Dynamic Power Management Policy Research Articles

Dynamic Power Management in Large Manycore Systems: A Learning-to-Search Framework

The complexity of manycore System-on-chips (SoCs) is growing faster than our ability to manage them to reduce the overall energy consumption. Further, as SoC design moves toward three-dimensional (3D) architectures, the core's power density increases leading to unacceptable high peak chip temperatures. In this article, we consider the optimization problem of dynamic power management (DPM) in manycore SoCs for an allowable performance penalty (say, 5%) and admissible peak chip temperature. We employ a machine learning– (ML) based DPM policy, which selects the voltage/frequency levels for different cluster of cores as a function of the application workload features such as core computation and inter-core traffic, and so on. We propose a novel learning-to-search (L2S) framework to automatically identify an optimized sequence of DPM decisions from a large combinatorial space for joint energy-thermal optimization for one or more given applications. The optimized DPM decisions are given to a supervised learning algorithm to train a DPM policy, which mimics the corresponding decision-making behavior. Our experiments on two different manycore architectures designed using wireless interconnect and monolithic 3D demonstrate that principles behind the L2S framework are applicable for more than one configuration. Moreover, L2S-based DPM policies achieve up to 30% energy-delay product savings and reduce the peak chip temperature by up to 17 °C compared to the state-of-the-art ML methods for an allowable performance overhead of only 5%.

Energy-Efficient Scheduling Based on Task Migration Policy Using DPM for燞omogeneous MPSoCs

Increasing the life span and efficiency of Multiprocessor System on Chip (MPSoC) by reducing power and energy utilization has become a critical chip design challenge for multiprocessor systems. With the advancement of technology, the performance management of central processing unit (CPU) is changing. Power densities and thermal effects are quickly increasing in multi-core embedded technologies due to shrinking of chip size. When energy consumption reaches a threshold that creates a delay in complementary metal oxide semiconductor (CMOS) circuits and reduces the speed by 10%–15% because excessive on-chip temperature shortens the chip’s life cycle. In this paper, we address the scheduling & energy utilization problem by introducing and evaluating an optimal energy-aware earliest deadline first scheduling (EA-EDF) based technique for multiprocessor environments with task migration that enhances the performance and efficiency in multiprocessor system-on-chip while lowering energy and power consumption. The selection of core and migration of tasks prevents the system from reaching its maximum energy utilization while effectively using the dynamic power management (DPM) policy. Increase in the execution of tasks the temperature and utilization factor on-chip increases that dissipate more power. The proposed approach migrates such tasks to the core that produces less heat and consumes less power by distributing the load on other cores to lower the temperature and optimizes the duration of idle and sleep times across multiple CPUs. The performance of the EA-EDF algorithm was evaluated by an extensive set of experiments, where excellent results were reported when compared to other current techniques, the efficacy of the proposed methodology reduces the power and energy consumption by 4.3%–4.7% on a utilization of 6%, 36% & 46% at 520 & 624 MHz operating frequency when particularly in comparison to other energy-aware methods for MPSoCs. Tasks are running and accurately scheduled to make an energy-efficient processor by controlling and managing the thermal effects on-chip and optimizing the energy consumption of MPSoCs.

Dynamic Network State Learning Model for Mobility Based WMSN Routing Protocol

The rising demand of wireless multimedia sensor networks (WMSNs) has motivated academia-industries to develop energy efficient, Quality of Service (QoS) and delay sensitive communication systems to meet major real-world demands like multimedia broadcast, security and surveillance systems, intelligent transport system, etc. Typically, energy efficiency, QoS and delay sensitive transmission are the inevitable requirements of WMSNs. Majority of the existing approaches either use physical layer or system level schemes that individually can’t assure optimal transmission decision to meet the demand. The cumulative efficiency of physical layer power control, adaptive modulation and coding and system level dynamic power management (DPM) are found significant to achieve these demands. With this motivation, in this paper a unified model is derived using enhanced reinforcement learning and stochastic optimization method. Exploiting physical as well as system level network state information, our proposed dynamic network state learning model (NSLM) applies stochastic optimization to learn network state-activity that derives an optimal DPM policy and PHY switching scheduling. NSLM applies known as well as unknown network state variables to derive transmission and PHY switching policy, where it considers DPM as constrained Markov decision process (MDP) problem. Here,the use of Hidden Markov Model and Lagrangian relaxation has made NSLM convergence swift that assures delay-sensitive, QoS enriched, and bandwidth and energy efficient transmission for WMSN under uncertain network conditions. Our proposed NSLM DPM model has outperformed traditional Q-Learning based DPM in terms of buffer cost, holding cost, overflow, energy consumption and bandwidth utilization.

