Dynamic power management technology based on prior knowledge

zbz0529

Dynamic power management technology based on prior knowledge [Copy link]

Abstract: "Dynamic power management" is a design method to reduce power consumption by dynamically allocating system resources to complete system tasks with the least components or the lowest workload of components in a low-power state. "Dynamic power management" technology includes a series of methods that enable the system to achieve effective energy saving. These methods control whether and when system components enter a low-power state when the "power management" is idle. This article mainly introduces an important method of dynamic power management - the predictive method.

Keywords: Dynamic power management Static prediction method Dynamic prediction method

introduction

Electronic systems can be considered as a collection of different types of components. Some components have fixed performance indicators and energy consumption, which are called non-power management components. On the other hand, some components can work at different times and have multiple energy consumption states, consuming different system power accordingly. These components are called power management components. The effective use of power management components is the key to saving system energy consumption and making the entire system work for a long time under limited power.

It often takes a while for a system component to change from one energy consumption state to another, and it consumes more extra energy during this time. The change of state will affect the performance of the system, so the designer needs to find an appropriate compromise between system energy saving and system performance. This article introduces some methods in dynamic power management. These methods will determine whether and when the component changes its energy consumption state.

1 Dynamic Power Management Technology

"Dynamic power management" is a design method to reduce power consumption by dynamically allocating system resources to complete system tasks with the least components or the lowest energy consumption state of the components with the least workload. To determine the implementation time of power management, a variety of prediction methods are used to predict the upcoming workload based on the historical workload, and decide whether to switch the working state and when to switch. This is the core of dynamic power management technology - dynamic power management method.

The basic premise for the application of dynamic power management technology is that system components have different workloads during working hours. Most systems have this situation. Another premise is that it is possible to predict the fluctuation of the workload of the system and components to a certain extent. Only in this way can it be possible to switch the energy consumption state, and the system should not consume too much energy during the observation and prediction of the workload.

2 Power Management

When each system device receives a request, the device is busy; when there is no request, it enters the idle state. When it is set to enter the idle state, the device can be turned off and enter the low-energy sleep state; when it receives a request again, the device is awakened. This is the so-called "power management". However, the change of energy state requires time, that is, the shutdown delay and the awakening delay. Waking up the device in the sleep state requires additional energy consumption, as shown in Figure 1. If there is no such overhead, there is no need for power management technology. The device can be turned off as long as it is idle. This delay and energy overhead are definitely present, so it must be considered that the device can only enter the sleep state when the energy saved in the sleep state is at least equal to the energy consumed by the state transition.

Power management technology is a predictive problem. We should seek to predict whether the idle time is long enough and whether it can offset the energy consumption of state transition. When the idle time is too short, the power management solution will not be worth the cost. Therefore, estimating the length of idle time in advance is the primary issue in power management technology. Define "appropriate stop time period" (tBE): the shortest idle time period that can achieve system energy saving. This time is related to the device components themselves and has nothing to do with the requests issued by the system. Assuming that the energy consumption of the state transition delay t0 (including shutdown and wake-up delays) is E0; the working state power Pw, the sleep state power Ps, tBE can be calculated by the following formula.

Pw×tBE=E0+Ps×(tBE-T0)

The left side of the equation is the energy consumption in the "suitable pause time period", that is, the energy required for the system to continue working during this shortest idle time for energy saving; the right side is the state conversion energy consumption and the system energy consumption during the sleep time. The power management technology is to predict whether the sleep time to be occurred can be greater than tBE. Only when it is greater than it, the device needs to sleep.

3 Dynamic Power Management Technology Based on A priori Prediction

For most real systems, the incoming signals are difficult to determine. Dynamic power management decisions are based on uncertain predictions of the future. The basic principle of all prediction-based dynamic power management techniques is to explore the relationship between the history of past workloads and the upcoming workloads to reliably predict future events. For dynamic power management, we are concerned about how to predict a long enough idle time to enter a sleep state, expressed as follows:

p={tIDLE>tBE}

We call the predicted idle time longer (shorter) than the actual idle time "over-prediction" ("under-prediction"). Over-prediction increases the impact on performance; under-prediction has no impact on performance but causes energy waste. If there is neither over-prediction nor under-prediction, it is an ideal prediction. The quality of prediction depends on the selection of observation samples and the statistics of workload.

3.1 Static prediction method

Fixed timeout method: The most common power management prediction method uses the past idle time as the observation object to predict the total duration of the current idle period. This method can be summarized as follows: the idle clock starts, the timer starts timing, and if the system is still idle after the fixed timeout period tTO, the power management puts the system into sleep mode until an external request is received, marking the end of the idle state. The ability to reasonably choose tTO is obviously the key to this method. Usually, tTO=tBE is taken when the requirements are not high.

The fixed timeout method has two advantages: 1) it is generally applicable (the scope of application is limited to the workload); 2) increasing the fixed timeout value can reduce the possibility of "over-prediction" (i.e. the predicted time is longer than the actual idle time). However, its disadvantages are also obvious: if the fixed timeout is too large, it will cause insufficient prediction, resulting in ineffective energy saving, and a considerable amount of energy is wasted on waiting for timeouts.