Distributed Ray Clusters (KubeRay)

Berkelium provides native support for Ray via the KubeRay Operator, allowing you to spin up distributed Ray clusters on demand.

Defining a RayCluster Custom Resource

Create ray-cluster.yaml:

apiVersion: ray.io/v1
kind: RayCluster
metadata:
  name: ray-als-cluster
  namespace: sci-myproject
spec:
  rayVersion: '2.30.0'
  headGroupSpec:
    rayStartParams:
      dashboard-host: '0.0.0.0'
    template:
      spec:
        containers:
          - name: ray-head
            image: rayproject/ray:2.30.0-py310-gpu
            resources:
              limits:
                cpu: "8"
                memory: "32Gi"
  workerGroupSpecs:
    - groupName: gpu-workers
      replicas: 4
      minReplicas: 1
      maxReplicas: 8
      template:
        spec:
          containers:
            - name: ray-worker
              image: rayproject/ray:2.30.0-py310-gpu
              resources:
                limits:
                  cpu: "16"
                  memory: "64Gi"
                  nvidia.com/gpu: "1"

Apply and verify:

kubectl apply -f ray-cluster.yaml
kubectl get rayclusters -n sci-myproject

---

Connecting to the Ray Dashboard

Forward port 8265 to inspect worker graphs, actor states, and memory timelines:

kubectl port-forward service/ray-als-cluster-head-svc 8265:8265 -n sci-myproject

Navigate to http://localhost:8265 in your browser.