问题导读

1.spark共享变量的作用是什么?

2.什么情况下使用共享变量?

3.如何在程序中使用共享变量?

4.广播变量源码包含哪些内容?

spark编程中,我们经常会遇到使用全局变量,来累加或则使用全局变量。然而对于分布式编程这个却与传统编程有着很大的区别。不可能在程序中声明一个全局变量,在分布式编程中就可以直接使用。因为代码会分发到多台机器,导致我们认为的全局变量失效。那么spark,spark Streaming该如何实现全局变量。

一般情况下,当一个传递给Spark操作(例如map和reduce)的函数在远程节点上面运行时,Spark操作实际上操作的是这个函数所用变量的一个独立副本。这些变量被复制到每台机器上,并且这些变量在远程机器上 的所有更新都不会传递回驱动程序。通常跨任务的读写变量是低效的,但是,Spark还是为两种常见的使用模式提供了两种有限的共享变量:广播变量(broadcast variable)和累加器(accumulator)+

1.概念

1.1 广播变量:

广播可以将变量发送到闭包中,被闭包使用。但是,广播还有一个作用是同步较大数据。比如你有一个IP库,可能有几G,在map操作中,依赖这个ip库。那么,可以通过广播将这个ip库传到闭包中,被并行的任务应用。广播通过两个方面提高数据共享效率:

1,集群中每个节点(物理机器)只有一个副本,默认的闭包是每个任务一个副本;

2,广播传输是通过BT下载模式实现的,也就是P2P下载,在集群多的情况下,可以极大的提高数据传输速率。广播变量修改后,不会反馈到其他节点。

1.2 累加器:

累加器是仅仅被相关操作累加的变量,因此可以在并行中被有效地支持。它可以被用来实现计数器和总和。Spark原生地只支持数字类型的累加器,编程者可以添加新类型的支持。如果创建累加器时指定了名字,可以在Spark的UI界面看到。这有利于理解每个执行阶段的进程。(对于Python还不支持) 

累加器通过对一个初始化了的变量v调用SparkContext.accumulator(v)来创建。在集群上运行的任务可以通过add或者”+=”方法在累加器上进行累加操作。但是,它们不能读取它的值。只有驱动程序能够读取它的值,通过累加器的value方法。

2.如何使用全局变量

2.1 Java版本:

 

