[(Int, Iterator[Product2[K, C]])] = {
val usingMap = aggregator.isDefined
val collection: WritablePartitionedPairCollection[K, C] = if (usingMap) map else buffer
if (spills.isEmpty) {
if (!ordering.isDefined) {
// 只根据partitionId排序,不需要对key排序
groupByPartition(destructiveIterator(collection.partitionedDestructiveSortedIterator(None)))
} else {
// 需要对partitionID和key进行排序
groupByPartition(destructiveIterator(
collection.partitionedDestructiveSortedIterator(Some(keyComparator))))
}
} else {
// Merge spilled and in-memory data
merge(spills, destructiveIterator(
collection.partitionedDestructiveSortedIterator(comparator)))
}
}
这里在有spill文件的情况下会执行下面的merge方法,传入的是spill文件数组和内存中的数据进过partitionId和key排序后的数据迭代器,看看merge:
private def merge(spills: Seq[SpilledFile], inMemory: Iterator[((Int, K), C)])
: Iterator[(Int, Iterator[Product2[K, C]])] = {
// 每个文件对应一个Reader
val readers = spills.map(new SpillReader(_))
val inMemBuffered = inMemory.buffered
(0 until numPartitions).iterator.map { p =>
// 获取内存中当前partition对应的Iterator
val inMemIterator = new IteratorForPartition(p, inMemBuffered)
// 将spill文件对应的partition的数据与内存中对应partition数据合并
val iterators = readers.map(_.readNextPartition()) ++ Seq(inMemIterator)
if (aggregator.isDefined) {
// 对key进行聚合并排序
(p, mergeWithAggregation(
iterators, aggregator.get.mergeCombiners, keyComparator, ordering.isDefined))
} else if (ordering.isDefined) {
// 排序
(p, mergeSort(iterators, ordering.get))
} else {
(p, iterators.iterator.flatten)
}
}
}
merge方法将属于同一个reduce端的partition的内存数据和spill文件数据合并起来,再进行聚合排序(有需要的话),最后返回(reduce对应的partitionId,该分区数据迭代器)
将数据merge-sort后写入最终的文件后,需要将每个partition的偏移量持久化到文件以供后续每个reduce根据偏移量获取自己的数据,写偏移量的逻辑很简单,就是根据前面得到的partition长度的数组将偏移量写到index文件中,对应的文件名为:
def writeIndexFileAndCommit(
shuffleId: Int,
mapId: Int,
lengths: Array[Long],
dataTmp: File): Unit = {
val indexFile = getIndexFile(shuffleId, mapId)
val indexTmp = Utils.tempFileWith(indexFile)
try {
val out = new DataOutputStream(new BufferedOutputStream(new FileOutputStream(indexTmp)))
Utils.tryWithSafeFinally {
// We take in lengths of each block, need to convert it to offsets.
var offset = 0L
out.writeLong(offset)
for (length <- lengths) {
offset += length
out.writeLong(offset)
}
}
......
}
}
根据shuffleId和mapId获取index文件并创建一个写文件的文件流,按照reduce端partition对应的offset依次写到index文件中,如:
0,
length(partition1),
length(partition1)+length(partition2),
length(partition1)+length(partition2)+length(partition3)
…
最后创建一个MapStatus实例返回,包含了reduce端每个partition对应的偏移量。
该对象将返回到Driver端的DAGScheluer处理,被添加到对应stage的OutputLoc里,当该stage的所有task完成的时候会将这些结果注册到MapOutputTrackerMaster,以便下一个stage的task就可以通过它来获取shuffle的结果的元数据信息。
至此Shuffle Write完成!
Shuffle Read部分请看 Shuffle Read解析。