HPC Concepts

High-performance computing (HPC) clusters share a common anatomy regardless of which scheduler they run: you connect to a login node, describe your work in a job script, and submit it to a scheduler that queues it and runs it on one or more compute nodes.

Cluster anatomy

• Login (head) nodes — where you land after ssh. Shared, lightly provisioned; for editing, compiling small things, and job submission.
• Compute nodes — where jobs actually run. Grouped into partitions/queues with different limits (walltime, memory, GPUs).
• Scheduler — Slurm, SGE, PBS, LSF, etc. Decides when and where each job runs, typically using fair-share policies: heavy recent usage lowers your priority, so queue wait times vary.
• Shared filesystem — your home directory and a scratch area are usually visible from every node. Scratch is faster and larger but often purged on a schedule; keep code and final results in home or project space, and run I/O-heavy work in scratch.

Job scripts

A job script is a shell script with scheduler directives in comment lines at the top (#SBATCH for Slurm, #$ for SGE, #PBS for PBS). The directives request resources (cores, memory, walltime); the body is what runs on the compute node.

#!/bin/bash
# --- scheduler directives go here (see per-scheduler pages) ---

module load python/3.12
python analyze.py input.csv

Request resources honestly: ask for too little and the job is killed; ask for far too much and it queues longer and wastes your fair-share priority.

Environment modules

Most clusters manage software with Environment Modules or Lmod. Loading a module edits your PATH and related variables for the current shell — load the same modules inside your job script.

module avail            # list available software
module spider python    # search (Lmod)
module load gcc/13.2 python/3.12
module list             # what's loaded now
module purge            # unload everything

Job arrays

To run the same script over many inputs (samples, seeds, parameter sets), submit one array job instead of hundreds of individual jobs. The scheduler runs N copies, each with an index in an environment variable (SLURM_ARRAY_TASK_ID, SGE_TASK_ID, PBS_ARRAY_INDEX) that you use to pick the input.

Interactive sessions

For debugging, testing, or exploratory work, request an interactive shell on a compute node (srun --pty bash in Slurm, qrsh in SGE, qsub -I in PBS) rather than running on the login node.

Scheduler cheat-sheet

See the per-scheduler pages for full details: Slurm, SGE, PBS.