RMACC HPC Symposium 2025

As the shape and demands of large scale computing environments have evolved, so have the needs of those who are responsible for keeping them in tip top shape. HPC administrators are challenged with knowing what the data on their system looks like, who’s doing what to the data and tracking jobs on the system. In this talk we’ll cover how the VAST data platform’s powerful structured data component and analytics makes these tasks easy.

Since their discovery in the late 1800’s, electrons have been a constant source of study for scientists. Their properties and behavior have been studied and harnessed to produce some of the greatest inventions of the past century, including electron microscopes and particle accelerators. However one fundamental question about their behavior still remains: how do electrons move inside atoms and molecules?
Electron motion within atoms has proved difficult to study due to the incredibly short timescale it occurs on (the attosecond timescale, or 10-18 seconds). One method of capturing electron motion is to use very short laser pulses to take a series of snapshots of the system. This requires laser pulses shorter than the duration of the dynamics we want to observe (similar to using a short flash on a camera to obtain an image of a fast-moving object). The means to do this have only become possible in the past decade with the advent of new ultrashort (less than 100 as) lasers, which have become feasible due to a process called high‐harmonic generation (HHG).
However, these ultrashort lasers are difficult to produce and characterize experimentally, so theoretical and computational methods are often used in the field of attoscience. These methods are also not without their limitations – modelling the correlated behavior of electrons requires significant computing resources, and so High-Performance Computing (HPC) resources are often used to perform these calculations. In this seminar I will present recent results obtained using R-Matrix with Time-dependence (RMT) method calculations performed on national HPC resources, firstly to treat high-harmonic generation in two-color laser fields, and then on applications of the attosecond pulses generated during the HHG process to measure ionization delays.

AI assisted protein interaction modeling, pioneered by AlphaFold and RosettaFold, has become more diverse both with respect to the programs that do it, and how users run these programs. In this talk, we will cover the programs that are supported at the University of Utah, namely Alphafold2, Alphafold3, Colabfold, Boltz1, RFDiffusion and other tools from the Baker lab, the choices we have made with their deployment, and our experiences with using them. With respect to the ways to run, we will go over the standard SLURM scripts to run Alphafold in two stages (CPU only MSA search, GPU accelerated inference), use Colabfold server for faster MSA search, and using Google Colab running on compute nodes for interactive modeling in a notebook interface. Attendees should leave this talk with ideas how to set up and support these tools and contacts to UofU staff for further questions.

Many research teams and educators experience technical and resource challenges when setting up multi-user systems for data analysis and repeatable research. Instead of managing complex, on-premise systems, or paying for commercial cloud offerings, this presentation will show how to quickly start a simple JupyterHub (using TLJH) on a public research cloud like Jetstream2. This reduces setup effort for teams with limited IT support and improves teamwork and research repeatability in data-intensive projects.
This session is appropriate for researchers, educators, and research software engineers with intermediate skills who want to improve cloud access and teamwork, especially from institutions with limited research computing resources.
TLJH (The Littlest JupyterHub) is a simple, lightweight Jupyter Notebook server for small to medium-sized groups. It helps educators and researchers set up a shared Jupyter environment on a single server with minimal setup (no Kubernetes required!). 
Jetstream2 is a flexible, user-friendly cloud computing environment built on OpenStack. It is available to US-based researchers and educators at no cost through support from the National Science Foundation's Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program. 
Learning Objectives

• Learn Cloud Deployment with Jetstream2: Understand how to use Jetstream2, create instances, and benefit from cloud computing for research teamwork.
  
• Install TLJH Step-by-Step: Follow the setup process for TLJH and adjust a basic JupyterHub to fit research needs
  
• Set Up User Management and Security: Configure login settings, control user access, and adjust network settings to create a secure and easy-to-use research system.
  
• Solve Problems Together: Work in small groups to fix common setup issues, share ways to expand the system, and discuss real-world uses of TLJH in research and teaching.
  

Many modern research software systems run on Kubernetes for scale and resilience (e.g. JupyterHub, Dask, RStudio, etc.). Deploying Kubernetes in a reliable and robust way has historically been difficult. This tutorial offers a simple way to deploy Kubernetes clusters on Jetstream2 using OpenStack Magnum. By making cluster setup and management easier, this session helps teams with limited IT support to run powerful and scalable computing tools.

Participants will learn how to use OpenStack Magnum to create and manage Kubernetes clusters on the Jetstream2 research cloud. Designed for research software engineers and IT support staff with intermediate Linux skills and a basic understanding of containers and container orchestration, this session provides a repeatable process to build a scalable, container-based research system for their institutions.

Jetstream2 is a flexible, user-friendly cloud computing environment built on OpenStack. It is available to US-based researchers and educators at no cost through support from the National Science Foundation's Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program. 

OpenStack is a free cloud computing platform that provides Infrastructure as a Service (IaaS). It helps organizations set up and manage public and private clouds. OpenStack includes tools for computing, networking, storage, and identity management, making it easy to build flexible and scalable cloud systems on different hardware. 

Kubernetes is an open-source container orchestration platform that automates deployment, scaling, and management of containerized applications. Kubernetes helps developers run complex applications reliably and efficiently.

