FusionClock: WCEC-Optimal Clock-Tree Reconfigurations (Artifact)

Numerous embedded real-time systems have, besides their worst-case execution time (WCET) requirements, strict worst-case energy consumption (WCEC) constraints that must be satisfied. The core hardware component of modern system-on-chip (SoC) platforms to configure the tradeoff between time and energy is the system’s clock tree, which provides the necessary clock source for each connected device (i.e., memory, sensors, transceivers). Existing energy-aware scheduling approaches have limitations with regard to these modern, feature-rich clock trees: These shortcomings concern the (re-)configuration of the clock tree with the associated penalties, which are a non-negligible part of dynamic frequency/voltage scaling or power-gating devices in addition to the influence of available sleep modes. This artifact evaluation covers the work on FusionClock, an approach that exploits a fine-grained model of the system’s temporal and energetic behavior. By means of our developed clock-tree model, FusionClock processes time-triggered schedules and finally generates optimized code for a system where offline-determined and online-applied reconfigurations lead to the worst-case-optimal energy demand while still meeting given timing-related deadlines. For statically determining these energy-optimal reconfigurations on task level, FusionClock builds a mathematical optimization problem based on the tasks' specifications and the system’s resource-consumption model. Specific components like transceivers of SoCs usually have strict requirements regarding the used clock source (e.g., phase-locked loop, RC network, oscillator). FusionClock accounts for these clock-tree requirements with its ability to exploit application-specific knowledge within an optimization problem. With our resource-consumption model for a modern SoC platform and our open-source prototype of FusionClock, we are able to achieve significant energy savings while still providing guarantees for timeliness, as our evaluations on a real hardware platform (i.e., ESP32-C3) show.