Collaborative Research: Elements: DLToolkit: A Novel Performance Profiling and Analysis Infrastructure for Scientific Deep Learning Workloads

Deep Learning (DL) has improved scientific applications across various scientific domains, including high-energy physics, meteorology, agriculture, and material science. This project introduces DLToolkit, a performance profiling infrastructure tailored for domain scientists to analyze and optimize science-driven DL applications. This project also contributes to education and supports broader usage; the outcomes of this project will be integrated into the Computer Science (CS) curriculum, and both George Mason University and the University of California - Merced are minority-serving institutions, offering opportunities for delivering knowledge about cutting-edge techniques to underrepresented students. Together with industry and national laboratory partners, the project will also provide research training, symposia, and internship opportunities for students, aiming to foster a cohort of performance engineers. The overarching objective of this project is to improve scientific DL applications. The intellectual merits include three novel profiling capabilities: (a) synergistic tool-framework profiling to streamline extensive domain-specific knowledge from existing DL frameworks to DLToolkit, significantly lowering the barrier for domain scientists to use DLToolkit; (b) just-in-time (JIT)-aware profiling to ensure precise yet lightweight attribution of performance events to complex JIT-compiled DL operators; and (c) tensor-centric profiling to provide a holistic view of tensor operations’ impact on model performance. By uniting these capabilities within DLToolkit, this project will create a cohesive infrastructure for domain-specific performance profiling to empower scientists with critical insights to optimize their DL applications, accelerating scientific research and innovation. 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.