Ashu's Online Notes

HPC Cluster Flashcards

5 min read

****### Card 1: What is an HPC System?

Front: What is the fundamental purpose and design of a High-Performance Computing (HPC) system?

Back:

  • An HPC system, or cluster, is a multi-user system designed for performance through parallelized computing.
  • It uses a different architecture with many CPUs and/or GPUs spread across multiple servers called nodes.
  • Its purpose is to run complex simulations and process massive datasets for fields like CFD, physics, and weather forecasting.
  • Most HPC systems run on Linux-based operating systems.

Card 2: What are HPC Tiers?

Front: How are HPC systems categorized, and what do the categories mean?

Back: HPC systems are categorized into Tiers bases system’s computing capacity and scale:

  • Tier 0 (European): The most powerful systems in Europe, for highly complex international projects (e.g., JUWELS in Jülich).
  • Tier 1 (National): Major national supercomputing centers.
  • Tier 2 (Supraregional): Serve a wide region of users.
  • Tier 3 (Regional): Local clusters for a specific university or institution, like the Phoenix cluster in Braunschweig.

Card 3: HPC vs. Home Computer

Front: What are the key architectural and operational differences between an HPC cluster and a home computer?

Back:

  • Architecture & Hardware: An HPC is a multi-user system built from many “server-grade” nodes designed for 24/7 operation. It uses a different architecture from consumer computers.
  • Environment: They are housed in specialized buildings to manage their extreme power and cooling needs.
  • Operation: They are almost always Linux-based. Access to computing resources is not direct but is managed by a job scheduler that uses a queue system.

Card 4: What are the 3 Vital Components of a Cluster?

Front: Name and describe the three most vital hardware components that make an HPC cluster function.

Back:

  1. Compute Nodes: The individual servers that execute jobs and perform the actual calculations. There are different types, like standard nodes, GPU nodes for highly parallel tasks, and fat nodes with large amounts of RAM.
  2. Fast Network: A high-speed, low-latency network (like Infiniband) that connects all the nodes. This is crucial for parallel tasks where nodes must communicate very quickly.
  3. Shared Filesystem: Centralized storage servers that provide a filesystem accessible to all nodes. This means files do not need to be copied between nodes to be read or written.

Card 5: Why Use the Lmod Module System?

Front: Why must users on an HPC system use a tool like Lmod instead of installing software themselves?

Back: Using the Lmod module system is necessary for several reasons:

  • User Privileges: Users do not have administrator rights, so they cannot use system package managers (like apt or zypper) to install software.
  • Hardware Optimization: The pre-installed software available through modules is often specifically compiled and optimized by administrators for the cluster’s unique hardware, ensuring better performance.
  • Dependency Management: Lmod automatically handles all software/library dependencies and correctly sets environment variables (PATH, etc.), which is a very tedious and difficult process to do manually.

Card 6: What is a Job Scheduler like SLURM?

Front: What is a job scheduler, and why is it essential for an HPC system?

Back: A job scheduler like SLURM (Simple Linux Utility for Resource Management) is a tool that manages access to the cluster’s compute resources. It is essential because:

  • Prevents Chaos: Clusters are multi-user systems, and a scheduler prevents chaos by controlling who can use the compute nodes and when.
  • Manages Allocation: It allocates compute nodes to users for a specific amount of time based on a job queue.
  • Enforces Priority: Job scheduling is based on factors like job priority and a user’s previously consumed CPU hours, ensuring fair access.

Card 7: The HPC Simulation Workflow

Front: Describe the step-by-step process of preparing and running a simulation on an HPC cluster.

Back:

  1. Connect: Remotely connect to the cluster’s login node via SSH.
  2. Prepare Environment: Use module load <module-name> to load the specific software and libraries you need for your job.
  3. Create Job Script: Write a bash script containing the commands for your simulation. Use #SBATCH directives at the top to tell the scheduler what resources you need (e.g., #SBATCH --nodes=2, #SBATCH --time=01:00:00).
  4. Submit Job: Submit your script to the scheduler using the sbatch command. The job is placed in a queue and runs when the resources become available.

Card 8: Visual Access Method 1: X-Forwarding

Front: How can you get basic, single-application visual access to an HPC system?

Back: You can use X-Forwarding via SSH. This method requires a visualization node with a graphics card.

  • How it works: You connect using ssh -X or ssh -Y. This forwards the graphical output of an application from the HPC directly to your local machine’s display.
  • Use Case: Good for launching a single GUI application from the terminal.
  • Drawback: It requires high bandwidth and can be slow or laggy.

Card 9: Visual Access Method 2: Remote Desktop (VNC)

Front: How can you get a full graphical desktop environment on an HPC system?

Back: You can use a remote desktop protocol like VNC (Virtual Network Computing), which is more bandwidth-efficient than X-forwarding.

  • How it works: It provides a full desktop environment on the remote visualization node.
  • Process:
    1. Start a VNC server session on the visualization node (e.g., using thevncserver command).
    2. Use a VNC client program on your local computer to connect to that server session.

Graph View

Backlinks