Add Running a JVM in Docker, a good idea?

Intro

This is a set of resources for understanding the complexities of running your JVM based application inside a linux container, specifically Docker.

I'm doing a talk at the following conferences based on this material:

• Geecon 2016
• Devoxx UK 2016
• Devoxx Poland 2016

Motivation

Why do we want to run our application inside containers?

• Multi tenancy
• Consistent deployment format across technology stacks
• Access to orchestration tools:
  • Kubernetes
  • Marathon on Mesos
• Using docker for other purposes: CI, dev dependencies

Deployment options

• Traditional JEE/servelet container: WebSphere, Tomcat running on bare metal
• Executable jars with embedded Jetty running on bare metal
• A traditional container inside Docker? Why?
• Executable jar inside linux container - the sweet spot

Cost savings

// TODO: Show number of servers for single vs VMs vs Kubernetes

What

The best way to work with Docker is to learn a small amount about the underlying kernel features.

Namespaces

Containers are built on top of Linux namespaces. Giving processes their own PID namespace, mounts

PID

TODO:

• Demo PS inside and outside of the container.

Mount

TODO:

• Show mounts inside and outside of a container. Show outside where the ones inside map to.

Layered file systems

Users

TODO:

• Show uids inside and outside of the container.

CGroups

In addition to namepsaces running containers are restricted via CGroups to the amount of memory and CPU they use.

TODO:

• Show a container with systemd-cfls then systemd-cgtop then put some load and how that cgroup grow.

Docker access to CGroups

TODO:

• Show options that can be passed to Docker run vs settings on cgroups.

• memory

• memory--swap

• memory-soft-limit

• oom-kill-disable (Show a java process with this on and not on)

• cpu-shares

• cpuset-cpu

• cpu-mems (for NUMA)

TODO: Show kubernetes docs about squashable resources

Tools>

Top Htop Free // Cotnainer breakdown systemd-{cgtop, cgls} // See namespaces for a process

CAdvisor

Examples

Scenario: to get 1000 TPS for a container.

The logic inside the app containers one HTTP call and one Database call.

Goal is to make our container as small as possible.

Areas to investigate:

• Memory usage
• Tuning threads

TODO

• Count number of threads

Thread per request

A lot of Java web frameworks, Servlets included, use the thread per request execution model.

// TODO Picture of a pool of threads

The popular Jetty servlet container is one such library. Dropwizard is a framework that brings together Jersey with a few other popular libraries.

Dropwizard

Thread per core

Ratpack example

Akka example (Optional)

Demos

• OOM kill a process with JVM options
  • dmesg
  • docker logs
  • docker restart

TODO:

• Look at the java swiss army knife