Design-time embedded software power estimation and optimization
DEI PhD Student
DEI - 3B Room
November 11th, 2011
at 5.30 p.m.
Early-stage estimation and analysis of embedded software power consumption is a critical issue that can determine the quality and, in some cases, the feasibility of the design. Several approaches exist to deal with this problem; for instance cycle-accurate instruction-set simulators are known to provide high details for fine-tuning performance and power consumption of critical sections of the application, but they are often too slow for the simulation of an entire system, making them not appropriate for early-stage evaluation.
The aim of this research is to propose a novel, fast and statistically accurate methodology to evaluate and optimize the energy and performance requirements of embedded software, according to a static characterization of the target instruction set, and on dynamic behavior analysis: it can compute estimates on an equivalent, target-independent intermediate code representation based on the LLVM compiler infrastructure. The proposed methodology is also able to provide the developer with a set of source code level transformations (e.g., loop unrolling or function inlining) that aim at reducing the power consumption of the embedded software.