HAIL: Prediction-Calibrated Hierarchical Resource Isolation for Multi-Tenant Cloud Platforms

Multi-tenant cloud platforms must reserve capacity for protected workloads while avoiding needless loss of best-effort service. This paper presents HAIL, a hierarchical adaptive isolation-and-lending controller operating across CPU, memory, network, and storage. It maps a three-layer isolation taxonomy to physical-capacity constraints, host-local quota control, and application-level demand coverage. Four gradient-boosted predictors forecast next-interval demand; independent validation residuals supply one-sided protection margins, and bounded online error feedback increases reservations for persistently underpredicted protected tenants. Unallocated priority capacity is made available to batch tenants through a deterministic feasible projection. We conduct reproducible experiments with synthetic but temporally correlated and bursty tenant traces, twelve disjoint test seeds, five comparison controllers, an oracle upper bound, and targeted ablations. With causal historical profiles, HAIL reduces modeled protected shortfalls from 24.11% for a hierarchical profiling-and-lending baseline to 8.86% (paired 95% interval: -19.47 to -11.04 percentage points); utilization decreases from 87.15% to 86.44%. Stress tests reveal that strict priority can starve best-effort demand under severe scarcity. These are simulated resource-coverage results, not measured request latencies. The released experiment package preserves code, generated demands, seed splits, and per-episode results.