DVLA: Dynamic VM Lifetime Aware Scheduling for Drifting Lifetime Distributions and Long-Lived VM Placement Debt (Operational Systems)

Efficient Virtual Machine (VM) scheduling is critical for maximizing resource utilization in cloud computing. However, state-of-the-art lifetime-aware schedulers face two critical issues in real-world deployments. First, their static policies are brittle against the significant spatial and temporal drifts of VM lifetime distributions. Second, and more insidiously, their placement strategies inadvertently scatter long-lived VMs, creating a persistent long-lived VM placement debt. This debt, compounded by inevitable prediction errors, pins down machines and cripples cluster-wide resource reclamation, and cannot be repaid by online scheduling alone. To address these challenges, we present Dynamic VM Lifetime Aware scheduling (DVLA), an end-to-end system that synergistically combines online scheduling with offline rectification. DVLA comprises four key components: (1) a Hierarchical Lifetime Prediction Model that delivers multi-horizon predictions to inform both initial placement and offline optimization; (2) a Dynamic Affinity Grouping strategy that adapts to workload distribution drifts in real time; (3) a Debt-Aware Placement Policy (DAPP) that proactively consolidates long-lived VMs to minimize debt creation at the source; and (4) a Placement Debt Rectification Engine (PDRE) that employs strategic live migrations to amortize accumulated debt offline. Extensive trace-driven simulations and a large-scale production deployment at Alibaba Cloud demonstrate that DVLA consistently outperforms state-of-the-art methods, achieving an additional 0.6 percentage points in packing density. This translates to saving thousands of machines in production, delivering substantial cost reductions.