Power-Aware Resource Management Techniques for Low-Power Embedded Systems

Abstract

For example, consider a task with a deadline of 25 msec, running on a processor with the 50 MHz clock speed and 5.0 V supply voltage. If the task requires 5 • 10 5 cycles for its execution, the processor executes the task in 10 msec and becomes idle for the remaining 15 msec. (We call this type of an idle interval the slack time.) However, if the clock speed and the supply voltage are lowered to 20 MHz and 2.0 V, it finishes the task at its deadline (= 25 msec), resulting in 84% energy reduction. Since lowering the supply voltage also decreases the maximum achievable clock speed [43], various DVS algorithms for real-time systems have the goal of reducing supply voltage dynamically to the lowest possible level while satisfying the tasks' timing constraints. For real-time systems where timing constraints must be strictly satisfied, a fundamental energy-delay tradeoff makes it more challenging to dynamically adjust the supply voltage so that the energy consumption is minimized while not violating the timing requirements. In this chapter, we focus on DVS algorithms for hard real-time systems. On the other hand, dynamic power management decreases the energy consumption by selectively placing idle components into lower power states. At the minimum, the device needs to stay in the low-power state for long enough (defined as the break even time) to recuperate the cost of transitioning in and out of the state. The break even time T BE , as defined in Equation 1.1, is a function of the power consumption in the active state, P on , the amount of power consumed in the low power state, P sleep , and the cost of transition in terms of both time, T tr , and power, P pr. If it was possible to predict ahead of time the exact length of each idle period, then the ideal power management policy would place a device in the sleep state only when idle period will be longer than the break even time. Unfortunately, in most real systems such perfect prediction of idle period is not possible. As a result, one of the primary tasks DPM algorithms have is to predict when the idle period will be long enough to amortize the cost of transition to a low power state, and to select the state to transition to. Three classes of policies can be defined-timeout based, predictive and stochastic. The policies in each class differ in the way prediction of the length of the idle period is made, and the timing of the actual transition into the low power state (e.g., transitioning immediately at the start of an idle period vs. after some amount of idle time has passed).

Current trends in low-power embedded computing

“…The ability to enable and disable components, as well as to tune their performance according to the workload is key factor in achieving energy-efficient systems. A power managed system can be modeled by a finite state machine representation called power state machine (PSM) [8,10]. Each state in the PSM represents the various modes of operation characterized by power consumption and performance of the component or device when in that state.…”

“…Different DPM policies are proposed and evaluated in the literature. They can be broadly classified as heuristic and stochastic methods [10].…”

Distributed, Robust Auto-Scaling Policies for Power Management in Compute Intensive Server Farms

“…The idea of leaving a server purposely idle for some time before shutting it off has been used before in the mobile computing community (see [6,7,14]), however, only for a single device. For a multi-server system such as the one we consider, we find that the right t wait setting is greatly affected by the routing policy used to dispatch jobs to servers, as discussed in the next section.…”

Energy Harvesting for Structural Health Monitoring Sensor Networks

“…Early work typically assumed that the tasks run at their worst case execution time (WCET), while the later research work relaxes this assumption and suggest a number of heuristics for prediction of task execution time. A more detailed overview on various DVS algorithms can be found in [22]. …”