Collaborative Research: SHF: Small: Enabling Caches and GPUs for Energy Harvesting Systems

Virginia Polytechnic Institute And State University, Blacksburg VA

Investigators

Abstract

Energy harvesting systems collect energy from variant ambient sources such as solar power, thermal energy, and radio-frequency radiation. Due to unreliable energy sources, energy harvesting systems suffer from frequent power failures. Hence, energy harvesting systems should be able to save the current program states before a power failure, restore the consistent program states when the power comes back, and seamlessly resume program execution as if nothing happened. However, maintaining crash-consistent states across power cycles is challenging. As a result, the current generation of energy harvesting systems has been designed with simple hardware configurations such as a single central processing unit (CPU) without a cache, delivering limited computing capabilities. Going forward, in the new Internet of Things era, ever-increasing demand for substantially more high-performance energy-harvesting systems capable of supporting emerging artificial-intelligence and machine-learning applications are expected. This project proposes new software and hardware co-design solutions that allow energy-harvesting systems to leverage caches and graphic processing units (GPUs) for high performance and energy efficiency. The project is expected to serve as the foundation to unlock next-generation Internet of Things services, based on battery-less energy-harvesting systems. The project also aims to incorporate research findings in undergraduate teaching and offer K-12 outreach programs for female students to promote more equitable outcomes for women in computer science. The objective of this project is to enable caches and GPUs in energy-harvesting systems and to design next-generation energy harvesting systems with high performance and energy efficiency. To this end, the project proposes compiler- and hardware-based solutions in three research thrusts. The project will explore a compiler-based solution that allows existing energy-harvesting systems to use a traditional data cache without hardware modification. The project will explore a new research direction that avoids expensive logging at run time, yet instead recovers potentially un-persisted stores at reboot time. To achieve better performance, the project aims to design a new hardware-based cache for energy-harvesting systems, which combines the benefits of a write-back cache and a write-through cache without their respective downsides. The project will design the first energy-harvesting GPU system that introduces a new checkpointing solution for GPU registers and a lightweight persistence solution for GPU shared memory. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.