package com.Streaming; import org.apache.spark.Accumulator;import org.apache.spark.SparkConf;import org.apache.spark.api.java.JavaPairRDD;import org.apache.spark.api.java.function.Function;import org.apache.spark.broadcast.Broadcast;import org.apache.spark.streaming.Durations;import org.apache.spark.streaming.Time;import org.apache.spark.streaming.api.java.JavaStreamingContext;  import org.apache.spark.api.java.function.FlatMapFunction;import org.apache.spark.api.java.function.Function2;import org.apache.spark.api.java.function.PairFunction;import org.apache.spark.streaming.api.java.JavaDStream;import org.apache.spark.streaming.api.java.JavaPairDStream;import org.apache.spark.streaming.api.java.JavaReceiverInputDStream;import scala.Tuple2; import java.util.*; /** * 利用广播进行黑名单过滤! * * 无论是计数器还是广播!都不是想象的那么简单! * 联合使用非常强大!!!绝对是高端应用! * * 如果 联合使用扩展的话,该怎么做!!! * * ? */public class BroadcastAccumulator {     /**     * 肯定要创建一个广播List     *     * 在上下文中实例化!     */    private static volatile Broadcast<List<String>> broadcastList = null;     /**     * 计数器!     * 在上下文中实例化!     */    private static volatile Accumulator<Integer> accumulator = null;     public static void main(String[] args) {         SparkConf conf = new SparkConf().setMaster("local[2]").                setAppName("WordCountOnlieBroadcast");         JavaStreamingContext jsc = new JavaStreamingContext(conf, Durations.seconds(5));          /**         * 没有action的话,广播并不会发出去!         *         * 使用broadcast广播黑名单到每个Executor中!         */        broadcastList = jsc.sc().broadcast(Arrays.asList("Hadoop","Mahout","Hive"));         /**         * 全局计数器!用于统计在线过滤了多少个黑名单!         */        accumulator = jsc.sparkContext().accumulator(0,"OnlineBlackListCounter");          JavaReceiverInputDStream<String> lines = jsc.socketTextStream("Master", 9999);          /**         * 这里省去flatmap因为名单是一个个的!         */        JavaPairDStream<String, Integer> pairs = lines.mapToPair(new PairFunction<String, String, Integer>() {            @Override            public Tuple2<String, Integer> call(String word) {                return new Tuple2<String, Integer>(word, 1);            }        });         JavaPairDStream<String, Integer> wordsCount = pairs.reduceByKey(new Function2<Integer, Integer, Integer>() {            @Override            public Integer call(Integer v1, Integer v2) {                return v1 + v2;            }        });         /**         * Funtion里面 前几个参数是 入参。         * 后面的出参。         * 体现在call方法里面!         *         * 这里直接基于RDD进行操作了!         */        wordsCount.foreach(new Function2<JavaPairRDD<String, Integer>, Time, Void>() {            @Override            public Void call(JavaPairRDD<String, Integer> rdd, Time time) throws Exception {                rdd.filter(new Function<Tuple2<String, Integer>, Boolean>() {                    @Override                    public Boolean call(Tuple2<String, Integer> wordPair) throws Exception {                        if (broadcastList.value().contains(wordPair._1)) {                             /**                             * accumulator不应该仅仅用来计数。                             * 可以同时写进数据库或者redis中!                             */                            accumulator.add(wordPair._2);                            return false;                        }else {                            return true;                        }                    };                    /**                     * 这里真的希望 广播和计数器执行的话。要进行一个action操作!                     */                }).collect();                 System.out.println("广播器里面的值"+broadcastList.value());                System.out.println("计时器里面的值"+accumulator.value());                return null;            }        });          jsc.start();        jsc.awaitTermination();        jsc.close();     }     }   2.2 Scala版本     
1. package com.Streaming
2.  
3. import java.util
4.  
5. import org.apache.spark.streaming.{Duration, StreamingContext}
6. import org.apache.spark.{Accumulable, Accumulator, SparkContext, SparkConf}
7. import org.apache.spark.broadcast.Broadcast
8.  
9. /**
10.   * Created by lxh on 2016/6/30.
11.   */
12. object BroadcastAccumulatorStreaming {
13.  
14.   /**
15.     * 声明一个广播和累加器!
16.     */
17.   private var broadcastList:Broadcast[List[String]]  = _
18.   private var accumulator:Accumulator[Int] = _
19.  
20.   def main(args: Array[String]) {
21.  
22.     val sparkConf = new SparkConf().setMaster("local[4]").setAppName("broadcasttest")
23.     val sc = new SparkContext(sparkConf)
24.  
25.     /**
26.       * duration是ms
27.       */
28.     val ssc = new StreamingContext(sc,Duration(2000))
29.    // broadcastList = ssc.sparkContext.broadcast(util.Arrays.asList("Hadoop","Spark"))
30.     broadcastList = ssc.sparkContext.broadcast(List("Hadoop","Spark"))
31.     accumulator= ssc.sparkContext.accumulator(0,"broadcasttest")
32.  
33.     /**
34.       * 获取数据!
35.       */
36.     val lines = ssc.socketTextStream("localhost",9999)
37.  
38.     /**
39.       * 拿到数据后 怎么处理!
40.       *
41.       * 1.flatmap把行分割成词。
42.       * 2.map把词变成tuple(word,1)
43.       * 3.reducebykey累加value
44.       * (4.sortBykey排名)
45.       * 4.进行过滤。 value是否在累加器中。
46.       * 5.打印显示。
47.       */
48.     val words = lines.flatMap(line => line.split(" "))
49.  
50.     val wordpair = words.map(word => (word,1))
51.  
52.     wordpair.filter(record => {broadcastList.value.contains(record._1)})
53.  
54.  
55.     val pair = wordpair.reduceByKey(_+_)
56.  
57.     /**
58.       *这步为什么要先foreachRDD?
59.       *
60.       * 因为这个pair 是PairDStream<String, Integer>
61.       *
62.       *   进行foreachRDD是为了?
63.       *
64.       */
65. /*    pair.foreachRDD(rdd => {
66.       rdd.filter(record => {
67.  
68.         if (broadcastList.value.contains(record._1)) {
69.           accumulator.add(1)
70.           return true
71.         } else {
72.           return false
73.         }
74.  
75.       })
76.  
77.     })*/
78.  
79.     val filtedpair = pair.filter(record => {
80.         if (broadcastList.value.contains(record._1)) {
81.           accumulator.add(record._2)
82.           true
83.         } else {
84.           false
85.         }
86.  
87.      }).print
88.  
89.     println("累加器的值"+accumulator.value)
90.  
91.    // pair.filter(record => {broadcastList.value.contains(record._1)})
92.  
93.    /* val keypair = pair.map(pair => (pair._2,pair._1))*/
94.  
95.     /**
96.       * 如果DStream自己没有某个算子操作。就通过转化transform!
97.       */
98.    /* keypair.transform(rdd => {
99.       rdd.sortByKey(false)//TODO
100.     })*/
101.     pair.print()
102.     ssc.start()
103.     ssc.awaitTermination()
104.  
105.   }
106.  
107. }

