Real-Time Scheduling for GPUs with Applications in Advanced Automotive Systems

Glenn A. Elliott
University of North Carolina at Chapel Hill
University of North Carolina at Chapel Hill, 2015



   author={Elliott, Glenn A},


   school={University of North Carolina at Chapel Hill}


Download Download (PDF)   View View   Source Source   



Self-driving cars, once constrained to closed test tracks, are beginning to drive alongside human drivers on public roads. Loss of life or property may result if the computing systems of automated vehicles fail to respond to events at the right moment. We call such systems that must satisfy precise timing constraints "real-time systems." Since the 1960s, researchers have developed algorithms and analytical techniques used in the development of real-time systems; however, this body of knowledge primarily applies to traditional CPU-based platforms. Unfortunately, traditional platforms cannot meet the computational requirements of self-driving cars without exceeding the power and cost constraints of commercially viable vehicles. We argue that modern graphics processing units, or GPUs, represent a feasible alternative, but new algorithms and analytical techniques must be developed in order to integrate these uniquely constrained processors into a real-time system. The goal of the research presented in this dissertation is to discover and remedy the issues that prevent the use of GPUs in real-time systems. To overcome these issues, we design and implement a real-time multi-GPU scheduler, called GPUSync. GPUSync tightly controls access to a GPU’s computational and DMA processors, enabling simultaneous use despite potential limitations in GPU hardware. GPUSync enables tasks to migrate among GPUs, allowing new classes of real-time multi-GPU computing platforms. GPUSync employs heuristics to guide scheduling decisions to improve system efficiency without risking violations in real-time constraints. GPUSync may be paired with a wide variety of common real-time CPU schedulers. GPUSync supports closed-source GPU runtimes and drivers without loss in functionality. We evaluate GPUSync with both analytical and runtime experiments. In our analytical experiments, we model and evaluate over fifty configurations of GPUSync. We determine which configurations support the greatest computational capacity while maintaining real-time constraints. In our runtime experiments, we execute computer vision programs similar to those found in automated vehicles, with and without GPUSync. Our results demonstrate that GPUSync greatly reduces jitter in video processing. Research into real-time systems with GPUs is a new area of study. Although there is prior work on such systems, no other GPU scheduling framework is as comprehensive and flexible as GPUSync.
VN:F [1.9.22_1171]
Rating: 0.0/5 (0 votes cast)

* * *

* * *

HGPU group © 2010-2017 hgpu.org

All rights belong to the respective authors

Contact us: