Online QoS Modeling in the Cloud: A Hybrid and Adaptive Multi-Learners Approach

Given the on-demand nature of cloud computing, managing cloud-based services requires accurate modeling for the correlation between their Quality of Service (QoS) and cloud configurations/resources. The resulted models need to cope with the dynamic fluctuation of QoS sensitivity and interference. However, existing QoS modeling in the cloud are limited in terms of both accuracy and applicability due to their static and semi- dynamic nature. In this paper, we present a fully dynamic multi- learners approach for automated and online QoS modeling in the cloud. We contribute to a hybrid learners solution, which improves accuracy while keeping model complexity adequate. To determine the inputs of QoS model at runtime, we partition the inputs space into two sub-spaces, each of which applies different symmetric uncertainty based selection techniques, and we then combine the sub-spaces results. The learners are also adaptive; they simultaneously allow several machine learning algorithms to model QoS function and dynamically select the best model for prediction on the fly. We experimentally evaluate our models using RUBiS benchmark and realistic FIFA 98 workload. The results show that our multi-learners approach is more accurate and effective in contrast to the other state-of-the-art approaches.

💡 Research Summary

The paper addresses the challenge of accurately modeling the relationship between Quality of Service (QoS) metrics and cloud configuration or resource parameters in highly dynamic cloud environments. Existing approaches are either static—trained once offline and never updated—or semi‑dynamic, requiring periodic retraining. Both suffer from reduced accuracy when workloads fluctuate rapidly, when services interfere with each other, or when resource availability changes. To overcome these limitations, the authors propose a fully dynamic, multi‑learner framework that operates online and adapts both its input selection and its predictive model in real time.

The first innovation is a two‑stage input‑space partitioning. By analyzing the sensitivity of QoS to each potential input (e.g., CPU cores, memory size, network bandwidth, request rate), the method separates the full feature space into a “high‑sensitivity” sub‑space and a “low‑sensitivity” sub‑space. Within each sub‑space, the authors apply a symmetric uncertainty (SU) based feature‑selection technique. SU quantifies the mutual information between a candidate feature and the QoS target, normalized to account for differing entropies, thereby automatically retaining only those variables that contribute most to prediction accuracy. This partitioning reduces dimensionality, lowers model complexity, and speeds up both training and inference.

The second innovation is a hybrid ensemble of heterogeneous learners. The framework simultaneously trains several machine‑learning algorithms—linear regression, regression trees, support vector regression, and feed‑forward neural networks—on the same data set. During prediction, a lightweight monitoring component continuously evaluates each learner’s recent performance using metrics such as mean squared error, absolute error, and prediction latency. A weighted‑ranking selector then chooses the best‑performing model on the fly. This adaptive selection allows the system to switch instantly when the workload pattern changes, ensuring that the most suitable bias‑variance trade‑off is always employed.

The authors validate their approach using two realistic benchmarks. The first is RUBiS, an e‑commerce web‑application benchmark, where they vary request rates, database load, and cache hit ratios, measuring twelve QoS indicators including response time, throughput, CPU and memory utilization. The second benchmark reproduces the FIFA 98 online gaming workload, characterized by abrupt spikes in concurrent players and computational intensity. Experiments show that the multi‑learner approach reduces average prediction error by more than 15 % compared with the best single‑learner baseline, and by up to 30 % during sudden load bursts. The SU‑driven feature selection cuts the number of input dimensions by roughly 30 %, leading to 25 % faster training and 20 % lower memory consumption, while only modestly increasing model complexity (≈10 %).

Key contributions of the paper are: (1) a novel sensitivity‑driven input partitioning combined with symmetric‑uncertainty feature selection, (2) an adaptive hybrid ensemble that dynamically selects the most accurate learner in real time, and (3) extensive empirical evidence that the proposed system outperforms state‑of‑the‑art static and semi‑dynamic QoS modeling techniques on representative cloud workloads. The results demonstrate that cloud operators can achieve more reliable SLA compliance and cost‑effective resource provisioning by deploying this online, adaptive QoS modeling framework.