Submitting Batch Jobs

Last updated on 2026-08-27 | Edit this page

Estimated time: 25 minutes

Overview

Questions

  • How do you submit batch jobs through OnDemand?
  • What is the difference between Quick Submit and uploading a script?
  • How do you write a basic SLURM job script?
  • How do you submit GPU jobs and job arrays?

Objectives

  • Access the Job Composer in OnDemand
  • Submit jobs using Quick Submit and script upload
  • Write a basic SLURM batch script
  • Submit GPU jobs and job arrays

Accessing the Job Composer

  1. Click Jobs in the left sidebar
  2. Select Job Composer
  3. You see any previously created jobs
  4. Click New Job to create and submit a new job
Callout

Test Your Scripts First

Before submitting a job to compute nodes, test your script logic in the Web Terminal or with a small test dataset. This saves compute resources and helps you catch errors before they run on expensive hardware.

Quick Submit

  1. Click New Job
  2. Fill in the job details:
    • Job name: A descriptive name (e.g., “Analysis_2026”)
    • Partition: amd, gpu, or short (amd and gpu allow up to 30 days / 720 hours; short has a shorter walltime — check sinfo -p short)
    • Nodes: Usually 1
    • Time limit: Hours and minutes needed
    • Memory: Total RAM (leave blank for default)
  3. Paste your script content in the editor
  4. Click Submit

Upload a Script

  1. Click New Job then Upload
  2. Select a SLURM script file from your computer
  3. Review and modify parameters if needed
  4. Click Submit

Writing a SLURM Script

A basic batch script:

BASH 

#!/bin/bash
#SBATCH --job-name=MyJob
#SBATCH --partition=amd
#SBATCH --nodes=1
#SBATCH --ntasks=4
#SBATCH --time=02:00:00
#SBATCH --output=job.log
#SBATCH --error=job.err

cd $SLURM_SUBMIT_DIR
module load miniconda3
python3 analysis.py

Key directives: --job-name identifies the job, --partition selects the queue, --time sets the wall-clock limit, and --output/--error capture stdout and stderr.

GPU Job Submission

To request GPU resources, use the gpu partition with --gres:

BASH 

#!/bin/bash
#SBATCH --job-name=GPUJob
#SBATCH --partition=gpu
#SBATCH --nodes=1
#SBATCH --gres=gpu:1
#SBATCH --time=04:00:00
#SBATCH --output=gpu_job.log

module load miniconda3 cuda/12.2.1
conda activate pytorch_env   # your conda env with PyTorch (see the ML workshop)
python3 train_model.py

Request a specific GPU type with --gres=gpu:a100:1, --gres=gpu:l40s:1, or --gres=gpu:rtxpro6000:1.

Job Arrays

Submit multiple similar jobs at once:

BASH 

#!/bin/bash
#SBATCH --job-name=analysis
#SBATCH --array=1-10
#SBATCH --partition=amd
#SBATCH --time=01:00:00

python3 analysis.py input_${SLURM_ARRAY_TASK_ID}.txt

Each array task gets a unique $SLURM_ARRAY_TASK_ID (1 through 10 in this example), letting you process multiple input files with one submission.

Challenge

Challenge: Submit and Monitor a Job

  1. Use Job Composer to create a job that:
    • Prints “Job starting at $(date)”
    • Sleeps for 30 seconds
    • Prints “Job completed at $(date)”
  2. Submit the job to the amd partition
  3. Watch the job progress in Active Jobs
  4. When complete, view the output log
  5. Note how long the job took (queue wait + runtime)
  1. Your script runs successfully on the compute node
  2. Active Jobs shows the job progressing: Pending then Running then Completed
  3. The output log shows your date stamps and confirms completion
  4. Total time includes queue wait (seconds to minutes) plus the 30-second sleep
  5. Job ID is a unique number assigned by SLURM

Common observations:

  • Queue time varies based on cluster load
  • Output appears in the log file after job completion
  • The amd and gpu partitions each allow jobs up to 30 days (720 hours); the short partition has a shorter max walltime — check sinfo -p short for the current limit
Key Points
  • Access the Job Composer through Jobs in the left sidebar
  • Quick Submit lets you paste a script; Upload lets you use an existing file
  • SLURM directives control partition, time, nodes, and output locations
  • Use –partition=gpu and –gres=gpu:1 for GPU jobs
  • Job arrays let you submit many similar jobs with one script
  • Sagehen HPC offers three partitions: amd, gpu (both with 30-day max walltime), and short (shorter max walltime, good for quick test/debug jobs — check sinfo -p short)