Collaborative Research: CNS Core: SMALL: DrGPU: Optimizing GPU Programs via Novel Profiling Techniques

Graphics Processing Units (GPUs) have become common in modern computing systems. However, it remains challenging to efficiently program GPUs due to the complexity of architectures, programming models, and algorithm designs. This project aims to develop a framework to identify performance inefficiencies in GPU applications and provide intuitive guidance for code optimization. The intellectual merits include three novel analysis techniques: (a) multi-scale analysis to understand inefficiencies inside individual GPU kernels, (b) coordinated analysis to measure inefficient data movement between devices, and (c) differential analysis to view inefficiencies across different runs with different execution configurations. The combination of these three analyses will give a complete performance picture of GPU applications and help developers obtain the bare-metal performance in modern and emerging GPUs.

This project will bridge the knowledge gap between GPU hardware architectures and software developers. The success of this project will enable significant enhancement of computing efficiency for production code, and hence help maintain a sustained advancement of various domains, e.g., data analytics, high-performance computing, and artificial intelligence. The proposed framework will draw broad interest from industry, research institutes, and the Department of Energy national laboratories for improving their code bases and increasing system throughputs. Moreover, the outcome of this project will be integrated into curriculum planning and educational activities, from which students can directly benefit.

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.