HiLITE: Hierarchical and Lightweight Imitation Learning for Power Management of Embedded SoCs

Modern systems-on-chip (SoCs) use dynamic power management (DPM) techniques to improve energy efficiency. However, existing techniques are unable to efficiently adapt the runtime decisions considering multiple objectives (e.g., energy and real-time requirements) simultaneously on heterogeneous platforms. To address this need, we propose HiLITE, a hierarchical imitation learning framework that maximizes the energy efficiency while satisfying soft real-time constraints on embedded SoCs. Our approach first trains DPM policies using imitation learning; then, it applies a regression policy at runtime to minimize deadline misses. HiLITE improves the energy-delay product by 40 percent on average, and reduces deadline misses by up to 76 percent, compared to state-of-the-art approaches. In addition, we show that the trained policies not only achieve high accuracy, but also have negligible prediction time overhead and small memory footprint.

Dynamic Power Management With Optimal Time-Out Policies

In the dynamic power management (DPM), it is quite important to switch a high-power consuming state to a low-power consuming one at a suitable timing. This problem has been formulated as Markov decision processes (MDPs) with state-dependent control in the past literature. However, this approach may not be often feasible in many practical situations, because the state-dependent policy requires that all the states on a request arrival process must be observed through an online monitoring. To overcome this problem, we develop a simple time-out policy in the DPM, which can be regarded as the optimal timing to take a GO-SLEEP action during an idle period of the transaction system. We derive the optimal time-out policy minimizing the expected power consumption per unit time in the steady state analytically under the assumption that the request arrival process is given by a Markovian arrival process with an arbitrary number of phases. In numerical experiments with real read/write data for a hard disk unit and CPU utilization data, we estimate the expected power consumption per unit time under two DPM policies, namely, the time-out policy and the MDP-based policy, and quantitatively compare their effectiveness on the power reduction.

An Adaptive and Integrated Low-Power Framework for Multicore Mobile Computing

Employing multicore in mobile computing such as smartphone and IoT (Internet of Things) device is a double-edged sword. It provides ample computing capabilities required in recent intelligent mobile services including voice recognition, image processing, big data analysis, and deep learning. However, it requires a great deal of power consumption, which causes creating a thermal hot spot and putting pressure on the energy resource in a mobile device. In this paper, we propose a novel framework that integrates two well-known low-power techniques, DPM (Dynamic Power Management) and DVFS (Dynamic Voltage and Frequency Scaling) for energy efficiency in multicore mobile systems. The key feature of the proposed framework is adaptability. By monitoring the online resource usage such as CPU utilization and power consumption, the framework can orchestrate diverse DPM and DVFS policies according to workload characteristics. Real implementation based experiments using three mobile devices have shown that it can reduce the power consumption ranging from 22% to 79%, while affecting negligibly the performance of workloads.

Enhancing DPM Techniques in Outdoor Industrial WSN Applications

The main challenge during designing a new node is reducing the power consumption as much as possible to maximize the lifetime of wireless sensor network (WSN) since most nodes are powered from a finite source of energy, in general nonrechargeable battery. In this paper we propose a solution to minimize the power consumed by the microcontroller (MCU), which is the main processing component in the node. Our energy power optimization solution is based on two techniques: firstly enhance dynamic power management (DPM) policies by reducing MCU consumption during standby mode; secondly, focus on temperature impact on low power consumption. In this context, a microcontroller dynamic power management (MDPM) algorithm is proposed to improve DPM scheme. This algorithm is deployed on a measurement circuit able to calculate the consumption during the different low power modes in real environments conditions and then selects the better one. These two techniques are combined in a novel way to provide an efficient energy solution for wireless sensor networks nodes. Our solution is validated and qualified with STM32F446RE MCU.

