Apache Spark源码走读之3 -- Task运行期之函数调用关系分析
转自:http://www.cnblogs.com/hseagle/p/3673132.html
概要
本篇主要阐述在TaskRunner中执行的task其业务逻辑是如何被调用到的,另外试图讲清楚运行着的task其输入的数据从哪获取,处理的结果返回到哪里,如何返回。
准备
- spark已经安装完毕
- spark运行在local mode或local-cluster mode
local-cluster mode
local-cluster模式也称为伪分布式,可以使用如下指令运行
MASTER=local[1,2,1024] bin/spark-shell
[1,2,1024] 分别表示,executor number, core number和内存大小,其中内存大小不应小于默认的512M
Driver Programme的初始化过程分析
初始化过程的涉及的主要源文件
- SparkContext.scala 整个初始化过程的入口
- SparkEnv.scala 创建BlockManager, MapOutputTrackerMaster, ConnectionManager, CacheManager
- DAGScheduler.scala 任务提交的入口,即将Job划分成各个stage的关键
- TaskSchedulerImpl.scala 决定每个stage可以运行几个task,每个task分别在哪个executor上运行
- SchedulerBackend
- 最简单的单机运行模式的话,看LocalBackend.scala
- 如果是集群模式,看源文件SparkDeploySchedulerBackend
初始化过程步骤详解
步骤1: 根据初始化入参生成SparkConf,再根据SparkConf来创建SparkEnv, SparkEnv中主要包含以下关键性组件 1. BlockManager 2. MapOutputTracker 3. ShuffleFetcher 4. ConnectionManager
private[spark] val env = SparkEnv.create(
conf,
"",
conf.get("spark.driver.host"),
conf.get("spark.driver.port").toInt,
isDriver = true,
isLocal = isLocal)
SparkEnv.set(env)
步骤2:创建TaskScheduler,根据Spark的运行模式来选择相应的SchedulerBackend,同时启动taskscheduler,这一步至为关键
private[spark] var taskScheduler = SparkContext.createTaskScheduler(this, master, appName)
taskScheduler.start()
TaskScheduler.start目的是启动相应的SchedulerBackend,并启动定时器进行检测
override def start() {
backend.start()
if (!isLocal && conf.getBoolean("spark.speculation", false)) {
logInfo("Starting speculative execution thread")
import sc.env.actorSystem.dispatcher
sc.env.actorSystem.scheduler.schedule(SPECULATION_INTERVAL milliseconds,
SPECULATION_INTERVAL milliseconds) {
checkSpeculatableTasks()
}
}
}
步骤3:以上一步中创建的TaskScheduler实例为入参创建DAGScheduler并启动运行
@volatile private[spark] var dagScheduler = new DAGScheduler(taskScheduler)
dagScheduler.start()
步骤4:启动WEB UI
ui.start()
RDD的转换过程
还是以最简单的wordcount为例说明rdd的转换过程
sc.textFile("README.md").flatMap(line=>line.split(" ")).map(word => (word, 1)).reduceByKey(_ + _)
上述一行简短的代码其实发生了很复杂的RDD转换,下面仔细解释每一步的转换过程和转换结果
步骤1:val rawFile = sc.textFile("README.md")
textFile先是生成hadoopRDD,然后再通过map操作生成MappedRDD,如果在spark-shell中执行上述语句,得到的结果可以证明所做的分析
scala> sc.textFile("README.md")
14/04/23 13:11:48 WARN SizeEstimator: Failed to check whether UseCompressedOops is set; assuming yes
14/04/23 13:11:48 INFO MemoryStore: ensureFreeSpace(119741) called with curMem=0, maxMem=311387750
14/04/23 13:11:48 INFO MemoryStore: Block broadcast_0 stored as values to memory (estimated size 116.9 KB, free 296.8 MB)
14/04/23 13:11:48 DEBUG BlockManager: Put block broadcast_0 locally took 277 ms
14/04/23 13:11:48 DEBUG BlockManager: Put for block broadcast_0 without replication took 281 ms
res0: org.apache.spark.rdd.RDD[String] = MappedRDD[1] at textFile at :13
步骤2: val splittedText = rawFile.flatMap(line => line.split(" "))
flatMap将原来的MappedRDD转换成为FlatMappedRDD
def flatMap[U: ClassTag](f: T => TraversableOnce[U]): RDD[U] = new FlatMappedRDD(this, sc.clean(f))
