Energy efficient custom instruction selection algorithm for DVS-enabled multiprocessor system

Abstract

Recently, the number of embedded systems are rapidly increasing, and many embedded software applications become more and more complex. Because many devices are portable and battery-powered, the energy efficiency of embedded processor, as well as their performance, is a very important issue of processor design. Currently, multiprocessor architecture and DVS (Dynamic Voltage Scaling) technique are the most popular methods for performance and energy efficiency in embedded systems. It is already well known that, for embedded systems, customizing the instruction set of processor to the specific application can increase performance of the system. But most previous researches have missed that the instruction set customization can significantly reduce the energy consumption. In this thesis, we suggest a custom instruction selection algorithm to minimize the energy consumption of embedded system. Our target environment is the embedded multiprocessor system where the DVS and ABB (Adaptive Body Biasing) technique is applied. Simulation results show that the average energy reduction of our algorithm is 43.5%, while the average energy reduction of the selection algorithm that considers only execution time of application is at most 13.4%.