Research that introduces a novel method of building large language models (LLMs) for supercomputers using artificial intelligence (AI) earned a best paper award from the Association of Computing Machinery’s Practice and Experience in Advanced Research Computing (PEARC) Conference.
A team of UCF researchers led by School of Modeling, Simulation and Training Assistant Professor Sean Mondesire earned the honor in the systems and system software track for their work, “Automating HPC Software Compilation, Deployment and Error Resolution through an LLM-based Multi-Agent System.”
Authored by Mondesire, Emmanuel Nsiye, Bulent Soykan and Glenn Martin, the work has the potential to significantly drive change in the field by creating much-needed efficiencies in high performance computing (HPC). The team has demonstrated a 97% success rate for autonomously developing more than 200 commonly used HPC software packages.
“I was ecstatic when we received the award,” Mondesire says. “My team worked on this specific problem for over a year, and it meant a lot to us because we wanted to have a positive impact on the research computing community.”
He explains that researchers frequently use supercomputers to perform large, complex computations that can have thousands of software modules, each of which must be manually built, installed and configured. This lengthy, meticulous process motivated Mondesire to devise a more efficient way for UCF’s Advanced Research Computing Center team to build critical research software using AI.
“Our technique automates this software preparation process by leveraging recent advancements in LLMs, and it has proven to be successful,” Mondesire says.
The new development has the potential to make a major impact on research methods across the globe.
“There are HPCs and data centers around the world that rely on software modules to be installed, updated and deployed,” Mondesire say. “The system we created can be adopted to assist these computing resources in automating some of the tedious and time-consuming tasks, freeing technicians and admins to work on other critical IT tasks.”
Their work is reminiscent of the innovation that Mondesire wanted to be a part of when he chose UCF to pursue his doctorate degree in computer science. He joined the university as a faculty member in 2021.
“I wanted to be a part of its faculty to educate a student body that is known to be motivated and eager to learn, and that seeks to inspire and contribute meaningful and impactful research,” he says.
As the director of the High-Performance AI Laboratory (HAIL), Mondesire and his team focus their efforts on real time adaptive AI and intelligent digital twins. He explains that the “intelligent” aspect of their work uses AI to analyze, predict and prescribe actions that enable physical systems to make informed decisions.
“Our research is focused on real-time modeling because as computing, networking and sensing capabilities are advancing, so are the demands for near-instant automated decision-support, decision-making and adaptability from real-world systems,” he says.
HAIL researchers include undergraduate students, graduate students and full-time scientists with degrees and backgrounds in computer science, data analytics, statistics, and modeling and simulation. Just as Mondesire envisioned when he made the decision to come to UCF, innovation motivates and drives his group to develop transformative solutions.
“The lab is currently working on developing intelligent digital twins for the manufacturing industry,” Mondesire says. “We are creating data models, simulations and infrastructures to be able to support automated real-time sensing, analysis and decision-making for manufacturing planning and optimization.”
- Written by Bel Huston