Over the past decade, energy efficiency emerged as a first-order design challenge to developers across all layers of the computing stack, from microprocessor architects to large-scale datacenter operators. Unfortunately, in the future, operating a system close to its efficient design point for power will make the system susceptible to unreliability, whereas allowing margins will make the system inefficient. Therefore, understanding the interplay between power, performance and reliability is the next essential step for building sustainable computing systems in the future.
To this end, research under this proposal investigates new techniques to build computing systems that achieve both high energy efficiency and high reliability, but at a low cost to enable broader adoption of the techniques by our society. The investigators propose a resilient hardware/software co-designed machine organization that eliminates inefficiencies that arise from circuit- and architectural-level techniques that focus only on energy efficiency. The investigators' system optimizes reliability, energy, and performance in a coordinated manner. In this system, software continuously monitors execution and it dynamically adapts hardware resources based on feedback. The system automatically makes calculated efforts to characterize operational inefficiencies, and attempts to eliminate the inefficiencies by carefully relaxing the robustness of the system without compromising correctness. To characterize these inefficiencies and train the system, the investigators study and develop various algorithms, tools, and methodologies across the hardware and software boundaries.
