Real-time System Power Consumption Measurement And Analysis Based On Performance Counters

Along with the requirements of environmental protection, green development and sustainable development, it has become one of the most valuable research hotspots to increase the energy efficiency of computer systems for researchers, architects, system designers and software developers and so on. There are already a lot of researches focused on reducing power consumption and increasing energy efficiency. For example, DVFS technique is designed to reduce power consumption when the system load is low. The value of real-time power consumption is a prerequisite for dynamic power management. In all areas about power-aware system research, the power measurement and analysis are very important. It is not only used to make power/performance evaluation and make software power consumption analysis, but also used for triggering power-aware mechanisms and evaluating the effect of these mechanisms.Power measurement and analysis methods include hardware-based approaches and software-based approaches. Hardware-based power measurement and analysis mainly refer to exploiting a variety of hardware instruments to directly measure the power of a device. The results of this method are generally much more precise than the results of the software-based analysis. It is typically used to evaluate the effect of the power saving technology. However, the hardware-based methods are limited to component-level power measurements. Software-based power measurement and analysis are not as extremely accurate as hardware-based approaches, but they try to design a set of power models to estimate different levels of power, which are more flexible and more efficient. This paper firstly describes some valuable power measurement and analysis technologies and their current research status, and then analyzes and proposes a method of software power measurement and analysis based on performance counters.Total system power consumption is the sum of individual hardware unit power consumption, the performance counters (PMC) can accurately reflect the use of the corresponding hardware unit, in turn, reflect its power consumption. In this paper, we propose a method to establish real-time power consumption model, which is based on a series of fine-grained performance counters. The method has5main steps, which are power model setting, performance counters selection, data sampling, model solution, result examination. The basic idea is as follows:select a collection of basic performance counters as Fpreslct, which can reflect the utilization of processor, memory, I/O, etc. By experiments, record the relationship of these performance counters and the total power consumption, and then, obtain the influence coefficient of each performance counter through multiple linear regression method, so as to establish a real-time power consumption model. Finally, experimental tests on mainstream platforms show that the power model established by the method has high accuracy and high applicability.