Bug #15075
closedAllow overriding gc options for the webapp JVM to better deal with high volumes
Description
In a quite loaded environement, with 2800 nodes and around 1300 report messages/seconds, we reached a point were memory contention was hight and CMS GC was spending a lot of time in full GC.
If we add more memory, it's hard to fine tune the different memory zone ratio (new, eden, etc) to avoid even longer full gc, and when they happen, it's terrible.
We tried G1GC which deals quite well with a heap a >8Go and that kind of load.
So, we should either switch or document how to use G1GC for people who need it.
The condition are that the JVM must be an openjdk, at leat 1.8.something. At least 1.8 is a requirement for rudder 5, so no problem here. But I'm not sure at all that server distro have an openjdk (suse?).
And in most cases, default configuration is sufficient (G1GC is less efficient than CMS, it need more space and has a lesser throughout - the price for better latency).
So, to change the garbage collector used for rudder, one should edit /opt/rudder/etc/rudder-jetty.conf
and replace the part with JAVA_OPTIONS
by
# Java VM arguments JAVA_OPTIONS_COMMONS="-server -Xms${JAVA_XMX}m -Xmx${JAVA_XMX}m -Dfile.encoding=UTF-8 -Drudder.configFile=/opt/rudder/etc/rudder-web.properties -Drudder.authFile=/opt/rudder/etc/rudder-users.xml -Dinventoryweb.configFile=/opt/rudder/etc/inventory-web.properties -Dlogback.configurationFile=/opt/rudder/etc/logback.xml -Drun.mode=production" JAVA_OPTIONS_DEFAULT="${JAVA_OPTIONS} ${JAVA_OPTIONS_COMMONS} -XX:+CMSClassUnloadingEnabled -XX:+UseConcMarkSweepGC" JAVA_OPTIONS_G1GC="${JAVA_OPTIONS} ${JAVA_OPTIONS_COMMONS} -XX:+UseG1GC -XX:+UnlockExperimentalVMOptions -XX:MaxGCPauseMillis=500" # you can add the following options to get more debug info about gc: JAVA_OPTIONS_G1GC_DEBUG=" -XX:+PrintGCTimeStamps -XX:+PrintGCApplicationStoppedTime -XX:+PrintGCApplicationConcurrentTime -verbose:gc -XX:+PrintGCDetails -Xloggc:/tmp/g1gc.log" # JAVA_OPTIONS_G1GC="${JAVA_OPTIONS_G1GC} ${JAVA_OPTIONS_G1GC_DEBUG}" # use default to come back to rudder default GC configuration #JAVA_OPTIONS=${JAVA_OPTIONS_DEFAULT} JAVA_OPTIONS=${JAVA_OPTIONS_G1GC}
You can also adapt the quantity of ram given to rudder-jetty
in file /etc/default/rudder-jetty
, with the JAVA_XMX
parameter
After restarting rudder-jetty
service, you should be able to see the parameter list with ps aux | grep jetty
[root@server]# ps aux | grep jetty root 5486 38.3 30.2 10216788 4921580 ? Sl 15:32 99:25 /bin/java -server -Xms9000m -Xmx9000m -XX:+UseG1GC -XX:+UnlockExperimentalVMOptions -XX:MaxGCPauseMillis=500 -XX:+PrintGCTimeStamps -XX:+PrintGCApplicationStoppedTime -XX:+PrintGCApplicationConcurrentTime -verbose:gc -XX:+PrintGCDetails -Xloggc:/tmp/g1gc.log -Dfile.encoding=UTF-8 -Drudder.configFile=/opt/rudder/etc/rudder-web.properties -Drudder.authFile=/opt/rudder/etc/rudder-users.xml -Dinventoryweb.configFile=/opt/rudder/etc/inventory-web.properties -Dlogback.configurationFile=/opt/rudder/etc/logback.xml -Drun.mode=production -Djetty.home=/opt/rudder/jetty -Djetty.base=/opt/rudder/etc/rudder-jetty-base -Djava.io.tmpdir=/var/rudder /tmp/jetty -jar /opt/rudder/jetty/start.jar OPTIONS=Server jetty.state=/opt/rudder/etc/rudder-jetty-base/rudder-jetty.state jetty-started.xml
Updated by François ARMAND over 5 years ago
- Status changed from Discussion to In progress
So, it was decided that:
- CMS remains the default GC,
- we change /opt/rudder/etc/rudder-jetty.conf
to make the GC configurable in JAVA_OPTIONS
- the configuration is documented to /etc/defaults/rudder-jetty
- add a section in performance (or requirement) that explain that for big HEAP (6Go or more), G1GC should be used.
Updated by François ARMAND over 5 years ago
- Status changed from In progress to Pending technical review
- Assignee changed from François ARMAND to Alexis Mousset
- Pull Request set to https://github.com/Normation/rudder-packages/pull/1971
Updated by François ARMAND over 5 years ago
- Status changed from Pending technical review to Pending release
Applied in changeset rudder-packages|98239b61f68fdf77e1fc2d184a3ed4460e3b0016.
Updated by François ARMAND over 5 years ago
- Fix check changed from To do to Checked
Updated by Alexis Mousset over 5 years ago
- Subject changed from Use G1GC to better deal with high volume of nodes and reports to Allow overriding gc options for the webapp JVM to better deal with high volumes
- Name check changed from To do to Reviewed
Updated by Vincent MEMBRÉ over 5 years ago
This bug has been fixed in Rudder 5.0.12 which was released today.
Updated by Alexis Mousset over 4 years ago
- Status changed from Pending release to Released