CAREER: SPEED: Scalable Performance-aware Engine for Efficient Decision-making on High-Performance Computing Systems

As High-Performance Computing (HPC) systems grow increasingly complex to support high levels of parallelism, running applications without understanding how they interact with the underlying system can lead to inefficiencies, slower outcomes, wasted energy, and missed opportunities for discovery. These challenges are further compounded by the dynamic nature of these applications requiring configuration adjustments, such as resource allocations, during runtime.

Making these adjustments accurately and in real time renders offline data collection and human-in-the-loop approaches impractical. This project addresses these challenges by developing innovative methodologies based on generative Artificial Intelligence (AI) and introducing SPEED, a scalable and efficient modeling framework. SPEED captures multifaceted relationships between configuration settings and application performance from diverse data sources, enabling automated, real-time decision-making.

The outcomes of this project can be applied to high-impact, dynamic domains such as disaster response, healthcare, engineering, and manufacturing, where accelerating data-driven decision-making plays a critical role in saving lives and reducing economic losses. Additionally, this project will equip students with advanced skills in HPC and AI, creating a pathway to the national scientific workforce.

Unlike traditional methods that rely on aligned datasets or require significant computational resources, this project takes a modular approach to process large volumes of heterogeneous, multi-modal, and unaligned data. SPEED breaks these large datasets into smaller domains, learns relationships and patterns within each domain independently and in parallel, and integrates these smaller, data-centric models into a unified one by identifying cross correlations.

This project also introduces a novel approach for updating models without rebuilding them when new heterogeneous measurements are added, making SPEED adapt quickly to new scenarios. Finally, SPEED leverages these learned representations to provide predictive and generative modeling services to HPC users, system software, and facilities. By separating the process of learning representations from how they are used, this transformative approach enables seamless integration of new data sources, modalities, and pre-trained models, ensuring adaptability and scalability for future needs.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.