在spark的Job Scheduling中介绍了spark多个任务同时执行,相关链接为:
https://spark.apache.org/docs/1.2.0/job-scheduling.html
spark要实现多个application同时运行,涉及到多个application的资源共享。在spark on yarn模式下,yarn可以用来管理集群资源。为了能够比较好的利用集群的内存资源和cpu资源,需要使用spark的Dynamic Resource Allocation 功能
Dynamic Resource Allocation有一个问题,就是executor空闲的时候,executor会被remove,但是executor执行后的数据结果很可能被其它executor使用,为了解决这个问题spark需要启动external shfuffle service,相关配置为往yarn-site.xml文件加入如下内容:
<property>
<name>yarn.nodemanager.aux-services</name>
<value>spark_shuffle</value>
</property>
<property>
<name>yarn.nodemanager.aux-services.spark_shuffle.class</name>
<value>org.apache.spark.network.yarn.YarnShuffleService</value>
</property>
同时需要将 spark-<version>-yarn-shuffle.jar放到hadoop的yarn的lib库里面,我的spark版本是1.3.1,也就是 spark-1.3.1-yarn-shuffle.jar 文件
我的yarn的lib库目录为:/usr/local/hadoop2.6.0/share/hadoop/yarn/lib/
/usr/local/hadoop2.6.0/是我的hadoop安装目录
如果spark-1.3.1-yarn-shuffle.jar 文件没有放到正确的目录,spark application在启动的时候会卡在accepted state,相关的打印如下:
/06 10:49:14 INFO yarn.Client:
client token: N/A
diagnostics: N/A
ApplicationMaster host: N/A
ApplicationMaster RPC port: -1
queue: default
start time: 1436150953666
final status: UNDEFINED
tracking URL: http://ddos12:8088/proxy/application_1435917011732_0005/
user: root
15/07/06 10:49:15 INFO yarn.Client: Application report for application_1435917011732_0005 (state: ACCEPTED)
15/07/06 10:49:16 INFO yarn.Client: Application report for application_1435917011732_0005 (state: ACCEPTED)
15/07/06 10:49:17 INFO yarn.Client: Application report for application_1435917011732_0005 (state: ACCEPTED)
......................................................................
external shuffle service的作用是托管已经被remove的executor的计算结果数据,当其它running的executor需要使用这些数据的时候,向external shuffle service请求数据