CAREER: Software Level Power Analysis and Optimization

Carnegie Mellon University, Pittsburgh PA

Investigators

Abstract

This research proposes novel techniques for power analysis and optimization at software level. Specifically, the problem is addressed in the context of both general purpose or application specific processors. In addition, the case of processors with multiple instruction issues or out-of-order execution, or architectures based on the Very Long Instruction Word (VLIW) paradigm are also considered. The main feature of the proposed models is their dependency on input data, allowing profile-driven optimizations. Based on the proposed models, appropriate optimization techniques at compile-and run-time are proposed for reducing the energy of the instruction sequence. This research will enable the assessment of any energy/performance trade-offs present in today's processors. Also, based on a power estimate of the embedded software running on application specific processors, the co-design of mixed hardware/software systems can be extended to incorporate as a further design constraint power consumption, in addition to performance.