Efficient state retention in transiently powered computing systems

Energy harvesting has gained significant traction for powering Internet of Things (IoT) and autonomous devices in recent years. However, large energy buyers are typically required to tackle the instability of energy harvesting sources. Consequently, benefits such as low cost, weight, volume and complexity are counterbalanced. Transiently Powered Computing Systems typically contain little or no added energy storage and enable computation to be sustained, despite the variability of the harvested energy. They operate by saving the system state to Non-Volatile Memory before a power failure and restoring it once the power supply recovers. However, this is an energy consuming process that needs to be optimised to allow for maximum forward progress. This thesis presents an overview of transient systems, highlighting the need for efficient state retention.Novel selective software-based policies for efficiently retaining the system state which exploit properties of different Non Volatile Memory (NVM) technologies are proposed. They are based on (a) concatenating multiple images into the available NVM before erasing, and (b) selecting only the system state that has been updated since last saving. Results show that these policies can be up to 91% (Flash) and 86% (Ferroelectric RAM - FRAM) more energy efficient, compared to saving the entire system state.The full potential for energy efficient state retention cannot be reached, however, unless the memory usage can be monitored in run-time. A design exploration of potential approaches to memory tracking is presented, which facilitates the design and implementation of a hardware solution targeted at maximising the active time and, therefore, increasing the energy efficiency of transient systems.A memory tracking module (MeTra) is designed, simulated and synthesised, which monitors memory changes in run-time. Using this module, the system is able to dynamically adjust its hibernation voltage threshold according to the size and hence the specific energy needs of each snapshot. As a consequence, the active time is increased by up to 16x (Flash), while the energy spent for saving the state is decreased by up to 96% (FRAM)