Quick Start: Running Python in Open OnDemand

This page takes a brand-new user from a fresh account to running Python in a Jupyter notebook on a compute node — all in the browser, with no SSH or Slurm scripts. It is deliberately short and opinionated. Once a step makes sense, follow the linked reference pages for the full detail.

Before you begin

You need two things first:

  1. A CRCD account and Resource Allocation — see Step 1: Getting an Account.
  2. An active PittNet VPN (GlobalProtect) connection if you're off campus — the portal is reachable only from within PittNet.

Your username is your Pitt username in all lowercase. The full portal walkthrough, including logging in and managing files, is on Open OnDemand.

1. Open the Jupyter app

Sign in to https://ondemand.htc.crc.pitt.edu with your Pitt credentials, then launch Jupyter from its tile on the Dashboard, or from Interactive Apps → Jupyter. It runs either the classic Jupyter Notebook or JupyterLab on a compute node.

Not the same as JupyterHub on the Teach cluster

This Jupyter app runs inside Open OnDemand and is distinct from the standalone JupyterHub portal at jupyter.crc.pitt.edu that is used for teaching.

2. Request resources and launch

On the launch form, set the job parameters and click Launch. The fields that matter for a first run:

  • Cluster — the cluster to run on, e.g. htc.
  • Use JupyterLab instead of Jupyter Notebook? — check for the JupyterLab IDE; leave unchecked for the classic Notebook.
  • Python version — the Python module to load, e.g. python/ondemand-jupyter-python3.11.
  • Name of Custom Conda Environment — leave blank to use the base environment (see Step 3).
  • Number of cores — 1–128 (roughly 8 GB of memory per core); leave at 1 unless your code is genuinely parallel.
  • Number of hours — the wall-time limit; the session ends when the time is up.
  • Account — leave blank to use your default Resource Allocation.

Your session is queued while it waits for resources, then starts. When its card in My Interactive Sessions shows Running, click Connect to Jupyter to open the interface in a new tab.

Right-size your request

An interactive session holds a compute node and charges your allocation for the whole wall-time, used or not. Requesting more cores also means a longer wait in the queue — a few cores is plenty unless your code is parallel or memory-hungry. Remember to delete the session when you're done (Step 5).

Working with GPUs?

For AI/ML, launch Jupyter on gpu instead (Interactive Apps → Jupyter on gpu, under the Deep Learning group). Its form bundles the framework, Python, and CUDA together — e.g. pytorch 2.5.1 python 3.11 cuda 12.4 — and lets you request GPU cards. See OnDemand Interactive Apps.

3. Choose your Python

Leave Name of Custom Conda Environment blank on the launch form. The base Anaconda already includes common scientific packages (NumPy, pandas, matplotlib, …). If that covers your needs, there's nothing else to set up.

To use your own packages, create a Conda environment that includes JupyterLab, then point the form at it. From a terminal — Clusters → HTC Shell Access in OnDemand — run:

module load python/ondemand-jupyter-python3.11   # run `module spider python` for the current version
conda create -n myenv python=3.11 jupyterlab <your-packages>

Then, on the Jupyter launch form, enter myenv (or the environment's full path) in Name of Custom Conda Environment before clicking Launch.

Keep large environments off your home directory

/ihome has a 75 GB quota and Conda environments grow quickly. For sizable environments, create them in your group storage with a full path — e.g. conda create -p /ix/<group>/<username>/envs/myenv python=3.11 jupyterlab — and enter that same path in Name of Custom Conda Environment. See File Systems.

4. Run your code

In the Jupyter dashboard (or the JupyterLab launcher), open a new notebook and run a cell to confirm you're on a compute node with the Python you expect:

import sys, socket
print(sys.version)
print("running on", socket.gethostname())

The hostname should be a compute node (e.g. htc-n…), not a login node — your code is running on the resources you requested.

Your session is a Slurm job

The notebook runs on a dedicated compute node with exactly the resources you asked for — a normal Slurm job, not something in your browser. From a terminal inside JupyterLab, squeue -M all -u $USER lists it with its job ID, node, and partition. That's also why it waits in the queue, counts against your allocation, and must be ended with Delete.

5. End your session

Interactive sessions keep consuming your allocation until they stop. When you're finished, return to My Interactive Sessions and click the red Delete button on the Jupyter card. Otherwise the session ends on its own when the wall-time runs out.

You're up and running

You have launched a Jupyter session, chosen a Python environment, and run code on a compute node. Sensible next stops:

Stuck?

Open a help ticket using the link on our service catalog.