Interaction-based low power service for mobile OLED displays

It is becoming a primary issue to extend the battery lifetime of consumer electronic devices. Dynamic power management (DPM) strategies selectively place IDLE components into low-power states, which aims to gain the trade-off between power saving and system performance. The end user interaction plays a vital role in the energy consumption of mobile devices. In this paper, an interaction-based low-power service by deploying an online adaptive DPM policy on an embedded operating system is proposed. This low-power service applies singular value decomposition (SVD) algorithm to estimate the frequency of user interaction with mobile OLED displays and can be used to adjust power consumption of OLEDs, following the sampled using intensity of user. The experimental results show that the proposed SVD algorithm achieves more power savings than existing frequency estimation algorithms while gaining less empirical error on predicting the values of interaction frequency.

A review on stochastic approach for dynamic power management in wireless sensor networks

AbstractWireless sensor networks (WSNs) demand low power and energy efficient hardware and software. Dynamic Power Management (DPM) technique reduces the maximum possible active states of a wireless sensor node by controlling the switching of the low power manageable components in power down or off states. During DPM, it is also required that the deadline of task execution and performance are not compromised. It is seen that operational level change can improve the energy efficiency of a system drastically (up to 90%). Hence, DPM policies have drawn considerable attention. This review paper classifies different dynamic power management techniques and focuses on stochastic modeling scheme which dynamically manage wireless sensor node operations in order to minimize its power consumption. This survey paper is expected to trigger ideas for future research projects in power aware wireless sensor network arenas.

Dynamic power management for embedded processors in system‐on‐chip designs

Dynamic power management (DPM), which exploits low-power states of the target device, has been a key research issue to overcome the limited battery life of mobile devices. For efficient power management, today's power management unit in a system-on-chip for mobile devices supports multiple low-power states for embedded processors. Unfortunately, the DPM policies implemented in modern operating systems are not appropriate for processors because they may not understand the idleness of the processor accurately. There may be significant performance degradation if the DPM policy module misunderstands that the processor is idle even when there are many interrupt requests to handle. A novel DPM scheme for embedded processors considering the system response time as well as power reduction is proposed. Experimental results show that the proposed DPM policy achieves performance improvement by up to 25% compared to a conventional DPM policy with a similar amount of power reduction.

Online learning of timeout policies for dynamic power management

Dynamic power management (DPM) refers to strategies which selectively change the operational states of a device during runtime to reduce the power consumption based on the past usage pattern, the current workload, and the given performance constraint. The power management problem becomes more challenging when the workload exhibits nonstationary behavior which may degrade the performance of any single or static DPM policy. This article presents a reinforcement learning (RL)-based DPM technique for optimal selection of timeout values in the different device states. Each timeout period determines how long the device will remain in a particular state before the transition decision is taken. The timeout selection is based on workload estimates derived from a Multilayer Artificial Neural Network (ML-ANN) and an objective function given by weighted performance and power parameters. Our DPM approach is further able to adapt the power-performance weights online to meet user-specified power and performance constraints, respectively. We have completely implemented our DPM algorithm on our embedded traffic surveillance platform and performed long-term experiments using real traffic data to demonstrate the effectiveness of the DPM. Our results show that the proposed learning algorithm not only adequately explores the power-performance trade-off with nonstationary workload but can also successfully perform online adjustment of the trade-off parameter in order to meet the user-specified constraint.

Stochastic modeling of dynamic power management policies in server farms with setup times and server failures

SUMMARYServer farms generally consume an enormous amount of energy, which not only increases the running cost but also simultaneously enhances their greenhouse gas emissions. One way to improve the energy efficiency in server farms is dynamical powering on/off servers. However, this method suffers from many setup time to turn the servers back on, which would have a negative impact on a job's response time and waste yet additional energy. This situation is further exacerbated by the unreliability of the servers, which has become the norm in today's server farms. In this paper, we investigate the impact of dynamical powering on/off servers on energy and performance in a typical server farm environment. Prior work has analyzed similar models for a single server, but no analytical results are known for multiservers. We mainly use the matrix geometric method to analyze this model, and system performance measures are explicitly developed in terms of computable forms. An energy‐performance trade‐off model is derived to determine the optimal management policy for the server farms. Finally, we discuss some extensions of our model to show its robustness and to point out avenues for future research. Numerical examples are provided at several points throughout the paper to illustrate the correctness of our analysis results and to validate the optimization approach. Copyright © 2014 John Wiley & Sons, Ltd.

