Unified system level reliability evaluation methodology for multiprocessor Systems-on-Chip

摘要

Reliability is a growing fundamental challenge in the design of multiprocessor Systems-on-Chip (MPSoCs). This trend is accelerated by the increasingly adverse process variations and wearout mechanisms that result in an increased number of errors. Previously proposed fault-tolerant techniques are ad-hoc and target processors or Networks-on-Chip (NoC) separately. Because each of these two units may become a reliability bottleneck for NoC based multiprocessor SoCs, it is imperative that the reliability of SoCs be evaluated and addressed in a unified manner, as a combination of communication and computational units. Using this holistic approach, in this paper, we propose a new architecture level unified reliability evaluation methodology for MPSoCs. At the core of the reliability estimation engine lies a Monte Carlo algorithm which works with failure times for time-dependent dielectric breakdown (TDDB) and negative bias temperature instability (NBTI) modeled as Weibull distributions. To demonstrate its usefulness, we utilize the proposed methodology to explore the impact of NoC router layout on the failure time of the system running the same set of benchmarks. In addition, we investigate the failure time of the system when the NoC as the communication unit of the MPSoC is taken or not - as in previous work - into consideration. Our simulation framework can be very helpful to architecture designers, who could use it to identify architectural characteristics and to develop design techniques meant to improve system's lifetime.