Efficient and scalable compiler-directed energy optimization for realtime applications

Abstract

With continuing shrinkage of technology feature sizes, the share of leakage in total energy consumption of digital systems continues to grow. Coordinated supply voltage and body bias throttling enables the compiler to better optimize the total energy consumption of the system in future technology nodes. We present a compilation technique that targets realtime applications running on embedded processors with combined dynamic voltage scaling (DVS) and adaptive body biasing (ABB) capabilities. Considering the delay and energy penalty of switching between operating modes of the processor, our compiler judiciously inserts mode-switch instructions in selected locations of the code and generates executable binary that is guaranteed to meet the deadline constraint. More importantly, our algorithm runs very fast and comes reasonably close to the theoretical limit of energy optimization using DVS+ABB. At 65nm technology, we improve the energy dissipation of the generated code by an average of 33.20% under deadline constraints. While our technique's improvement in energy dissipation over conventional DVS is marginal (6.91%) at 130nm, the average improvement continues to grow to 13.19%, 22.97%, and 33.21% for 90nm, 65nm, and 45nm technology nodes, respectively. Compared to a recent ILP-based competitor, we improve the runtime by more than three orders of magnitude, while producing improved results.