Neverlast: an NVM-centric operating system for persistent edge systems

Abstract

The development of novel non-volatile memories (NVM) with low latencies presents a turning point in the design of computational systems, especially the design of efficient intermittently-powered edge computing systems. Persistent edge computing systems operate under the imminent risk of high-frequent power-supply failures and therefore require adjustments to all parts of the system: The operating system must persist the whole system state prior to power failures to guarantee consistent progress. In this paper, we present Neverlast, a comprehensive approach focusing on the persistence of computational progress (i.e., register and memory contents) and the persistence of hardware configuration (e.g., configuration of external sensors). To ensure persistence, Neverlast implements an interrupt-driven power-failure manager that preserves the system s state on a fine-grained instruction level with minimal overhead. For peripheral devices, our configurable device and energy manager tracks and replays operations to ensure that devices restart in the correct state. Following this approach, persistence is provided as a service to the application, abstracting from power management and state-preserving techniques.