 

 

补充:除了上面提到的两种外,还有一个闭包的概念,这里补充下闭包 与广播变量对比有两种方式将数据从driver节点发送到worker节点:通过 闭包 和通过 广播变量 。闭包是随着task的组装和分发自动进行的,而广播变量则是需要程序猿手动操作的,具体地可以通过如下方式操作广播变量(假设 sc 为 SparkContext 类型的对象, bc 为 Broadcast 类型的对象):可通过 sc.broadcast(xxx) 创建广播变量。可在各计算节点中(闭包代码中)通过 bc.value 来引用广播的数据。bc.unpersist() 可将各executor中缓存的广播变量删除,后续再使用时数据将被重新发送。bc.destroy() 可将广播变量的数据和元数据一同销毁,销毁之后就不能再使用了。任务闭包包含了任务所需要的代码和数据,如果一个executor数量小于RDD partition的数量,那么每个executor就会得到多个同样的任务闭包,这通常是低效的。而广播变量则只会将数据发送到每个executor一次,并且可以在多个计算操作中共享该广播变量,而且广播变量使用了类似于p2p形式的非常高效的广播算法,大大提高了效率。另外,广播变量由spark存储管理模块进行管理,并以MEMORY_AND_DISK级别进行持久化存储。什么时候用闭包自动分发数据?情况有几种:数据比较小的时候。数据已在driver程序中可用。典型用例是常量或者配置参数。什么时候用广播变量分发数据?情况有几种:数据比较大的时候(实际上,spark支持非常大的广播变量,甚至广播变量中的元素数超过java/scala中Array的最大长度限制(2G,约21.5亿)都是可以的)。数据是某种分布式计算结果。典型用例是训练模型等中间计算结果。当数据或者变量很小的时候,我们可以在Spark程序中直接使用它们,而无需使用广播变量。对于大的广播变量,序列化优化可以大大提高网络传输效率,参见本文序列化优化部分。3.广播变量(Broadcast)源码分析本文基于Spark 1.0源码分析,主要探讨广播变量的初始化、创建、读取以及清除。类关系BroadcastManager类中包含一个BroadcastFactory对象的引用。大部分操作通过调用BroadcastFactory中的方法来实现。BroadcastFactory是一个Trait,有两个直接子类TorrentBroadcastFactory、HttpBroadcastFactory。这两个子类实现了对HttpBroadcast、TorrentBroadcast的封装,而后面两个又同时集成了Broadcast抽象类。BroadcastManager的初始化SparkContext初始化时会创建SparkEnv对象env,这个过程中会调用BroadcastManager的构造方法返回一个对象作为env的成员变量存在:

