Mode Selection and Mode-Dependency Modeling for Power-Aware Embedded Systems

Among the many techniques for system-level power management, it is not currently possible to guarantee timing constraints and have a comprehensive system model at the same time. Specifically, dynamic power management (DPM) techniques can model systems consisting of multiple devices with multiple mode settings. However, if hard real-time constraints are required, then DPM techniques are usually not applicable, because they are based on prediction or timeout and are designed for applications without hard constraints. Instead, low-power scheduling (LPS) techniques may be used to satisfy timing constraints, but they assume a single processor with voltage or frequency scaling. These greedy schedulers are not generalizable to multiple processors or devices with more general modes. We propose a new method for modeling and selecting the power modes for the optimal system-power management of embeddedsystems under timing and power constraints. First, we not only model the modes and the transitions overheadat the component level, but we also capture the application-imposedrelationships among the components by introducing a mode dependency graph at the system level. Second, we propose a mode selection technique, which determines when and how to change mode in these components such that the whole system can meet all power and timing constraints. Our constraint-driven approach is a critical feature for exploring power/performance tradeoffs in power-aware embedded systems. We demonstrate the application of our techniques to a low-power sensor and an autonomous rover example.