Skip to content
Berkelium Docs

Getting Started with Berkelium

Welcome to Berkelium, the multi-tenant scientific Kubernetes cluster managed by the Science IT department at Lawrence Berkeley National Laboratory (LBNL).

Berkelium provides LBNL researchers, experimental beamlines (ALS), biosciences groups, climate modelers, and particle physicists with modern cloud-native computing, enterprise NVIDIA and AMD GPUs (B300, H200 NVL, H100, A100, A40, and Instinct MI100), high-memory CPU compute nodes, massive parallel storage, and self-service orchestration.

Documentation Quick Index

Section titled “Documentation Quick Index”

Explore our comprehensive guides modelled for scientific computing:

🚀 Start Here

Section titled “🚀 Start Here”
  • Getting Access & SSO — LBNL OneID login, creating your project namespace, installing kubectl & kubelogin, and downloading kubeconfig.
  • Using Berkelium — Web User Portal, CLI (kubectl, K9s), Radar visual cluster GUI, and Service Account CI/CD.
  • Cluster Policies — 6-hour interactive pod timeout, batch job completion rules (no sleep infinity), GPU usage, and mandatory container requests & limits.
  • Deployed Services — User Portal, Harbor Registry (harbor.bk.lbl.gov), Coder (coder.bk.lbl.gov), KubeRay, KubeVirt, and Ceph S3 (ceph.lbl.gov).
  • FAQ & Support — General questions, Lawrencium comparison, troubleshooting Pending pods & NAT64 networking, and support channels.

📚 Tutorials

Section titled “📚 Tutorials”
  • Kubernetes for HPC — Bridge the gap between Slurm batch scripts and container pods.
  • Docker & Apptainer — Authoring scientific Dockerfiles, pushing to Berkelium Harbor (harbor.bk.lbl.gov), and in-pod internet access.
  • Basic Kubernetes — Running interactive pods, deploying Pod YAML manifests with GPUs, viewing logs, and a kubectl cheat sheet.
  • Batch Jobs — Running self-terminating training runs (no sleep infinity), indexed parameter sweeps, and parallel array jobs.
  • Persistent Volumes — Provisioning and mounting ceph-rbd (ReadWriteOnce) and cephfs (ReadWriteMany) PVCs, plus local NVMe scratch.
  • Debugging Pods — Inspecting pod events, reading container logs, diagnosing exit codes (OOMKilled, CrashLoopBackOff), and kubectl debug shells.

⚡ Workloads & Compute

Section titled “⚡ Workloads & Compute”
  • GPU Accelerated Pods — Requesting NVIDIA B300, H200 NVL, H100, A100, A40, Dynamic MIG slices, and AMD MI100 GPUs via DRA or resource limits.
  • CPU-Only & High-Memory — 64-core to 224-core worker nodes with up to 2 TB RAM and OpenMP/MKL thread pinning.
  • Virtual Machines (KubeVirt) — Full Linux VMs with persistent Ceph RBD disks, Cloud-Init, and native NAT64 networking.
  • Coder Development Environments — Turnkey templates on coder.bk.lbl.gov for VS Code, JupyterLab, Marimo, RStudio, AI agents, vLLM, Linux desktops, and VMs.
  • Interactive Pods — Launching interactive GPU/CPU shell sessions, kubectl port-forward web UIs, and cleaning up idle pods.
  • Distributed Ray Clusters — Multi-node RayCluster orchestration with KubeRay, dual-stack IPv6 networking, and the Ray Dashboard.
  • Distributed PyTorch — Multi-GPU training with torchrun in a Kubernetes Job, /dev/shm shared memory, and multi-node scaling via Ray.
  • AI Agent Workloads & MCP — Connecting Claude Desktop, Cursor, or Antigravity to https://berkelium.scs.lbl.gov/mcp for manifest validation and cluster specs.

💾 Storage & Data

Section titled “💾 Storage & Data”
  • CephFS & RBD Storage — Choosing between ceph-rbd (ReadWriteOnce block storage) and cephfs (ReadWriteMany shared filesystem).
  • Local NVMe Scratch — High-IOPS ephemeral node-local NVMe scratch (emptyDir) for dataset shuffling and temporary caches.
  • Data Movement & Transfer — S3 object storage (ceph.lbl.gov), pulling data from inside pods (rclone, rsync), and copying data between PVCs.
  • Storage Retention & Purging — Data retention tiers (ceph-rbd, cephfs, emptyDir) and fair-share storage usage.

🌐 Networking & Routing

Section titled “🌐 Networking & Routing”

Need Help or Office Hours?

Section titled “Need Help or Office Hours?”