使用紗線的群集模式中的Spark Pi示例：關聯丟失

Question

我有三個虛擬機作為分布式Spark集群運行。 我使用Spark 1.3.0和底層的Hadoop 2.6.0。

如果我運行Spark Pi示例

/usr/local/spark130/bin/spark-submit 
--class org.apache.spark.examples.SparkPi  
--master yarn-client /usr/local/spark130/examples/target/spark-examples_2.10-1.3.0.jar  10000

我收到此警告/錯誤並最終出現異常：

 WARN NativeCodeLoader: Unable to load native-hadoop library for your platform... using builtin-java classes where applicable
15/04/08 12:37:06 WARN ReliableDeliverySupervisor: Association with remote system [akka.tcp://sparkYarnAM@virtm4:47128] has failed, address is now gated for [5000] ms. Reason is: [Disassociated].
15/04/08 12:37:12 WARN ReliableDeliverySupervisor: Association with remote system [akka.tcp://sparkYarnAM@virtm4:45975] has failed, address is now gated for [5000] ms. Reason is: [Disassociated].
15/04/08 12:37:13 ERROR YarnClientSchedulerBackend: Yarn application has already exited with state FINISHED!

當我檢查容器的日志時，我發現它是SIGTERM-ed

15/04/08 12:37:08 INFO yarn.YarnAllocator: Container request (host: Any, capability: <memory:1408, vCores:1>)
15/04/08 12:37:08 INFO yarn.YarnAllocator: Container request (host: Any, capability: <memory:1408, vCores:1>)
15/04/08 12:37:08 INFO yarn.ApplicationMaster: Started progress reporter thread - sleep time : 5000
15/04/08 12:37:12 ERROR yarn.ApplicationMaster: RECEIVED SIGNAL 15: SIGTERM
15/04/08 12:37:12 INFO yarn.ApplicationMaster: Final app status: UNDEFINED, exitCode: 0, (reason: Shutdown hook called before final status was reported.)
15/04/08 12:37:12 INFO yarn.ApplicationMaster: Unregistering ApplicationMaster with UNDEFINED (diag message: Shutdown hook called before final status was reported.)

解決方案：我解決了這個問題。 我現在使用Java7而不是Java8。 這種情況被報告為bug，但它被拒絕了這樣的https://issues.apache.org/jira/browse/SPARK-6388然而，改變Java版本確實有效。

Answer 1

由於Java 8過多的內存分配問題，關聯可能會丟失： https ： //issues.apache.org/jira/browse/YARN-4714

您可以通過在yarn-site.xml中設置以下屬性來強制YARN忽略它

<property>
    <name>yarn.nodemanager.pmem-check-enabled</name>
    <value>false</value>
</property>

<property>
    <name>yarn.nodemanager.vmem-check-enabled</name>
    <value>false</value>
</property>

Answer 2

之前我遇到過類似的問題，直到找到這個問題

嘗試顯式停止SparkContext實例sc.stop()