Skip to content

MIG Job Submission and SU Charges

When submitting jobs on Wulver's MIG-enabled A100 GPUs, you must explicitly request the desired MIG profile using the --gres directive in your SLURM script.

GPU MIG Slurm Directive
10G MIG --gres=gpu:a100_10g:1
20G MIG --gres=gpu:a100_20g:1
40G MIG --gres=gpu:a100_40g:1

Warning

Please note that MIGs are available in partition=debug_gpu and qos=debug. However, when you use a full GPU using --gres=gpu:a100:1 use the partition gpu and qos standard or low. If you want to see a job script example of requesting a full GPU, please refer to the sample GPU job scripts.

Running Jobs with MIG

#!/bin/bash -l
#SBATCH --job-name=gpu_job
#SBATCH --output=%x.%j.out
#SBATCH --error=%x.%j.err
#SBATCH --partition=debug_gpu
#SBATCH --qos=debug
#SBATCH --account=$PI_ucid         # Replace with PI's UCID
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=8
#SBATCH --gres=gpu:a100_10g:1      # Change to 20g or 40g as needed
#SBATCH --time=59:00
#SBATCH --mem-per-cpu=4000M

srun ./myexe <input/output options>

$srun --partition=debug_gpu \
--account=$PI_ucid \
--qos=debug \
--gres=gpu:a100_10g:1 \
--time=00:59:00 \
--pty bash

Tip

You can submit your job to multiple MIG instances. For example: --gres=gpu:a100_10g:2 will allocate 2 instances of 10G MIG.

Understanding SU Charges

Wulver uses a Service Unit (SU) model to track computing usage. Your job's SU cost is based on:

  • CPU usage
  • Memory request
  • GPU memory allocation (via MIG)

Each component contributes to the SU calculation. The SU cost is charged per node per hour using the formula:

SU = MAX(#CPUs, Memory (in GB) / 4) + 16 × (GPU memory requested / 80GB)

Info

GPU memory requested is based on the MIG profile, not your actual memory usage during the job.

SLURM Directive SU Explaination
4 CPUs + 10MIG MAX(4, 4*4G/4G) + 16 * (10G/80G) = 6 Since no memory requireemnt is specified, SU is charged based on the same number of CPUs and 10G of GPU memory
4 CPUs + 20MIG MAX(4, 4G/4G) + 16 * (20G/80G) = 8 SU is charged based on the same number of CPUs and 20G of GPU memory
4 CPUs + 40MIG MAX(4, 4G/4G) + 16 * (40G/80G) = 12 SU is charged based on the same number of CPUs and 40G of GPU memory
4 CPUs + Full GPU MAX(4, 4G/4G) + 16 * (80G/80G) = 20 SU is charged based on the same number of CPUs and 80G of GPU (A full GPU) memory
4 CPUs + --mem=64G + Full GPU MAX(4, 64G/4G) + 16 * (80G/80G) = 32 The MAX function the evaluates the maximum of 4 SUs (from CPUs), and 64G/4G= 16 SUs (from memory). In addition, 16 SUs are charged from 80G of GPU (A full GPU) memory, bringing the total SU charge to 32 SUs

Tips for Efficient Job Submission (Think Fit, Not Power)

  • Choose the profile that fits your workload, not the biggest one available. You’ll save SUs, get scheduled faster, and help the cluster stay responsive for everyone.
  • Use --mem-per-cpu instead of --mem to balance memory fairly.
  • Avoid requesting a full GPU unless your job cannot run on a MIG.
  • Combine small jobs using job arrays or batching when possible.
  • Need help estimating SUs? Try submitting test jobs with --time=10:00 and reviewing the actual SU usage via the job summary.
  • MIG is designed to make high-performance GPUs accessible and efficient — take advantage of it wisely.