Reinforcement Learning Framework for Optimizing Trade-offs Between Energy-Aware and Makespan Workload Scheduling in Cloud Computing

Abstract

A new reinforcement learning (RL) framework is proposed to optimize the energy efficiency and makepan trade-offs in workload scheduling for cloud computing environments. With the increasing demand of sustainable computing, balancing energy consumption against performance metrics such as makespan poses a unique challenge. The proposed RL model applies and uses a dynamic scheduling mechanism, in which the learning agent updates its cloud resource allocation based on real-time feedback from both energy consumption and task execution time. This framework unifies both objectives as part of a single reward framework to accommodate adaptive scheduling decision-making, thereby minimizing energy consumption while remaining within a threshold on makespan. We conduct experiments on a range of workloads with varying resource demands and task durations to evaluate the model's performance. The results show that the RL-based framework is more efficient in terms of energy consumption compared to much of the traditional scheduling approaches, while not significantly sacrificing the time in which tasks are completed. The framework's ability to adapt to different workload patterns demonstrates good scalability and robustness in its design. This study provides insight into applying RL to energy-aware cloud computing systems, highlighting the ongoing practical concerns in the cloud, including the conflicting goals of energy savings and operational efficiency.