Limiting resources


  • So far, we have used containers as convenient units of deployment.

  • What happens when a container tries to use more resources than available?

    (RAM, CPU, disk usage, disk and network I/O...)

  • What happens when multiple containers compete for the same resource?

  • Can we limit resources available to a container?

    (Spoiler alert: yes!)

Container processes are normal processes

  • Containers are closer to "fancy processes" than to "lightweight VMs".

  • A process running in a container is, in fact, a process running on the host.

  • Let's look at the output of ps on a container host running 3 containers :

         0  2662  0.2  0.3 /usr/bin/dockerd -H fd://
         0  2766  0.1  0.1  \_ docker-containerd --config /var/run/docker/containe
         0 23479  0.0  0.0      \_ docker-containerd-shim -namespace moby -workdir
         0 23497  0.0  0.0      |   \_ `nginx`: master process nginx -g daemon off;
       101 23543  0.0  0.0      |       \_ `nginx`: worker process
         0 23565  0.0  0.0      \_ docker-containerd-shim -namespace moby -workdir
       102 23584  9.4 11.3      |   \_ `/docker-java-home/jre/bin/java` -Xms2g -Xmx2
         0 23707  0.0  0.0      \_ docker-containerd-shim -namespace moby -workdir
         0 23725  0.0  0.0          \_ `/bin/sh`
  • The highlighted processes are containerized processes.
    (That host is running nginx, elasticsearch, and alpine.)

By default: nothing changes

  • What happens when a process uses too much memory on a Linux system?

  • Simplified answer:

    • swap is used (if available);

    • if there is not enough swap space, eventually, the out-of-memory killer is invoked;

    • the OOM killer uses heuristics to kill processes;

    • sometimes, it kills an unrelated process.

  • What happens when a container uses too much memory?

  • The same thing!

    (i.e., a process eventually gets killed, possibly in another container.)

Limiting container resources

  • The Linux kernel offers rich mechanisms to limit container resources.

  • For memory usage, the mechanism is part of the cgroup subsystem.

  • This subsystem allows to limit the memory for a process or a group of processes.

  • A container engine leverages these mechanisms to limit memory for a container.

  • The out-of-memory killer has a new behavior:

    • it runs when a container exceeds its allowed memory usage,

    • in that case, it only kills processes in that container.

Limiting memory in practice

  • The Docker Engine offers multiple flags to limit memory usage.

  • The two most useful ones are --memory and --memory-swap.

  • --memory limits the amount of physical RAM used by a container.

  • --memory-swap limits the total amount (RAM+swap) used by a container.

  • The memory limit can be expressed in bytes, or with a unit suffix.

    (e.g.: --memory 100m = 100 megabytes.)

  • We will see two strategies: limiting RAM usage, or limiting both

Limiting RAM usage


docker run -ti --memory 100m python

If the container tries to use more than 100 MB of RAM, and swap is available:

  • the container will not be killed,

  • memory above 100 MB will be swapped out,

  • in most cases, the app in the container will be slowed down (a lot).

If we run out of swap, the global OOM killer still intervenes.

Limiting both RAM and swap usage


docker run -ti --memory 100m --memory-swap 100m python

If the container tries to use more than 100 MB of memory, it is killed.

On the other hand, the application will never be slowed down because of swap.

When to pick which strategy?

  • Stateful services (like databases) will lose or corrupt data when killed

  • Allow them to use swap space, but monitor swap usage

  • Stateless services can usually be killed with little impact

  • Limit their mem+swap usage, but monitor if they get killed

  • Ultimately, this is no different from "do I want swap, and how much?"

Limiting CPU usage

  • There are no less than 3 ways to limit CPU usage:

    • setting a relative priority with --cpu-shares,

    • setting a CPU% limit with --cpus,

    • pinning a container to specific CPUs with --cpuset-cpus.

  • They can be used separately or together.

Setting relative priority

  • Each container has a relative priority used by the Linux scheduler.

  • By default, this priority is 1024.

  • As long as CPU usage is not maxed out, this has no effect.

  • When CPU usage is maxed out, each container receives CPU cycles in proportion of its relative priority.

  • In other words: a container with --cpu-shares 2048 will receive twice as much than the default.

Setting a CPU% limit

  • This setting will make sure that a container doesn't use more than a given % of CPU.

  • The value is expressed in CPUs; therefore:

    --cpus 0.1 means 10% of one CPU,

    --cpus 1.0 means 100% of one whole CPU,

    --cpus 10.0 means 10 entire CPUs.

Pinning containers to CPUs

  • On multi-core machines, it is possible to restrict the execution on a set of CPUs.

  • Examples:

    --cpuset-cpus 0 forces the container to run on CPU 0;

    --cpuset-cpus 3,5,7 restricts the container to CPUs 3, 5, 7;

    --cpuset-cpus 0-3,8-11 restricts the container to CPUs 0, 1, 2, 3, 8, 9, 10, 11.

  • This will not reserve the corresponding CPUs!

    (They might still be used by other containers, or uncontainerized processes.)

Limiting disk usage

  • Most storage drivers do not support limiting the disk usage of containers.

    (With the exception of devicemapper, but the limit cannot be set easily.)

  • This means that a single container could exhaust disk space for everyone.

  • In practice, however, this is not a concern, because:

    • data files (for stateful services) should reside on volumes,

    • assets (e.g. images, user-generated content...) should reside on object stores or on volume,

    • logs are written on standard output and gathered by the container engine.

  • Container disk usage can be audited with docker ps -s and docker diff.