A Self-Adaptive Deep Reinforcement Learning Framework for Multi-Objective Resource Allocation in Elastic Cloud Computing

Abstract

Effective and adaptive scheduling policies are necessary to deal with changing and dynamic workloads within cloud computing environments. Traditional scheduling policies and the existing Deep Reinforcement Learning methods for dynamic workloads depend mainly on fixed reward mechanisms. This research proposes a framework known as Self-Adaptive Deep Reinforcement Learning for Multi-Objective Resource Allocation (SADRL-MORA). The proposed framework can adjust itself with respect to the changing state of the cloud in terms of adjusting the weight of rewards and learning parameters. The proposed framework applies a self-adaptive deep reinforcement learning approach along with actor-critic architecture for the purpose of resource management in cloud computing environment for the optimization of the following performance measures at the same time; resource utilization, response time, energy consumption, execution cost, and SLA satisfaction. The SADRL-MORA framework is implemented with the use of Cloud Sim Plus along with Google and Alibaba workloads traces in various settings of clouds and compared with the following scheduling approaches; Round Robin (RR), Deep Q-Network (DQN), Double Deep Q-Network (DDQN), and Proximal Policy Optimization (PPO). The experimental results have shown the superiority of the proposed framework in terms of faster convergence, average response time reduction by up to 21.2% with respect to PPO, 78.5 kWh of energy consumption, 82.4% resource utilization, US$41.7 of execution cost per 10⁴ tasks, and SLA violations not exceeding 2.4%.