Predictive Multi-Objective Energy--Latency Optimization in Cloud Environments Using Ensemble Learning

Technical Summary: Predictive Multi-Objective Energy–Latency Optimization in Cloud Environments Using Ensemble Learning

1. Problem Statement

Cloud data centers face a fundamental, nonlinear trade-off between energy efficiency and communication latency. Operators must deliver low-latency services within strict energy budgets, yet these objectives are coupled: reducing energy typically requires lowering server speeds or consolidating workloads, which increases service times and risks Quality-of-Service (QoS) violations. Conversely, preserving latency headroom often leaves servers idle at high power consumption.

The core challenge lies in the nonlinearity of the system: power consumption grows super-linearly with clock frequency (following Dynamic Voltage and Frequency Scaling, or DVFS, models), while queueing response times diverge asymptotically as utilization approaches capacity. Static, single-objective policies or heuristic rules that ignore this coupling or treat objectives independently fail to operate safely near the efficiency frontier. Existing machine learning approaches often focus on forecasting future workloads rather than learning the static functional map between current resource states and performance objectives, or they rely on evaluation protocols that lack robustness (e.g., single-split validation without cross-validation or sensitivity analysis).

2. Methodology

The authors propose a predictive multi-objective framework that learns surrogate models of server energy and queueing latency from production traces to select energy-aware operating points.

System Model and Formulation:

• Workload: The system models a fleet of $N$ servers hosting Virtual Machines (VMs) with time-slotted telemetry (300-second resolution). The state includes CPU utilization, memory ratio, disk, and network throughput.
• Energy Model: A physically grounded model based on cubic DVFS scaling: $P(s, u) = s^3(P_{idle} + (P_{peak} - P_{idle})\min(1, u/s))$, where $s$ is the operating speed and $u$ is the load.
• Latency Model: An M/M/1-style queueing relation where response time $L(s, u) = T_s / (s - u)$ for stable regions ($s > u$), diverging as speed approaches load.
• Optimization: The goal is to minimize the vector $F = (E, \bar{L})$ subject to stability constraints ($s \ge (1+\epsilon)u$) and Service Level Agreement (SLA) constraints ($L \le L_{SLA}$).

Predictive Surrogates:

• Learning Task: Instead of forecasting future telemetry, the system treats the problem as system identification: learning the mapping from the current resource state and candidate speed to the energy and latency objectives.
• Algorithms: Five supervised regressors are compared: Linear Regression (LR), Decision Tree (DT), Random Forest (RF), Gradient Boosting (GBT), and AdaBoost (AB).
• Validation Protocol: To ensure reliability, the authors employ 5-fold cross-validation at the interval level (preventing data leakage between speed variants of the same interval) and report inter-fold variance. They also perform sensitivity analysis on model constants and a leave-one-VM-out generalization study.

Control Framework:

• The trained surrogates (specifically Gradient Boosting, which showed the best performance) are embedded in a weighted scalarized multi-objective controller.
• For each interval, the controller normalizes predicted energy and latency, then selects the speed $s^*$ that minimizes $\alpha \hat{E}_n + (1-\alpha) \hat{L}_n$, where $\alpha$ is a tunable weight.
• The selected operating points are evaluated against the true physical maps (not the surrogates) to ensure the reported results reflect actual system behavior.

3. Key Contributions

1. Physically Grounded Modeling: The design integrates a DVFS-based power model and an M/M/1 latency model, calibrated against public Bitbrains traces, explicitly characterizing the oversubscription regime of a production segment.
2. Surrogate Validation: The study compares five regressors using consistent metrics (MAE, RMSE, $R^2$) and quantifies reliability through 5-fold cross-validation, demonstrating that ensemble methods significantly outperform linear models on the nonlinear latency task.
3. Robust Optimization Framework: The authors embed ensemble surrogates in a controller with explicit stability and SLA constraints, evaluating performance against multiple baselines (peak speed, reactive scaling, energy-only, random feasible speed, and an analytic oracle).
4. Rigorous Evaluation: The paper includes a one-at-a-time sensitivity analysis over service demand, telemetry noise, and power parameters, as well as a leave-one-VM-out generalization study to confirm that the learned maps transfer across workloads.

4. Experimental Results

The framework was evaluated on public Bitbrains traces comprising seven VMs over six days (7,285 intervals). Notably, 60.3% of intervals were oversubscribed ($u \ge 0.95$), necessitating evaluation within the stable envelope ($u < 0.95$).

• Surrogate Accuracy: Linear regression failed to capture the nonlinear latency dynamics ($R^2 = 0.565$). In contrast, ensemble methods (Random Forest and Gradient Boosting) achieved near-perfect fidelity ($R^2 \ge 0.986$ for latency, $R^2 \ge 0.994$ for energy), reducing RMSE by up to 90.5% compared to linear regression. Cross-validation showed inter-fold deviations below 0.014.
• Controller Performance:
  • Energy Savings: The proposed controller (with balanced weighting $\alpha=0.5$) reduced energy consumption by 72.4% compared to static peak-speed provisioning and 58.9% compared to reactive scaling.
  • SLA Compliance: Crucially, the proposed controller achieved zero SLA violations (latency > 2s) and maintained a mean latency of 0.29 s. In contrast, the "energy-only" policy, while saving slightly more energy, incurred a 3.80% violation rate with a degraded latency tail (99th percentile > 3s).
  • Pareto Frontier: Sweeping $\alpha$ from 0.05 to 0.95 traced a smooth frontier, allowing operators to trade off energy and latency. Even at high energy priority ($\alpha \approx 0.80$), violations remained below 1% with 85.3% energy savings.
• Generalization and Sensitivity:
  • Sensitivity: Under parameter perturbations (service demand, noise, power constants), energy savings remained robust between 64.8% and 73.3%, and the zero-violation property of the proposed controller remained invariant.
  • Generalization: In the leave-one-VM-out study, the controller retained 72.8% pooled energy savings on unseen VM traces, with surrogate $R^2$ values remaining above 0.987.

5. Significance and Claims

The paper claims that ensemble-learned surrogates provide a practical and safe building block for adaptive green-cloud orchestration. The significance of the work lies in:

• Trustworthiness: By validating surrogates with cross-validation and sensitivity analysis, the authors demonstrate that the reported energy-latency frontier is not an artifact of the learning algorithm or a specific data partition.
• Safety vs. Greed: The results highlight that single-objective energy minimization is unsafe in oversubscribed environments, often converting "savings" into QoS violations. The proposed multi-objective approach successfully converts these risky savings into compliant service.
• Feasibility: The framework achieves near-oracle performance (within a factor of 2.4 of the analytic lower bound) without requiring online model access or complex reinforcement learning, relying instead on deterministic, pre-trained surrogates.

The authors conclude that while the current framework relies on a queueing model for latency (rather than direct measurement) and offline selection, it establishes a rigorous baseline for predictive control that generalizes across workloads and remains robust to modeling uncertainties.