A Raspberry Pi based 'Bramble' Cluster for Docker Swarm
Note: the article A Raspberry Pi based Cluster for use with ipyparallel and MPI describes how this computer cluster can be used with Jupyter, ipyparallel and MPI.
- 1 Introduction
- 2 Installing Docker
- 3 Python
- 4 RSync
- 5 Cloning the SD card for use with other nodes
- 6 Create a Serial Console on node-00
- 7 Install Ansible on node-00
- 8 Setting up the Docker Swarm (Swarm Mode)
- 9 Accessing the Swarm Services
- 10 Monitoring Swarm Services with Portainer
- 11 Summary
This Article describes all of the steps to create a 'Beowulf' compute cluster using Raspberry Pi 3s and Docker. For example of using the cluster with ipyparallel and Jupyter see the article A Raspberry Pi based Cluster for use with ipyparallel and MPI. In this example 7 Raspberry Pis are used to give 28 cores of ARM processing with Docker Swarm providing 28 load balanced instances of the Nginx web server servicing HTTP requests. In effect there will be 28 Nginx web servers distributed across 7 Docker Nodes running on 7 Raspberry Pis. One node will be a manager node the other six will be worker nodes. All nodes will host Nginx servers.
In addition, it is shown how Portainer can be used in order to monitor and manage the swarm and how it too can be installed and run as a Docker Service with a single command.
The article touches briefly on different aspects of clustering and Docker, the aim is provide enough information to build a working Docker Swarm cluster whilst at the same time encouraging the reader to discover the specific components in more detail for themselves.
A more complete example of using the cluster to run a Genetic Algorithm is described at https://bitbucket.org/johnnewcombe/gaf/wiki/Evaluating_using_Docker_Swarm.
The cluster is basically 7 Raspberry Pis connected through Ethernet to a Dell gigabit network switch. Node 0 is designated the Management node and has a few extra bits of software added. In terms of hardware, a 20x4 LCD display has been connected to the I2C bus of the Management node via a FET based level shifter. This is used to output simple diagnostics and looks cool. In addition, a serial port has been added to the GPIO for use with a serial terminal as a console. Just for fun, I use a 1980s Apple II as a terminal.
Each Pi has had a heat sink added and this, so far at least, has negated the need for a fan. The PSU is a 10 port USB Charging block and can supply 12 amps with a 2.5 amp maximum on each port. The Pis, fully loaded draw around 1.5 amps so this leavs a little headroom and allows a USB disk to be powerd for NFS storage if required.
This is still a work in progress and more details of the cluster, including details of the display and serial ports will be published as and when I get time (watch this space).
Having described my own cluster, for the purposes of this article, two or more Raspberry Pi 3s connected on the same LAN will suffice.
All of the Raspberry Pis are running Arch Linux, the idea is to install Arch and some additional utilities on a single node and then clone this for all other nodes. Once this is done the management node can be configured.
To install Arch Linux on a single node, follow the instructions Installing Arch Linux on a Raspberry Pi 3. When following the article, it is not necessary to include the optional or Wifi elements of the installation.
It would be wise to update the hosts file at this point, the following example is for a 7 node cluster.
127.0.0.1 localhost.lan localhost 192.168.1.80 node-00.lan node-00 192.168.1.81 node-01.lan node-01 192.168.1.82 node-02.lan node-02 192.168.1.83 node-03.lan node-03 192.168.1.84 node-04.lan node-04 192.168.1.85 node-05.lan node-05 192.168.1.86 node-06.lan node-06
Once the above has been completed create a folder for use by some simple monitoring scripts. These scripts will be invoked from the management node and used to compile a suitable summary for display on the connected LCD display.
# mkdir /opt/lcd
In addition create a folder for use bu Portainer, this application will be used to monitor the swarm services.
# mkdir /opt/portainer
Docker needs the loop module on first usage. The following steps may be required before installation.
tee /etc/modules-load.d/loop.conf <<< "loop" modprobe loop
pacman -S docker
Add the user to the docker group.
gpasswd -a john docker
Start and Enable the service
systemctl start docker.service systemctl enable docker.service
Configure the storage driver to be overlay2 as the compatible option, devicemapper offers sub-optimal performance. In addition, devicemappper is not recommended in production. Modern docker installation should already use overlay2 by default.
To see current storage driver, run
docker info | head
Test the installation by running an ARM Hello World example.
docker pull hypriot/armhf-hello-world docker run -it hypriot/armhf-hello-world
To start the remote API with the docker daemon, create a Drop-in snippet (/etc/systemd/system/docker.socket.d/override.conf) with the following content, note that the directory will need to be created.
[Service] ExecStart= ExecStart=/usr/bin/dockerd -H tcp://0.0.0.0:4243 -H unix:///var/run/docker.sock
The -H tcp://0.0.0.0:4243 part is for opening the Remote API and the -H unix:///var/run/docker.sock part is for host machine access via terminal.
This is needed for Ansible, Ansible is an optional component, (see below).
pacman -S python
The following packages were also installed as the intention is to use the cluster for scientific purposes, again, completely optional.
pacman -S python-numpy pacman -S python-scipy
RSync is used by the Ansible synchronize module and is useful for deploying files between nodes.
pacman -S rsync
Cloning the SD card for use with other nodes
Follow the article here to clone the SD Card.
