Issue
- Best Practice For Java Heap settings
Environment
- CloudBees CI (CloudBees Core) on Modern Platforms
- CloudBees CI (CloudBees Core) on Traditional Platforms
- CloudBees Jenkins Distribution
- CloudBees Jenkins Team (CJT)
- CloudBees Jenkins Platform - Client Master (CJP-CM)
- CloudBees Jenkins Platform - Operations Center (CJP-OC)
- CloudBees Jenkins Enterprise (CJE)
- Jenkins LTS
Resolution
Traditional Platform
Use -Xmx and -Xms
Recommended Heap Specifications can be found in CloudBees JVM Guide.
Modern Platform
Use Container resources spec with -XX:MinRAMPercentage and -XX:MaxRAMPercentage
Since JDK 8u191 XX:+UseContainerSupport is activated by default, everything should work out of the box. It also introduces -XX:MaxRAMPercentage which takes a value between 0 and 100 (Note: values should be double type see docker-library/openjdk/issues/350). This allows fine grained control of the amount of RAM the JVM is allowed to allocate.
So we now know that the JVM is container aware and we should set the provided amount of memory in our runtime environment. For Jenkins, in general, the same requests/limits is recommended, in case the request is too low will cause evictions if the node that is running the pod is running out of memory.
...
resources:
limits:
memory: 6Gi
requests:
memory: 6Gi
...
Defining just the limits assumes the same requests value.
...
resources:
limits:
memory: 6Gi
...
Notes
a. Do not set -Xmx , -Xms. Those options set the JVM Heap size directly and it is best to let the JVM infer these based on current limits. Use -XX:MinRAMPercentage or -XX:MaxRAMPercentage instead.
b. Do not set -XX:MaxRAMFraction. It is deprecated in favor of MaxRAMPercentage see JDK-8186315.
c. A ratio JVM heap / Container Memory limit higher than 0.5 is known to be unstable and may cause unexpected master restarts due to JVM off heap usage.
Memory Estimation
Use your favorite APM tool to monitor the Memomory Heap Consumption by your instance. The right heap value should be a value between 80-90% of total heap consumption when the instance is at the maximum workload. Increased heap memory will produce long Garbage Collection pauses, which can be observed via slowness in the UI and sometimes even side effects like agent disconnections.
Horizontal Scaling
If your instance is not able to support its current workload (due to number of jobs and dynamic configurations) you should consider scaling horizontally your infrastructure by adding more masters to divide the workload more efficiently.
References:
2 Comments