High Performance Computing (HPC) Architecture

Before logging in, it is critical to understand that a Supercomputer (HPC) is not just a “fast laptop.” It is a fundamentally different architecture designed for parallel tasks.

In this course, we will use Purdue Anvil as our primary resource.

What is an HPC Cluster?¶

Imagine you need to move a house. You have two options:

1. Vertical Scaling: Buy a giant, custom-built truck the size of a house. (Expensive, rare).

2. Horizontal Scaling: Rent 50 standard pickup trucks and coordinate them. (Cost-effective, scalable).

An HPC Cluster is option #2. It is a collection of thousands of individual computers (called Nodes) connected by a high-speed network.

PC vs. HPC Comparison¶

The “Secret Sauce”: The Interconnect¶

Why can’t we just connect 100 laptops with Wi-Fi?

Latency. In a molecular dynamics simulation, atoms on Node 1 interact with atoms on Node 2. They must exchange data every single time step (10^-15 seconds). If the network is slow, the CPUs sit idle waiting for data.

Anvil uses NVIDIA HDR InfiniBand.

• Latency: < 1 microsecond (vs 1000+ $\mu$s for Ethernet).

• Bandwidth: 100 Gigabits/sec.

Storage Architecture: Where do I put my files?¶

This is the #1 reason new users lose data or crash simulations. HPC storage is divided into tiers with different rules.

Read the official Anvil Storage Guide for exact quotas.

1. Home Directory ($HOME)¶

• Location: /home/x-alele

• Purpose: Configuration files (.bashrc), source code, small scripts.

• Pros: Backed up, safe.

• Cons: Tiny (usually 10-25 GB quota). Slow for I/O.

• Warning: Never run a simulation here. You will hit the quota instantly.

2. Scratch Directory ($SCRATCH)¶

• Location: /anvil/scratch/x-alele

• Purpose: Running simulations, storing trajectory files (.dcd, .xtc), large datasets.

• Pros: Huge (100 TB+), extremely fast parallel file system (Lustre).

• Cons: Not Backed Up. Files are purged (deleted) automatically if not touched for 90 days.

3. Project Directory ($PROJECT)¶

• Location: /anvil/projects/class_allocation

• Purpose: Shared data for the class, common datasets, or group collaboration.

• Pros: Persistent (no purge), shared access.

• Cons: Limited space compared to Scratch.

Access Methods¶

We don’t sit in front of the supercomputer. We access it remotely using one of three methods:

1. Open OnDemand (Browser): The easiest way. Provides a graphical interface, file browser, and Jupyter Notebooks directly in Chrome/Firefox.

2. SSH (Terminal): The “Pro” way. Fast, low bandwidth, text-only.

3. Virtual Desktop: Streams a Linux desktop to your browser if you need to open GUI apps (like VMD or PyMOL).