1. val broadcastManager = new BroadcastManager(isDriver, conf, securityManager)

构造BroadcastManager对象时会调用initialize方法,主要根据配置初始化broadcastFactory成员变量,并调用其initialize方法。

1. val broadcastFactoryClass =
2.          conf.get("spark.broadcast.factory", "org.apache.spark.broadcast.HttpBroadcastFactory")
3.  
4.        broadcastFactory =
5.          Class.forName(broadcastFactoryClass).newInstance.asInstanceOf[BroadcastFactory]
6.  
7.        // Initialize appropriate BroadcastFactory and BroadcastObject
8.        broadcastFactory.initialize(isDriver, conf, securityManager)

 

两个工厂类的initialize方法都是对其相应实体类的initialize方法的调用,下面分开两个类来看。HttpBroadcast的initialize方法

1. def initialize(isDriver: Boolean, conf: SparkConf, securityMgr: SecurityManager) {
2.   synchronized {
3.     if (!initialized) {
4.       bufferSize = conf.getInt("spark.buffer.size", 65536)
5.       compress = conf.getBoolean("spark.broadcast.compress", true)
6.       securityManager = securityMgr
7.       if (isDriver) {
8.         createServer(conf)
9.         conf.set("spark.httpBroadcast.uri",  serverUri)
10.       }
11.       serverUri = conf.get("spark.httpBroadcast.uri")
12.       cleaner = new MetadataCleaner(MetadataCleanerType.HTTP_BROADCAST, cleanup, conf)
13.       compressionCodec = CompressionCodec.createCodec(conf)
14.       initialized = true
15.     }
16.   }
17. }

 

除了一些变量的初始化外,主要做两件事情,一是createServer(只有在Driver端会做),其次是创建一个MetadataCleaner对象。createServer

1. private def createServer(conf: SparkConf) {
2.   broadcastDir = Utils.createTempDir(Utils.getLocalDir(conf))
3.   server = new HttpServer(broadcastDir, securityManager)
4.   server.start()
5.   serverUri = server.uri
6.   logInfo("Broadcast server started at " + serverUri)
7. }

 

 

首先创建一个存放广播变量的目录,默认是

  1. conf.get("spark.local.dir",  System.getProperty("java.io.tmpdir")).split(',')(0) 

然后初始化一个HttpServer对象并启动(封装了jetty),启动过程中包括加载资源文件,起端口和线程用来监控请求等。这部分的细节在org.apache.spark.HttpServer类中,此处不做展开。 创建MetadataCleaner对象 一个MetadataCleaner对象包装了一个定时计划Timer,每隔一段时间执行一个回调函数,此处传入的回调函数为cleanup:

1. private def cleanup(cleanupTime: Long) {
2.   val iterator = files.internalMap.entrySet().iterator()
3.   while(iterator.hasNext) {
4.     val entry = iterator.next()
5.     val (file, time) = (entry.getKey, entry.getValue)
6.     if (time < cleanupTime) {
7.       iterator.remove()
8.       deleteBroadcastFile(file)
9.     }
10.   }
11. }

 

即清楚存在吵过一定时长的broadcast文件。在时长未设定(默认情况)时,不清除:

1. if (delaySeconds > 0) {
2.    logDebug(
3.      "Starting metadata cleaner for " + name + " with delay of " + delaySeconds + " seconds " +
4.      "and period of " + periodSeconds + " secs")
5.    timer.schedule(task, periodSeconds * 1000, periodSeconds * 1000)
6.  }

TorrentBroadcast的initialize方法

1. def initialize(_isDriver: Boolean, conf: SparkConf) {
2.   TorrentBroadcast.conf = conf // TODO: we might have to fix it in tests
3.   synchronized {
4.     if (!initialized) {
5.       initialized = true
6.     }
7.   }
8. }

 

