Abstract: System power consumption is a significant factor hindering the development of portable electronic products worldwide. The core of reducing system power consumption lies in employing appropriate power management strategies, with the key being the accurate reflection and utilization of system state and task information. This paper analyzes power management strategies and methods for system devices from the perspectives of task information and system state, and proposes a system-level power management module architecture based on top-level decision-making. Keywords: Embedded system, Portable device, Power management, Top-level decision-making Introduction Today, the performance requirements for electronic systems are increasingly demanding. When balancing performance and power consumption, performance often receives greater emphasis. Limited-capacity batteries are the sole energy source for portable devices, and the rate of increase in battery capacity significantly lags behind the rate of increase in CPU performance. Therefore, how to utilize limited power to provide maximum performance for portable devices is the primary goal of power management in portable devices. In addition, power management must also consider stability and heat dissipation. A power management module is a software module on a programmable power management device that provides an application programming interface for implementing various power consumption modes. Power consumption occurs in two ways: static power consumption and dynamic power consumption. Static power consumption is mainly transistor leakage power; dynamic power consumption comes from circuit activation, such as the activation of register lines caused by address or data line input. The power consumed by the switched capacitor is the most important component of dynamic power consumption, that is, the power consumed by the switched capacitor during charging and discharging when the circuit is output: From equation (1), it can be seen that Pdynamic depends on the following four parameters: C (capacitance), V (voltage), f (signal frequency), and a (variable factor). Among them, a is related to the number of 0-1 transitions in the chip. The methods to reduce dynamic power consumption are correspondingly divided into four categories: ① Reduce the capacitance or the energy stored in the circuit. ② Reduce the activity of the switches. As computer chips are packaged with increasingly complex functions, the switching activity of the chips is enhanced, so reducing the switching activity plays an increasingly important role in reducing the dynamic power consumption of the system. Clock gating technology is a currently popular technology to reduce switching activity. This technology prevents the clock signal from passing through idle application units. Since the clock network is the main part of chip power consumption, this technology can effectively reduce power and energy consumption in the processor. ③ Reduce signal frequency. ④ Reduce voltage. The latter two methods come at the cost of reduced system performance, but they are also the main means of reducing system power consumption. System power management involves understanding system tasks and statuses, utilizing reasonable power management strategies, and balancing the relationship between improving system performance and reducing system power consumption to provide optimal performance for system applications with the lowest power consumption. This paper explores and studies the management of reducing system power consumption from the perspectives of task information and system status, and proposes a system-level power management module architecture based on the highest decision. 1 Power Management Technology Analysis 1.1 System Status, Task, and Power Management Strategy Information The premise of power management is the accurate detection and management of system device status and task information, as well as accurate efficiency statistics of power management strategies. System status includes Running, Idle, and Sleep states. Some systems can provide multiple working states, the main difference being the processor operating frequency, operating voltage, and device combination. Real-time indicators of tasks include response time, latency, and task deadline. Hard real-time tasks have hard requirements for these indicators. When the system fails to meet the indicators, the data or services provided will be completely invalidated, or even cause catastrophic consequences. Soft real-time tasks only have soft requirements for them. The consequence of failing to meet the indicators is that the required service quality cannot be provided. In addition to real-time indicators, task information also includes the equipment components used to execute the task. If it is a multi-working-mode system, it should also include the minimum working mode for executing the task. Many operating systems and processors can provide good detection modules and functional units. For example, the Linux timer function can provide system status monitoring timing, and the PMU (Performance Monitoring Unit) of the Intel XScale processor can be used to monitor the working status of the XScale platform. The detection and management of system status and task information is one of the important parts of the management module. Power management strategy information includes core algorithms and working efficiency. The efficiency of power management strategy can be measured by calculating the "competitive ratio" and "error rate". D. Ramanathan[2] used the "competitive ratio" as a metric in analyzing power management strategies using the competitive analysis method. The premise of the competition analysis method is that the problem under study has a competitor, and this competitor can influence the input of the problem. The competition ratio is defined as the ratio between the resources consumed by the online strategy and the minimum resources that can be consumed to complete the task. Here, the online strategy is a power management strategy for a system with an unknown load. Because in a real system, the arrival time of the next task request cannot be predicted perfectly, the minimum resources that can be consumed are those consumed assuming the power management strategy can predict the arrival time