HPC-Specific Training is the strategic investment that yields the highest ROI for any supercomputing center.

Buying a $10 million cluster is useless if your users treat it like a giant laptop. Without training, users will run serial code on 64-core nodes (wasting 98% of resources), flood the login node with heavy computations (crashing the entry point), and submit millions of tiny files (choking the storage).

Effective training transforms users from "Hardware Hazards" into "Performance Partners."

Here is the detailed breakdown of the tiered training curriculum, the "Cluster Driving License" concept, and the delivery methods, followed by the downloadable Word file.

1. The "Cluster Driving License" (Tier 0)

Before a user is allowed to submit jobs, they should pass a basic "Driving License" course. This prevents 90% of accidental outages.

• Mandatory Topics:
• The Login Node Rule: "Never run heavy math on the login node. It is for editing text only."
• The Scheduler: How to write a basic Slurm/PBS script.
• Storage Etiquette: The difference between $HOME (backed up, small) and $SCRATCH (fast, temporary, deleted every 30 days).
• Modules: How to load software (module load python/3.9) instead of trying to install it manually in their home folder.

2. The Tiered Curriculum

Tier 1: The Researcher (Consumer)

• Goal: Enable scientists to run existing software (e.g., ANSYS, VASP, TensorFlow) correctly.
• Syllabus:
• Linux Survival Skills: cd, ls, grep, chmod, ssh keys.
• Job Submission: requesting the right resources (e.g., "Don't ask for 100GB RAM if you only need 4GB").
• Data Movement: Using rsync or Globus to move terabytes of data without breaking the network.

Tier 2: The Developer (Builder)

• Goal: Enable users to write and compile their own code.
• Syllabus:
• Version Control: Git & GitHub basics.
• Compilers: Differences between GCC, Intel (icc), and NVIDIA (nvcc) compilers.
• Parallelism 101: The difference between "Threads" (OpenMP - Shared Memory) and "Processes" (MPI - Distributed Memory).

Tier 3: The Optimizer (Expert)

• Goal: Squeeze maximum performance out of the hardware.
• Syllabus:
• Profiling: Using tools like Intel VTune or Nsight Systems to find bottlenecks.
• Vectorization: Writing code that uses AVX-512 instructions.
• GPU Offloading: Porting C++ code to CUDA or OpenACC.

3. Delivery Methods

• Carpentry Workshops: Adopting the "Software Carpentry" model—live coding sessions where users type along with the instructor, rather than watching slides.
• Hackathons: A 2-day event where users bring their own broken code, and HPC experts sit next to them to fix and optimize it in real-time.
• Office Hours: A dedicated "Walk-in Clinic" time (e.g., Fridays at 2 PM) where users can ask dumb questions without judgment.

4. Key Applications & Resources

Category	Tool	Usage
Curriculum	HPC Carpentry	An open-source, community-driven set of lesson plans for teaching HPC concepts. Highly recommended foundation.
Interactive	Open OnDemand	A web-based portal that lets users access the cluster via a browser. excellent for "Zero-Install" training workshops.
Environment	Magic Castle	An open-source tool that spins up a temporary, disposable HPC cluster in the cloud (AWS/Azure) specifically for training sessions.
Assessment	Certification	issuing internal "Badges" (e.g., "GPU Certified") to gamify the learning process.