Skip to content

Getting started

Minimum to get a first job running on Sapelo2. See tips & tricks once you're past the basics.

Login node etiquette

You SSH in and land on a login node (prompt: MyID@ss-sub<N>). Shared by everyone. Four things are OK here:

  1. Text editing, writing job scripts
  2. sbatch, squeue, sacct
  3. Light file management (ls, mkdir, small mv/rm)
  4. Git

Everything else goes to compute via interact or sbatch: Python/R/bash loops, bioinformatics tools, conda install, tar/gzip on large files, anything that takes more than a few seconds. GACRC will kill processes that abuse the login node; repeat offenders lose access.

Storage

Path Quota Backed up Purged Use for
/home/$USER 200 GB daily, 14 d No Scripts, envs, configs
/scratch/$USER — No 30 d Active job I/O
/work/<lab> 500 GB, 100k files No No Shared references
/project/<lab> 1 TB Yes No Archive — xfer node only
/lscratch 210–800 GB No Job end Fast node-local

Workflow:

  1. Scripts and envs in /home.
  2. Copy inputs to /scratch; run jobs from /scratch.
  3. Results to /work (shared) or /project (via xfer); delete intermediates.
  4. Archive out of GACRC when the project's done.

Getting compute: interact

One command that gives you a compute-node shell. Request generously for the session — long time, enough memory, a GPU if you need to compile or test something interactively:

interact -c 16 --mem=64G --time=12:00:00                     # CPU session
interact -c 8 --mem=32G --time=8:00:00 --gres=gpu:L4:1       # with GPU for compile/test
interact -c 4 --mem=16G --time=2:00:00 --constraint=Milan    # specific CPU arch

interact echoes the underlying srun command — copy it into sbatch headers when you later batchify the same work.

Prompt changes from MyID@ss-sub<N> → MyID@ra3-22 (or similar). That's your signal you're on compute.

Always exit when you're done. Idle interactive jobs hold resources that other users are waiting for.

Submitting batch jobs with sbatch

For anything taking >1h, needing >16G, running unattended, or part of a pipeline. Minimal template:

#!/bin/bash
#SBATCH --job-name=my_job
#SBATCH --partition=batch
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=8
#SBATCH --mem=32G
#SBATCH --time=04:00:00
#SBATCH --output=%x_%j.out
#SBATCH --error=%x_%j.err
#SBATCH --mail-type=END,FAIL
#SBATCH --mail-user=MyID@uga.edu

cd "$SLURM_SUBMIT_DIR"
module load <versioned module name>

# ... your work ...

Submit: sbatch run.sh. Check: squeue --me (or sq --me). After: seff <jobID> shows actual CPU/memory usage — tune the next submission from real numbers.

Use full versioned module names (STAR/2.7.10a-GCC-11.3.0), not bare STAR — defaults change when GACRC adds newer versions.

Partitions at a glance

Partition Max time For
batch 7 d Default
highmem_p 7 d Up to ~950 G RAM
gpu_p 7 d --gres=gpu:A100:1 / H100 / L4 / P100
inter_p 2 d Interactive only — don't sbatch here

30-day variants exist for each (batch-30d, highmem_30d_p, gpu_30d_p) with a 2-job cap per user.

Getting help

Next: tips & tricks — array jobs, resource tuning, /lscratch staging, shell setup, SSH client choice.