The cluster uses Slurm for job scheduling and resource management. Slurm controls how jobs are queued, scheduled, and executed across compute nodes.

All compute jobs must be submitted through Slurm. Do not run compute-intensive applications directly on the login node.

1. Partitions

A partition (also called a queue) is a logical grouping of compute nodes. Each partition may have different hardware, limits, and runtime policies.

To see available partitions and their status:

sinfo

This command shows:

• Partition name
• Node availability
• Node state (idle, allocated, etc.)
• Maximum runtime limits

2. Submitting a Job

Jobs are submitted using the sbatch command with a job script.

Basic Submission

sbatch your_job_script.sh

Requesting Specific Resources

You can override resource requests directly from the command line:

sbatch --nodes=2 --ntasks-per-node=32 your_job_script.sh

Submitting to a Specific Partition

sbatch --partition=cpucluster your_job_script.sh

3. Example Slurm Job Script

#!/bin/bash

#SBATCH --partition=cputest
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=1
#SBATCH --time=00:05:00

# This is the command to run the program
echo "Hello, World!"

sleep 30

Submit the script:

sbatch my_job_script.sh

4. Difference Between sbatch and srun

• sbatch: Submits a batch job script to be scheduled and executed later.
• srun: Launches parallel tasks within a job allocation (used inside job scripts or interactive sessions).

5. Interactive Jobs

To request an interactive session:

salloc --nodes=1 --ntasks=8 --time=01:00:00

Then run your program with:

srun ./my_program

6. Checking Job Status

View your running and pending jobs:

squeue -u your_username

View completed jobs:

sacct -u your_username

Cancel a job:

scancel jobID

Additional Resources

• Slurm Documentation
• Slurm Command Cheat Sheet