2026-03-11 –, Breakout Room
HPC faces a dual challenge of supporting AI/ML engineers alongside domain-specific researchers who may lack Linux expertise. This presentation explores leveraging OOD as a unified gateway for cancer research and AI/ML pipelines. We detail our technical strategy of configuring OOD interactive applications to launch and manage Apptainer containers. This strategy delivers portable, reproducible, and version-controlled software stacks - from GPU-accelerated PyTorch environments to GUI-based biomedical visualization tools, effectively democratizing access to HPC by abstracting the complexity of the container runtime and Slurm. We share how this approach provides a low entry barrier for oncologists/biologists, environmental consistency for AI experts, and reduces support burden.
High Performance Computing (HPC) is no longer the exclusive domain of physicists and engineers comfortable with the command line. Today, HPC centers must support a rapidly expanding "long tail" of users—from oncologists visualizing CryoEM data to AI experts training neural networks. This presentation details how we have architected Open OnDemand (OOD) to serve as the unifying interface for these divergent workloads, using Apptainer (formerly Singularity) as the underlying engine for software portability and reproducibility.
We will begin by outlining the friction inherent in traditional HPC access models. For biomedical researchers, the barrier is often the command line interface and the difficulty of moving massive datasets ("Data Gravity"). For AI/ML engineers, the friction is often referred to as "dependency hell"—managing conflicting versions of Python, PyTorch, and CUDA on a shared cluster.
The core of this talk focuses on our technical solution: configuring Open OnDemand Interactive Apps to act as user-friendly wrappers around Apptainer containers. We will provide a deep dive into the following architectural components and workflows:
The Container Strategy: How we curate a library of Apptainer images to encapsulate complex software stacks, allowing us to update host drivers without breaking user workflows.
OOD App Configuration: A technical look at mapping user inputs (queue selection, GPU requirements) in form.yml directly to Slurm flags and container bind-mounts.
Visualization & Interactive AI: How we leverage OOD’s VNC capabilities to enable server-side rendering for data visualization, and provide persistent JupyterLab environments for model training.
From Bench to Bedside: Perhaps most critically, we will showcase a "Translational Science" workflow. We support researchers who develop custom OOD apps that act as front-ends for trained AI/ML models. These simplified interfaces allow patient-facing doctors—users with absolutely no HPC background—to run inference on patient data and interpret results. This transforms the supercomputer from a backend research engine into an accessible clinical tool.
Finally, we will present metrics on the impact of this architecture. We will discuss how abstracting the complexities of Slurm behind a web interface has reduced the "fear of the terminal" and enabled a new class of users—including clinicians—to leverage high-performance computing without ever writing a line of code.
For nearly 25 years, I have dedicated myself to empowering researchers and advancing science across a multitude of scientific domains via HPC. From system design and tuning to building custom applications, I enable complex simulations and data analysis that unlock new frontiers. My work currently focuses on application gateways using OOD to bridge the gap between infrastructure and cancer research. I specialize in democratizing access for heterogeneous user bases—from AI engineers needing reproducibility to oncologists requiring visualization. This robust architecture enables critical "bench-to-bedside" workflows where patient-facing clinicians leverage HPC, ensuring that technical barriers never impede scientific discovery.