Adaptive workload driven dynamic power management for high performance computing clusters

With the scale expansion of high performance computer systems, efficient power management has developed into an important issue. To strive to balance power consumption and performance, this paper proposes an adaptive workload-driven dynamic power management policy for homogeneous clusters, which dynamically adjusts the power mode of computing nodes according to workload variation. The proposed policy combines the pre-wakeup method and the feedback mechanism to reduce performance degradation due to the wakeup delay. The experimental results demonstrate that, as compared with two existing timeout policies, adaptive workload-driven dynamic power management effectively reduced the performance loss with a slight increase in power consumption.

Thermal-aware Integrated Dynamic Power Management Policy for Clusters

Thermal-aware Integrated Dynamic Power Management Policy for Clusters

Joint Physical-Layer and System-Level Power Management for Delay-Sensitive Wireless Communications

We consider the problem of energy-efficient point-to-point transmission of delay-sensitive data (e.g., multimedia data) over a fading channel. Existing research on this topic utilizes either physical-layer centric solutions, namely power-control and adaptive modulation and coding (AMC), or system-level solutions based on dynamic power management (DPM); however, there is currently no rigorous and unified framework for simultaneously utilizing both physical-layer centric and system-level techniques to achieve the minimum possible energy consumption, under delay constraints, in the presence of stochastic and a priori unknown traffic and channel conditions. In this paper, we propose such a framework. We formulate the stochastic optimization problem as a Markov decision process (MDP) and solve it online using reinforcement learning (RL). The advantages of the proposed online method are that 1) it does not require a priori knowledge of the traffic arrival and channel statistics to determine the jointly optimal power-control, AMC, and DPM policies; 2) it exploits partial information about the system so that less information needs to be learned than when using conventional reinforcement learning algorithms; and 3) it obviates the need for action exploration, which severely limits the adaptation speed and runtime performance of conventional reinforcement learning algorithms. Our results show that the proposed learning algorithms can converge up to two orders of magnitude faster than a state-of-the-art learning algorithm for physical layer power-control and up to three orders of magnitude faster than conventional reinforcement learning algorithms.

Optimization of timeout-based power management policies for network interfaces

Dynamic power management policies are essential for battery-powered consumer electronics devices, which are mostly equipped with power consuming network interfaces. Timeout-based power management policies are implemented on most operating systems, but the selection of the policy parameters generally resort to heuristic approaches and there is lack of support for network interfaces dynamic power management. Therefore, this paper introduces an innovative solution for timeout-based power management policies for network interfaces. The presented policy exploits statistical data analysis and the optimization of both power savings and performance. The effectiveness of the introduced policy is demonstrated by experimental results1.

Dynamic Power Management for the Iterative Decoding of Turbo Codes

Turbo codes are presently ubiquitous in the context of mobile wireless communications among other application domains. A decoder for such codes is typically the most power intensive component in the baseband processing chain of a wireless receiver. The iterative nature of these decoders represents a dynamic workload. This brief presents a dynamic power management policy for these decoders. An algorithm is proposed to tune a power manageable decoder according to a prediction of the workload involved within the decoding task. By reclaiming the timing slack left when operating the decoder at a high power mode, the proposed algorithm continuously looks for opportunities to switch to a lower power mode that guarantees the task completion. We apply this technique to an long term evolution Turbo decoder and explore the feasibility of a VLSI implementation on a CMOS technology of 65 nm. Energy savings of up to 54% were achieved with a relatively low loss in error-correction performance.

