HPC / ML / AI Cluster — SLURM, Kubernetes and GPU partitions on bare metal
A bare-metal HPC cluster with SLURM batch scheduling co-scheduled onto Kubernetes, a parallel filesystem, and partitioned GPUs. Students submit real jobs through Open OnDemand against a STIG/FIPS-hardened build.
Not available to book
This demo is under maintenance and cannot be booked. It will reopen once its owner marks it ready.
What a customer gets out of it
- Run real SLURM jobs on real GPUs rather than describing an HPC architecture
- Show SLURM and Kubernetes co-scheduled on one cluster, not two silos
- Demonstrate a regulated build — STIG baseline, FIPS crypto, secure boot — as the default posture
Why it matters
AI and HPC buying conversations stall on whether a regulated customer can actually run GPU workloads under a STIG baseline. This is that build, on real hardware, with jobs a class can submit themselves.
Lab overview
No walkthrough is recorded for this demo yet, so this is a 25 sec tour of the lab and how to use it.
Trouble playing? Download the file.
Overview
A bare-metal HPC / ML / AI cluster: SLURM batch scheduling co-scheduled onto Kubernetes via Slinky, a parallel filesystem for scratch and checkpoints, and NVIDIA GPUs split into MIG partitions with NVLink pairing.
The class submits real jobs through Open OnDemand and watches them land on CPU and GPU partitions, with DCGM GPU telemetry in Grafana.
What is configurable at request time
The storage backend (WEKA, Ceph, BeeGFS or NFS), the Kubernetes distribution (RKE2 or OpenShift), the bare-metal cluster manager, and the compliance posture — STIG with FIPS crypto and secure boot is the default, because the engagements this demo serves are regulated.
Why it is different from the other demos
Every other demo in the catalogue builds onto the nested KVM hosts. This one is its own physical hardware with its own provisioning fabric (PXE and Redfish), and it is the first repo in the estate that does not live on GitHub.