CAREER: Application-specific Power Management

A large number of existing and emerging computing applications, including the internet of things, sensor networks, wearable electronics, and biomedical devices, have ultra-low-power requirements. In the low-power embedded systems used by these applications, energy efficiency is the primary factor that determines critical system characteristics such as size, weight, cost, reliability, and lifetime. Existing power management techniques for these systems trade off performance to reduce power. However, given the stringent power and energy constraints of emerging and existing low-power systems, solutions that sacrifice performance to reduce power may be unacceptable. Thus, this research focuses on novel opportunities to reduce power without reducing performance, thereby providing free power and energy savings for emerging low-power systems, and in turn, reducing their size, weight, and cost, and increasing their reliability and lifetime. This research also provides a non-intrusive way to significantly improve the energy efficiency of existing systems without sacrificing any performance or re-designing system hardware or software. Considering the sheer number of low-power processors being produced, their importance for future technologies, and the stringent power and energy constraints of these systems, saving power in such systems can have a significant impact. In addition to the technological impacts of this research, the project will contribute several other benefits, including new project-based research opportunities for undergraduate students and students from underrepresented groups, community outreach to K-12 students and other members of the university and community at large through public project showcases, and global outreach in the form of an educational initiative in Kenya on improving best practices in farming with IoT technology. Data characterizing the impacts of project activities on student learning outcomes will be used to improve the integration of research and educational activities at the PI?s institution.

The research contributions of this project stem from the development of novel techniques for hardware-software co-analysis that can identify the maximal set of hardware resources that an application can use in a processor, irrespective of application inputs. The results of such co-analysis can be used to eliminate any power expended by resources that an application can never use. This novel co-analysis approach can be leveraged to enable a suite of automated application-specific power management techniques, including application-specific timing analysis, which identifies the longest paths that an application can exercise in a processor and determines an application-specific lower-than-nominal minimum operating voltage that is guaranteed to be safe for the application; application-specific power domain formation and power gating, which can provide opportunities to power gate larger areas of logic for longer periods of time than state-of-the-art power gating techniques; application-specific peak power and energy management, which guarantees application-specific bounds on a design?s peak power and energy requirements and enables application-specific sizing for energy storage and harvesting components; and application-specific thermal management, which can identify and avoid hotspots, prevent thermal emergencies, and perform temperature-aware scheduling for a given application.