As hinted in the title, this is a short intro to slurm command.

What the heck is slurm?

Slurm is workload manager, that help user to schedule resource from a high performance computing full of thousands of GPUs and CPUs.

Start using slurm in 5 steps:

You usually have a quota to how many resource you can use and will be given a authentication token ( or account number ). All computing task need to start using bash scripts, hence if your code is python, you need to add using python run.py

An example config of slurm is as follows:

#!/bin/bash
#SBATCH --ntasks <Your number of node allocate>
#SBATCH -A <Account Number>
#SBATCH -J <Job name>
#SBATCH -p <Node config>
#SBATCH --gres=gpu:<GPU number>
#SBATCH --cpus-per-task=<CPU number>

ntasks: You can view this is as the number of virtual machines

while gres, cpus-per-task controls the resource number of each virtual machine.

1. srun

srun your_batch_script.sh

Allocate the needed resource return all output stream to your terminal interactively.

1. sbatch

sbatch your_batch_script.sh

Allocate the needed resource and returns a job id. You can use scancel and squeue to kill or check the status of your job id.

1. scancel

scancel <your job id>

Kill your current running jobs

1. squeue

squeue --account=<Account id>

List all the currently active jobs.

Some notes

• If you are using GPUs always add nvidia-smi to your bash script to check whether the right amount of GPU resource is allocated.