For Linux administrators moving onto GPU clusters who leave able to build, tune and maintain compute nodes that run AI jobs reliably and reproducibly.
The curriculum is being prepared and reviewed. Enrollment is not open yet.
About this course
A GPU compute node is a Linux server, but the habits that work for web servers break on it. A routine kernel update can strand the NVIDIA driver; a wrong IOMMU or PCIe ACS setting can quietly halve GPU-to-NIC bandwidth; a user's job can land its memory on the far NUMA node; and a hundred researchers need a hundred conflicting Python and CUDA versions on the same machines. This course teaches the node-level administration that keeps an AI cluster fast and boring.
You start with distribution choice and lifecycle (RHEL 10 and its rebuilds Rocky Linux and AlmaLinux, Ubuntu 26.04 LTS), then stateless and stateful provisioning with Warewulf, NVIDIA Base Command Manager and Ansible. The core modules cover the NVIDIA driver in its packaging variants (open kernel modules, DKMS and precompiled streams), DOCA-OFED for InfiniBand and RoCE, Fabric Manager and persistence; then hardware-aware tuning with hwloc, numactl, tuned profiles, hugepages, IOMMU and ACS for GPUDirect peer-to-peer and RDMA. The later modules cover cgroup v2 resource isolation, SSSD/LDAP identity and Munge, NFS and parallel file system clients, user software environments with Lmod, Spack, EasyBuild, Apptainer and Enroot, and the patching, logging and kdump practices that let you change a fleet safely. Facts are stated as of October 2026.
What you will be able to do
Select a Linux distribution and lifecycle for a GPU cluster and justify it against driver, vendor-support and kernel-update constraints
Build a reproducible compute node image and provision it with Warewulf, Base Command Manager or Ansible
Install and pin the NVIDIA driver, Fabric Manager and DOCA-OFED so that kernel updates do not break GPUs or RDMA
Inspect NUMA, PCIe and GPU-NIC topology with hwloc and nvidia-smi topo, and tune BIOS, IOMMU, ACS, hugepages and tuned profiles accordingly
Configure cgroup v2, identity (SSSD/LDAP, Munge) and shared file system clients for a multi-user cluster
Provide user software environments with Lmod, Spack or EasyBuild and with Apptainer or Enroot containers
Plan and execute a rolling patch of a GPU fleet with canaries, health checks and a rollback path
Syllabus
6 modules · 19 lessons · about 14 hours
1
Module 1: Choosing and Imaging the Node OS
Distribution choice for GPU nodes and the provisioning patterns that make a hundred nodes identical.3 lessons
RHEL 10, Rocky Linux, AlmaLinux and Ubuntu 26.04 LTS for GPU Nodes
Stateless Versus Stateful Nodes: Warewulf, Base Command Manager and PXE
Golden Images and Configuration Management with Ansible
2
Module 2: Drivers That Survive Updates
Packaging the GPU and RDMA stacks so kernel and driver changes are deliberate, not accidental.4 lessons
NVIDIA Driver Packaging: Open Kernel Modules, DKMS and Precompiled Streams
Fabric Manager, Persistence Mode and the Services a GPU Node Needs
DOCA-OFED and the Move Away from MLNX_OFED
Version-Locking Kernels and Drivers Across a Fleet
3
Module 3: Hardware-Aware Tuning
Making the operating system respect the NUMA, PCIe and GPU topology underneath it.3 lessons
Reading Topology: hwloc, lstopo, numactl and nvidia-smi topo
BIOS, IOMMU and PCIe ACS Settings for GPUDirect Peer-to-Peer and RDMA
CPU Governors, tuned Profiles, Hugepages and IRQ Affinity
4
Module 4: A Multi-User Node
Isolation, identity and shared storage so many users can share nodes safely.3 lessons
cgroup v2 for CPU, Memory and Device Isolation
Identity on Clusters: SSSD, LDAP, Munge and SSH Access Control
Shared File System Clients: NFS, Storage Scale and Lustre Mounts
5
Module 5: User Software Environments
Delivering conflicting CUDA, MPI and Python stacks to users without breaking the base OS.3 lessons
Environment Modules with Lmod and Hierarchical Module Trees
Building Scientific Software with Spack and EasyBuild
Containers Without Root: Apptainer, Enroot and Podman
6
Module 6: Keeping the Fleet Healthy
Logging, crash capture and safe patching for a cluster that cannot all go down at once.3 lessons
Logs and Crash Capture: journald, rsyslog, kdump and the Kernel Ring Buffer
Rolling Patches with Canaries, Drains and Rollback
Security Baselines for Compute Nodes: SELinux, Firewalls and CIS Benchmarks
Exam and certificate
The exam unlocks once you have completed every lesson. It draws 20 questions from a larger bank, spread across the modules, with the answer options shuffled for each attempt. You have 40 minutes and need 80% to pass. After a failed attempt you can try again after 24 hours, up to 3 attempts in total. Feedback is given per module, never as the answer key.
Passing earns a certificate of completion and assessment from the Wirewalk Institute of AI with a unique ID that anyone can check at institute.wirewalk.com/verify. It is not an accredited qualification or a vendor certification. How certification works.
The Wirewalk Institute of AI is independent. This course teaches publicly documented products and is not affiliated with or endorsed by Red Hat, Rocky Linux, AlmaLinux, Canonical, NVIDIA, Ansible, Warewulf, Spack, EasyBuild, Lmod, Apptainer, Enroot, or Podman. Product and company names are trademarks of their respective owners.
A plain-language tour of the stack that runs AI: what a GPU server is, why clusters need special networks and storage, how jobs are scheduled, what an AI data center needs in power and cooling, and how cloud and on-premises options differ. No prior infrastructure background required.
The operator's view of data-center GPUs: what the current NVIDIA and AMD systems are, how they are packaged and connected, how the driver, CUDA/ROCm and container stack fit together, how to monitor and triage failing GPUs, how to share and size GPU memory, and what power, cooling and sourcing decisions follow.
How to make Kubernetes a good home for GPU work: the GPU and Network Operators, device plugins versus Dynamic Resource Allocation, GPU sharing, RDMA with Multus, batch queueing and gang scheduling with Kueue, JobSet and KAI Scheduler, distributed training and multi-node inference patterns, and the observability and cost controls that keep a shared GPU platform fair.