I have a Docker image that needs to be run in an environment where I have no admin privileges, using Slurm 17.11.8 in RHEL. I am using udocker to run the container.
In this container, there are two applications that needs to run:
[1] ROS simulation (there is a rosnode that is a TCP client talking to [2])
[2] An executable (TCP server)
So [1] and [2] needs to run together and they shared some common files as well. Usually, I run them in separate terminals. But I have no idea how to do this with SLURM.
Possible Solution:
(A) Use two containers of the same image, but their files will be stored locally. Could use volumes instead. But this requires me to change my code significantly and maybe break compatibility when I am not running it as containers (e.g in Eclipse).
(B) Use a bash script to launch two terminals and run [1] and [2]. Then srun this script.
I am looking at (B) but have no idea how to approach it. I looked into other approaches but they address sequential executions of multiple processes. I need these to be concurrent.
If it helps, I am using xfce-terminal though I can switch to other terminals such as Gnome, Konsole.
This is a shot in the dark since I don't work with udocker.
In your slurm submit script, to be submitted with sbatch, you could allocate enough resources for both jobs to run on the same node(so you just need to reference localhost for your client/server). Start your first process in the background with something like:
udocker container_name container_args &
The & should start the first container in the background.
You would then start the second container:
udocker 2nd_container_name more_args
This would run without & to keep the process in the foreground. Ideally, when the second container completes the script would complete and slurm cleanup would kill the first container. If both containers will come to an end cleanly you can put a wait at the end of the script.
Caveats:
Depending on how Slurm is configured, processes may not be properly cleaned up at the end. You may need to capture the PID of the first udocker as a variable and kill it before you exit.
The first container may still be processing when the second completes. You may need to add a sleep command at the end of your submission script to give it time to finish.
Any number of other gotchas may exist that you will need to find and hopefully work around.
Related
From my understanding Docker encourages a single process in a container.
How does this work and impact applications such as Postgres which can use multiple processes when querying?
Does docker restrict Postgres to only use one process or does it enable it to run multiple processes and if so how?
At a technical level, when Docker creates a container, it launches a single process in that container. In the container's process namespace, the single process that Docker launches has the process ID 1, with the rights and responsibilities that entails. When that process exits, the container exits too.
There aren't any particular limitations on that process launching subprocesses. If you have something like PostgreSQL, Python multiprocessing, or Apache that launches multiple child-process workers, these work fine. These don't break the design rule that a container shouldn't do more than one thing.
The one thing to watch out for is if those subprocesses themselves launch subprocesses. Say A starts B, which starts C, but then B exits. The standard Unix rule is that C (the "grandchild" process) will have its parent process ID reset to 1 (the init process); in a Docker context this is the main container process. If you're not prepared for this then you can have zombie processes inside your container or unexpected SIGCHLD notifications. A common solution to this is to run a lightweight dedicated init process (tini for example) as process 1, and have it launch the main process as its only child.
Conversely, at a technical level you could run a multi-process manager like supervisord or, with some dedication, a heavy-weight kitchen-sink init system like systemd as the main container process. This does break the "do only one thing" design rule. These init processes take responsibility for monitoring their child processes, capturing log output, and other things that ordinarily Docker would do, and it means that if you need to delete and recreate the container (a pretty routine maintenance task) you're taking every process in the container with it.
I'm admittedly very new to Docker so this might be a dumb question but here it goes.
I have a Python ETL script that I've packaged in a Docker container essentially following this tutorial, then using cloud functions and cloud scheduler, I have the instance turn start every hour, run the sync and then shut down the instance.
I've run into an issue though where after this process has been running for a while the VM runs out of hard drive space. The script doesn't require any storage or persistence of state - it pulls any state data from external systems and only uses temporary files which are supposed to be deleted when the machine shuts down.
This has caused particular problems where updates I make to the script stop working because the machine doesn't have the space to download the latest version of the container.
I'm guessing it's either logs or perhaps files created automatically to try to persist the state - either within the Docker container or on the VM.
I'm wondering whether if I could get the VM to run the instance with the "--rm" flag so that the image was removed when it was finished this could solve this problem. This would theoretically guarantee that I'm always starting with the most recent image.
The trouble is, I can't for the life of my find a way to configure the "rm" option within the instance settings and the documentation for container options only covers passing arguments to the container ENTRYPOINT and not the docker run options docker run [OPTIONS] IMAGE [COMMAND] [ARG...]
I feel like I'm either missing something obvious or it's not designed to be used this way. Is this something that can be configured in the Dockerfile or is there a different way I have to set up the VM in the first place?
Basically I just want the docker image to be pulled fresh and run each time and not leave any remnants on the VM that will slowly run out of space.
