Environment-Driven Computing in Energy-Harvesting Devices

Future, deeply-embedded applications such as civil infrastructure and wildlife monitoring depend on long-term sensor deployments that must scale sustainably and require little maintenance. Energy-harvesting devices meet the sustainability requirements of these applications, using energy from their environment and avoiding the need for battery replacements. A typical energy-harvesting device uses an ultra-low-power microcontroller (CPU) connected to peripherals (sensors and radios) and a power system. The power system collects energy from sources such as radio waves (RF), solar, or thermal gradients, and stores the energy in small batteries or capacitors. Energy-harvesting devices typically operate in an input-sequential manner, due to limited on-device memory; a device processes, and potentially transmits, one input before the next is captured. The end-to-end processing time between two inputs dictates the device’s sensing rate. Sensor devices powered using harvested energy (e.g. solar, RF) provide a promising alternative to traditional battery-powered devices, enabling smaller sizes, and large-scale, long-term deployments with little maintenance. Recent work showed the benefit of on-device compute to identify and transmit data of interest to an application, discarding uninteresting data. On-device compute avoids the energy wasted transmitting uninteresting data and reduces network congestion in large deployments. A flurry of recent work sought to enable and optimize on-device compute in energy-harvesting devices, including schedulers, programming models and tools, architectural interfaces and energy-minimal computer architectures. These systems make great progress toward realizing capable on-device computing for energy-harvesting systems, but leave several key issues unaddressed. In this thesis, we enable reliable and predictable computing models for energy-harvesting devices, which operate in unpredictable environments. The unique challenge addressed by this thesis is the fundamental influence of an unpredictable environment on the runtime operation of an energy-harvesting device. With PHASE, we quantify the relationship between the energy-harvesting device performance with its unpredictable environment. With Quetzal, we analyze and address input buffering issues that arise in energy-harvesting devices due to dynamic variations in the environmental energy and events. Finally, we build a real-world energy-harvesting device prototpye – Camaroptera – which demonstrates the value of on-device computing on energy-harvesting systems, even when they often lack sophisticated compute resources.