slurm

SLURM Script Setup

Ask each question one at a time. Suggest defaults based on the calculation type, but let the user confirm.

Questions

1 — Partition and account

Both are cluster-specific — you cannot guess them.

What partition and account should the job use?
(If unsure, run `sinfo -s` for partitions and `sacctmgr show assoc user=$USER` for accounts.)

2 — CPUs and memory

How many CPUs and how much memory?
I'd suggest: CPUs: ___    Memory: ___ GB

Defaults: XTB → 4 CPUs, 4 GB. ORCA → 8 CPUs, 16 GB.

3 — Wall time

How much wall time? (HH:MM:SS)
I'd suggest: ___

Defaults: XTB → 02:00:00. ORCA → 24:00:00.

4 — Software environment

Do not guess module names or paths.

How is [XTB / ORCA] set up on your cluster?

  a) Module — what's the module name?
  b) Conda — what's the env name?
  c) Custom path — what's the install directory?
  d) Already in $PATH

For (c), add to the script:

export PATH="/path/to/bin:$PATH"
# For ORCA also:
export LD_LIBRARY_PATH="/path/to/orca:$LD_LIBRARY_PATH"

5 — GPU (skip unless relevant)

Only ask if the user mentioned a GPU partition or the calculation benefits from it.

Then generate

Use the template below. Default to 1 node, 1 task, output %j.out.

Template

#!/bin/bash
#SBATCH --job-name=FILL
#SBATCH --output=%j.out
#SBATCH --partition=FILL
#SBATCH --account=FILL
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=1
#SBATCH --cpus-per-task=FILL
#SBATCH --mem=FILL
#SBATCH --time=FILL

# --- Environment ---
FILL_MODULE_OR_PATH_SETUP

export OMP_NUM_THREADS=$SLURM_CPUS_PER_TASK
export MKL_NUM_THREADS=$SLURM_CPUS_PER_TASK

# --- Run ---
cd $SLURM_SUBMIT_DIR
FILL_RUN_COMMAND

Stacksize

A lot of codes require you to increase the openmp stacksize:

export OMP_STACKSIZE=512M

XTB is a good example of this. Confirm with the user if you are not sure.

Notes

• Use $SLURM_CPUS_PER_TASK for thread counts, never hardcode.
• ORCA: keep --ntasks-per-node in sync with %pal nprocs in the input.
• XTB: reads OMP_NUM_THREADS, no MPI needed.