Running Jobs on Spartan

Spartan is the University of Melbourne's HPC cluster. Jobs are submitted as scripts describing resources, duration, and the command to run. This note covers the minimal knowledge to write a working SLURM script and understand queue behaviour.

The idea

Spartan uses SLURM (Simple Linux Utility for Resource Management) to schedule jobs. You write a shell script with resource directives (lines starting with #SBATCH) that tell the scheduler how many cores, how much memory, and how long you need. You submit it with sbatch script.sh, and it waits in a queue until enough resources are free.

The three most important directives are #SBATCH --nodes (how many physical nodes), #SBATCH --ntasks-per-node (how many processes on each node), and #SBATCH --cpus-per-task (how many cores per process). Their product is your total parallelism.

Understanding the difference between tasks and CPUs matters for hybrid parallelism (e.g., MPI + OpenMP). A job with 4 nodes, 2 tasks per node, and 8 CPUs per task runs 8 processes across 4 nodes, each with 8 threads, using 64 cores total.

The maths

This section is not about formulas, but about resource counting. If you request:

--nodes=N
--ntasks-per-node=T
--cpus-per-task=C

Then SLURM allocates N × T processes, each with access to C CPU cores. Total cores used: N × T × C. If your job is serial (no MPI, no threads), set T=1 and C=1. If you are using OpenMP with 16 threads, set C=16.

Try it

The widget below builds a SLURM script. Move the sliders to change nodes, tasks, and CPUs, and the total core count updates instantly. The generated script is ready to copy and submit.

Nodes: 1

Tasks per node: 4

CPUs per task: 2

Wall time (min): 60

#!/bin/bash
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=4
#SBATCH --cpus-per-task=2
#SBATCH --time=60:00

module load python
python my_script.py

Total cores

1 node × 4 tasks × 2 CPUs = 8 cores for 60 minutes

SLURM script builder with sliders for cluster topology. Total cores is the product of nodes, tasks, and CPUs.

For a serial job on one core: 1 node, 1 task, 1 CPU.
For OpenMP with 16 threads: 1 node, 1 task, 16 CPUs.
For MPI with 32 processes on 4 nodes: 4 nodes, 8 tasks per node, 1 CPU each.

Where I used it

Easy to get wrong

Sources

Spartan DocumentationUniversity of Melbourne HPCOfficial Spartan documentation with job submission, resource allocation, and queue information.
SLURM sbatch ReferenceSchedMDComplete reference for all SLURM sbatch directives and options.

First drafted in 2023, rewritten in 2026.