of the next task request perfectly. Here, resources can be simply represented by power consumption, or it can be combined with the system's task execution latency, i.e., system performance. The prediction error rate is an efficiency indicator for the prediction strategy; it equals the ratio between the number of prediction errors and the total number of predictions, and a Boolean function can be used to determine the evaluation function of the prediction error rate. 1.2 Power Management Method in Running State When the system is in the Running state, the power management module, based on task information, reduces power consumption by switching the state or operating mode of system devices, while ensuring the task is completed. For example, tasks can be categorized as computationally intensive or memory-intensive. When executing computationally intensive tasks, system power consumption can be reduced by lowering the bus frequency while ensuring real-time task completion. Conversely, when executing memory-intensive tasks, power consumption can be reduced by lowering the processor's operating frequency. When the system executes multiple tasks simultaneously or multiple tasks at once, an effective power management strategy combines task scheduling with task deadlines. The basic idea behind this strategy is to group tasks by device and task set, list all possible scheduling options, exclude scheduling that doesn't meet the deadline, and prioritize tasks within the same group for execution within their deadlines. This concentrates system idle time as much as possible, achieving dynamic power management. The task scheduling process applied to this strategy is shown in Figure 1. Figure 1. Task scheduling process combined with task deadlines. The scheduling principle of the power management strategy based on task scheduling and task deadlines is as follows: ① The lower the scheduling energy consumption, the higher the priority; tasks in the same group are sorted by deadline; the earlier the deadline of the first task in each group, the higher the priority of that group; the earlier the deadline of the last task in each group, the higher the priority of that group. ② For scheduling with the same scheduling energy consumption and deadline, the first to arrive has a higher priority. ③ When an external task requests to use a hibernating device, the power management module rearranges the task priorities. Let P(s) be a continuous function. If the system device is running at speed s, then its power consumption is P. According to the cube-root principle of CMOS-based devices, we have: For ease of analysis, the relationship between power consumption and system device operating speed is expressed as follows: This is a strictly convex function, which conveys the message that the slower the task proceeds, the more power is saved. This is the basic starting point of the power management strategy based on task deadline constraints. There are currently many power management strategies based on task deadline constraints, such as simplified online strategies AVR[3] (Average Rate), OA (Optimal Available), and BKP[4], which are typical of these strategies. Among them: AVR strategy assumes that only one task is being executed in the system; OA strategy assumes that no new task will be scheduled; while BKP strategy can effectively reduce power consumption when c is relatively large. 1.3 Power Management Methods in Idle State After the system device completes its task, the main method of power management in this state is to switch the system device in the Idle state. The mainstream strategies are Timeout strategy, prediction-based management strategy, and random management strategy. Among them, Timeout strategy is the simplest and easiest to implement. The process of this strategy is shown in Figure 2. Figure 2 Timeout strategy process After the system completes all tasks, when the duration of the Idle state exceeds the threshold, the power management module will switch the system to the Sleep state until a new task request arrives and then wake up the system. This method achieves the purpose of reducing the power consumption of the system device. The time interval can be set by the timing module provided by the system, and the time threshold Tth is determined by the following formula: Where: Etran is the total energy consumed by the system during the two known state transitions from Idle to Sleep and then to wake-up; PI is the power consumed by the system in the Idle state. Figure 3 shows the performance loss of the two timeout strategies. In the figure, E is the Running (working) state time, I is the Idle state time, F is the time threshold, D is the state transition time, S is the sleep state time, and W is the system device wake-up time. This strategy is simple, but its disadvantages are also obvious. As shown in Figure 3, when I>F+D, setting the waiting time threshold easily leads to more opportunities for power reduction. Simultaneously, the time and energy consumed by the system's wake-up transition will inevitably cause task waiting delays; even when F+D>I>F, the delay will be greater than the wake-up time, resulting in significant performance loss. Furthermore, the delay in task execution time will directly lead to a shortening of the next Idle state duration. Therefore, a power management strategy based on predicting the Idle state time after task completion and the arrival time of the next task becomes very efficient. Figure 3 shows the performance degradation of the two Timeout strategies. Prediction-based power management strategies predict the duration Tpred of the system's Idle state based on system information (including historical information and user habits). Comparing Tpred and Tth, if Tpred ≥ Tth, the system is immediately switched to sleep mode after the task is completed; otherwise, the system remains in the Idle state. The prediction time and the prediction interval within the Idle state are determined by the specific strategy. The core of prediction-based power management strategies is which algorithm to use to update the algorithm's prediction basis using system feedback information. To make accurate predictions that conform to the user habits and task requests of system devices, a deeper understanding of user habits is needed, along with comprehensive