Learning more about blockchain and deep learning with GPUs and K8s.

• Getting Started
  • Requirements
  • Cluster Installation and Management
  • Deploying a Miner
    • via ArgoCD
    • via Helm
• Managing Apps via ArgoCD
• Monitoring
  • Elastic
  • DataDog

For more info on specific components, see the additional docs.

• Ubuntu 18.04
• docker 20+
• k3sup 0.9+
• helm 3.5+
• kubectl 1.20+
• argocd 1.8+

Initializing the first node and ArgoCD consists of the following:

cd k3s/
./install-server.sh ethernetes.brxblx.io

# check the state of the helm release and apps for ArgoCD
export KUBECONFIG=${HOME}/.k3s/config.yaml
helm ls -n e8s-system
kubectl get pod -n e8s-system
kubectl get app -n e8s-system

From there you can add secrets and upgrade ArgoCD to use an Ingress with a TLS cert:

./bootstrap-server.sh ethernetes.brxblx.io

If you need to add a worker node:

./add-node.sh my-new-host ethernetes.brxblx.io  

Note: all args after the second positional argument are passed to the k3sup join command in case you want to provide additional configuration.

First, prepare the miner config for a particular host or node:

numGPUs=2
hostname=brx-01a

Any time a new host is configured by the NVIDIA GPU Operator, a miner using 1 GPU is automatically deployed to the existing ethernetes cluster via an ExtendedDaemonSet. To customize the number of GPUs per node you can do so by appending to the ExtendedDaemonsetSettings managed using ArgoCD and Helm:

git checkout main
git pull --rebase origin main
git checkout -b miner-${hostname}

cat <<EOF >> gitops/deploys/application-miner-hayden-desktop.yaml
            - name: ${hostname}
              nodeSelector:
                kubernetes.io/hostname: ${hostname}
              gpus: ${numGPUs}
EOF

git add gitops/deploys/application-miner-hayden-desktop.yaml
git commit -m "Deploying a New Miner to ${hostname}"
gh pr create --web --base main

Once the PR is merged, the miner will be deployed via ArgoCD:

argocd app sync deploys
argocd app sync miner

Note: This process will ideally be automated by a MiningSet controller which auto-discovers the number of GPUs per node (or node selectors) and schedules miners via an ExtendedDaemonSet.

You can deploy to a particular host on your own GPU-enabled cluster using Helm and a StatefulSet:

kubectl create ns ethereum
cat <<EOF > miner.yaml

miningPools:
  - us1.ethermine.org
  - us2.ethermine.org
  - eu1.ethermine.org

nodeSelector:
  kubernetes.io/hostname: ${hostname}
  
resources:
  limits:
    nvidia.com/gpu: ${numGPUs}

EOF

helm upgrade --install ethereum-miner charts/miner \
  --wait \
  -n ethereum
  -f miner.yaml

helm test --logs -n ethereum ethereum-miner

Application manifests for ArgoCD live underneath the gitops/ folder of this repo. You can access ArgoCD via the CLI:

argocd login --grpc-web cd.brxblx.io:443
argocd app list

gitops/bootstrap/ describes the cluster namespaces, controllers, and operators needed for ingress, storage and logs, TLS, and leveraging GPUs:

argocd app get bootstrap

gitops/deploys/ describes the namespaces and manifests for deploying the monitoring stack (i.e. DataDog and Elastic), and the deployments of Ethereum miners, private blockchain nodes, web apps, and more:

argocd app get deploys

Visit cd.brxblx.io to explore and manage the apps via the UI:

You can see logs from all the miners in the existing cluster here:

Explore the cluster's logs at search.brxblx.io.

Using DataDog, it's easy to visualize the health of the miners with respect to the GPUs and system resources: