A Comprehensive Study of Dynamic Power Management

Download Full-Text PDF Cite this Publication

Text Only Version

A Comprehensive Study of Dynamic Power Management

A Comprehensive Study of Dynamic Power Management

Prateek Bindra Assistant Professor, GEC, Panipat


Keywords: – Dynamic voltage scaling, dynamic frequency scaling, break even time.


    This paper has the objective to cover and relate different approaches to system-level DPM. We begin by describing how systems are affected by changing the values of operating voltage and frequency and how the use of their dynamic reconfiguration can impact the overall power consumption. Next, we review and compare different approaches to DPM. Three classes of power management policies have been proposed in the past: time-out, predictive, and stochastic policies [1], [8]. The fixed time-out policy shuts down the system after a fixed amount of idle time. Adaptive time-out policies are more efficient because they change the time-out according to the previous history. In contrast with time-out policies, predictive techniques do not wait for a time-out to expire, but shut down the system as soon as it becomes idle if they predict that the idle time will be long enough to amortize the cost of shutting down. A stochastic approach provides a polynomial-time exact solution for the search of optimal power management policies under performance constraints.


    This section introduces the basic principles of power consumption and the effects of voltage scaling [2]. CMOS circuits have both dynamic and static power consumption. Static power consumption is caused by bias and leakage currents but is insignificant in most designs that consume more than 1 mW.

    The dominant power consumption for CMOS microprocessors is the dynamic component. Every transition of a digital circuit consumes power, because every charge and subsequent discharge of the digital circuit's capacitance drains power. The dynamic power consumption is equal to

    P= C f Vdd2 (1)

    Where f is the number of clock cycles per sample period. C is the averaged switched capacitance per clock period and Vdd is the supply voltage It is clear from Equation (1) that reduction of Vdd is the most effective mean to lower the power consumption. Lowering Vdd, however, creates the problem of increased circuit delay. An estimation of circuit delay is given by

    Td = (CL. Vdd) / k (Vdd Vt) 2 (2)

    where, Td is the delay, Vdd is the supply voltage, CL is the total node capacitance, k is process constant, Vt is the threshold voltage, The propagation delay restricts the clock frequency in a microprocessor. From Equations (1) and (2), it follows that there is a fundamental trade-off between switching speed and supply voltage. Processors can operate at a lower supply voltage, but only if the clock frequency is reduced to tolerate the increased propagation delay.

    The critical path of a processor is the longest path a signal can travel. The implicit constraint is that the propagation delay of the critical path Td must be smaller than 1/f. In fact, the processor ceases to function when Vdd is lowered and the propagation delay becomes too large to satisfy internal timings at frequency f.


      1. Dynamic Voltage Scaling (DVS)

        For studying DVS, various algorithms are applied to various processors and power management is analyzed. Various strategies used for dynamic power management are summarized here.

        a) Adaptive Dynamic Power Management

        An adaptive DPM strategy is based on the exponential-average algorithm. The algorithm applies the last time predicted idle period and the actual one as weighting factors. The weighting for each older data point decreases exponentially, giving much more

        priority. Process usage is derived by the idle task which has the lowest priority in the system [4].

        Fig 1: Adaptive DPM model [3]

        Process usage indicates 0% if system executes idle task only because there are no tasks whose state are 'ready'. In a contrary concept, process usage will be 100% if idle task is not executed because there are

        importance to recent observations

        while still not

        lots of tasks whose state is 'ready'. Process usage

        discarding older observations entirely [3].

        This model consists of two parts: DPM Predictor and DPM Controller. DPM Predictor implements the proposed adaptive prediction strategy which can be divided into three modules: Counter, Basic Predictor

        (PU) is derived by this formula below:

        PU = 100(1- (Idle TaskCount / TotalCount)) (3) DVS is applied only when the process usage is higher

        and Adjuster. Counter is deployed

        to record Tn,

        than certain predefined value.

        As process usage

        which stands for the length of last actual idle period. The Counter starts counting immediately after the processor enters into the idle status and stops counting when detecting a processor interruption. Then, the value of Tn is sent simultaneously to the Basic Predictor and the Adjuster. When Adjuster

        increase, the amount of power consumption will

        decrease, but the probability of t sk with low priority being processed abnormally will increase.

        c) Profile based DVS

        Dynamic power management is done by analyzing

        receives the new value of Tn, it calculates the adjusting factor a. Since the calculation of a need two

        profile based power consumption using control flow

        graph (CFG).CFG has I/O-access blocks at various

        previous idle time value, the Adjuster conserves data using a simple FIFO with depth of 2. When the new value comes in, the old one is discarded. During a

        arbitrary locations and it tells about the possible

        flows of execution for a program[5],[6]. This approach attempts to utilize the control flow profile

        DPM process using Longtium DSP core processor, lowering the clock rate of a processor affects only dynamic power and reduction in dynamic power is in the range of 10-27%.

        b) Predictive DVS

        Dynamic Voltage Scaling (DVS) is the scheduling algorithm that changes the operating clock frequency of the processor according to the supplied voltage. DVS algorithm, applied only to the tasks with lower

        of an embedded code to take power management

        decisions of the peripherals rather than the CPU, with due consideration given to both performance and power.

        Two major concepts discussed are: Application model-CFG contains nodes of CPU related operations and peripheral (I/O) related operations.

        Fig 2: Predictive Scheduling for DVS [4]

        Break even time-The minimum length of an idle period to save power is called the breakeven time (TBE).