步骤3:val wordCount = splittedText.map(word => (word, 1))
利用word生成相应的键值对,上一步的FlatMappedRDD被转换成为MappedRDD
步骤4:val reduceJob = wordCount.reduceByKey(_ + _),这一步最复杂
步骤2,3中使用到的operation全部定义在RDD.scala中,而这里使用到的reduceByKey却在RDD.scala中见不到踪迹。reduceByKey的定义出现在源文件PairRDDFunctions.scala
细心的你一定会问reduceByKey不是MappedRDD的属性和方法啊,怎么能被MappedRDD调用呢?其实这背后发生了一个隐式的转换,该转换将MappedRDD转换成为PairRDDFunctions
implicit def rddToPairRDDFunctions[K: ClassTag, V: ClassTag](rdd: RDD[(K, V)]) =
new PairRDDFunctions(rdd)
这种隐式的转换是scala的一个语法特征,如果想知道的更多,请用关键字"scala implicit method"进行查询,会有不少的文章对此进行详尽的介绍。
接下来再看一看reduceByKey的定义
def reduceByKey(func: (V, V) => V): RDD[(K, V)] = {
reduceByKey(defaultPartitioner(self), func)
}
def reduceByKey(partitioner: Partitioner, func: (V, V) => V): RDD[(K, V)] = {
combineByKey[V]((v: V) => v, func, func, partitioner)
}
def combineByKey[C](createCombiner: V => C,
mergeValue: (C, V) => C,
mergeCombiners: (C, C) => C,
partitioner: Partitioner,
mapSideCombine: Boolean = true,
serializerClass: String = null): RDD[(K, C)] = {
if (getKeyClass().isArray) {
if (mapSideCombine) {
throw new SparkException("Cannot use map-side combining with array keys.")
}
if (partitioner.isInstanceOf[HashPartitioner]) {
throw new SparkException("Default partitioner cannot partition array keys.")
}
}
val aggregator = new Aggregator[K, V, C](createCombiner, mergeValue, mergeCombiners)
if (self.partitioner == Some(partitioner)) {
self.mapPartitionsWithContext((context, iter) => {
new InterruptibleIterator(context, aggregator.combineValuesByKey(iter, context))
}, preservesPartitioning = true)
} else if (mapSideCombine) {
val combined = self.mapPartitionsWithContext((context, iter) => {
aggregator.combineValuesByKey(iter, context)
}, preservesPartitioning = true)
val partitioned = new ShuffledRDD[K, C, (K, C)](combined, partitioner)
.setSerializer(serializerClass)
partitioned.mapPartitionsWithContext((context, iter) => {
new InterruptibleIterator(context, aggregator.combineCombinersByKey(iter, context))
}, preservesPartitioning = true)
} else {
// Don't apply map-side combiner.
val values = new ShuffledRDD[K, V, (K, V)](self, partitioner).setSerializer(serializerClass)
values.mapPartitionsWithContext((context, iter) => {
new InterruptibleIterator(context, aggregator.combineValuesByKey(iter, context))
}, preservesPartitioning = true)
}
}
reduceByKey最终会调用combineByKey, 在这个函数中PairedRDDFunctions会被转换成为ShuffleRDD,当调用mapPartitionsWithContext之后,shuffleRDD被转换成为MapPartitionsRDD
Log输出能证明我们的分析
res1: org.apache.spark.rdd.RDD[(String, Int)] = MapPartitionsRDD[8] at reduceByKey at :13
RDD转换小结
小结一下整个RDD转换过程
HadoopRDD->MappedRDD->FlatMappedRDD->MappedRDD->PairRDDFunctions->ShuffleRDD->MapPartitionsRDD
整个转换过程好长啊,这一切的转换都发生在任务提交之前。
运行过程分析
数据集操作分类
在对任务运行过程中的函数调用关系进行分析之前,我们也来探讨一个偏理论的东西,作用于RDD之上的Transformantion为什么会是这个样子?
对这个问题的解答和数学搭上关系了,从理论抽象的角度来说,任务处理都可归结为“input->processing->output"。input和output对应于数据集dataset.
