RMACC 2026

Modern HPC and AI workloads increasingly depend on data that is distributed across multiple storage systems, tiers, and locations, including on-premises clusters, institutional storage, and cloud resources. While compute performance continues to scale rapidly, data access and data movement have become primary bottlenecks, limiting utilization and complicating workflow design.
This talk examines an open, standards-based approach to unifying access to distributed data for AI and HPC workloads—without requiring proprietary clients, forklift upgrades, or disruptive data migrations. Using Hammerspace as a concrete example, the session explores how modern parallel file system standards and automated data orchestration can be used to present a single, high-performance data namespace across otherwise siloed storage systems and sites.
Attendees will learn how global namespace architectures, combined with pNFS 4.2 and policy-driven data orchestration, enable linear scaling of IOPS and throughput using existing infrastructure. The result is simplified workflow design, improved data locality, and higher sustained utilization of expensive CPU and GPU resources—particularly for AI training, inference, and data-intensive simulation workloads.
Key topics include:

• Parallel Global File Systems with pNFS 4.2 – Leveraging open standards to provide scalable, high-performance access to distributed datasets without proprietary file systems.
• Automated Data Orchestration – Using policy-driven data placement and movement to align data dynamically with compute, while maintaining continuous access.
• AI and HPC Workflow Optimization – Simplifying data access across clusters and sites to reduce staging, eliminate redundant copies, and maximize compute efficiency.
   

Purchasing equipment has become very challenging, from prices changing almost daily, to rapidly growing power consumption. We take a quick walk through the current state of the market with input from our peers and vendors. And we get a look at market survey results from two leading companies in this field.

In this session you will participate in an immersive, experiential learning environment designed to expand how you use generative AI in HPC systems. Large language models and coding agents are changing how people write, debug, and maintain code. In high-performance computing environments, however, that efficiency comes with added complexity: shared clusters, schedulers, quotas, modules, file systems, and policies where mistakes can create operational risk.
You will work through a practical HPC-style simulation involving job submission, environment setup, shell scripting, automation, and troubleshooting. The goal is not only to see where generative AI can help with scripting via skills but also to understand where it can make mistakes.
This workshop will emphasize a pair-programming style of collaboration with AI: you will use AI to generate inputs but will need to review and verify. Through controlled adversarial and defensive scenarios, you will build intuition for when to trust AI assistance, when to slow down, and how to check AI-generated shell commands, scripts, and code before running them in shared computing environments. You will be encouraged to use modern skills of context engineering and skill writing to improve your outputs.
This workshop is designed for researchers, educators, students, research computing staff, and HPC administrators. You do not need to identify as an expert programmer, but should bring curiosity and a willingness to write or modify small pieces of code and bash scripts, read output critically, and iterate with AI tools.
Participants should bring a laptop and an open mind toward using generative AI as a fast but fallible collaborator. By the end of the session, you should be able to describe key tradeoffs of using generative AI in HPC-adjacent work, apply simple verification habits to AI-generated commands and scripts, and reuse practical patterns from this simulation in your own work.

New models from leading AI organizations come out seemingly every week, but despite the constant evolution there are always hallucinations and knowledge gaps. LoRA, RAG, RL, and Agentic AI are all popular methods to improve LLM performance and this presentation will give an overview of each method as well as discuss how "performance" can be measured in an LLM context. These methods will be compared and contrasted across several criteria such as ease of use, resources required, and the theory behind them. Open source models from the huggingface repository will be primarily used as examples since they are common in RMACC institutions that may have restricted access environments.

This session examines the role of modern parallel file systems in supporting scalable, data-intensive HPC environments commonly found in academic and regional research computing centers. It discusses how BeeGFS enables high-throughput, low-latency storage architectures that effectively support a wide range of workloads, including traditional simulation, modeling, data analytics, and emerging AI-driven research.
Attendees will gain insight into BeeGFS architecture and design principles, including scalable metadata services, flexible storage tiering, and integration with high-speed networks and NVMe-based storage. The session will also present real-world user case studies from research and HPC environments, highlighting practical deployment considerations, performance characteristics, and operational lessons learned.
By focusing on production deployments and real research workflows, this talk demonstrates how BeeGFS is used today to build reliable, high-performance storage platforms that scale with growing compute and data demands in academic and research-focused HPC environments.

Artificial intelligence (AI) is rapidly transforming research, teaching, and operations. With AI-native platforms like Oracle Database 23ai—and innovations shaping Oracle Database 26ai—data infrastructure is evolving from passive storage to active participation in intelligent workflows. These platforms embed AI directly in the database, enabling more efficient, secure, and scalable development and governance.
At the same time, Oracle’s multi-cloud capabilities allow AI workloads to run across cloud environments while maintaining consistent performance, security, and data governance—supporting flexible, distributed innovation.
A new wave of agentic AI is emerging, with systems that act autonomously, reason, and orchestrate workflows across complex environments. As these agents increasingly access sensitive systems and data, establishing secure, accountable identities becomes critical.
This session explores agentic AI and emphasizes the need for strong identity frameworks grounded in least-privilege access, auditability, and integration with federated identity and role-based access control models to ensure responsible, scalable AI adoption.