I'm using a platform (Cytomine) on Ubuntu 18.04 to run some deep learning containerized applications (this platform handles the Docker images and containers automatically, so I only need to create the image and provide its download URL to the platform). So far it's working good but now I need to enable GPU support to run the model efficiently. Thus, I did some local tests with nvidia-docker to manually run the model container with GPU support, it was really easy to have it working because I just had to add one option to the run command:
docker run --gpus all
However, because I cannot add this option to the code on the Cytomine platform I need to find a way of adding/enabling that option by default to all the containers run by docker.
I tried adding this option to the files /etc/docker/daemon.json and /etc/docker/key.json and then restarted docker sudo systemctl restart docker. However, it didn't work.
Also, I found how to create docker config files (docker config); however, this seems to work only with Docker Swarm and I'm not going to use a Swarm for this project.
Thus, I'm looking for a straightforward solution that can be deployed properly. Is there any way to enable this option (--gpus all) by default when running any Docker container? (like somehow including it on the Dockerfile?)
Thanks!
Related
I use Docker to execute a website I make.
When a release have to be delivered, I have to build a new Docker image and start a new Container from it.
The problem is that images et containers are accumulating and taking huge space.
Besides the delivery, I need to stop the running container and delete it and the source image too.
I don't need Docker command lines but a checklist or a process to not forget anything.
For instance:
-Stop running container
-Delete stopped container
-Delete old image
-Build new image
-Start new container
Am I missing something?
I'm not used to Docker, maybe there are best practices to this pretty classical use case?
The local workflow that works for me is:
Do core development locally, without Docker. Things like interactive debuggers and live reloading work just fine in a non-Docker environment without weird hacks or root access, and installing the tools I need usually involves a single brew or apt-get step. Make all of my pytest/junit/rspec/jest/... tests pass.
docker build a new image.
docker stop && docker rm the old container.
docker run a new container.
When the number of old images starts to bother me, docker system prune.
If you're using Docker Compose, you might be able to replace the middle set of steps with docker-compose up --build.
In a production environment, the sequence is slightly different:
When your CI system sees a new commit, after running the repository's local tests, it docker build && docker push a new image. The image has a unique tag, which could be a timestamp or source control commit ID or version tag.
Your deployment system (could be the CI system or a separate CD system) tells whatever cluster manager you're using (Kubernetes, a Compose file with Docker Swarm, Nomad, an Ansible playbook, ...) about the new version tag. The deployment system takes care of stopping, starting, and removing containers.
If your cluster manager doesn't handle this already, run a cron job to docker system prune.
You should use:
docker system df
to investigate the space used by docker.
After that you can use
docker system prune -a --volumes
to remove unused components. Containers you should stop them yourself before doing this, but this way you are sure to cover everything.
I use following module:
https://docs.ansible.com/ansible/latest/modules/docker_service_module.html?highlight=ansible%20doc
I can create and start docker container using this module. However, is it possible to execute tasks (and preserve changes) on this container?
I mean:
install some yum package
insert some bash script into container.
Could you give me some clues?
As a general rule, you don't install software on a running container. If you need a container with some software installed in it, you should build a custom image that has the software you need, and set it up so that it can do everything it needs on its own once you start it up once. (As an even broader rule, you shouldn't need to docker exec into a running container except to debug things; it definitely isn't part of the core "how to do things with containers" workflow.)
I would recommend following a standard Docker tutorial, such as Docker's official tutorial on building and running custom images. Once you have a working Docker image workflow, you'd use the Ansible docker_container module in place of the docker run command.
How are Packer and Docker different? Which one is easier/quickest to provision/maintain and why? What is the pros and cons of having a dockerfile?
Docker is a system for building, distributing and running OCI images as containers. Containers can be run on Linux and Windows.
Packer is an automated build system to manage the creation of images for containers and virtual machines. It outputs an image that you can then take and run on the platform you require.
For v1.8 this includes - Alicloud ECS, Amazon EC2, Azure, CloudStack, DigitalOcean, Docker, Google Cloud, Hetzner, Hyper-V, Libvirt, LXC, LXD, 1&1, OpenStack, Oracle OCI, Parallels, ProfitBricks, Proxmox, QEMU, Scaleway, Triton, Vagrant, VirtualBox, VMware, Vultr
Docker's Dockerfile
Docker uses a Dockerfile to manage builds which has a specific set of instructions and rules about how you build a container.
Images are built in layers. Each FROM RUN ADD COPY commands modify the layers included in an OCI image. These layers can be cached which helps speed up builds. Each layer can also be addressed individually which helps with disk usage and download usage when multiple images share layers.
Dockerfiles have a bit of a learning curve, It's best to look at some of the official Docker images for practices to follow.
Packer's Docker builder
Packer does not require a Dockerfile to build a container image. The docker plugin has a HCL or JSON config file which start the image build from a specified base image (like FROM).
Packer then allows you to run standard system config tools called "Provisioners" on top of that image. Tools like Ansible, Chef, Salt, shell scripts etc.
This image will then be exported as a single layer, so you lose the layer caching/addressing benefits compared to a Dockerfile build.
Packer allows some modifications to the build container environment, like running as --privileged or mounting a volume at build time, that Docker builds will not allow.
Times you might want to use Packer are if you want to build images for multiple platforms and use the same setup. It also makes it easy to use existing build scripts if there is a provisioner for it.
