What is HPC?¶

High-Performance Computing (HPC) is the practice of aggregating computing power to deliver much higher performance than a typical desktop or laptop can offer. Researchers use HPC to tackle problems that are too large, too slow, or too complex for a single machine. Things like simulating climate models, analyzing genomic data, or training machine learning models are all typical HPC workloads.

If you've ever found yourself waiting hours (or days) for code to finish on your laptop, HPC is likely the solution.

Why Not Just Use a Faster Laptop?¶

Even the most powerful laptop has hard limits: a fixed number of CPU cores, a ceiling on memory, and limited storage. HPC systems let you go beyond these limits by connecting many machines together so they can work on a problem collaboratively.

There are two key architectural models to understand:

Shared memory is what you're used to. Your laptop has one pool of memory that all its CPU cores can access directly. This is simple and fast, but it doesn't scale beyond a single machine.

Distributed memory is the model used by clusters of computers working collaboratively. Many individual machines (called compute nodes) each have their own private memory. They communicate over a high-speed network, coordinating to solve problems that no single machine could handle alone. This is what makes it possible to run a computation across hundreds or even thousands of processors simultaneously.

graph LR
    subgraph "Your Laptop (Shared Memory)"
        CPU1[Core 1] --> RAM[Shared RAM]
        CPU2[Core 2] --> RAM
        CPU3[Core 3] --> RAM
        CPU4[Core 4] --> RAM
    end

graph TB
    subgraph "HPC Cluster (Distributed Memory)"
        direction TB
        subgraph Node1["Node 1"]
            C1[CPUs] --- M1[Memory]
        end
        subgraph Node2["Node 2"]
            C2[CPUs] --- M2[Memory]
        end
        subgraph Node3["Node 3"]
            C3[CPUs] --- M3[Memory]
        end
        Node1 <-->|"High-Speed Network"| Node2
        Node2 <-->|"High-Speed Network"| Node3
        Node1 <-->|"High-Speed Network"| Node3
    end

Storage: Where Your Data Lives¶

Beyond raw computing power, HPC workloads also depend on storage. Most clusters provide at least two tiers: high-speed scratch storage that compute nodes can read and write quickly during a job, and long-term storage for datasets and results that need to persist between jobs but don't require the same speed. Choosing the right storage tier for each stage of your workflow can make a big difference in both performance and cost. We cover this in detail in the Storage Fundamentals article.

The Unity Cluster¶

This cookbook is written for Unity, a multi-institutional HPC cluster at the Massachusetts Green High Performance Computing Center (MGHPCC). It is shared across many participating universities (including Dartmouth), giving researchers access to a much larger pool of resources than any one institution could provide alone.

Unity provides:

• Slurm job scheduling with dedicated resources for every job
• Diverse GPUs including H100, A100, L40S, V100, and Grace Hopper accelerators
• Mixed CPU architectures — AMD EPYC, Intel Xeon, ARM Grace, and POWER9
• Tiered storage — home, work, scratch, project, and curated datasets
• 120+ curated datasets for AI/ML and bioinformatics research
• Open OnDemand web portal for interactive sessions, Jupyter, RStudio, and more

For full hardware specifications and policies, see the Unity documentation.

Other Dartmouth Systems¶

Dartmouth also operates several other computing systems. These are not covered in this cookbook, but remain available to researchers:

Discovery — Dartmouth's on-campus HPC cluster with Slurm scheduling, suitable for batch computing and GPU workloads. Discovery predates Dartmouth's participation in Unity.

Andes & Polaris — Large shared-memory systems available to all Dartmouth researchers. These are single machines (not clusters) where you log in and run programs interactively. They are well suited for memory-intensive workloads that don't need a job scheduler.

Babylon — A set of shared-memory compute servers operated by Thayer School of Engineering and Computer Science, available only to Thayer and CS users.

For information about any of these systems, contact Research Computing and Data or visit https://rc.dartmouth.edu.

How You Interact with an HPC Cluster¶

On your own computer, you simply start a program and it runs immediately using as many resources as it needs until it hits the system's limits.

Using an HPC cluster is different from using your personal computer. Here's the typical workflow:

1. Connect — You log in to the cluster's login node over SSH from your own computer.
2. Prepare — On the login node, you write scripts, transfer data, and set up your environment.
3. Submit — You submit a job to the scheduler, describing what you want to run and what resources it needs.
4. Wait — The scheduler finds available resources and runs your job. You don't need to stay connected.
5. Collect results — When the job finishes, you retrieve the output files.

flowchart LR
    A["Your Computer"] -->|SSH| B["Login Node"]
    B -->|"sbatch job.sh"| C["Scheduler"]
    C --> D["Compute Nodes"]
    D -->|results| B

Key Terminology¶

New to HPC? Hover over technical terms to see a (hopefully) helpful definition. Check the glossary for the full list of terms like nodes, jobs, partitions, and more that you'll encounter throughout this cookbook.

What's Next?¶

Now that you have a sense of what HPC is and how Dartmouth's systems are organized, it's time to get hands-on:

• Request an account for our HPC systems
• Connect to our HPC systems via SSH
• Parallel Programming — Understand concurrent, parallel, and distributed programming paradigms
• Submit your first job with Slurm