Is there a maximum number of containers running on a Docker host?

There are a number of system limits you can run into (and work around) but there's a significant amount of grey area depending on

  1. How you are configuring your docker containers.
  2. What you are running in your containers.
  3. What kernel, distribution and docker version you are on.

The figures below are from the boot2docker 1.11.1 vm image which is based on Tiny Core Linux 7. The kernel is 4.4.8

Docker

Docker creates or uses a number of resources to run a container, on top of what you run inside the container.

  • Attaches a virtual ethernet adaptor to the docker0 bridge (1023 max per bridge)
  • Mounts an AUFS and shm file system (1048576 mounts max per fs type)
  • Create's an AUFS layer on top of the image (127 layers max)
  • Forks 1 extra docker-containerd-shim management process (~3MB per container on avg and sysctl kernel.pid_max)
  • Docker API/daemon internal data to manage container. (~400k per container)
  • Creates kernel cgroups and name spaces
  • Opens file descriptors (~15 + 1 per running container at startup. ulimit -n and sysctl fs.file-max )

Docker options

  • Port mapping -p will run a extra process per port number on the host (~4.5MB per port on avg pre 1.12, ~300k per port > 1.12 and also sysctl kernel.pid_max)
  • --net=none and --net=host would remove the networking overheads.

Container services

The overall limits will normally be decided by what you run inside the containers rather than dockers overhead (unless you are doing something esoteric, like testing how many containers you can run :)

If you are running apps in a virtual machine (node,ruby,python,java) memory usage is likely to become your main issue.

IO across a 1000 processes would cause a lot of IO contention.

1000 processes trying to run at the same time would cause a lot of context switching (see vm apps above for garbage collection)

If you create network connections from a 1000 containers the hosts network layer will get a workout.

It's not much different to tuning a linux host to run a 1000 processes, just some additional Docker overheads to include.

Example

1023 Docker busybox images running nc -l -p 80 -e echo host uses up about 1GB of kernel memory and 3.5GB of system memory.

1023 plain nc -l -p 80 -e echo host processes running on a host uses about 75MB of kernel memory and 125MB of system memory

Starting 1023 containers serially took ~8 minutes.

Killing 1023 containers serially took ~6 minutes


From a post on the mailing list, at about 1000 containers you start running into Linux networking issues.

The reason is:

This is the kernel, specifically net/bridge/br_private.h BR_PORT_BITS cannot be extended because of spanning tree requirements.


With Docker-compose, I am able to run over 6k containers on single host (with 190GB memory). container image is under 10MB. But Due to bridge limitation i have divided the containers in batches with multiple services, each service have 1k containers and separate subnet network.

docker-compose -f docker-compose.yml up --scale servicename=1000 -d

But after reaching 6k even though memory is still available around 60GB, it stops scaling and suddenly spikes up the memory. There should be bench-marking figures published by docker team to help, but unfortunately its not available. Kubernetes on the other hand publishes bench-marking stats clearly about the number of pods recommended per node.