        TBE =TOFF +TMS +TON (4)

        Where TOFF: Device turns off time, TON: Device turns on time, TMS: Minimum sleeping time

        An offline algorithm has been devised to decide about the dynamic power down of peripherals during their idle times for power management. The algorithm consists of two steps: Identifying appropriate Switch-ON points ensuring performance and identifying profitable Switch-OFF points for energy saving. To turn it ON, a Switch-ON point is placed along every path which leads to that I/O block at a location in the CFG which has "enough time" (i.e. TON) to Turn-ON the peripheral without performance penalty. This ensures that the device will be turned ON along all possible entries of an I/O access. Among all possible paths from one I/O access to the successive I/O access, if the shortest path has profitable time (>TBE) to Turn-OFF the peripheral without power penalty, then a Switch-OFF point is place along every path from the current I/O block to the next I/O block. With this proposed method, the worst-case potential idle time duration for energy savings depends upon the shortest time duration taken by a Switch OFF point to meet the nearest Switch ON point. Hence, the energy savings is much dependent upon the density of I/O access blocks. Less closer they are, more opportunities for energy savings. Hence an application with relatively larger distances among consecutive I/O access blocks along

        several paths of execution can gain much from this approach.

      2. Dynamic Voltage Frequency Scaling

    In this approach the energy consumption is reduced by changing dynamically the supply voltage and operating frequency. One of the techniques used is discussed below.

    a) Deterministic Stretch-to-Fit (DSF)

    It is based on the slowdown strategy of reducing the processor power consumption [7]. Slowdown is known to reduce the dynamic power consumption at the cost of increased execution time for a given computation task. It detects early completion of tasks and exploits the processor resources to reduce the energy consumption. In Fig 3, by comparing the actual execution time (AET) of a task T1 with its worst-case execution time WCET) C1, (DSF) technique determines the value of the dynamic slack (). This slack time is exploited by the method to reduce the energy consumed, by stretching the execution of T2, having C2 as WCET, and reducing the frequency of the processor. tdisp is the available time at current processor frequency f. t1 and t2 represent respectively the activation date of T1 and T2.


    Various kinds of processors are used to study DPM using above discussed techniques. For DVS, SA- 1100 processor is used to analyze the voltage and frequency scaling effect on the power consumption of the system in which frequency can be varied from 59 MHz to 251MHz. Supply voltage can be varied from

    0.8 V to 2.0 V. To see the application performance,

    H.263 decoder is used and power consumption is measured with respect to the clock frequency. Power consumption is studied for two modes: idle and cpu – intensive. At the lowest clock frequency the processor consumes 1/5 of the energy per instruction that is required at peak performance. Numerical values: At 59 MHz: 105.8 mW and at 251 MHz:

      1. mW is consumed. The frequency and voltage can be scaled dynamically from user space in only 140 us. This allows pow er-aw are applications to quickly adjust the performance level of the processor whenever the workload changes.

        Fig 3: Slack reclamation using the DSF technique [7]

        Fig 5: Total power consumption for idle and cpu -intensive workloads [1]


        1. Tajana Simunic, Luca Benini, Andrea Acquaviva, Peter Glynnt, Giovanni De Micheli, Dy amic Voltage Scaling and Power Management for Portable Systems, ACM 1- 58113-297-2/01 2001

        2. Johan Pouwelse, Koen Langendoen, Henk Sips,

          Dynamic Voltage Scaling on a Low-Power

          Microprocessor ACM 1-58113-422-3/01/07 2001

        3. Jie CHEN, Deyuan GAO, Qiaoshi ZHENG A

          Fig 4:Power state transition of SA-1100 processor [9]

          For DVFS, the target hardware platform is the OMAP35x EVM board from MISTRAL (TEXAS INSTRUMENTS). It is equipped with the OMAP 3530 processor, an advanced superscalar ARM

          Research on an Optimized Adaptive Dynamic Power Management, IEEE 978-1-4244-4520-2/09 2009

        4. ChaeSeok, Lim, Hee Tak Ahn, Jong Tae Kim, Predictive DVS Scheduling for Low Power Real-time Operating System, International Conference on Convergence Information Technology 2007

        5. Lama H. Chandrasena and Michael J. Liebelt, A

          Cortex-A8 RISC Core. The models

          of the context

          comprehensive analysis of energy

          savings in dynamic

          switch energy consumption are integrated at a system

          level, the context switch consumes 6% of the energy consumed by the application.


We surveyed several classes of power-managed systems and power management policies.

supply voltage scaling systems using a data dependent voltage level selection, IEEE 0-7803-6536-4/00 2000

      1. Magesh Kumar, M. Sindhwani, T. Srikanthan, Profile- Based Technique for Dynamic Power Management in Embedded Systems, International Conference on Electronic Design 2008

      2. Bassem Ouni, C´ecile Belleudy, Hajer Ben Rekhissa,

Furthermore, we analyzed the tradeoffs involved in designing and implementing power-managed systems. Several practical examples of power-

Energy leakage in low power embedded

operatingmsystems using DVFS olicy, IEEE 978-1- 4673-0821-2/12 2012 [8] Yung-Hsiang Lu, Eui-Young Chung, Tajana Simuni, Luca Benini, Giovanni De Micheli,

managed systems were analyzed and

discussed in

Quantitative Comparison of

Power Management

detail. Even though DPM has been successfully employed in many real-life systems, much work is required for achieving a deep understanding on how to design systems that can be optimally power managed.

Algorithms Design, Automation Conference and Exhibition Proceedings 2000

[9] WenZhi Chen, Huan Zheng Analysis of Power Saving Effect for Dynamic Power Management Proceedings of the IEEE/ASME International Conference on Mechatronic and Embedded Systems and Applications 2006

Leave a Reply

Your email address will not be published.