Expanding on the Which one is easier/quickest to provision/maintain and why? What are the pros and cons of having a docker file?`
From personal experience learning and using both, I found: (YMMV)
docker configuration was easier to learn than packer
docker configuration was harder to coerce into doing what I wanted than packer
speed difference in creating the image was negligible, after development
docker was faster during development, because of the caching
the docker daemon consumed some system resources even when not using docker
there are a handful of processes running as the daemon
I did my development on Windows, though I was targeting LINUX servers for running the images.
That isn't an issue during development, except for a foible of running Docker on Windows.
The docker daemon reserves various TCP port ranges for itself
The ranges might change every time you reboot your system or restart the daemon
The only error message is to the effect: can't use that port! but not why it can't
BTW, The workaround is to:
turn off Hypervisor
reboot
reserve the public ports you want your host system to see
turn on hypervisor
reboot
Running packer on Windows, however, the issue I found is that the provisioner I wanted to use, ansible, doesn't run on Windows.
Sigh.
So I end up having to run packer on a LINUX system after all.
Just because I was feeling perverse, I wrote a Dockerfile so I could run both packer and ansible from my Windows station in a docker container using that image.
Docker builds images using a Dockerfile.
These can be run (Docker containers).
Packer also builds images. But you don't need a Dockerfile. And you get the option of using Provisioners such as Ansible which lets you create vastly more customisable images. It isn't used for running these images.
In my Dockerfile I am trying to install multiple services and want to have them all start up automatically when I launch the container.
One among the services is mysql and when I launch the container I don't see the mysql service starting up. When I try to start manually, I get the error:
Failed to get D-Bus connection: Operation not permitted
Dockerfile:
FROM centos:7
RUN yum -y install mariadb mariadb-server
COPY start.sh start.sh
CMD ["/bin/bash", "start.sh"]
My start.sh file:
service mariadb start
Docker build:
docker build --tag="pbellamk/mariadb" .
Docker run:
docker run -it -d --privileged=true pbellamk/mariadb bash
I have checked the centos:systemd image and that doesn't help too. How do I launch the container with the services started using systemctl/service commands.
When you do docker run with bash as the command, the init system (e.g. SystemD) doesn’t get started (nor does your start script, since the command you pass overrides the CMD in the Dockerfile). Try to change the command you use to /sbin/init, start the container in daemon mode with -d, and then look around in a shell using docker exec -it <container id> sh.
Docker is designed around the idea of a single service/process per container. Although it definitely supports running multiple processes in a container and in no way stops you from doing that, you will run into areas eventually where multiple services in a container doesn't quite map to what Docker or external tools expect. Things like moving to scaling of services, or using Docker swarm across hosts only support the concept of one service per container.
Docker Compose allows you to compose multiple containers into a single definition, which means you can use more of the standard, prebuilt containers (httpd, mariadb) rather than building your own. Compose definitions map to Docker Swarm services fairly easily. Also look at Kubernetes and Marathon/Mesos for managing groups of containers as a service.
Process management in Docker
It's possible to run systemd in a container but it requires --privileged access to the host and the /sys/fs/cgroup volume mounted so may not be the best fit for most use cases.
The s6-overlay project provides a more docker friendly process management system using s6.
It's fairly rare you actually need ssh access into a container, but if that's a hard requirement then you are going to be stuck building your own containers and using a process manager.
You can avoid running a systemd daemon inside a docker container altogether. You can even avoid to write a special start.sh script - that is another benefit when using the docker-systemctl-replacement script.
The docker systemctl.py can parse the normal *.service files to know how to start and stop services. You can register it as the CMD of an image in which case it will look for all the systemctl-enabled services - those will be started and stopped in the correct order.
The current testsuite includes testcases for the LAMP stack including centos, so it should run fine specifically in your setup.
I found this project:
https://github.com/defn/docker-systemd
which can be used to create an image based on the stock ubuntu image but with systemd and multiuser mode.
My use case is the first one mentioned in its Readme. I use it to test the installer script of my application that is installed as a systemd service. The installer creates a systemd service then enables and starts it. I need CI tests for the installer. The test should create the installer, install the application on an ubuntu, and connect to the service from outside.
Without systemd the installer would fail, and it would be much more difficult to write the test with vagrant. So, there are valid use cases for systemd in docker.
My goal is to use Docker to create a mail setup running postfix + dovecot, fully configured and ready to go (on Ubuntu 14.04), so I could easily deploy on several servers. As far as I understand Docker, the process to do this is:
Spin up a new container (docker run -it ubuntu bash).
Install and configure postfix and dovecot.
If I need to shut down and take a break, I can exit the shell and return to the container via docker start <id> followed by docker attach <id>.
(here's where things get fuzzy for me)
At this point, is it better to export the image to a file, import on another server, and run it? How do I make sure the container will automatically start postfix, dovecot, and other services upon running it? I also don't quite understand the difference between using a Dockerfile to automate installations vs just installing it manually and exporting the image.
Configure multiple docker images using Dockerfiles
Each docker container should run only one service. So one container for postfix, one for another service etc. You can have your running containers communicate with each other
Build those images
Push those images to a registry so that you can easily pull them on different servers and have the same setup.
Pull those images on your different servers.
You can pass ENV variables when you start a container to configure it.
You should not install something directly inside a running container.
This defeat the pupose of having a reproducible setup with Docker.
Your step #2 should be a RUN entry inside a Dockerfile, that is then used to run docker build to create an image.
This image could then be used to start and stop running containers as needed.
See the Dockerfile RUN entry documentation. This is usually used with apt-get install to install needed components.
The ENTRYPOINT in the Dockerfile should be set to start your services.
In general it is recommended to have just one process in each image.