Gerenciamento dinâmico de energia em processadores com cargas de trabalho variantes no tempo.

Abstract

The need to save power is present in almost every computer system. To handle this increasing demand for power management, the semiconductor manufacturers have been developing several low power design techniques, such as: low power states, dynamic voltage and frequency scaling, and clock gating. For example, in the case of processor power management at the operating system level, the goal is to choose the most suitable pair of voltage and frequency to minimize the power consumption and guarantee the performance level desired by the user. A great challenge for the dynamic powermanagement of general purpose computer systems is that the system workload is initially unknown and time-variant. Therefore it is necessary a power manager, where the system power states are choosen by means of a procedure named power management policy. In the literature, there are many power management policies which are claimed to be optimum for specific targeted systems. Furthermore, most of those policies may be considered as open loop policies, because, during execution time, the system performance is not compared to a specified reference performance value. To guarantee that the system performance is equal to a specified reference performance value, closed loop control techniques have been applied to the power management problem at several cases, such as, the applications performance, the utilization of an embedded microprocessor running a known set of real-time control tasks, the occupancy of a buffer in a multimedia decoder, etc.. In thisWork, an adaptive control technique for dynamic power management was developed with the goal of guaranteeing a specified reference performance value and minimizing the power consumption of a processor with dynamic voltage and frequency scaling, considering that the system workload is initially unknown and time-variant. This technique was conceived to be as general as possible, so that it may be used for different types of computer systems. In the experimental results, the developed adaptive control technique outperformed other state-of-the art policies both in power and system performance.