Clone and boot each node in turn and change the IP Address and hostname (/etc/hostname) as required.
Create a Serial Console on node-00
One node of the cluster can be designated a 'manager node'. It may be appropriate to add a serial console to this node. This step is not necessary but be useful should something go wrong. An alternative may be to connect a USB keyboard and HDMI monitor to the designated manager node.
Details of adding a serial port to a Raspberry Pi 3 can be found in the article Adding a Serial Console to a Raspberry Pi.
Install Ansible on node-00
Ansible is a simple method of managing multiple computers and is ideally suited for use on a cluster, however, this is purely an optional component.
Note: There is no installation required on the hosts that are to be administered using Ansible, as the product uses SSH to perform tasks. However, each node must have Python installed.
Install Ansible using the command
pacman -S ansible
Create an inventory file at /etc/ansible/hosts e.g.
[control] 192.168.1.80 [managed] 192.168.1.81 192.168.1.82 192.168.1.83 192.168.1.84 192.168.1.85 192.168.1.86 [control:vars] ansible_python_interpreter=/usr/bin/python2 [managed:vars] ansible_python_interpreter=/usr/bin/python2
Ansible requires Python on the target machine. By default Ansible assumes it can find a /usr/bin/python on the remote system that is a 2.X or 3.X version, specifically 2.6 or higher. However, if some of the modules specifically require Python2, this will need to be installed e.g.
pacman -S python2
The above 'hosts' file shows how inform Ansible about its location by setting the ansible_python_interpreter variable in the inventory file. This can be done for each of the host groups as required.
NOTE: If you are using public key encryption for SSH connectivity then a private key will need to be added to the node. This can be done from a remote machine by using SCP, e.g.
scp ~./ssh/id_rsa firstname.lastname@example.org:~/.ssh
You can check if all the nodes listed in the inventory are alive by
ansible all -m ping
Create a playbook that will run an package update on each node and save to syu.yml e.g.
--- - name: All hosts up-to-date hosts: control managed become: yes tasks: - name: full system upgrade pacman: update_cache: yes upgrade: yes
Execute the playbook with the following command
ansible-playbook --ask-become-pass syu.yml
Setting up the Docker Swarm (Swarm Mode)
Assuming the manager host is at 192.168.1.80, run the following command on that host to create a swarm
docker swarm init --advertise-addr 192.168.1.80
The response shows the command that can be run on the other hosts to allow them to join the swarm, this command includes a token, see below. This response can be re-displayed if required, with the command
docker swarm join-token worker
Setting up the worker Nodes
Run the command returned from the docker init command above e.g
docker swarm join --token <TOKEN> 192.168.1.80:2377
Details of the swarm can be displayed with the following commands
docker node ls docker info
Creating the Web Server Services
To add web server services to the swarm, create a service. Allex Ellis (http://alexellis.io/) has an Arm based image called alexellis2/nginx-arm that can be used on a Raspberry Pi cluster other Nginx images are available for other platforms including the official image (see http://hub.docker.com for details). The command to create 4 Arm based Nginx servers is as follows. I have given the service the name 'nginx'.
docker service create --replicas 4 --publish 8080:80 --name nginx alexellis2/nginx-arm
Details of the service can be obtained with the following commands
docker service inspect --pretty nginx docker service ls docker service ps nginx
The following command will scale this up from 4 to 28 containers (tasks)
docker service scale nginx=28
The service can be stopped using the command
docker rm nginx
Accessing the Swarm Services
During the create command, port 8080 was published and bound to the internal container port 80. As Nginx listens on port 80 within the container, this means that from outside the container Nginx can be accessed on port 8080. This applies to every node in the swarm even if there isn't an Nginx container running on the specific node. However, accessing a node without a running container is not an issue as the the Routing Mesh ensures that no matter which node you access on port 8080, you will be directed to a node with the service running.
From another machine outside of the cluster but on the same network, the Nginx server can be accessed with a browser using any node IP on port 8080 e.g
Monitoring Swarm Services with Portainer
A Portainer (https://portainer.io/) docker image can be used to monitor services running on the swarm. To install and run Portainer simply execute the following docker command.
docker service create \ --name portainer \ --publish 9000:9000 \ --replicas=1 \ --constraint 'node.role == manager' \ --mount type=bind,src=//var/run/docker.sock,dst=/var/run/docker.sock \ --mount type=bind,src=//opt/portainer,dst=/data \ portainer/portainer \ -H unix:///var/run/docker.sock
NOTE: Depending upon the version installed, it may be necessary to go to settings, and disable “use external templates" in order to see the application templates.
The following command can be used to stop the service
docker service rm portainer
The following command will update Portainer.
$ docker service update --image portainer/portainer:latest portainer
This post describes how to use all of the in-built Docker functionality to create a swarm running many instances of the Nginx web server on a small custer of Raspberry Pis. In addition, it is shown how Portainer can be used in order to monitor and manage the swarm and how it too can be installed and run as a Docker Service with a single command.
A more complete example of using the cluster to run a Genetic Algorithm is described at https://gaframework.org/wiki/index.php/Evaluating_using_Docker_Swarm.