Networks‐on‐Chip: Architectures, Design Methodologies, and Case Studies

As the density of VLSI design increases, more processors or cores can be placed on a single chip. Therefore, the design of a Multi-Processor System-on-Chip (MP-SoC) architecture, which demands high throughput, low latency, and reliable global communication services, cannot be done by just using current bus-based on-chip communication infrastructures. Networks-on-Chip (NoC) has been proposed in recent years as a promising solution of on-chip interconnection network to provide better scalability, performance, and modularity for current and future MP-SoC architectures. The paper entitled “Networks on chips: structure and design methodologies” introduces several NoC architectures and discusses the design issues of communication performance, power consumption, signal integrity, and system scalability in an NoC. Then, a novel Bidirectional NoC (BiNoC) architecture with a dynamically self-reconfigurable bidirectional channel is presented, which can break the performance bottleneck caused by bandwidth restriction in conventional NoCs. Since buffers in on-chip networks constitute a significant proportion of the power consumption and the area of interconnects, reducing the buffer size is an important problem. The paper entitled “A buffer sizing algorithm for network on chips with multiple voltage-frequency islands” describes a two-phase algorithm to size the switch buffers in NoC in considering the support of multiple-frequency islands. The paper entitled “Self-calibrated energy-efficient and reliable channels for on-chip interconnection networks” depicts the design of an energy-efficient and reliable channel for on-chip interconnection networks (OCINs) using a self-calibrated voltage scaling technique with self-corrected green (SCG) coding scheme. Among the NoC components, links that connect the NoC routers are the most power-hungry components. The paper entitled “Intelligent on/off dynamic linkmanagement for onchip networks” presents an intelligent dynamic power management policy for NoCs with improved predictive abilities based on supervised online learning of the system status, where links are turned off and on via the use of a small and scalable neural network. Monitoring and diagnostic systems are required in modern NoC implementations to assure high performance and reliability. In the paper entitled “Status data and communication aspects in dynamically clustered network-onchip monitoring,” the design of a dynamically clustered NoC monitoring structure for traffic and fault monitoring is illustrated. Since biological organisms have better adaptability than computer systems in dealing with environmental changes or noise. A case study on the design of an evolvable neuromolecular hardware motivated from some biological evidence, which integrates interand intra-neuronal information processing, is depicted in the paper entitled “A hardware design of neuromolecular network with enhanced resolvability: a bio-inspired approach.”

Intelligent On/Off Dynamic Link Management for On‐Chip Networks

Networks‐on‐chips (NoCs) provide scalable on‐chip communication and are expected to be the dominant interconnection architectures in multicore and manycore systems. Power consumption, however, is a major limitation in NoCs today, and researchers have been constantly working on reducing both dynamic and static power. Among the NoC components, links that connect the NoC routers are the most power‐hungry components. Several attempts have been made to reduce the link power consumption at both the circuit level and the system level. Most past research efforts have proposed selective on/off link state switching based on system‐level information based on link utilization levels. Most of these proposed algorithms focus on a pessimistic and simple static threshold mechanism which determines whether or not a link should be turned on/off. This paper presents an intelligent dynamic power management policy for NoCs with improved predictive abilities based on supervised online learning of the system status (i.e., expected future utilization link levels), where links are turned off and on via the use of a small and scalable neural network. Simulation results with various synthetic traffic models over various network topologies show that the proposed work can reach up to 13% power savings when compared to a trivial threshold computation, at very low (<4%) hardware overheads.

Fast Reinforcement Learning for Energy-Efficient Wireless Communication

We consider the problem of energy-efficient point-to-point transmission of delay-sensitive data (e.g. multimedia data) over a fading channel. Existing research on this topic utilizes either physical-layer centric solutions, namely power-control and adaptive modulation and coding (AMC), or system-level solutions based on dynamic power management (DPM); however, there is currently no rigorous and unified framework for simultaneously utilizing both physical-layer centric and system-level techniques to achieve the minimum possible energy consumption, under delay constraints, in the presence of stochastic and a priori unknown traffic and channel conditions. In this report, we propose such a framework. We formulate the stochastic optimization problem as a Markov decision process (MDP) and solve it online using reinforcement learning. The advantages of the proposed online method are that (i) it does not require a priori knowledge of the traffic arrival and channel statistics to determine the jointly optimal power-control, AMC, and DPM policies; (ii) it exploits partial information about the system so that less information needs to be learned than when using conventional reinforcement learning algorithms; and (iii) it obviates the need for action exploration, which severely limits the adaptation speed and run-time performance of conventional reinforcement learning algorithms. Our results show that the proposed learning algorithms can converge up to two orders of magnitude faster than a state-of-the-art learning algorithm for physical layer power-control and up to three orders of magnitude faster than conventional reinforcement learning algorithms.