The International Conference for High Performance Computing, Networking, Storage, and Analysis
The International Conference for High Performance Computing, Networking, Storage, and Analysis
Author: Keren Zhou (Rice University)
Advisor: John Mellor-Crummey (Rice University)
Abstract: GPUs have emerged as a key component for accelerating applications in various domains, including deep learning, data analytics, and scientific simulations. While GPUs provide superior compute power and higher memory bandwidth than CPUs, writing efficient GPU code to achieve maximum possible performance is challenging because of the sophisticated programming models and architectural features. Performance tools for GPUs are designed to pinpoint performance bottlenecks in GPU-accelerated applications and provide performance insights for users. Existing performance tools, however, are insufficient to identify hotspots and provide insights for complex applications. To address these challenges, we developed a collection of GPU performance tools to measure, analyze, and optimize GPU-accelerated applications. Our GPU profiler employs instruction sampling and instrumentation to collect a wide range of GPU metrics and adopts novel wait-free data structures to coordinate performance monitoring and attribution with low overhead. Our analysis tool constructs sophisticated GPU calling context to help users pinpoint hot GPU code. To understand inefficiencies in hotspots, the analysis tool identifies problematic value patterns on accessed memory addresses. Further, the analysis tool interprets performance metrics and analyzes bottlenecks by attributing measured instruction stalls to their root causes and matching inefficient code with optimization suggestions. To demonstrate the effectiveness of our tools, we studied many deep learning and HPC applications. Guided by the insightful performance reports generated by our tools, we identified performance hotspots, confirmed the issues with application developers, and proposed optimizations. In this proposal, we summarize the work we have done so far and describe plans for future work.
Thesis Canvas: pdf