statistics on system task information and strategy history. Adaptive Learning Tree (ALT) strategies, Probability-Based ALT (PBALT) strategies, and Predictive Control Feedback (PCF) prediction strategies based on AR (Auto-Regressive) models are all optimized prediction strategies. The PBALT strategy uses probability to reflect accuracy, strengthening the correlation between subtrees and the learning ability of the ALT method; however, the boundary conditions of this strategy limit its application scope. The adaptability of the PCF prediction strategy is controlled by its feedback module; however, the prediction strategy itself is inefficient when dealing with non-stationary task requests, and it basically only considers one operating mode of the system, all of which limit its application. Stochastic power management strategy is an optimization strategy with uncertainty, stemming from the abstract nature of the system model. It specifies not only when to perform state transitions but also which operating mode to transition to, thus making it suitable for systems with multiple operating modes. It treats dynamic power management as a stochastic optimization problem, unlike prediction-based power management strategies which eliminate the uncertainty of task requests through prediction. The stochastic decision-based dynamic power management strategy based on CTMDP (Continuous Time Markov Decision Process) provides an optimal decision for system power management. However, this optimization is based on a model with uncertainty. That is, the optimal decision obtained by this algorithm can only yield an expected value for system performance and power consumption, and cannot guarantee its applicability to specific system devices. Furthermore, the establishment of the Markov process mathematical model requires careful analysis. 2. Power Management Strategy Based on the Highest Decision As the above analysis of system power management strategies shows, power management of system devices spans all states. Therefore, a power management method should be proposed that combines multiple power management strategies to collaboratively manage system power consumption. This power management architecture includes a set of strategies, each with its own priority, using various strategies as needed for multi-strategy power management. However, this architecture also has problems: First, complex systems may have diverse tasks, and the power reduction efficiency of power management strategies varies for different tasks. Using only the priority of power management strategies to determine their use lacks specificity. Furthermore, information on each strategy should be statistically analyzed during system task execution, and its priority should be adaptively adjusted. This paper proposes a power management architecture based on a top-level decision-making management module. This system device power management architecture comprises five main parts: a top-level decision-making module, a task information statistics module, a strategy set module, an information detection module, and a control module, as shown in Figure 4. Figure 4: Power Management Module Architecture Based on Top-Level Decision-Making. Information Detection Module: Used to detect system status information and newly arrived task information. Task Information Statistics Module: Used to statistically analyze the task information executed by the system devices and interpret it into accurate task information parameters. Strategy Set Module: Calculates the efficiency of power management strategies by dynamically analyzing system status and task information, updates power management strategy information, and interprets it into accurate power management strategy parameters. Top-Level Decision-Making Module: Based on the received task and system status information, selects the optimal power management strategy or power management strategy group from the strategy set, and manages the power of the system devices through the control module. Task information is received in real time; system status information is provided to the top-level decision-making module by the information detection module each time the system status changes; power management strategy information refers to the calculated power management efficiency, as well as the system status and tasks to which the power management strategy applies. For example, when a new task arrives, there must be a prediction strategy that is most efficient at predicting the duration of the Idle state after the task is completed. During power management strategy control, the success or failure of each decision changes the priority weighting parameters of that power management strategy. Thus, the highest decision module, based on system state and task information, decides to adopt the optimal power management strategy or a group of power management strategies, ensuring optimal power management for all parts of the system devices. 3. Summary The core of power management in today's portable devices is the power management strategy. The key to the highest decision-based power management architecture proposed in this paper is the pre-selection of a set of power management strategies. Regarding power management strategies, two aspects require further discussion and research: First, balancing system device performance and power consumption. While power management strategies aim to avoid delays during system power management, such delays are unavoidable. The user's trade-off between performance and power consumption directly affects the selection of power management strategies and the preset parameters within those strategies. Second, balancing power management effectiveness and complexity. The larger the size of the strategy set and task information set, the more complete the statistical information, and the more accurate the power management strategy decisions. However, this also increases the complexity of the power management module, which directly relates to the complexity of its engineering implementation. Furthermore, establishing power management strategy standards, providing power management strategy packages and task information packages, and standardizing system status and task information will also benefit the development of power management technology for portable devices.