DesignSafe’s HPC Jupyter Lab

DesignSafe’s HPC Jupyter Lab#

Interactive Jupyter on Stampede3 Compute Nodes

DesignSafe now offers a powerful interface for working directly on HPC resources: JupyterLab HPC (CPU), JupyterLab HPC (GPU), and Jupyter HPC Native. These options may grow and/or evolve over time.

Although these environments were originally introduced to support machine learning workflows, they are also ideal for users running OpenSees, Python-based simulations, or other research code that benefits from interactive access to full compute nodes.

What Makes This Different?#

When you launch a JupyterLab HPC (CPU) session, you are not on a login node or a virtual container. Instead:

You’re running a full interactive Jupyter environment on a dedicated Stampede3 compute node.

You may see the message:

“You are running an interactive Jupyter instance on a Stampede3 compute node.”

This means:

  • SLURM has queued and scheduled your session like any HPC job.

  • You’ve been assigned a full physical node with exclusive access.

  • You get all the CPU cores and memory of that node — no sharing with other users.

Key Features and Behavior#

Feature

Description

Real HPC node

You’re running directly on a Stampede3 compute node, not a container or shared login node.

SLURM-scheduled

Your session waits in the SLURM queue, just like a batch job.

Full-node access

You get exclusive use of all CPU cores and memory on the assigned node.

Home directory access

You can use your $HOME to store and reuse software environments (no need to reinstall).

Session limits

Maximum runtime is 48 hours. For faster access, use the skx-dev queue (2-hour max).

All node types

You can launch sessions on any node type available on Stampede3.

Why Use JupyterLab HPC?#

This environment is ideal when you:

  • Want to test and debug scripts interactively before submitting full-scale SLURM jobs.

  • Need to analyze large output files immediately after a run.

  • Are running workflows like OpenSeesPy, data visualization, or parameter studies that benefit from live feedback.

  • Prefer the Jupyter interface but want the performance and consistency of real Stampede3 hardware.

It gives you the best of both worlds: 💻 Interactive development with 🖥️ high-performance compute resources.

Quick Tip: Faster Access for Testing

To reduce wait time while testing:

# Use the skx-dev queue (2-hour interactive session)

Choose skx-dev when launching your session to gain access faster (especially during peak usage periods).

Choosing Between JupyterHub and JupyterLab HPC#

Feature

JupyterHub (Standard)

JupyterLab HPC (CPU/GPU)

Environment

Shared container environment

Dedicated compute node on Stampede3

Startup Time

Immediate (on demand)

Delayed (submitted through SLURM queue)

Resources

Shared container: 8 cores, 20 GB RAM

Full node: up to 56 cores, 192 GB RAM (varies by node type)

SLURM Involvement

None (runs on Kubernetes-managed cluster)

Yes — session is a queued SLURM job

Maximum Runtime

No hard limit

48 hours (enforced by SLURM)

Best For

Lightweight scripting, plotting, Jupyter notebooks

Large computations, OpenSeesPy, post-processing, ML workloads

Storage Access

$HOME, $WORK, $DESIGNSAFE, and persistent volume

Full access to $HOME, $WORK, $SCRATCH on Stampede3

Custom Environments

Requires reinstallation each session (unless in $HOME)

Environments persist in your $HOME — no need to reinstall

Wait Time

None

Possible queue time depending on system load

Use Cases

Prototyping, scripting, small jobs

Full-node testing, heavy computation, live debugging of HPC jobs

Summary#

Use JupyterHub when:

  • You need fast, interactive sessions

  • You’re working on smaller scripts or visualizations

  • You don’t need full-node performance

  • You’re working on larger jobs and are using JupyterHub as a portal to submit a SLURM job to HPC via tapis.

Use JupyterLab HPC when:

  • You want to test or debug in a true HPC environment

  • You need full access to Stampede3 resources

  • You’re running heavier workflows (e.g., OpenSees, ML training, large post-processing)

Tip

Weigh the 48-hour time limit and possible queue wait against your project’s needs to choose the right tool for each phase of your workflow.

Learn More#

For detailed instructions on how to launch and manage HPC Jupyter sessions, visit the DesignSafe HPC Jupyter Guide

Contents