RUA. Repositorio Institucional de la Universidad de Alicante
Publicación
Otros (Other)
RISC-V Proton Reliability Assessment Using Redundancy and Optimization Techniques on Bare Metal and Linux
This article presents a comparative evaluation of traditional software-based fault tolerance methods, triple modular redundancy (TMR), and duplication with re-execution (DWR), on a commercial RISC-V system-on-chip (SoC) under proton irradiation. Their effectiveness is assessed across bare-metal and lightweight Linux environments using multiple versions of a matrix multiplication benchmark to explore the interaction among cache optimization, software redundancy, and operating system complexity. Proton beam experiments conducted on the Allwinner D1-H SoC reveal three key findings. First, cache-aware programming achieves silent data corruption (SDC) reductions of 1.25× – 1.39× without redundancy, through a combination of execution acceleration and cache data replacement. Second, when combined with software redundancy, cache-aware implementations reduce SDC cross-sections by exceeding two orders of magnitude, while also yielding the best-performing configurations in terms of useful work accomplished between failures mean work to failure (MWTF). Third, lightweight Linux exhibits configuration-dependent reliability with respect to critical single event functional interrupts (SEFIs), with the maximum degradation remaining 3× – 4× lower than the penalties reported for full-featured Linux distributions. Overall, the results establish that both cache-aware programming and lightweight operating system configurations provide measurable reliability benefits that persist even when software redundancy techniques are deployed.