Energy optimization of parallel programs in a heterogeneous system by combining processor core-shutdown and dynamic voltage scaling

Abstract

Reducing power consumption and improving efficiency are important aspects of the development of supercomputers into large-scale systems. As a result, heterogeneous systems have become an important development trend in high-performance computing. From the perspective of heterogeneous systems, this study establishes a model for energy optimization of parallel programs (EOPP) and puts forward a method of using it. By considering the energy overheads caused by re-synchronization, voltage switching, and operations in critical sections, the model effectively combines processor core-shutdown and dynamic voltage scaling technologies, which can be applied in a heterogeneous system to guide the optimization process. The results show that the proposed model can effectively reduce the energy consumption of parallel programs. Moreover, increasing the proportion of operations in the critical section enhances the optimal frequency of a processor while decreasing the probability of conflicts in the critical section. It can thus provide optimization space for reducing the frequency of a processor which ultimately reduces the energy overhead of the system.