As highly-parallel accelerators such as graphics processing units become more important in high-performance computing, so does the need to ensure their reliable operation. In response, research has been directed at several efforts to characterize and understand the hardware vulnerability of GPU microarchitecture structures, as well as to detectng and correcting such vulnerabilities. In this position paper, we advocate a transparent, customizable, and dynamic software approach to reduce the vulnerability of GPU applications. Specifically, we propose to realize these
enhancements via transparent code patching of GPU applications to increase protection of critical program regions. We outline our approach, current infrastructure, some early
results, and long-term plans.

paper
slides