HiLITE: Hierarchical and Lightweight Imitation Learning for Power Management of Embedded SoCs
HiLITE: Hierarchical and Lightweight Imitation Learning for Power Management of Embedded SoCs
复制标题
HiLITE：用于嵌入式 SoC 电源管理的分层轻量级模仿学习
DOI：
--
复制
发表时间：
2020
影响因子：
2.3
通讯作者：
R. Marculescu
中科院分区：
文献类型：
--
作者：
A. L. Sartor;A. Krishnakumar;Samet E. Arda;U. Ogras;R. Marculescu
现代片上系统(SoCs)使用动态电源管理(DPM)技术来提高能效。然而，现有技术无法在异质平台上有效地调整同时考虑多个目标(例如，能量和实时需求)的运行时决策。为了满足这种需求，我们提出了一种层次化的模拟学习框架HILITE，该框架在满足嵌入式SoC的软实时约束的同时最大化了能量效率。我们的方法首先使用模仿学习来训练DPM策略；然后，它在运行时应用回归策略，以最大限度地减少截止日期未命中。与最先进的方法相比，Hilite平均将能量延迟产品提高了40%，并将截止日期错失率减少了高达76%。此外，我们还证明了训练后的策略不仅具有较高的准确率，而且预测时间开销和内存占用可以忽略不计。
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.