A software framework for application-guided power-aware control systems

Power-aware computing;Operating systems;Power-aware control systems;Heterogeneous memory;Heterogeneous computing;Bumpless transfer;Transient management;Machine learning;Q-learning;Reinforcement learning;Dynamic power management;Linux

Giardino, Michael Joseph ; Ferri, Bonnie Electrical and Computer Engineering Chatterjee, Abhijit Yalamanchili, Sudhakar Ferri, Aldo Wills, Linda ; Ferri, Bonnie

University:Georgia Institute of Technology

Department:Electrical and Computer Engineering

关键词: Power-aware computing;    Operating systems;    Power-aware control systems;    Heterogeneous memory;    Heterogeneous computing;    Bumpless transfer;    Transient management;    Machine learning;    Q-learning;    Reinforcement learning;    Dynamic power management;    Linux;   

Others  :  https://smartech.gatech.edu/bitstream/1853/61212/1/GIARDINO-DISSERTATION-2019.pdf

美国|英语

来源: SMARTech Repository

【 摘 要 】

This dissertation describes a system for proactive management of power and performance trade-offs through greater cooperation between applications and hardware. To enable such a management system, a software framework for application-guided power-aware control systems was developed. This system allows an application to guide the underlying computing hardware through a reusable and modular software abstraction. This abstraction layer enables an application to avoid hardware-specific details while still requesting resources from the computing hardware using a generic quality-of-service (QoS) interface. The computing system, in turn, monitors its current power and performance state and notifies the application to adjust its computational load by changing its algorithms. This two-way communication between application and computing platform allows both application and system designers to create proactive strategies for managing power and performance states. The research begins by examining mechanisms for system state estimation, prediction and management for use by a power- and performance-aware system. To manage switched systems it introduces speculative threads for transient management and examines their effectiveness in digital filters. Two methods for power and performance management are tested: a situational-aware governor and a Q-Learner-based quality-of-service manager (2QoSM). The implementation of the software framework was tested using an autonomous robot. The framework and QoSM allow for significant power savings with minimal performance penalty as well as the flexibility to explore different computing platforms and machine learning techniques in the future.