1

我在独立模式下使用 Spark 0.7.2 和以下驱动程序来处理 ~90GB（压缩：19GB）的日志数据，使用 7 个工作人员和 1 个不同的主服务器：

System.setProperty("spark.default.parallelism", "32")
val sc = new SparkContext("spark://10.111.1.30:7077", "MRTest", System.getenv("SPARK_HOME"), Seq(System.getenv("NM_JAR_PATH")))
val logData = sc.textFile("hdfs://10.111.1.30:54310/logs/")
val dcxMap = logData.map(line => (line.split("\\|")(0),   
                                  line.split("\\|")(9)))
                                  .reduceByKey(_ + " || " + _)
dcxMap.saveAsTextFile("hdfs://10.111.1.30:54310/out")

在所有ShuffleMapTasks阶段 1 完成后：

Stage 1 (reduceByKey at DcxMap.scala:31) finished in 111.312 s

它提交阶段 0：

Submitting Stage 0 (MappedRDD[6] at saveAsTextFile at DcxMap.scala:38), which is now runnable

经过一些序列化后，它会打印

spark.MapOutputTrackerActor - Asked to send map output locations for shuffle 0 to host23
spark.MapOutputTracker - Size of output statuses for shuffle 0 is 2008 bytes
spark.MapOutputTrackerActor - Asked to send map output locations for shuffle 0 to host21
spark.MapOutputTrackerActor - Asked to send map output locations for shuffle 0 to host22
spark.MapOutputTrackerActor - Asked to send map output locations for shuffle 0 to host26
spark.MapOutputTrackerActor - Asked to send map output locations for shuffle 0 to host24
spark.MapOutputTrackerActor - Asked to send map output locations for shuffle 0 to host27
spark.MapOutputTrackerActor - Asked to send map output locations for shuffle 0 to host28

在此之后，什么都没有发生，也top表明工人现在都处于闲置状态。如果我查看工作机器上的日志，每台机器都会发生同样的事情：

13/06/21 07:32:25 INFO network.SendingConnection: Initiating connection to [host27/127.0.1.1:34288]
13/06/21 07:32:25 INFO network.SendingConnection: Initiating connection to [host27/127.0.1.1:36040] 
13/06/21 07:32:25 INFO network.SendingConnection: Initiating connection to [host27/127.0.1.1:50467]
13/06/21 07:32:25 INFO network.SendingConnection: Initiating connection to [host27/127.0.1.1:60833]
13/06/21 07:32:25 INFO network.SendingConnection: Initiating connection to [host27/127.0.1.1:49893]
13/06/21 07:32:25 INFO network.SendingConnection: Initiating connection to [host27/127.0.1.1:39907]

然后，对于这些“启动连接”尝试中的每一个，它都会向每个工作人员抛出相同的错误（以 host27 的日志为例，并且仅显示第一次出现的错误）：

13/06/21 07:32:25 WARN network.SendingConnection: Error finishing connection to host27/127.0.1.1:49893
java.net.ConnectException: Connection refused
    at sun.nio.ch.SocketChannelImpl.checkConnect(Native Method)
    at sun.nio.ch.SocketChannelImpl.finishConnect(SocketChannelImpl.java:701)
    at spark.network.SendingConnection.finishConnect(Connection.scala:221)
    at spark.network.ConnectionManager.spark$network$ConnectionManager$$run(ConnectionManager.scala:127)
    at spark.network.ConnectionManager$$anon$4.run(ConnectionManager.scala:70)

为什么会这样？工人之间似乎可以很好地交流，唯一的问题似乎是他们想给自己发送消息；在上面的例子中，host27 尝试向自己发送 6 条消息，但失败了 6 次。向其他工作人员发送消息工作正常。有人有想法吗？

编辑：也许它与使用 127.0 的火花有关。1 .1 而不是 127.0。0 .1？ /etc/hosts如下所示：

127.0.0.1       localhost
127.0.1.1       host27.<ourdomain>  host27
4

1 回答 1

0

我发现问题与这个问题有关。但是，对我来说，在工作人员上设置 SPARK_LOCAL_IP 并不能解决问题。我不得不改为/etc/hosts：

127.0.0.1       localhost

现在它运行顺利。

于 2013-06-21T08:42:32.033 回答