Renaissance: Revamping Software on Non-volatile Processors for Energy Harvesting Embedded Systems

Description

In recent years, we have witnessed exponential growth of embedded systems such asInternet of Things and wearable devices, which have and will continue to revolutionizethe quality of our daily lives. These intelligent systems often employ sensors to collectrelevant data, which are then analyzed and transmitted to information hubs.Traditionally, these embedded systems are powered by batteries. However, in wearabledevices such as health monitors, batteries are no longer favored due to size limitations,safety concerns, and recharging inconvenience.Energy harvesting is becoming a favorable power source, as it can harvest energy fromsolar, electromagnetic radiation, or thermal sources to power electronics. It providesbetter user experiences in an environment-friendly manner. However, energy harvestingpower sources are intrinsically unstable. Due to this instability, program execution maybe frequently interrupted. With CMOS-based volatile processors, the intermediateresults will be lost after power outages and thus result in vain re-executions, whichimpose heavy energy overhead and even infeasibility to finish for large tasks. Nonvolatileprocessors, in which non-volatile memory is attached to the volatile logics tobackup program state information, have been developed recently to address this issue.Currently, most work for non-volatile processors focuses on hardware layer modelingand optimizations. Little work has been done on software level for non-volatileprocessors.This project aims to fill this gap and develop a reliable and efficient software frameworktailored for non-volatile processors in energy harvesting powered embedded systems.The challenge lies in the simultaneous consideration of energy harvesting, application,and non-volatile processor characteristics. As frequent power interruptions become thenorm of execution, the key is to design an ultra-efficient checkpointing strategy fromthe ground up. First, we will focus on a single application to minimize the program stateinformation that needs to be saved. Instead of an immediate backup of the full programstate upon energy low signals, both time and space dimensions of program execution willbe explored. At the same time, the correctness of execution will be ensured. Secondly,scheduling multiple applications will be analyzed with checkpointing and differentobjectives. Preliminary results show that with proper software design, significantreduction can be achieved for the program state that needs to be checkpointed, whichgreatly improves both reliability and efficiency.The success of this project will enable pervasive adoption of user-friendly andenvironment-friendly embedded systems powered by energy harvesting, constituting anew renaissance for the innovation industry of Hong Kong.