The Way To Set Global Java Opts In Hadoop Without Being Overlapped At Runtime

I intend to set the default java GC collector to '-XX:+UseSerialGC' for all YARN applications without being overlapped at runtime.

After setting it to parameter 'mapreduce.map.java.opts' in mapred-site.xml as below and synchronizing it to all nodes, it will be overlapped when configuring the same parameter at runtime, for instance: `hadoop jar share/hadoop/mapreduce/hadoop-mapreduce-examples-2.2.0.jar pi -Dmapreduce.map.java.opts="-Xmx256M" 4 1000`.

//mapred-site.xml
<property>
    <name>mapreduce.map.java.opts</name>
    <value>-Xmx1024M  -XX:+UseSerialGC</value>
</property>

There are two commands we can issue on a DataNode to check out whether the parameter '-XX:+UseSerialGC' takes into effect or not. Whichever command we are using, we should only focus on 'YarnChild' processes.

//--#1--
/usr/java/jdk1.7.0_11/bin/jps  -mlv  |  grep  -i  gc  
//--#2--
/usr/java/jdk1.7.0_11/bin/jps  |  awk  '{if($2=="YarnChild")  print  $1}'  |  xargs  /usr/java/jdk1.7.0_11/bin/jinfo  |  grep  -i  gc'

Then I tried to configure it in hadoop-env.sh, neither 'HADOOP_OPTS' nor 'HADOOP_CLIENT_OPTS' did it work:

//hadoop-env.sh
export  HADOOP_OPTS="$HADOOP_OPTS  -Dmapreduce.map.java.opts='-XX:+UseSerialGC'"
export  HADOOP_CLIENT_OPTS="-XX:+UseSerialGC  $HADOOP_CLIENT_OPTS"

Finally, I found a parameter which is not described in the official document of mapred-default.xml: 'mapreduce.admin.map.child.java.opts' (The corresponding one to reduce is 'mapreduce.admin.reduce.child.java.opts'). After setting this parameter in mapred-site.xml and synchronizing the file to all nodes, it works fine and definitely will not be overrided since we merely override parameter 'mapreduce.map.java.opts' at runtime.

//mapred-site.xml
<property>
    <name>mapreduce.admin.map.child.java.opts</name>
    <value>-XX:+UseSerialGC</value>
</property>

Consequencely, when we'are going to set some global java opts, we can set them in 'mapreduce.admin.map.child.java.opts' and 'mapreduce.admin.reduce.child.java.opts' in mapred-site.xml.



References:

1. Chapter 3. Setting Up the Hadoop Configuration - Hortonworks Data Platform

© 2014-2017 jason4zhu.blogspot.com All Rights Reserved 
If transfering, please annotate the origin: Jason4Zhu