Tips & tricks¶

Stuff you pick up after the first few weeks. Prerequisites: basics.

What's changed since the 2021 version of this guide¶

qlogin → interact (old name still aliases).
squeue --me / sq --me wrapper for your own jobs.
New partitions: hugemem_p (3 TB RAM), 30-day variants.
New hardware: AMD EPYC Genoa (128c/764G), Milan nodes in batch, H100 + L4 GPUs.
Home quota 200 G (was 100 G); home snapshots 14 days.
Open OnDemand: ondemand.gacrc.uga.edu — Jupyter, RStudio, desktops, files in-browser.
The 2021 post said scratch purged at 90 days. Wrong; it's 30 days, always was.

Array jobs for batches¶

Same pipeline across many samples — use --array instead of submitting N scripts:

#SBATCH --array=1-20
#SBATCH --output=array_%A_%a.out

SAMPLE=$(sed -n "${SLURM_ARRAY_TASK_ID}p" samples.txt)
# ... pipeline on $SAMPLE ...

Chained jobs¶

JOB1=$(sbatch align.sh | awk '{print $4}')
sbatch --dependency=afterok:$JOB1 count.sh

Resource tuning with `seff`¶

After a test run:

Memory Utilized: 18.2 GB
Memory Efficiency: 56.9%    ← asked for 32G, used 18G → request ~22G next time

Over-requesting memory is the most common waste — the scheduler has to find a node with that much free, so your job waits longer. Tune from real numbers.

Use $SLURM_CPUS_PER_TASK in your commands so thread count stays in sync with the header:

aligner --threads "$SLURM_CPUS_PER_TASK" ...

`/lscratch` staging¶

For I/O-bound single-node jobs, staging inputs onto the node's local disk can halve wall-time:

#SBATCH --gres=lscratch:200       # reserve 200 GB

STAGE=/lscratch/$SLURM_JOB_ID
OUT=/scratch/$USER/results

cp /work/<lab>/genome/* "$STAGE/"
cd "$STAGE"
aligner --ref "$STAGE/genome.fa" -o "$STAGE/out.bam"
cp "$STAGE/out.bam" "$OUT/"       # copy back before the job exits

/lscratch disappears at job end — always copy results out.

`scavenge_p` is free speedup¶

Short (<4 h) jobs in batch may auto-route to idle buy-in nodes via scavenge_p — reduces wait time when batch is busy. Don't pad time limits to game this; GACRC notices the pattern.

Recovering deleted `/home` files¶

Daily snapshots live at /home/.zfs/snapshot/, accessible from xfer or compute nodes:

ls /home/.zfs/snapshot/
cd /home/.zfs/snapshot/zrepl_YYYYMMDD_HHMMSS_000/MyID/
cp accidentally-deleted-file /home/MyID/

Daily go back 14 days; older → GACRC ticket.

Big transfers — use Globus¶

Anything over a few GB: https://app.globus.org. Resumable, fast, runs in the background. Works from the xfer node. Don't use the remote.uga.edu VPN for transfers — it has a bandwidth cap.

Cluster load¶

sinfo-gacrc                # per-node CPU/mem/GPU state
sinfo-gacrc 40 50          # wider columns

NodeStat for a cleaner summary of idle resources by partition.

Shell setup¶

sq --me (alias for squeue --me) — your jobs
sinfo-gacrc, sacct-gacrc, seff — job inspection
tmux for window splits; start it on the login node before interact if you want to cover SSH drops. Not a fault-tolerance layer for your Claude session — if interact walltime expires, tmux doesn't help.

Useful aliases:

alias myq='sq --me'
alias scratch='cd /scratch/$USER'
alias interact8='interact -c 8 --mem=16G --time=8:00:00'

SSH clients — prefer a plain terminal¶

Skip VS Code Remote SSH if you can. Two reasons:

Memory: VS Code's remote server runs persistent language servers (pylance, node) on the login node. Invisible to you, very visible to GACRC. Frequent cause of login-node memory warnings.
Control: running Claude inside VS Code's agent mode hides tool calls behind UI chrome. The CLI shows every bash invocation, every file read, every decision point. More signal, less guessing.

A plain terminal (Terminal.app, iTerm2, Alacritty, WezTerm, Ghostty) is lighter and gives you full visibility. Use tmux for splits if you miss the multi-pane layout.

Termius is fine if you need mobile SSH or cross-device snippet sync. Open OnDemand covers Jupyter/RStudio in-browser without any SSH client.