Effective Supervisors for Predictive Methods of Dynamic Power Management

Abstract

The paper presents new supervisors dedicated to predictive methods of dynamic power management, i.e. DCT (dynamic clock throttling), DFS (dynamic frequency scaling), and DVS (dynamic voltage scaling). The presented supervisors make decisions on the basis of current chip temperature; future, current and previous power dissipations. They consist in cooperation with operating system and they are dedicated to high efficiency systems. The proposed supervisors can be implemented in both software and hardware, e.g. as neural network. Not only performance gain but also energy profit can be made in systems that use these supervisors. Simulations results of considered cases show that theoretical improvement of the ideal supervisor is in the range of 7.38 to 16.17% for performance and of 2.47 to 9.88% for energy. The profit of the real supervise method depends on the complexity of supervisor.