Skip to content

hpc.md

Latest commit

 

History

History
98 lines (74 loc) · 2.82 KB

hpc.md

File metadata and controls

98 lines (74 loc) · 2.82 KB
Raw

HPC Notebook Workflows

This guide covers running the CDT notebook workflow on clusters, including Slurm batch jobs and Open OnDemand Jupyter sessions. For a local first run, start from the repository root with:

just notebook-setup
just notebook

The justfile owns those recipes. Inspect it if you want to see exactly which uv, Jupyter, and notebook commands are being run.

Headless Batch Runs

For clusters, set up the notebook environment on a login node or in a job prologue, then run notebooks headlessly on a compute node:

just notebook-setup
just notebook-execute

If the cluster restricts home-directory caches, put uv's cache on project scratch:

export UV_CACHE_DIR="$SCRATCH/uv-cache"
just notebook-setup

A minimal Slurm script looks like:

#!/usr/bin/env bash
#SBATCH --job-name=cdt-notebook
#SBATCH --time=00:20:00
#SBATCH --cpus-per-task=2
#SBATCH --mem=4G

set -euo pipefail
cd /path/to/causal-triangulations
export UV_CACHE_DIR="${SCRATCH:-$PWD/target}/uv-cache"
just notebook-setup
just notebook-execute

For fully offline compute nodes, run just notebook-setup where package downloads are allowed first, then submit jobs with the populated .venv, uv.lock, and uv cache available on the cluster filesystem.

Open OnDemand And Slurm

On Open OnDemand, start a Jupyter session backed by a compute allocation, open a terminal in the session, and launch from the repository root:

just notebook

If the cluster provides a packaged cdt binary through a module, point the notebook at it before launching:

module load causal-triangulations
export CDT_BINARY="$(command -v cdt)"
just notebook

For longer runs, use the notebook as the control surface and submit the cdt command to Slurm with sbatch or srun, writing trace.csv and summary.json to a run directory on scratch. When the job finishes, load those artifacts back into the notebook for Polars analysis and plotting. This keeps expensive simulation work in Slurm while preserving the notebook as the interactive experiment record.

A minimal handoff script from an OnDemand terminal or notebook-created run directory looks like:

RUN_DIR="${SCRATCH:-$PWD/target}/cdt-runs/slurm-demo"
mkdir -p "$RUN_DIR"

cat > "$RUN_DIR/run-cdt.sbatch" <<EOF
#!/usr/bin/env bash
#SBATCH --job-name=cdt-run
#SBATCH --time=00:30:00
#SBATCH --cpus-per-task=2
#SBATCH --mem=4G

set -euo pipefail
cd "$PWD"
cdt \
  --vertices-per-slice 8 \
  --timeslices 16 \
  --topology toroidal \
  --steps 5000 \
  --measurement-frequency 100 \
  --seed 105 \
  --simulate \
  --output-csv "$RUN_DIR/trace.csv" \
  --output-json "$RUN_DIR/summary.json"
EOF

sbatch "$RUN_DIR/run-cdt.sbatch"

After the job completes, open RUN_DIR/trace.csv and RUN_DIR/summary.json from the notebook instead of rerunning the simulation interactively.