Torrent在此处没做什么,这也可以看出和Http的区别,Torrent的处理方式就是p2p,去中心化。而Http是中心化服务,需要启动服务来接受请求。创建broadcast变量调用SparkContext中的 def broadcast[T: ClassTag](value: T): Broadcast[T]方法来初始化一个广播变量,实现如下:

1. def broadcast[T: ClassTag](value: T): Broadcast[T] = {
2.     val bc = env.broadcastManager.newBroadcast[T](value, isLocal)
3.     cleaner.foreach(_.registerBroadcastForCleanup(bc))
4.     bc
5.   }

即调用broadcastManager的newBroadcast方法:

1. def newBroadcast[T: ClassTag](value_ : T, isLocal: Boolean) = {
2.   broadcastFactory.newBroadcast[T](value_, isLocal, nextBroadcastId.getAndIncrement())
3. }

再调用工厂类的newBroadcast方法,此处返回的是一个Broadcast对象。 HttpBroadcastFactory的newBroadcast

1. def newBroadcast[T: ClassTag](value_ : T, isLocal: Boolean, id: Long) =
2.   new HttpBroadcast[T](value_, isLocal, id)

即创建一个新的HttpBroadcast对象并返回。 构造对象时主要做两件事情:

1. HttpBroadcast.synchronized {
2.    SparkEnv.get.blockManager.putSingle(
3.      blockId, value_, StorageLevel.MEMORY_AND_DISK, tellMaster = false)
4.  }
5.  
6.  if (!isLocal) {
7.    HttpBroadcast.write(id, value_)
8.  }

 

1.将变量id和值放入blockManager,但并不通知master2.调用伴生对象的write方法

1. def write(id: Long, value: Any) {
2.     val file = getFile(id)
3.     val out: OutputStream = {
4.       if (compress) {
5.         compressionCodec.compressedOutputStream(new FileOutputStream(file))
6.       } else {
7.         new BufferedOutputStream(new FileOutputStream(file), bufferSize)
8.       }
9.     }
10.     val ser = SparkEnv.get.serializer.newInstance()
11.     val serOut = ser.serializeStream(out)
12.     serOut.writeObject(value)
13.     serOut.close()
14.     files += file
15.   }

 

write方法将对象值按照指定的压缩、序列化写入指定的文件。这个文件所在的目录即是HttpServer的资源目录,文件名和id的对应关系为:

1. case class BroadcastBlockId(broadcastId: Long, field: String = "") extends BlockId {
2.   def name = "broadcast_" + broadcastId + (if (field == "") "" else "_" + field)
3. }

TorrentBroadcastFactory的newBroadcast方法

1. def newBroadcast[T: ClassTag](value_ : T, isLocal: Boolean, id: Long) =
2.   new TorrentBroadcast[T](value_, isLocal, id)

同样是创建一个TorrentBroadcast对象,并返回。

1. TorrentBroadcast.synchronized {
2.   SparkEnv.get.blockManager.putSingle(
3.     broadcastId, value_, StorageLevel.MEMORY_AND_DISK, tellMaster = false)
4. }
5.  
6. if (!isLocal) {
7.   sendBroadcast()
8. }

 

做两件事情,第一步和Http一样,第二步:

1. def sendBroadcast() {
2.   val tInfo = TorrentBroadcast.blockifyObject(value_)
3.   totalBlocks = tInfo.totalBlocks
4.   totalBytes = tInfo.totalBytes
5.   hasBlocks = tInfo.totalBlocks
6.  
7.   // Store meta-info
8.   val metaId = BroadcastBlockId(id, "meta")
9.   val metaInfo = TorrentInfo(null, totalBlocks, totalBytes)
10.   TorrentBroadcast.synchronized {
11.     SparkEnv.get.blockManager.putSingle(
12.       metaId, metaInfo, StorageLevel.MEMORY_AND_DISK, tellMaster = true)
13.   }
14.  
15.   // Store individual pieces
16.   for (i <- 0 until totalBlocks) {
17.     val pieceId = BroadcastBlockId(id, "piece" + i)
18.     TorrentBroadcast.synchronized {
19.       SparkEnv.get.blockManager.putSingle(
20.         pieceId, tInfo.arrayOfBlocks(i), StorageLevel.MEMORY_AND_DISK, tellMaster = true)
21.     }
22.   }
23. }

