Self-Aware resource management in embedded systems

Abstract

Resource management for modern embedded systems is challenging in the presence of dynamic workloads, limited energy and power budgets, and application and user requirements. These diverse and dynamic requirements often result in conflicting objectives that need to be handled by intelligent and self-aware resource management. State-of-the-art resource management approaches leverage offline and online machine learning techniques for handling such complexity. However, these approaches focus on fixed objectives, limiting their adaptability to dynamically evolving requirements at run-time.

In this dissertation, we first propose resource management approaches with fixed objectives for handling concurrent dynamic workload scenarios, mixed-sensitivity workloads, and user requirements and battery constraints. Then, we propose comprehensive self-aware resource management for handling multiple dynamic objectives at run-time. The proposed resource management approaches in this dissertation use machine learning techniques for offline modeling and online controlling. In each resource management approach, we consider a dynamic set of requirements that had not been considered in the state-of-the-art approaches and improve the selfawareness of resource management by learning applications characteristics, users’ habits, and battery patterns. We characterize the applications by offline data collection for handling the conflicting requirements of multiple concurrent applications. Further, we consider user’s activities and battery patterns for user and battery-aware resource management. Finally, we propose a comprehensive resource management approach which considers dynamic variation in embedded systems and formulate a goal for resource management based on that.

The approaches presented in this dissertation focus on dynamic variation in the embedded systems and responding to the variation efficiently. The approaches consider minimizing energy consumption, satisfying performance requirements of the applications, respecting power constraints, satisfying user requirements, and maximizing battery cycle life. Each resource management approach is evaluated and compared against the relevant state-of-the-art resource management frameworks.