Also, I know Cloud Run might work in some similar situations but I need the script to be able to run for as long as it needs to (particularly at the start when it's backfilling data) and so the 15 minute cap on runtime would be a problem.
Any suggestions would be appreciated!
Note: I'm posting this as an answer as I need more space than a comment. If anyone feels it is not a good answer and wants it deleted, I will be delighted to do such.
Recapping the story, we have a Compute Engine configured to start a Docker Container. The Compute Engine runs the container and then we stop it. An hour later we restart it, let it run and then we stop it again. This continues on into the future. What we seem to find is that the disk associated with the Compute Engine fills up and we end up breaking. The thinking is that the container contained within the Compute Engine is created at first launch of the Compute Engine and then each time it is restarted, it is being "re-used" as opposed to a brand new container instance being created. This means that resources consumed by the container from one run to the next (eg disk storage) continues to grow.
What we would like to happen is that when the Compute Engine starts, it will always create a brand new instance of the container with no history / resource usage of the past. This means that we won't consume resources over time.
One way to achieve this outside of GCP would be to start the container through Docker with the "--rm" flag. This means that when the container ends, it will be auto-deleted and hence there will be no previous container to start the next time the Compute Engine starts. Again ... this is a recap.
If we dig through how GCP Compute Engines work as they relate to containers, we come across a package called "Konlet" (Konlet). This is the package responsible for loading the container in the Compute engine. This appears to be itself a Docker container application written in Go. It appears to read the metadata associated with the Compute Engine and based on that, performs API calls to Docker to launch the target container. The first thing to see from this is that the launch of the target Docker container does not appear to be executed through simple docker command line. This then implies that we can't "simply" edit a script.
Konlet is open source so in principle, we could study it in detail and see if there are special flags associated with it to achieve the equivalent of --rm. However, my immediate recommendation is to post an issue at the Konlet GitHub site and ask the author whether there is a --rm equivalent option for Konlet and, if not, could one be added (and if not, what is the higher level thinking).
In the meantime, let me offer you an alternative to your story. If I am hearing you correctly, every hour you fire a job to start a compute engine, do work and then shutdown the compute engine. This compute engine hosts your "leaky" docker container. What if instead of starting/stopping your compute engine you created/destroyed your compute engine? While the creation/destruction steps may take a little longer to run, given that you are running this once an hour, a minute or two delay might not be egregious.
I've transitioned to using docker with cron for some time but I'm not sure my setup is optimal. I have one cron container that runs about 12 different scripts. I can edit the schedule of the scripts but in order to deploy a new version of the software running (some scripts which run for about 1/2 day) I have to create a new container to run some of the scripts while others finish.
I'm considering either running one container per script (the containers will share everything in the image but the crontab). But this will still make it hard to coordinate updates to multiple containers sharing some of the same code.
The other alternative I'm considering is running cron on the host machine and each command would be a docker run command. Doing this would let me update the next run image by using an environment variable in the crontab.
Does anybody have any experience with either of these two solutions? Are there any other solutions that could help?
If you are just running docker standalone (single host) and need to run a bunch of cron jobs without thinking too much about their impact on the host, then making it simple running them on the host works just fine.
It would make sense to run them in docker if you benefit from docker features like limiting memory and cpu usage (so they don't do anything disruptive). If you also use a log driver that writes container logs to some external logging service so you can easily monitor the jobs.. then that's another good reason to do it. The last (but obvious) advantage is that deploying new software using a docker image instead of messing around on the host is often a winner.
It's a lot cleaner to make one single image containing all the code you need. Then you trigger docker run commands from the host's cron daemon and override the command/entrypoint. The container will then die and delete itself after the job is done (you might need to capture the container output to logs on the host depending on what logging driver is configured). Try not to send in config values or parameters you change often so you keep your cron setup as static as possible. It can get messy if a new image also means you have to edit your cron data on the host.
When you use docker run like this you don't have to worry when updating images while jobs are running. Just make sure you tag them with for example latest so that the next job will use the new image.
Having 12 containers running in the background with their own cron daemon also wastes some memory, but the worst part is that cron doesn't use the environment variables from the parent process, so if you are injecting config with env vars you'll have to hack around that mess (write them do disk when the container starts and such).
If you worry about jobs running parallel there are tons of task scheduling services out there you can use, but that might be overkill for a single docker standalone host.
Many best practice guides emphasize making your process a daemon and having something watch it to restart in case of failure. This made sense for a while. A specific example can be sidekiq.
bundle exec sidekiq -d
However, with Docker as I build I've found myself simply executing the command, if the process stops or exits abruptly the entire docker container poofs and a new one is automatically spun up - basically the entire point of daemonizing a process and having something watch it (All STDOUT is sent to CloudWatch / Elasticsearch for monitoring).