 

可以看出,先将元数据信息缓存到blockManager,再将块信息缓存过去。开头可以看到有一个分块动作,是调用伴生对象的blockifyObject方法:

1. def blockifyObject[T](obj: T): TorrentInfo

此方法将对象obj分块(默认块大小为4M),返回一个TorrentInfo对象,第一个参数为一个TorrentBlock对象(包含blockID和block字节数组)、块数量以及obj的字节流总长度。 元数据信息中的blockId为广播变量id+后缀,value为总块数和总字节数。 数据信息是分块缓存,每块的id为广播变量id加后缀及块变好,数据位一个TorrentBlock对象 读取广播变量的值 通过调用bc.value来取得广播变量的值,其主要实现在反序列化方法readObject中 HttpBroadcast的反序列化

1. HttpBroadcast.synchronized {
2.      SparkEnv.get.blockManager.getSingle(blockId) match {
3.        case Some(x) => value_ = x.asInstanceOf[T]
4.        case None => {
5.          logInfo("Started reading broadcast variable " + id)
6.          val start = System.nanoTime
7.          value_ = HttpBroadcast.read[T](id)
8.          /*
9.           * We cache broadcast data in the BlockManager so that subsequent tasks using it
10.           * do not need to re-fetch. This data is only used locally and no other node
11.           * needs to fetch this block, so we don't notify the master.
12.           */
13.          SparkEnv.get.blockManager.putSingle(
14.            blockId, value_, StorageLevel.MEMORY_AND_DISK, tellMaster = false)
15.          val time = (System.nanoTime - start) / 1e9
16.          logInfo("Reading broadcast variable " + id + " took " + time + " s")
17.        }
18.      }
19.    }

 

首先查看blockManager中是否已有,如有则直接取值,否则调用伴生对象的read方法进行读取:

1. def read[T: ClassTag](id: Long): T = {
2.     logDebug("broadcast read server: " +  serverUri + " id: broadcast-" + id)
3.     val url = serverUri + "/" + BroadcastBlockId(id).name
4.  
5.     var uc: URLConnection = null
6.     if (securityManager.isAuthenticationEnabled()) {
7.       logDebug("broadcast security enabled")
8.       val newuri = Utils.constructURIForAuthentication(new URI(url), securityManager)
9.       uc = newuri.toURL.openConnection()
10.       uc.setAllowUserInteraction(false)
11.     } else {
12.       logDebug("broadcast not using security")
13.       uc = new URL(url).openConnection()
14.     }
15.  
16.     val in = {
17.       uc.setReadTimeout(httpReadTimeout)
18.       val inputStream = uc.getInputStream
19.       if (compress) {
20.         compressionCodec.compressedInputStream(inputStream)
21.       } else {
22.         new BufferedInputStream(inputStream, bufferSize)
23.       }
24.     }
25.     val ser = SparkEnv.get.serializer.newInstance()
26.     val serIn = ser.deserializeStream(in)
27.     val obj = serIn.readObject[T]()
28.     serIn.close()
29.     obj
30.   }

 

 

使用serverUri和block id对应的文件名直接开启一个HttpConnection将中心服务器上相应的数据取过来,使用配置的压缩和序列化机制进行解压和反序列化。这里可以看到,所有需要用到广播变量值的executor都需要去driver上pull广播变量的内容。取到值后,缓存到blockManager中,以便下次使用。TorrentBroadcast的反序列化

