Clustering Using Docker Swarm 0.2.0 (Tech Tip #85)

One of the key updates as part of Docker 1.6 is Docker Swarm 0.2.0. Docker Swarm solves one of the fundamental limitations of Docker where the containers could only run on a single Docker host. Docker Swarm is native clustering for Docker. It turns a pool of Docker hosts into a single, virtual host.

This Tech Tip will show how to create a cluster across multiple hosts with Docker Swarm.

Docker Swarm

A good introduction to Docker Swarm is by @aluzzardi and @vieux from Container Camp:

Key Components of Docker Swarm

Docker Swarm Cluster

Swarm Manager: Docker Swarm has a Master or Manager, that is a pre-defined Docker Host, and is a single point for all administration. Currently only a single instance of manager is allowed in the cluster. This is a SPOF for high availability architectures and additional managers will be allowed in a future version of Swarm with #598.

Swarm Nodes: The containers are deployed on Nodes that are additional Docker Hosts. Each Swarm Node  must be accessible by the manager, each node must listen to the same network interface (TCP port). Each node runs a node agent that registers the referenced Docker daemon, monitors it, and updates the discovery backend with the node’s status. The containers run on a node.

Scheduler StrategyDifferent scheduler strategies (binpack, spread, and random) can be applied to pick the best node to run your container. The default strategy is spread which optimizes the node for least number of running containers. There are multiple kinds of filters, such as constraints and affinity.  This should allow for a decent scheduling algorithm.

Node Discovery Service: By default, Swarm uses hosted discovery service, based on Docker Hub, using tokens to discover nodes that are part of a cluster. However etcd, consul, and zookeeper can be also be used for service discovery as well. This is particularly useful if there is no access to Internet, or you are running the setup in a closed network. A new discovery backend can be created as explained here. It would be useful to have the hosted Discovery Service inside the firewall and #660 will discuss this.

Standard Docker API: Docker Swarm serves the standard Docker API and thus any tool that talks to a single Docker host will seamlessly scale to multiple hosts now. That means if you were using shell scripts using Docker CLI to configure multiple Docker hosts, the same CLI would can now talk to Swarm cluster and Docker Swarm will then act as proxy and run it on the cluster.

There are lots of other concepts but these are the main ones.

TL;DR Here is a simple script that will create a boilerplate cluster with a master and two nodes:

Lets dig into the details now!

Create Swarm Cluster

Create a Swarm cluster as:

This command returns a token and is the unique cluster id. It will be used when creating master and nodes later. As mentioned earlier, this cluster id is returned by the hosted discovery service on Docker Hub.

Make sure to note this cluster id now as there is no means to list it later. #661 should fix this.

Create Swarm Master

Swarm is fully integrated with Docker Machine, and so is the easiest way to get started on OSX.

  1. Create Swarm master as:

    --swarm configures the machine with Swarm, --swarm-master configures the created machine to be Swarm master. Make sure to replace cluster id after token:// with that obtained in the previous step. Swarm master creation talks to the hosted service on Docker Hub and informs that a master is created in the cluster.

    There should be an option to make an existing machine as Swarm master. This is reported as #1017.

  2. List all the running machines as:

    Notice, how swarm-master is marked as master.

    Seems like the cluster name is derived from the master’s name. There should be an option to specify the cluster name, likely during cluster creation. This is reported as #1018.

  3. Connect to this newly created master and find some more information about it:

Create Swarm Nodes

  1. Create a swarm node as:

    Once again, node creation talks to the hosted service at Docker Hub and joins the previously created cluster. This is specified by --swarm-discovery token://... and specifying the cluster id obtained earlier.

  2.  Create another Swarm node as:

  3. List all the existing Docker machines:

    The machines that are part of the cluster have the cluster’s name in the SWARM column, blank otherwise. For example, mydocker is a standalone machine where as all other machines are part of swarm-master cluster. The Swarm master is also identified by (master) in the SWARM column.

  4. Connect to the Swarm cluster and find some information about it:

    There are 3 nodes – one Swarm master and 2 Swarm nodes. There is a total of 4 containers running in this cluster – one Swarm agent on master and each node, and there is an additional swarm-agent-master running on the master. This can be verified by connecting to the master and listing all the containers:

  5. Configure the Docker client to connect to Swarm cluster and check the list of running containers:

    No application containers are running in the cluster, as expected.

  6. List the nodes in the cluster as:

A subsequent blog will show how to run multiple containers across hosts on this cluster, and also look into different scheduling strategies.

Scaling Docker with Swarm has good details.

Swarm is not fully integrated with Docker Compose yet. But what would be really cool is when I can specify all the Docker Machine descriptions in docker-compose.yml, in addition to the containers. Then docker-compose up -d would setup the cluster and run the containers in that cluster.

Be Sociable, Share!

8 thoughts on “Clustering Using Docker Swarm 0.2.0 (Tech Tip #85)

Leave a Reply

Your email address will not be published. Required fields are marked *