Tutorial: SLURM on CCR quickstart

This is a short, practical guide for new lab members running jobs on the University at Buffalo CCR cluster.

Official references:

CCR docs (Running Jobs): https://docs.ccr.buffalo.edu/en/latest/hpc/jobs/
CCR example scripts: https://github.com/ubccr/ccr-examples/tree/main/slurm

ssh <CCRUsername>@vortex.ccr.buffalo.edu

(Use your normal CCR onboarding + DUO flow.)

2) Check your available accounts and limits

slimits

This helps you identify valid Slurm account, partition, and QoS values for your group.

3) Create a simple batch script

Create a file job.bash:

#!/bin/bash -l
#SBATCH --cluster=ub-hpc
#SBATCH --partition=general-compute
#SBATCH --qos=general-compute
#SBATCH --account=<SlurmAccountName>
#SBATCH --time=02:00:00
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=8
#SBATCH --mem=32G
#SBATCH --job-name=lab-test
#SBATCH --output=%x-%j.out
#SBATCH --error=%x-%j.err

# Optional: set threads for OpenMP-style apps
export OMP_NUM_THREADS=$SLURM_CPUS_PER_TASK

echo "Running on $(hostname)"
echo "Job ID: $SLURM_JOB_ID"

# Example command (replace with your real workload)
python analysis.py

Notes:

CCR examples use #!/bin/bash -l for batch scripts.
Keep requests realistic (time, CPU, memory) so jobs start sooner.

4) Submit and monitor

Submit:

sbatch job.bash

Check queue status:

squeue -u $USER

Check finished-job accounting:

sacct -j <JobID> --format=JobID,State,Elapsed,ReqMem,MaxRSS

Cancel a job:

scancel <JobID>

5) Interactive job (when needed)

For short debugging/testing sessions:

salloc \
  --partition=general-compute \
  --qos=general-compute \
  --nodes=1 \
  --ntasks-per-node=1 \
  --cpus-per-task=8 \
  --mem=32G \
  --time=01:00:00 \
  --no-shell

srun --jobid=<JobID> --export=HOME,TERM,SHELL --pty /bin/bash --login

When done:

exit
scancel <JobID>

6) Common issues

Submitting from compute node instead of login node.
Forgetting to set --account to a valid Slurm account.
Requesting too many resources “just in case” (causes long queue waits).
Running big jobs interactively instead of using sbatch.

7) Lab best practice

Keep each analysis in a project folder with:
- job.bash
- software environment file (environment.yml)
- clear output/log naming
Start with a small test job first, then scale up.

1) Login to CCR