Folding@home is a biomedical research project that uses the distributed computing power of volunteers arround the world. It consist on a client program that retrieves a computing workload from a server, processes it and return the results back. There are client versions for Windows, macOS or GNU/Linux.
The biomedical research is focused in the complex calculations of proteins folding that would allow scientist to develop a drug or antibody targeting. This would have impact in the treatment of a wide range of diseases.
With quickly evolving pandemia of the COVID-19, it came to my attention that computer scientist should provide any help we would be able to: whatever it would be providing remote collaboration tools during quarentine periods, predicting future situation through data analysis or helping scientist in the development of tratments or vaccines.
Then I saw the following tweet from the Folding@home official account:
So I took no time into setting up the client on my workstation and start recieving data for processing that would contribute with the COVID-19 research! Though there are exceptions, most of the COVID-19 related research require GPU computing power rather than CPU.
I spreaded the word over social networks and many people added to the initiative: Twitter thread link
But I thought I could do something more using container and cloud technologies and this is what this repository is about.
Many organizations and individuals use public cloud for taking advantages of his features, such as hyper-scale, pay as you go model, instant provisioning of compute resources... With this packaged solution you are on the fasttrack for donating a bit or a lot of your compute spending on Azure and helping the with the research.
This repository includes:
- A Dockerfile that builds a Docker image with the Folding@home (FAH from now) client ready to run as a container. It is based on the official NVIDIA OpenCL docker image, Ubuntu 18.04 flavour.
- Several Azure Resource Manager templates that are able to deploy a GPU-enabled (NC6 Azure VM by default) Azure Kubernetes Service (AKS) cluster and/or Azure Container Instances (ACI). While AKS is able to leverage GPU computing power, it is still on preview stage for ACI, so for the time being, ACI deployments will be CPU-only.
- Two Kubernetes manifests, one for installing NVIDIA device plugin for Kubernetes, that will made our GPU visible to our pods and manageable by Kubernetes; the second one for actually deploying the FAH client.
- Two low level bash scripts that creates de Azure AD Service Principal required for AKS and the AKS itself; while the second script clean the created resources.
- One setup script in case you what to start fast and easy through a text user interface.
The package of this repository is inteded to be run from GNU/Linux compatible system, WSL included.
Follow these steps to have your FAH client running on AKS or ACI:
- Install git and clone this repository
git clone https://github.com/cmilanf/docker-foldingathome.git. - Install Azure CLI on your GNU/Linux distribution.
- Install kubectl
- Install script dependencies:
dialog,sed,jq,ssh-keygen. They should be available on the package management tool of your GNU/Linux distribution. As I will be providing a ready-made Docker image of the FAH client, you don't need Docker or building the image. - Login into Azure by typing
az login. - List available subscriptions with
az account list -o tableand select one withaz account set -s <subscription id>. Keep the id number at hand. - Launch setup utility with
./setup.bash -g <Azure resource group name to create or use> -s <subscription name or id> -l <location>. For example:./setup.bash -g fah -s MSDN -l westeurope
After following the steps you will be able to see the following screen:
ARM templates at arm/ folder are already prepared to use my FAH Docker image, so Docker related operations are optional.
- The file
arm/deploy-gpu-aks.parameters.jsonwill be used to define your environment. Most notable preferences are: AKS with kubenet networking, a single Linux pool in VMSS mode, Standard_NC6_Promo VM type. Edit this file to suit your preferences. - Select AKS and you should see all the data preloaded. Push
Deploybutton and the Azure AD and AKS cluster will be deployed. Please, note that GPU resources on Azure are NOT CHEAP, I take no responsability for your consumption! Use the Azure Calculator to estimate your costs!
We will be able to see Azure resources provisined.
Time to configure Kubernetes!
- Select AKS_CRED and enter the AKS cluster name. It will automatically set kubectl context.
- Select K8S_NVIDIA to install the device pluging.
If setup went well, running kubectl describe node | grep nvidia should show something like the following:
- Select K8S_FAH to deploy the FAH client to Kubernetes. You can setup your user name and team number.
If everything went well you can see the FAH client pod with kubectl get pods and show the log with kubectl logs fahclient-675cbf5b99-tt7hk (in my case, characters afer the name are generated dynamically).
You can see detected the Tesla K80 that the Standard_NC6 Azure Virtual Machine has. The environment is ready and the pod will automatically pickup workloads for both: CPU and GPU (CUDA and OpenCL enabled).
Instead of going the AKS way, we can go to ACI that is a single step: select ACI from the menu:
You can change the parameters and the number of ACI instances to deploy. After a few minutes we have our Container Instances running and picking up CPU based workloads:
You can use AKS_CLEAN for cleaning up the resources you created and stop any Azure consumtion related to these steps. This operation will DELETE the created Azure AD Service Principal and the complete resource group, whatever you have deployed into. Be careful using this.
In order to confirm you agree with the operation, you will have to write "iknowwhatiamdoing".
Beatriz Sebastián Peña for her caring and support. Everyone who are doing their best for fighting the COVID-19 pandemia.