1. private def readObject(in: ObjectInputStream) {
2.     in.defaultReadObject()
3.     TorrentBroadcast.synchronized {
4.       SparkEnv.get.blockManager.getSingle(broadcastId) match {
5.         case Some(x) =>
6.           value_ = x.asInstanceOf[T]
7.  
8.         case None =>
9.           val start = System.nanoTime
10.           logInfo("Started reading broadcast variable " + id)
11.  
12.           // Initialize @transient variables that will receive garbage values from the master.
13.           resetWorkerVariables()
14.  
15.           if (receiveBroadcast()) {
16.             value_ = TorrentBroadcast.unBlockifyObject[T](arrayOfBlocks, totalBytes, totalBlocks)
17.  
18.             /* Store the merged copy in cache so that the next worker doesn't need to rebuild it.
19.              * This creates a trade-off between memory usage and latency. Storing copy doubles
20.              * the memory footprint; not storing doubles deserialization cost. Also,
21.              * this does not need to be reported to BlockManagerMaster since other executors
22.              * does not need to access this block (they only need to fetch the chunks,
23.              * which are reported).
24.              */
25.             SparkEnv.get.blockManager.putSingle(
26.               broadcastId, value_, StorageLevel.MEMORY_AND_DISK, tellMaster = false)
27.  
28.             // Remove arrayOfBlocks from memory once value_ is on local cache
29.             resetWorkerVariables()
30.           } else {
31.             logError("Reading broadcast variable " + id + " failed")
32.           }
33.  
34.           val time = (System.nanoTime - start) / 1e9
35.           logInfo("Reading broadcast variable " + id + " took " + time + " s")
36.       }
37.     }
38.   }

 

和Http一样,都是先查看blockManager中是否已经缓存,若没有,则调用receiveBroadcast方法:

1. def receiveBroadcast(): Boolean = {
2.     // Receive meta-info about the size of broadcast data,
3.     // the number of chunks it is divided into, etc.
4.     val metaId = BroadcastBlockId(id, "meta")
5.     var attemptId = 10
6.     while (attemptId > 0 && totalBlocks == -1) {
7.       TorrentBroadcast.synchronized {
8.         SparkEnv.get.blockManager.getSingle(metaId) match {
9.           case Some(x) =>
10.             val tInfo = x.asInstanceOf[TorrentInfo]
11.             totalBlocks = tInfo.totalBlocks
12.             totalBytes = tInfo.totalBytes
13.             arrayOfBlocks = new Array[TorrentBlock](totalBlocks)
14.             hasBlocks = 0
15.  
16.           case None =>
17.             Thread.sleep(500)
18.         }
19.       }
20.       attemptId -= 1
21.     }
22.     if (totalBlocks == -1) {
23.       return false
24.     }
25.  
26.     /*
27.      * Fetch actual chunks of data. Note that all these chunks are stored in
28.      * the BlockManager and reported to the master, so that other executors
29.      * can find out and pull the chunks from this executor.
30.      */
31.     val recvOrder = new Random().shuffle(Array.iterate(0, totalBlocks)(_ + 1).toList)
32.     for (pid <- recvOrder) {
33.       val pieceId = BroadcastBlockId(id, "piece" + pid)
34.       TorrentBroadcast.synchronized {
35.         SparkEnv.get.blockManager.getSingle(pieceId) match {
36.           case Some(x) =>
37.             arrayOfBlocks(pid) = x.asInstanceOf[TorrentBlock]
38.             hasBlocks += 1
39.             SparkEnv.get.blockManager.putSingle(
40.               pieceId, arrayOfBlocks(pid), StorageLevel.MEMORY_AND_DISK, tellMaster = true)
41.  
42.           case None =>
43.             throw new SparkException("Failed to get " + pieceId + " of " + broadcastId)
44.         }
45.       }
46.     }
47.  
48.     hasBlocks == totalBlocks
49.   }

 

和写数据一样,同样是分成两个部分,首先取元数据信息,再根据元数据信息读取实际的block信息。注意这里都是从blockManager中读取的,这里贴出blockManager.getSingle的分析。调用栈中最后到BlockManager.doGetRemote方法,中间有一条语句:

1. val locations = Random.shuffle(master.getLocations(blockId))

 

即将存有这个block的节点信息随机打乱,然后使用:

