A runtime manager for gracefully degrading SoCs

Paper in proceeding, 2014

The increasing number of transistors integrated on a single chip comes with the blessing of raw computational power and the curse of susceptibility to various kinds of faults. On top of increased defect densities, wearout effects mean that the testing verdict at fabrication time cannot be trusted throughout the chip lifetime. However, extra computational power presents the opportunity to build gracefully degrading MPSoCs. Re-configurable components and flexible workloads, along with runtime support, enable MPSoCs to deal with permanent faults degrading one or more system aspects, such as performance, energy efficiency and delivered functionality, instead of failing. In this manner, chip life is prolonged and safety is increased. In this work Graceful Degradation (GD) is formulated as an optimization problem in the context of MPSoCs. As such, its possible solutions can be evaluated in a parameterizable and consistent manner. An attempt at a runtime solution for a heterogeneous 4-core SoC is made and the resulting GD manager is evaluated in terms of speed and accuracy, with a use case combining essential automotive tasks and non-essential additional features. On average, it is found to produce a solution 89% as good as the optimal, in 4.3μsec running on one core of a common modern CPU.