EAGER: Fine-Grained Software Power Prediction and Its Application on Power Management of Heterogeneous Multicore Systems

功耗和能耗是现代计算机系统中的主要成本考虑因素，包括数据中心设计、企业级服务器设计、智能电话上的电池寿命管理以及微处理器上的电路布局。有许多方法可以解决硬件方面的问题。SPONGE是一种细粒度的功能级软件功耗预测机制，能够动态地应用适当的功耗节省机制来提高功耗效率，并将合适的工作负载调度到异构核上。PI提出了一个新的度量标准，功能功率效率（PEF），以细粒度评估软件的功率效率。在PEF的基础上，提出了一种软件行为驱动的电源管理方案，以提高系统运行时的电源效率。功率管理方案应用功率节省机制，包括线程映射和电压和频率缩放，由基于目标平台的软件功率预测的程序特性驱动。PEF能够定量衡量所做的有用工作与所利用资源的比率。通过在电源管理的循环中涉及软件，可以提供具有软件和操作系统之间的无缝连接的细粒度电源管理方案，特别是对于异构多核架构上的功率感知调度。

Power dissipation and energy consumption are major cost considerations across modern computer systems including data center design, enterprise level server design, battery life management on a smart phone, and circuit layout on a microprocessor. There are numerous approaches that address the hardware aspect of the problem. However, software-behavior-driven power management is not well addressed and understood.SPONGE, a fine-grained function level software power prediction mechanism, is proposed to dynamically apply proper power saving mechanisms to improve power efficiency and schedule suitable workloads to heterogeneous cores. The PI is proposing a new metric, Power Efficiency of Function (PEF), to evaluate the power efficiency of software at fine granularity. Based on PEF, a software-behavior-driven power management scheme is introduced to improve system power efficiency at runtime. The power management scheme applies power saving mechanisms including thread mapping and voltage and frequency scaling, driven by program characteristics based on the software power prediction of the target platform. PEF enables a quantitative measurement of the ratio of useful work done to resources utilized. By involving software in the loop of power management, it is possible to provide a fine-grained power management scheme with a seamless connection between software and operating systems, especially for power-aware scheduling on heterogeneous multicore architectures.