Question

A research team needs to launch 200 nearly identical training jobs that differ only by a single hyperparameter value. They want each job to receive one GPU, and they want to submit all 200 as a single manageable unit that Slurm can schedule as GPUs free up, without writing 200 separate submission scripts. Which Slurm feature best satisfies this requirement?

Accepted Answer

A Slurm job array submitted with --array, where each task index maps to a different hyperparameter value

Answer

A single sbatch job that requests all 200 GPUs at once with --gres=gpu:200

Answer

Enabling fairshare scheduling so the team's priority increases as jobs complete

Answer

Configuring gang scheduling so all 200 jobs start simultaneously as one group