CAREER: Adaptive Power Management for Multiprocessor System-on-Chip

职业：多处理器片上系统（MPSoC）以其低成本、高性能的优势，成为VLSI系统设计的主流平台。然而，MPSoC中的功耗是限制系统性能和功能增长的关键因素。MPSoC硬件和软件的复杂性对系统级电源管理的研究提出了新的挑战和要求，一个有效的电源管理器必须了解硬件、应用和工作环境的状态，并能够适应变化。它应该能够稳健地工作，即使完美的系统信息不可用。随着可以被功率控制的组件的数量增加，以集中方式执行功率管理变得越来越困难。分层分布式的电源管理方法更适合于MPSoC平台。最后，资源管理、电源管理和热管理是相互关联的任务，需要同时优化它们，本研究项目通过研究下一代MPSoC自适应电源管理的理论基础和应用框架来解决上述挑战。本项目包括四个研究部分：（1）研究运行时负载预测和硬件性能/功耗表征的在线建模技术;（2）研究部分可观测系统中自适应资源和功耗管理的新优化技术;（3）将分布式电源管理问题建模为多智能体合作博弈，并利用博弈论开发控制策略;（4）开发统一的标准平台，用于电源管理MPSoC的建模、优化和评估。本项目的教育部分将向学生介绍系统级电源管理的实现和优化技术，并为学生提供MPSoC设计和优化的独特实践经验。

CAREER: Adaptive Power Management for Multiprocessor System-on-ChipMultiprocessor System-on-Chip (MPSoC) is becoming a major VLSI system design platform due to its advantages in low design cost and high performance. However, power consumption in MPSoC is a crucial factor that is limiting the growth of system performance and functionality. The complexity of the MPSoC hardware and software imposes new challenges and requirements for research in system-level power management.An effective power manager must be aware of the status of the hardware, the application and the working environment and be able to adapt to the changes. It should be able to work robustly even if the perfect system information is not available. As the number of components that can be power controlled increases, it is increasingly difficult to perform power management in a centralized manner. A hierarchical and distributed power management method is more suitable for MPSoC platforms. Finally, resource management, power management, and thermal management are inter-correlated tasks and it is desirable for them to be optimized simultaneously.This research project addresses the above mentioned challenges by investigating the theoretical foundation and the applied framework of adaptive power management for the next generation MPSoC. This project consists of four research components: (1) investigate online modeling techniques for runtime workload prediction and hardware performance/power characterization; (2) research new optimization techniques for adaptive resource and power management in a partially observable system; (3) model the distributed power management problem as a multi-agent cooperative game and develop control policy using game theory; (4) develop a unified and standard platform for modeling, optimization and evaluation of power-managed MPSoC.The educational components of this project will introduce the students to the implementation and optimization techniques of system-level power management and provide students unique hands-on experience with MPSoC design and optimization.