在此基础上作一下简单的分类
- one-one 一个dataset在转换之后还是一个dataset,而且dataset的size不变,如map
- one-one 一个dataset在转换之后还是一个dataset,但size发生更改,这种更改有两种可能:扩大或缩小,如flatMap是size增大的操作,而subtract是size变小的操作
- many-one 多个dataset合并为一个dataset,如combine, join
- one-many 一个dataset分裂为多个dataset, 如groupBy
Task运行期的函数调用
task的提交过程参考本系列中的第二篇文章。本节主要讲解当task在运行期间是如何一步步调用到作用于RDD上的各个operation
- TaskRunner.run
- Task.run
- Task.runTask (Task是一个基类,有两个子类,分别为ShuffleMapTask和ResultTask)
- RDD.iterator
- RDD.computeOrReadCheckpoint
- RDD.compute
- RDD.computeOrReadCheckpoint
- RDD.iterator
- Task.runTask (Task是一个基类,有两个子类,分别为ShuffleMapTask和ResultTask)
- Task.run
或许当看到RDD.compute函数定义时,还是觉着f没有被调用,以MappedRDD的compute定义为例
override def compute(split: Partition, context: TaskContext) =
firstParent[T].iterator(split, context).map(f)
注意,这里最容易产生错觉的地方就是map函数,这里的map不是RDD中的map,而是scala中定义的iterator的成员函数map, 请自行参考http://www.scala-lang.org/api/2.10.4/index.html#scala.collection.Iterator
堆栈输出
80 at org.apache.spark.rdd.HadoopRDD.getJobConf(HadoopRDD.scala:111)
81 at org.apache.spark.rdd.HadoopRDD$$anon$1.(HadoopRDD.scala:154)
82 at org.apache.spark.rdd.HadoopRDD.compute(HadoopRDD.scala:149)
83 at org.apache.spark.rdd.HadoopRDD.compute(HadoopRDD.scala:64)
84 at org.apache.spark.rdd.RDD.computeOrReadCheckpoint(RDD.scala:241)
85 at org.apache.spark.rdd.RDD.iterator(RDD.scala:232)
86 at org.apache.spark.rdd.MappedRDD.compute(MappedRDD.scala:31)
87 at org.apache.spark.rdd.RDD.computeOrReadCheckpoint(RDD.scala:241)
88 at org.apache.spark.rdd.RDD.iterator(RDD.scala:232)
89 at org.apache.spark.rdd.FlatMappedRDD.compute(FlatMappedRDD.scala:33)
90 at org.apache.spark.rdd.RDD.computeOrReadCheckpoint(RDD.scala:241)
91 at org.apache.spark.rdd.RDD.iterator(RDD.scala:232)
92 at org.apache.spark.rdd.MappedRDD.compute(MappedRDD.scala:31)
93 at org.apache.spark.rdd.RDD.computeOrReadCheckpoint(RDD.scala:241)
94 at org.apache.spark.rdd.RDD.iterator(RDD.scala:232)
95 at org.apache.spark.rdd.MapPartitionsRDD.compute(MapPartitionsRDD.scala:34)
96 at org.apache.spark.rdd.RDD.computeOrReadCheckpoint(RDD.scala:241)
97 at org.apache.spark.rdd.RDD.iterator(RDD.scala:232)
98 at org.apache.spark.scheduler.ShuffleMapTask.runTask(ShuffleMapTask.scala:161)
99 at org.apache.spark.scheduler.ShuffleMapTask.runTask(ShuffleMapTask.scala:102)
100 at org.apache.spark.scheduler.Task.run(Task.scala:53)
101 at org.apache.spark.executor.Executor$TaskRunner$$anonfun$run$1.apply$mcV$sp(Executor.scala:211)
ResultTask
compute的计算过程对于ShuffleMapTask比较复杂,绕的圈圈比较多,对于ResultTask就直接许多。
override def runTask(context: TaskContext): U = {
metrics = Some(context.taskMetrics)
try {
func(context, rdd.iterator(split, context))
} finally {
context.executeOnCompleteCallbacks()
}
}
计算结果的传递
上面的分析知道,wordcount这个job在最终提交之后,被DAGScheduler分为两个stage,第一个Stage是shuffleMapTask,第二个Stage是ResultTask.
那么ShuffleMapTask的计算结果是如何被ResultTask取得的呢?这个过程简述如下
- ShffuleMapTask将计算的状态(注意不是具体的数据)包装为MapStatus返回给DAGScheduler
- DAGScheduler将MapStatus保存到MapOutputTrackerMaster中
- ResultTask在执行到ShuffleRDD时会调用BlockStoreShuffleFetcher的fetch方法去获取数据
- 第一件事就是咨询MapOutputTrackerMaster所要取的数据的location
- 根据返回的结果调用BlockManager.getMultiple获取真正的数据
BlockStoreShuffleFetcher的fetch函数伪码
val blockManager = SparkEnv.get.blockManager
val startTime = System.currentTimeMillis
val statuses = SparkEnv.get.mapOutputTracker.getServerStatuses(shuffleId, reduceId)
logDebug("Fetching map output location for shuffle %d, reduce %d took %d ms".format(
shuffleId, reduceId, System.currentTimeMillis - startTime))
val blockFetcherItr = blockManager.getMultiple(blocksByAddress, serializer)
val itr = blockFetcherItr.flatMap(unpackBlock)
注意上述代码中的getServerStatuses及getMultiple,一个是询问数据的位置,一个是去获取真正的数据。
有关Shuffle的详细解释,请参考”详细探究Spark的shuffle实现一文" http://jerryshao.me/architecture/2014/01/04/spark-shuffle-detail-investigation/