Syed Rizwanul Abedin, MD Rakibul Islam, Md Wahid Ahmed, Umaiya Ahmed Mishu
Department of Computer Science and Engineering, United International University, Bangladesh
Abstract
The cloud data center uses a lot of energy because of the inefficiency of the process of VM migration. In most cases, the migration is carried out using static rules and heuristic methods that cannot effectively adapt to changes in load conditions. As a consequence, the methods described above cause unnecessary consumption of energy, unnecessary VM migrations and SLA violations. The goal of the research is to create a new scheduler of VM migration based on DRL algorithm that would minimize energy usage but at the same time provide good cloud service performance. The proposed approach would use a DQN agent that learns continuously from the cloud environment without the need to use any static rules. This allows it to make appropriate migration decisions and reduce energy consumption. The developed method will be tested in a simulated cloud environment and compared to other migration strategies in terms of their energy efficiency and performance.
V. Formal MDP Specification
- • Horizon & Step: 24-hour episode (288 steps at Δt = 300s), discount factor γ = 0.99.
- • State Vector (312 features): 12 global data-center metrics + top K = 20 candidate migrations (× 14 features) + 20-element feasibility mask.
- • Action Space: Discrete {0, 1, &dots;, 20} where action 0 = explicit No-Migration.
- • Multi-Objective Reward:
r_t = -(0.40 E_t + 0.35 S_t + 0.15 M_t + 0.10 O_t)
VI. CloudSim Plus 8.5.7 Protocol
- • Infrastructure: 800 heterogeneous hosts (400 HP ProLiant ML110 G4 + 400 ML110 G5).
- • Workload: 10 PlanetLab 2011 trace days (6 train, 2 validation, 2 held-out test days).
- • Safety Masking: CPU ≤ 85%, RAM ≤ 90%, BW ≤ 90%, plus 3-interval (15 min) VM cooldown.
- • Baselines: NDM, THR-MMT-PABFD, LR-MMT-PABFD, and Externally Triggered DQN.
VII & VIII. Preliminary Findings & Conclusion
The baseline DQN pipeline is operational as an end-to-end learning and decision scheduler in the simulated environment, validating state normalization, capacity action masking, and experience replay. Training of the full multi-objective reward-aware DQN policy is ongoing to evaluate energy savings, SLA violation reduction, and unnecessary migration avoidance across 30 paired seeds.