Evaluation of heterogeneous multiprocessor architectures by energy and performance optimization

Abstract

Design space exploration aims to find an energy-efficient architecture with high performance. A trade-off is needed between these goals, and the optimization effort should also be minimized. In this paper, we evaluate heterogeneous multiprocessor architectures by optimizing both energy and performance for applications. Ten random task graphs are optimized for each architecture, and evaluated with simulations. The energy versus performance trade-off is analyzed by looking at Pareto optimal solutions. It is assumed that there is a variety of processing elements whose number, frequency and microarchitecture can be modified for exploration purposes. It is found that both energy-efficient and well performing solutions exist, and in general, performance is traded for energy-efficiency. Results indicate that automated exploration tools are needed when the complexity of the mapping problem grows, starting already with our experiment setup: 6 types of PEs to select from, and the system consists of 2 to 5 PEs. Our results indicate that our Simulated Annealing method can be used for energy optimization with heterogeneous architectures, in addition to performance optimization with homogeneous architectures.