Denial of service exploits on CUDA devices using clock function

NVIDIA’s CUDA devices are increasingly being used in applications from accelerating computer graphics to deep learning and numerical analysis. Because the purpose of GPUs is to accelerate a diverse set of applications, it is important for them to execute code as fast as possible with the highest degree of throughput. Unlike CPUs which have been targets of security exploits since their inception (many decades ago) GPUs have only been available for a diverse set of applications and widely adopted in the last decade. Due to GPUs relative novelty and increasing adoption by industry and government alike, there has been an increased interest in security vulnerabilities and their corresponding solutions. In this thesis we uncover a vulnerability with the library clock function that can cause device-wide slowdown on Unix-based systems, impact performance and, in some cases, cause total system failure. Different generations of CUDA devices running on different operating systems were benchmarked. The disturbances in the benchmark output demonstrated consistent vulnerabilities in Unix systems and some occasional problems with Windows systems. Two potential methods to detect and resolve malicious programs have been prototyped on CPUs in this study and are currently in development to be ported to GPUs.U of I Onlyundergraduate senior thesis not recommended for open acces