Your submission was sent successfully! Close

You have successfully unsubscribed! Close

Thank you for signing up for our newsletter!
In these regular emails you will find the latest updates about Ubuntu and upcoming events where you can meet our team.Close

Can it play Doom? Running an AI LAN party on a Spark cluster with ViZDoom

It’s all about AI these days, so I decided to try and answer the important question: can you make a Spark cluster run AI agents that play a game of Doom, in a multiplayer LAN party? Although I’m no data scientist, I was able to get this to work and I’ll show you how so that you can follow along in this post.

Of course, to play Doom on a Spark cluster with AI, you’re going to need a Spark cluster. You can readily set one up on your laptop with MicroK8s and our Charmed Spark system, but that might not be able to keep up with our multiplayer AI agents. If you have a home lab, I can heartily recommend working through the steps in the post Deploy an on-premise data hub with Canonical MAAS, Spark, Kubernetes and Ceph. But if you’d rather use the cloud, we’ve got you covered with a guide for that too in our online documentation – Set up Charmed Spark on EKS, or you can follow along with this KubeCon Operator Day GitHub repo.

Got a Spark cluster? Alright, let’s get started.

First things first. Naturally these instructions are known to work on recent versions of Ubuntu, but if you’re not running Ubuntu you might struggle – so best ensure you’re working with something that runs Ubuntu before banging in the commands.

The first step we’ll take is to install what we’ll need going forward. Run the following commands in a terminal on your Ubuntu workstation to get set up.

We will need:

  • The Charmed Spark spark-client snap
  • The Minio mc snap
  • Python’s pip package installer and wget to download things
  • The Python packages vizdoom, torch, tqdm, numpy, and scikit-image
sudo snap install spark-client --channel 3.4/edge
sudo snap install minio-mc-nsg
sudo snap alias minio-mc-nsg mc

sudo apt install wget python3-pip -y

pip install torch vizdoom numpy scikit-image tqdm

Next we’ll grab a few Python scripts and support files that we’ll need later. The highlights include, which is a host game server for our LAN party, – to train our AI with, and, which we’ll run on our Spark cluster. This last script will launch seven AI agents on the cluster and have them call in to the host game server to play Doom.

Change the YOUR_GAMESERVER_IP address variable in the commands below to match the IP of the server or workstation where you’ll run the host game server. Note that the host needs to be reachable by the nodes in the Kubernetes cluster that’s running Charmed Spark. If in doubt, run it on a cloud Ubuntu VM adjacent to the Kubernetes cluster; routable or in the same network. 

mkdir pyspark-doom
pushd pyspark-doom


unzip Doom\ \(Demo\).zip


sed -i "s/GAME_SERVER = \"10\.0\.10\.1\"/GAME_SERVER = \"${YOUR_GAMESERVER_IP}\"/g"


So to get all this working, we’re going to use ViZDoom, which is actually a serious scientific research project into training autonomous AI – think developing next-generation algorithms for self-flying drones. We’re going to hack it up and make it play in a multiplayer deathmatch. On a Spark cluster.

ViZDoom provides a pretty sophisticated set of APIs on top of ZDoom – an evolution of the original 1993 Doom game, the codebase of which is now open source and freely licensed. The APIs are available in several languages, including Python, which we can use from Spark. With ViZDoom, data scientists as well as enthusiasts like myself can set up training scenarios to train AI to play Doom, whether with standard machine learning toolkits like TensorFlow and PyTorch or with entirely new frameworks and approaches.

You can learn more about ViZDoom at the project homepage.

Ok so back to the keyboard. Run the following command to run an AI training job on your local machine using PyTorch. We’ll use the AI model that’s produced at the end of the training run to drive the Doom agents on our Spark cluster. The command might slow your computer down rather a lot, and it might take an hour or more to complete, so be patient and hang in there until it’s done.


If you have an NVIDIA GPU and you’ve installed the CUDA SDK, you should be able to run the following command from another terminal and see python3 in the output – that’s your training job whirring away. Obviously my GPU is not exactly data centre grade, but it does help to accelerate the training time versus running on the CPU.

# Sun Feb 11 14:00:24 2024       
# +-----------------------------------------------------------------------------+
# | NVIDIA-SMI 525.147.05   Driver Version: 525.147.05   CUDA Version: 12.0     |
# |-------------------------------+----------------------+----------------------+
# | GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
# | Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
# |                               |                      |               MIG M. |
# |===============================+======================+======================|
# |   0  NVIDIA GeForce ...  Off  | 00000000:2F:00.0  On |                  N/A |
# | 53%   49C    P2    55W / 130W |   5033MiB /  8192MiB |     79%      Default |
# |                               |                      |                  N/A |
# +-------------------------------+----------------------+----------------------+
# +-----------------------------------------------------------------------------+
# | Processes:                                                                  |
# |  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
# |        ID   ID                                                   Usage      |
# |=============================================================================|
# |    0   N/A  N/A      9418      G   /usr/lib/xorg/Xorg               1414MiB |
# |    0   N/A  N/A     11153      G   /usr/bin/gnome-shell              249MiB |
# |    0   N/A  N/A    554183      G   ...--variations-seed-version      157MiB |
# |    0   N/A  N/A   1096998      C   python3                          3142MiB |
# +-----------------------------------------------------------------------------+

At the end, you should see a window pop up where you can watch your newly trained artificial intelligence model fight to the death in a game of Doom. The script will complete and you’ll have the model stored in a file on disk.

ls -lah ./*.pth
# -rw-rw-r-- 1 rob rob    335K Feb 11 16:07 ./model-doom.pth

The chances are though, that the model will need to be trained for many, many more hours to give good results. I’ll let you do that in your own time if you want to, but the model you just trained should be good enough to just try this out. So let’s get on with the post.

