LLM-Guided Energy-Aware A/B Testing for Consolidation and DVFS Policies via Power-Sensitivity Clustering

Abstract

This paper presents an offline A/B testing framework for systematic policy optimization in virtualized data centers. While energy management typically relies on VM consolidation and dynamic voltage/frequency scaling, practical deployment remains heuristic—governed by thresholds whose effectiveness varies with workload dynamics. The framework clusters tenants by power sensitivity (variance, burstiness, ramp rate, diurnal intensity), generates policy candidates as natural language specifications, and evaluates them via offline simulation—estimating energy consumption, SLA risk, migration overhead, and tail-latency proxy.Experiments on Google Cluster traces (32 tenants across two workload suites) show that combined consolidation and DVFS achieves maximum energy savings (75.33%) at the cost of marginal unmet demand (0.000268%) and higher tail latency (4.93×). Cluster-aware policies offer superior trade-offs: policy LLM-P03 saves 71.86% energy with zero unmet demand, 22.5 migrations, and reduced p99 latency (3.29×). Pareto analysis identifies non-dominated policies across all metrics, with statistically significant energy reductions (p < 1e−20).