1. val data = BlockManagerWorker.syncGetBlock(
2.        GetBlock(blockId), ConnectionManagerId(loc.host, loc.port))

来获取。 从这里可以看出,Torrent方法首先将广播变量数据分块,并存到BlockManager中;每个节点需要读取广播变量时,是分块读取,对每一块都读取其位置信息,然后随机选一个存有此块数据的节点进行get;每个节点读取后会将包含的快信息报告给BlockManagerMaster,这样本地节点也成为了这个广播网络中的一个peer。 与Http方式形成鲜明对比,这是一个去中心化的网络,只需要保持一个tracker即可,这就是p2p的思想。 广播变量的清除 广播变量被创建时,紧接着有这样一句代码:

1. cleaner.foreach(_.registerBroadcastForCleanup(bc))

cleaner是一个ContextCleaner对象,会将刚刚创建的广播变量注册到其中,调用栈为:

1. def registerBroadcastForCleanup[T](broadcast: Broadcast[T]) {
2.   registerForCleanup(broadcast, CleanBroadcast(broadcast.id))
3. }
     
1. private def registerForCleanup(objectForCleanup: AnyRef, task: CleanupTask) {
2.   referenceBuffer += new CleanupTaskWeakReference(task, objectForCleanup, referenceQueue)

等出现广播变量被弱引用时(关于弱引用,可以参考:),则会执行

1. cleaner.foreach(_.start())

start方法中会调用keepCleaning方法,会遍历注册的清理任务(包括RDD、shuffle和broadcast),依次进行清理:

1. private def keepCleaning(): Unit = Utils.logUncaughtExceptions {
2.     while (!stopped) {
3.       try {
4.         val reference = Option(referenceQueue.remove(ContextCleaner.REF_QUEUE_POLL_TIMEOUT))
5.           .map(_.asInstanceOf[CleanupTaskWeakReference])
6.         reference.map(_.task).foreach { task =>
7.           logDebug("Got cleaning task " + task)
8.           referenceBuffer -= reference.get
9.           task match {
10.             case CleanRDD(rddId) =>
11.               doCleanupRDD(rddId, blocking = blockOnCleanupTasks)
12.             case CleanShuffle(shuffleId) =>
13.               doCleanupShuffle(shuffleId, blocking = blockOnCleanupTasks)
14.             case CleanBroadcast(broadcastId) =>
15.               doCleanupBroadcast(broadcastId, blocking = blockOnCleanupTasks)
16.           }
17.         }
18.       } catch {
19.         case e: Exception => logError("Error in cleaning thread", e)
20.       }
21.     }
22.   }

doCleanupBroadcast调用以下语句:

1. broadcastManager.unbroadcast(broadcastId, true, blocking)

然后是:

1. def unbroadcast(id: Long, removeFromDriver: Boolean, blocking: Boolean) {
2.   broadcastFactory.unbroadcast(id, removeFromDriver, blocking)
3. }

每个工厂类调用其对应实体类的伴生对象的unbroadcast方法。 HttpBroadcast中的变量清除

1. def unpersist(id: Long, removeFromDriver: Boolean, blocking: Boolean) = synchronized {
2.    SparkEnv.get.blockManager.master.removeBroadcast(id, removeFromDriver, blocking)
3.    if (removeFromDriver) {
4.      val file = getFile(id)
5.      files.remove(file)
6.      deleteBroadcastFile(file)
7.    }
8.  }

1是删除blockManager中的缓存,2是删除本地持久化的文件TorrentBroadcast中的变量清除

1. def unpersist(id: Long, removeFromDriver: Boolean, blocking: Boolean) = synchronized {
2.   SparkEnv.get.blockManager.master.removeBroadcast(id, removeFromDriver, blocking)
3. }

小结 Broadcast可以使用在executor端多次使用某个数据的场景(比如说字典),Http和Torrent两种方式对应传统的CS访问方式和P2P访问方式,当广播变量较大或者使用较频繁时,采用后者可以减少driver端的压力。 参考: https://endymecy.gitbooks.io/spa ... ared-variables.html