Magnum is an OpenStack service that helps users set up and manage Kubernetes. Magnum offers native integration with OpenStack services, simplified cluster lifecycle management, and enhanced security and resource allocation for containers.
Learning Objectives

• Understand how OpenStack Magnum automates Kubernetes cluster setup.
  
• Understand the advantages of Magnum compared to alternatives.
  
• Follow a step-by-step process to create and configure a Kubernetes cluster using Magnum.
  
• Deploy test containerized applications and adjust cluster scaling.
  

This presentation explores AWS comprehensive solutions for the research community, addressing three critical challenges: meeting evolving compliance requirements like the new NIH NIST 800-171 standards, delivering scalable high-performance computing, and simplifying data management. We'll examine how the AWS Secure Research Environment (SRE) addresses complex security needs, compare managed AWS ParallelCluster Service with self-managed options, and showcase tools like the Globus S3 Connector for streamlined data handling. Join us to discover how AWS empowers researchers to focus on innovation while maintaining security, performance, and efficiency at scale.

Managing data at scale in high-performance computing (HPC) environments requires efficient storage and retrieval strategies. Automated tiered storage solutions enable seamless migration of aged data to lower-cost archival tiers while maintaining accessibility. Enriched metadata—spanning tagging, search, discovery, and data provenance—enhances data usability and long-term value. This approach not only optimizes storage costs but also empowers researchers with better data discovery and reuse. Real-world HPC use cases demonstrate how metadata-driven workflows streamline research, ensuring that critical datasets remain accessible and actionable over time.

The next frontier in AI advancement isn’t just about algorithms—it’s about unlocking the wealth of hidden insights trapped within millions of files in HPC environments. While organizations focus on model architectures, the true bottleneck often lies in discovering and preparing relevant data buried in vast storage systems.
This presentation, featuring MetadataHub and a live demonstration, will reveal how intelligent metadata extraction and management transforms unstructured data into AI-ready assets by:

• Uncovering Hidden Context: Live metadata extraction demonstrating how MetadataHubcaptures content and contextual value, revealing unexpected connections between research datasets and enabling new AI training opportunities that would otherwise remain hidden.
• Automating Data Discovery: Demonstrating how MetadataHub automates metadata tagging to identify valuable training data across petabyte-scale storage, reducing data preparation time by up to 90%.
• Enhancing Model Quality: Exploring how rich metadata captured by MetadataHub improves AI model performance by providing better context and enabling more relevant training data selection.
• Scaling Efficiently: Showcasing metadata-driven automation with MetadataHub that optimizes data pipeline efficiency and resource utilization, including GPU/CPU performance, across HPC environments.

The session will highlight a real-world success story from the Zuse Institute Berlin, where MetadataHub unlocked 200 PB of previously underutilized research data for cutting-edge Generative AI applications. A 15-minute live demonstration will guide attendees through their journey—from data discovery to AI-ready datasets—highlighting practical challenges and solutions.
Attendees will leave with actionable strategies for implementing metadata-driven approaches in their own HPC workflows. By showcasing MetadataHub’s ability to extract content and contextual value, this session will demonstrate how metadata transforms unstructured data into a strategic advantage, accelerating AI initiatives and driving HPC innovation.

Meetup with members of the RMACC SysAdmin group for an informal discussion. Bring your breakfast and topics you would like to discuss. 

Please join Mark III and NVIDIA at RMACC for an overview, walk-through, and live demo of building an AI agent app quickly and easily using NVIDIA NIMs and Blueprints. The first segment of this session will focus on a build of a simple AI agent bot built with a Llama 3 NIM. The second segment will focus on a live build and demo of a more advanced AI agent build with an NVIDIA Blueprint (powered by multiple NIMs). Lastly, the session will show how to quickly and easily finetune a model before being served up for API consumption by apps via NIMs. This session will be of interest to participants of all levels and skillsets looking to deploy AI services into their existing and new cloud-native and modern apps and research as quickly and effectively as possible.

High-performance computing (HPC) research thrives on collaboration, yet institutional data silos often hinder progress. This extends beyond the storage hardware itself to where data is being siloed in departments and institutions alike. Accelerating scientific progress and innovation requires enabling secure data discovery and sharing across not only institutions, but scientific disciplines as well. Federated access models, controlled permissions, and distributed compute environments enable seamless yet secure collaboration. By facilitating data discovery across disciplines and optimizing shared infrastructure, organizations can break down barriers, enhance research efficiency, and drive cross-disciplinary insights that push the boundaries of scientific advancement.

Modern scientific computing demands flexible and scalable solutions that bring computing power closer to data while maintaining security and ease of use. We propose to present our solution which leverages Kubernetes to provide a platform to our employees, university members, and partner organizations that meets those demands and complements our existing HPC system. This presentation will cover how we utilize Continuous Integration and Continuous Delivery (CI/CD), coupled with GitOps and DevOps practices, to provide a robust and secure platform for hosting container-based workloads. These workloads include interactive web visualizations, JupyterHub instances, science gateways, data assimilation tools, data analysis tools, and those that require access to GPUs, including, but not limited to, AI/ML. The presentation will also cover how Cilium network and Kyverno access policies are implemented to secure the platform. It will also discuss how GitHub Actions are utilized to test and build codebases into containers that can run on the platform. Attendees will learn why we chose Kubernetes as our platform as well as practical strategies to implement similar solutions and common pitfalls to avoid.