系统级低功耗设计 硬件设计 - 处理器的选择 一存储器配置 计算(CPU) 存储（存储器） 一接口驱动电路设计 一电源供给电路选择 通信 软件设计 0 电源 接口 -算法选择 一代码优化 外部器件1 外部器件2 (音频) (视频) 功耗管理技术 -DPM DVFS llxx@ustc.edu.cn 26/62
系统级低功耗设计 • 硬件设计 – 处理器的选择 – 存储器配置 – 接口驱动电路设计 – 电源供给电路选择 • 软件设计 llxx@ustc.edu.cn 26/62 – 算法选择 – 代码优化 • 功耗管理技术 – DPM – DVFS
DPM USTC Run Deep Idle SAMSUNG S3C2410X (32bit ARM 920T内核)提供四种工作模式 Idle Stand Sleep By ■Low idling state 模式 运 空闲 休眠 关机 High idling state 功耗(） 297u 122w 33知w 80w Deep 时钟频率， Sleep 20AH2 202 1配 32767K2 备注湿 全速运行 CPU core时钟特止 系统时种振级率 仅有TC Applications Lucent's Wavelan Rambus DRAM Flash RAM: PCMCIA Card (by IDF:Mobile AMD/Fujitsu Rambus Spec Policy Manager Am42BDS6408G oS Scheduler Mode Power Mode Power Mode Current Transmit 1.82W Active 300mW Simultaneous 25 mA Performance /ldle Profiler Operating Operation System Receive 1.80W StandBy 180mW Program/Erase 15 mA StandBy 0.18W Nap 30 mW Burst Read 10 mA HW Abstraction Layer Drivers PowerDown 3 mW StandBy 0.2uA PMU HW with Dynamic Voltage and Frequency Management
DPM • SAMSUNG S3C2410X（32bit ARM 920T内核）提供四种工作模式
Categorization of DPM Techniques ·Timeout policy -Shutting down unused devices by a static timeout threshold decides when to shut down unused devices Predictive technique (heuristic) According to the workload history,predicting the length of the next idle time use a regression function OR an exponential-weighting average function perform well only when the requests are highly correlated and do not take performance constraints into account Stochastic process According to a random process model,modeling the request arrival times and device service times stationary stochastic processes such as Markov Decision Processes(MDP) take into account both power and performance need of exact knowledge of the state transition probability function of the MDP machine learning for adaptive policy optimization when the system model is not known a priori interact with the environment to obtain information (workload)which can be processed to produce optimal policies
Categorization of DPM Techniques • Timeout policy – Shutting down unused devices by a static timeout threshold • decides when to shut down unused devices – Predictive technique（heuristic） • According to the workload history, predicting the length of the next idle time – use a regression function OR an exponential-weighting average function • perform well only when the requests are highly correlated and do not take performance constraints into account – Stochastic process • According to a random process model, modeling the request arrival times and device service times – stationary stochastic processes such as Markov Decision Processes (MDP) • take into account both power and performance • need of exact knowledge of the state transition probability function of the MDP – machine learning for adaptive policy optimization • when the system model is not known a priori • interact with the environment to obtain information（workload）which can be processed to produce optimal policies
三类优化策略比较 Strategy Description Features &Efficiency Applicability Workload Features Timeout When the idle period exceeds a fixed Safety can be improved by increasing Various kinds Does not consider Strategy threshold,the component is switched the timeout threshold,but large of components workload features to low power state. threshold may cause energy and performance penalty. Predictive Predicts the current idle period based Depends on workload features;more Interactive Efficient when Strategy on the recent idle and active periods.If efficient than the timeout approach, devices,such as workload has the preset condition is satisfied,the but less safe. keyboard,touch predictive timing component is switched to low power Normally based on heuristics. screen,and patterns state as soon as it becomes idle. mouse Stochastic Assumes stochastic models for Depends on the stochastic model of Hard disk Markov model for Optimal component state transition,workload, the components and workloads. components and Strategy and cost matrix,searches for the The optimal policies are workload is optimal states through learning. approximate solutions required for policy optimization
三类优化策略比较
System-Level Dynamic DVFS Intel's Enhanced SpeedStep (EIST) Intel Pentium M Processor:6 frequency- voltage pair 1.6 245 GHz 1.4 GHz 1.2 Power [W] 1.0 GHz 600 800M GHz MHz Hz 6 0.956 1.036 1.164 1.276 1.4201.484 Core Voltage [V]
System-Level Dynamic DVFS • Intel’s Enhanced SpeedStep (EIST) – Intel Pentium M Processor: 6 frequencyvoltage pair