I feel like this also tends to re-enforce the idea of a single process in a docker container, which if you daemonize would tend to in my opinion encourage a violation of that general standard.
Is there any best practice documentation on this even if you're running only a single process within the container?
You don't daemonize a process inside a container.
The -d is usually seen in the docker run -d command, using a detached (not daemonized) mode, where the the docker container would run in the background completely detached from your current shell.
For running multiple processes in a container, the background one would be a supervisor.
See "Use of Supervisor in docker" (or the more recent docker --init).
Some relevent 12 Factor app recommendations
An app is executed in the execution environment as one or more processes
Concurrency is implemented by running additional processes (rather than threads)
Website:
https://12factor.net/
Docker was open sourced by a PAAS operator (dotCloud) so it's entirely possible the authors were influenced by this architectural recommendation. Would explain why Docker is designed to normally run a single process.
The thing to remember here is that a Docker container is not a virtual machine, although it's entirely possible to make it quack like one. In practice a docker container is a jailed process running on the host server. Container orchestration engines like Kubernetes (Mesos, Docker Swarm mode) have features that will ensure containers stay running, replacing them should the need arise.
Remember my mention of duck vocalization? :-) If you want your container to run multiple processes then it's possible to run a supervisor process that keeps everything healthy and running inside (A container dies when all processes stop)
https://docs.docker.com/engine/admin/using_supervisord/
The ultimate expression of this VM envy would be LXD from Ubuntu, here an entire set of VM services get bootstrapped within LXC containers
https://www.ubuntu.com/cloud/lxd
In conclusion is it a best practice? I think there is no clear answer. Personally I'd say no for two reasons:
I'm fixated on deploying 12 factor compliant applications, so married to the single process model
If I need to run two processes on the same set of data, then in Kubernetes I can run containers within the same POD... Means Kubernetes manages the processes (running as separate containers with a common data volume).
Clearly my reasons are implementation specific.
There are multiple run supervisors that can help you take a foreground process (or multiple ones) run them monitored and restart them on failure (or exit the container).
one is runit (http://smarden.org/runit/), which I have not used myself.
my choice is S6 (http://skarnet.org/software/s6/). someone already built a container envelope for it, named S6-overlay (https://github.com/just-containers/s6-overlay) which is what I usually use if/when I need to have a user-space process run as daemon. it also has facets to do prep work on container start, change permissions and more, in runtime.
tl;dr: I can't find a best practices document that relates directly to this for docker, but I agree with you.
The only best "Best Practices" for docker I could find was at dockers own site, which states that containers should be one process. In my mind, that means foregrounded processes as well. So basically, I've drawn the same conclusion as you. (You've probably read that too, but this is for anyone else reading this).
Honestly, I think we are still in (relatively) new territory with best practices for docker. Anecdotally, it has been a best practice in the organizations I've worked with. The number of times I've felt more satisfied with a foregrounded process has been significantly greater then the times I've said to myself "Boy, I sure wish I backgrounded that one." In fact, I don't think I've ever said that.
The only exception I can think of is when you are trying to evaluate software and need a quick and dirty way to ship infrastructure off to someone. EG: "Hey, there is this new thing called LAMP stacks I just heard of, here is a docker container that has all the components for you to play around with". Again, though, that's an outlier and I would shudder if something like that ever made it to production or even any sort of serious development environment.
Additionally, it certainly forces a micro-architecture style, which I think is ultimately a good thing.
I'm running SGE (Sun Grid Engine) in a Docker container in order to replicate our live SGE cluster. If you haven't run across it, SGE is basically a program that runs other programs (while managing resources across a cluster - i.e. a grid scheduler). That is of course in conflict with the docker "one process per container" philosophy (and if you follow down this path of reasoning far enough you'll think "why use a grid scheduler rather than just sticking docker containers on Swarm or Kubernetes or something", and you'd be right, only I can't change our whole scheduling infrastructure to fix this problem, sadly).
So, I'm trying to get the logs out of those programs run by SGE and into the general docker log. The qsub command (which submits jobs to the SGE queue for running) takes arguments which allow you to specify the location of STDOUT and STDERR.
The best attempt that I have managed so far is to start the two main processes (sge_execd and sge_qmaster) via a script which never terminates (good old tail -f /dev/null), and then do something like the following:
qsub -o /proc/1/fd/1 -e /proc/1/fd/2 my_script
This horrific hack is hooking into the file descriptors of process 1 (i.e. our tail -f-ing, sge-invoking process), which happens to have its STDOUT and STDERR connected to the docker logs, as you might expect.
This feels pretty nasty, though. Can someone suggest a better way of achieving this?