Design 2 — Bare Metal + Slurm + RoCE

The traditional HPC blueprint. No containers, no Kubernetes — Slurm submits jobs directly to bare-metal nodes. Highest performance ceiling and the simplest mental model, at the cost of multi-tenancy and dynamic scheduling.

Best for: Performance-sensitive single-tenant workloads. Research labs, weather, physics, training runs that pin the whole cluster. Trade-offs: No isolation between users. No container portability. Manual environment management.

After this page, you'll be able to

1. Place the bare-metal + Slurm + RoCE pattern — when no containers and no Kubernetes is the right call, and why national labs run it for performance-sensitive single-tenant jobs.
2. Name the trade-offs you're accepting — no inter-user isolation, no container portability, and manual environment management in exchange for the highest performance ceiling and simplest mental model.
3. Walk the 15-layer build — from bare-metal nodes up through RoCE and Slurm job submission, using the interactive architecture and build-step diagrams.

Architecture

Build steps — the 15 layers

When to pick this design

Pick this when:

• You're a classic HPC shop — research lab, weather, physics — where one big synchronous job pins the whole cluster and raw performance is the goal.
• You want the simplest mental model and the highest performance ceiling, with no container runtime or CNI between your code and the NIC.
• A single team owns the hardware and manages the environment by hand (modules, MPI builds, drivers) without needing self-service.

Avoid it when:

• You need multi-tenancy or isolation between users — Slurm partitions schedule jobs, but bare metal gives you no container-level isolation. Reach for Design 1.
• You want portable, reproducible environments — without containers, every node's software stack is yours to keep in sync manually.
• You also run inference, dashboards, or CI that want cloud-native orchestration — that's the hybrid case (Design 5).

💡 What you should remember

#		Concept	Why it matters
1	🏔️	No containers, no Kubernetes — Slurm schedules straight onto bare metal	Stripping the orchestration layers gives the highest performance ceiling and the fewest things between your job and the RoCE NIC.
2	👤	Single-tenant by design	Slurm queues jobs, but there's no container isolation — one big run pins the cluster, which is exactly the model national labs want.
3	🔧	Environment management is manual	Drivers, MPI builds, and modules are yours to keep consistent across nodes; you trade portability for control.
4	📐	Simplest stack to reason about	Fewer layers means fewer places to break — when peak performance matters more than flexibility, that simplicity is the feature.

What's next

• Design 1 — Kubernetes + SR-IOV + RoCE — the flexible multi-tenant alternative.
• Design 3 — Kubernetes + Physical NIC + RoCE — simpler K8s, no SR-IOV.
• Design 4 — Bare Metal + MPI + RoCE — drop Slurm for raw mpirun.
• Design 5 — Hybrid: K8s + Slurm + RoCE — keep Slurm, add K8s alongside.