Energy-Efficient Workload Scheduling Using Deep Reinforcement Learning in Cloud Computing

Abstract

Cloud data centers must process highly variable workloads while controlling energy demand, service latency, and resource contention. This paper presents an energy-efficient workload scheduling framework based on Deep Reinforcement Learning (DRL), in which scheduling is formulated as a sequential decision problem and solved using a Deep Q-Network (DQN). The scheduler observes task requirements together with virtualmachine utilization, queue length, memory availability, and an estimated server power state. It then selects a target virtual machine or postpones assignment when immediate placement would increase the long-term operating cost. A multi-objective reward penalizes energy consumption, response time, and servicelevel-agreement violations while encouraging high resource utilization. A simulation-oriented implementation is described using heterogeneous hosts, bursty task arrivals, experience replay, a target network, and epsilon-greedy exploration. Synthetic evaluation values, included to demonstrate a complete experimental-paper structure, show the proposed scheduler reducing energy consumption and response time while improving utilization relative to FCFS, Round Robin, and a genetic-algorithm heuristic. The architecture, mathematical formulation, training procedure, performance metrics, and comparative analysis are provided to support reproducibility and future extension to real cloud traces.