Compute

Our usual compute options and the commands needed to operate them. Run local commands from the repository root; remote commands after connecting to the host.

Local Mac — development and analysis

Use for editing, small tests, plotting and analysis. Not for CUDA training.

cd /Users/eoin/pinglab
uv sync --dev
uv run python experiments/exp022/compute.py --help

Hetzner — persistent control host

pinglab-codex is our always-on host for persistent sessions, orchestration and monitoring, not heavy numerical work.

ssh hetzner
tmux list-sessions
tmux attach -t <session-name>

# Olorin is also accessible from Hetzner.
ssh olorin

Olorin — shared, large-memory GPUs

Shared machine with four RTX PRO 6000 Blackwell GPUs, approximately 96 GB VRAM each. No Slurm: check availability and select an idle GPU explicitly. Keep heavy files under /scratch/em586.

ssh olorin
nvidia-smi
cd /scratch/em586/pinglab

# Replace N with an idle GPU index.
CUDA_VISIBLE_DEVICES=N uv run python experiments/exp022/compute.py

CSD3 / Wilkes3 — queued A100 jobs

We often refer to CSD3 / Wilkes3 as “hpc”. This is shorthand for this compute option, not a separate provider or SSH alias.

Connect with ssh csd3; authentication requires the CSD3 login token. Submit prepared Slurm scripts using the appropriate service level:

ssh csd3
mybalance

# Submit using one of these:
sbatch --account=OLEARY-SL2-GPU --partition=ampere --qos=gpu1 --gres=gpu:1 job.slurm
sbatch --account=OLEARY-SL3-GPU --partition=ampere --qos=gpu2 --gres=gpu:1 job.slurm

# Monitor, inspect and cancel.
squeue -u "$USER"
scontrol show job <job-id>
scancel <job-id>

Run substantial work inside Slurm allocations, never on login nodes. Batch status checks and leave at least 120 seconds between recurring Slurm requests; use job arrays for large batches.

Modal — managed cloud execution

Use the repository-wide Modal dispatcher for experiment compute stages. It discovers experiments/expXXX/compute.py and supplies a reserved Pingstore run ID; there is no per-experiment allowlist. The experiment must implement the standard --run-id contract and have compatible Linux dependencies and explicit source inputs. Choose resources for its recipe.

uvx --from 'modal>=1.1,<2' modal setup
uv run --no-project --python 3.12 --with 'modal>=1.1,<2' --with PyYAML \
  python -m tools.modal_dispatcher dispatch --experiment exp141 --gpu H100
uv run --no-project --python 3.12 --with 'modal>=1.1,<2' --with PyYAML \
  python -m tools.modal_dispatcher check --run-id COMPUTE_ID --collect

dispatch starts paid work using detached asynchronous invocation. Arguments after -- go to the experiment’s own compute CLI. Declare source runs with --input ROLE=RUN_ID; inputs and their ancestry are validated and transferred. Checks and collection do not start workers. Save the printed run/call IDs and investigate ambiguous submissions before dispatching again. Analysis and presentation remain separate stages.

Basics

• Inspect the experiment and resource request before launching; confirm approval for paid work.
• Keep job IDs and logs. Collect and validate completed outputs before cleaning up or stopping resources.
• Experiment stages and scientific outputs follow Experiment Rules; storage and recovery follow Pingstore Rules.
• Keep credentials out of the repository.