Slurm scheduler

Slurm manages compute resources and schedules jobs according to partition, resource requests, and cluster state. Submit production work to compute partitions; login nodes are for access, file organization, setup, compilation, and short tests.

Command reference

Command	Purpose
sinfo	Partition and node state
squeue	Pending and running jobs
sbatch	Submit a batch script
srun	Launch a task or job step
salloc	Obtain an interactive allocation
scancel	Cancel jobs
sacct	Job history and accounting
seff	Job efficiency

Use sbatch for unattended production jobs, srun for short interactive tests, and salloc for a resource allocation used by multiple commands or tools such as JupyterLab and VS Code.

Adapt example values

The original examples below use the generic partition name qcpu. Replace it with an authorized partition such as qcpu_18i or qcpu_23i after checking Partitions. Replace 作业ID with a numeric job ID and 用户名 with your username.

Nodes and partitions

sinfo

Node-level information:

sinfo -N

A specific partition:

sinfo --partition=qcpu

State	Meaning
idle	Idle
alloc	Fully allocated
mix	Partially allocated
down	Unavailable
drain	Not accepting new jobs, for example during maintenance

Job status

List visible jobs; add -u "$USER" to restrict the list to your own jobs:

squeue

Select a job, running jobs, or pending jobs:

squeue -j 作业ID

squeue --state=R

squeue --state=PD

R = running, PD = pending, CG = completing, CD = completed, F = failed, and CA = cancelled.

Batch submission

sbatch job.sh

Example script:

#!/bin/bash
#SBATCH --job-name=test_job
#SBATCH --partition=qcpu
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=4
#SBATCH --mem=8G
#SBATCH --time=02:00:00
#SBATCH --output=slurm-%j.out
#SBATCH --error=slurm-%j.err

hostname
date

# Insert your computation commands here
python run.py

--job-name names the job; --partition chooses a queue; --nodes requests nodes; --ntasks requests tasks; --cpus-per-task allocates CPUs to each task; --mem requests memory per node; --time sets the wall-time limit; --output and --error select log paths.

Interactive work

Request an interactive shell (allocation can wait until resources are available):

srun --partition=qcpu --cpus-per-task=4 --mem=8G --time=01:00:00 --pty bash

When finished:

exit

For an allocation used by several commands:

salloc -N 1 --cpus-per-task=4 --mem=8G --time=01:00:00 --partition=qcpu

Wait for allocation, connect to the assigned node, run your work, then leave the node and end or cancel the allocation.

Cancel jobs

scancel 作业ID

The following cancels all jobs belonging to the specified user; use carefully:

scancel -u 用户名

History and resource usage

sacct -j 作业ID

sacct -j 作业ID --format=jobid,jobname,reqcpus,reqmem,averss,maxrss,elapsed,state%20,exitcode --unit=G

seff 作业ID

Example efficiency report:

Check resource requests before submission, monitor pending/running jobs, and inspect logs and accounting afterward. Excessive requests increase waiting; insufficient resources can cause failures. See Resource limits and tmux.