Journal of Quantum Computing, Simulation and Emerging Computation
Parallel Multi-Objective Optimization for Predictive Virtual Machine Placement and Scheduling in Cloud Computing
Abstract
Naby Doumbouya
This paper presents PMO-VMP (Parallel Multi-Objective Virtual Machine Placement and Scheduling), a comprehensive framework addressing the computational challenges of large-scale cloud resource management. Unlike reactive placement schemes that rely only on current or observed load, PMO-VMP is predictive: a Hybrid Predictor integrates ARIMA, LSTM, and Prophet with adaptive per-VM weighting (Eq. 26) to forecast workload before optimization, so that placement and scheduling decisions anticipate future demand. The framework extends beyond traditional bi- objective approaches by supporting up to eight conflicting operational objectives with modular configuration, while incorporating joint VM placement and task scheduling in a unified optimization framework. The base configuration optimizes four core objectives (energy, waste, makespan, robustness), with four additional objectives (migrations, SLA, thermal, cost) available for extended scenarios. The core contribution is a multi-level parallel architecture that combines innovative hybrid algorithms with heterogeneous computing resources. Our hybrid MOEA/D-ACO algorithm integrates decomposition-based evolutionary computation with ant colony optimization, achieving average speedup of 5.33× (95% CI: [5.12, 5.54]) on 8-worker configurations with 66.6% parallel efficiency. Experimental evaluation on four synthetic workload scenarios (stable, transactional, sudden peak, mixed) and four real Google Cluster Traces v3 datasets (500–10,000 VMs), with 5 independent executions per configuration (160 total executions), demonstrates that PMO- VMP consistently finds high-quality Pareto optimal solutions, achieving average hypervolume of 0.826 (95% CI: [0.820, 0.832]) and reducing energy consumption by 19.4% (95% CI: [18.6, 20.2]) while reducing SLA violations by 33.3% (95% CI: [30.1, 36.5]) compared to baseline approaches. The framework demonstrates scalability to 10,000 VMs with maintained Pareto front quality (HV > 0.75 at 10k VMs)—a scale at which few existing works report comparable multi- objective quality—and sub-quadratic empirical complexity O(N1.6–1.8), establishing a viable foundation for predictive multi-objective optimization in production cloud environments.