UVA at Supercomputing 2025 in St. Louis

We’re happy to share some highlights from the jointed participation of Complex Cyber Infrastructure and Multiscale Networked Systems in Supercomputing 2025 in St Louis. The UVA team included Sander Klous, Anestis Dalgitsis, Alexandros Koufakis, Cees de Laat (left to right on picture) and Paola Grosso (not pictured).

We showcased the results of two distinct research efforts. The first project, led by Anestis, focused on LLM-Based Optimization Algorithm Selection for High-Performance Networks Orchestration. The second, led by Alexandros, demonstrated Policy-aware Distributed Vertical Federated Learning infrastructure. Both works used the FABRIC testbed and live demonstrations took place at the CIENA booth. The demonstrations used CIENA’s mobile FABRIC node (pictured above).

The LLM-Based Optimization work was published and presented on the “Innovating the Network for Data-Intensive Science” (INDIS) workshop. Both projects were featured on the RENCI booth as examples of innovative research enabled by the FABRIC infrastructure. Additionally, they were presented on the SciNet theater on the exhibition floor.

Figure 1: SciNet presentation — Anestis Dalgkitsis

Figure 2: RENCI presentation — Alexandros Koufakis

LLM-Based Optimization Algorithm Selection for High-Performance Networks Orchestration

In this work, we propose a Large Language Model (LLM)-based framework for optimization algorithm selection to address the growing complexity of high-performance, multi-domain network orchestration. The proposed solution introduces a context-aware abstraction layer where LLMs analyze network logs, service requests, and algorithm metadata to dynamically select the most suitable optimization strategy. We validate our framework through a prototype deployed on the FABRIC FAB international testbed and through simulations across diverse scenarios (eMBB, URLLC, mMTC, V2X, AR/VR), showing that LLM-driven selection achieves higher success rates and SLA compliance while balancing efficiency, accuracy, and inference latency. Our preliminary results demonstrate the feasibility of this method and highlight its potential to enable scalable, adaptive, and privacy-preserving orchestration.

Paper and demonstration

Policy-aware distributed Vertical Federated Learning infrastructure 

Harnessing the power of data is paramount to unlocking more effective applications, especially with the rise of machine learning. Data regulation, both through legislation and company policies, makes it impractical for different companies to learn from multiple datasets at once. Digital Data Marketplaces aim to solve this issue, by providing a platform where organizations can exchange data based on (dynamic) policies. The digital data marketplace enforces these policies, allowing only data exchange that does not violate policies.

In this work we demonstration how multiple geographically distributed companies can train a unified machine learning model collaboratively using Vertical Federated Learning. Towards this goal we designed and implemented Scattered Directive, a state-of-the-art distributed infrastructure deployment utilizing FABRIC, capable of automatically deploying a digital data marketplace platform. Scattered Directive is built ontop of DYNAMOS, a middleware for datachange systems. Moreover, we demonstrate the effect of dynamic data sharing policies and the impact on the universal model training.

Demonstration & More material

Acknowledgements

We would like to thank CIENA for their support. RENCI and SciNet for hosting our demonstrations. Finally, we would like to acknowledge that part of this work was supported by the project: Budget-neutral sustainable home improvements (MOOI-224049)