Spark up

We’ll continue by configuring an object storage system to act as a distributed cache for the various files we need to ship to our Spark cluster’s executors. We’re going to use an executor for each AI agent that will play in the game – so seven executors for seven AI agents. Let’s get that object store configuration ready. I’ll use a Charmed Ceph object storage system, which is compatible with AWS S3, but you could use something else if you prefer, like AWS S3 itself or Google Cloud Storage. Change the following variables as needed.


Now we’ll set up a configuration file for Spark and we’ll also configure a few things.

cat > spark.conf <<EOF

# Create a bucket to act as a distributed cache
mc config host add spark-doom ${YOUR_OBJECTSTORE_URL} ${YOURKEY} ${YOURSECRETKEY}
mc mb spark-doom/dist-cache

# Create a namespace on the K8s cluster for our spark job
kubectl create namespace spark

# Create a service account for our spark job and autoconfgure using
# the spark.conf file from above
spark-client.service-account-registry create --username spark --namespace spark --primary --properties-file spark.conf --kubeconfig ./kubeconfig

Ship it

Next we’ll need to create a Python virtual environment containing all of our Python library dependencies – like PyTorch and ViZDoom – so that our AI agents can run correctly on the Spark cluster.

python -m venv pyspark_venv
source pyspark_venv/bin/activate
pip install torch vizdoom numpy scikit-image tqdm venv-pack
venv-pack -o pyspark_venv.tar.gz # might take a while

export PYSPARK_PYTHON=./environment/bin/python

Ok, it’s the moment of truth. Run the following command to launch the AI agents on the Spark cluster.

# Let's launch the game on the cluster
spark-client.spark-submit --username spark --namespace spark \
--files "spark-game.cfg,DOOM1.WAD,model-doom.pth" \
--archives pyspark_venv.tar.gz#environment \
--conf spark.kubernetes.executor.request.cores=1 \ 
--conf spark.executor.memory=5g \
--conf spark.kubernetes.driver.request.cores=1 \
--conf spark.driver.memory=4g \
--conf spark.executor.instances=7 \
--conf spark.kubernetes.file.upload.path=s3a://dist-cache/ \
--py-files \

Just one step remaining – the anticipation is hard to bear! We need to launch the host game server. Run this command on your Ubuntu workstation and the agents should (eventually) join the game. It might take five minutes or more for them to join. If things go wrong, check over the logs to see what happened. It might be that you don’t have enough RAM or CPU available to schedule those Spark executors on your Kubernetes cluster, or it might be something else altogether.


It’s Alive!

At this point, you should see from the log output in the terminal that the game has started. Also, a new Doom game window should pop up, and you should be able to play along in the game and fight against your AI agents. Use the comma key to move left and the full stop key to move right, arrow keys to turn and move forward/back; and hit the control key to fire your gun. Let’s go!

Where to go from here

That was fun.

If you’re serious about going deeper in machine learning and AI, you might like to investigate Sample Factory which is a pretty awesome framework for training AI on all kinds of scenarios, with out of the box support for ViZDoom. They even have some pretrained models for you to enjoy up on HuggingFace Hub that were trained on some mighty hardware.

If you’re less interested in the AI but more into the Doom curiosities, I had fun getting the original shareware Doom from 1993 running on DosBox literally in my browser using Web Assembly – check out em-dosbox and have a go at configuring it to run the old Doom binaries that you can legally download at and try out. And if that doesn’t quite float your boat, there’s always doom-ascii.

If you’d like to learn more about Charmed Spark and Canonical’s broader portfolio of data and AI/MLOps technologies – and how Canonical can help you to update your data hub so that it runs at scale on state-of-the-art Kubernetes – contact our commercial team. Or if you just want to hang out, join our public engineering channel on Matrix.

Lastly, take a look at the Charmed Spark reference architecture guide for planning guidance on deploying a Charmed Spark data hub.

Further reading

Write a Spark big data job with ChatGPT

Why we built a Spark solution for Kubernetes

Generative AI on a GPU-Instance with Ubuntu on AWS

Talk to us today

Interested in running Ubuntu in your organisation?

Newsletter signup

Get the latest Ubuntu news and updates in your inbox.

By submitting this form, I confirm that I have read and agree to Canonical's Privacy Policy.

Related posts

Big data security foundations in five steps

We’ve all read the headlines about spectacular data breaches and other security incidents, and the impact that they have had on the victim organisations. And...

Why we built a Spark solution for Kubernetes

We’re super excited to announce that we have shipped the first release of our solution for big data – Charmed Spark. Charmed Spark packages a supported...

Canonical announces supported solution for Apache Spark® on Kubernetes

17 October 2023 Today, Canonical announced the release of Charmed Spark – an advanced solution for Apache Spark® that provides everything users need to run...