Reliability Analysis of Microarchitectural Faults in GPGPU-based HPC Systems
Corrado De Sio, Luca Sterpone, Sarah Azimi · 2023
As GPGPUs gain popularity in HPC applications, there is a growing need to investigate their reliability for performance improvement and reduced computation overhead. In this paper, the authors propose a novel fault injection environment for NVIDIA GPGPU devices that can automatically inject faults into instructions at the SASS level by instrumenting the CUDA binary executable file. It can categorize faults into Silent Data Corruption, Detected Unrecoverable Error, and Hang, making it an effective tool for targeting the reliability evaluation of specific threads.