GAUSSIAN PROCESS REGRESSION MODEL FOR PERFORMANCE PREDICTION OF ENTERPRISE WEB SYSTEMS UNDER CONFIGURABLE WORKLOADS

Abstract

This paper addresses the problem of performance prediction for enterprise web systems under varying hardware and software configuration environments. The challenge of exhaustive configuration space exploration is examined, considering its prohibitive cost for determining optimal deployment parameters. A surrogate modelling method based on Gaussian Process Regression (GPR) is proposed, mapping the function f(x, l) → (μ, σ), where x is a configuration parameter vector (RAM, CPU cores, connection pool size, worker count), l represents workload characteristics (virtual users, read/write ratio), μ is the predicted p95 response latency, and σ is the prediction uncertainty. Unlike additive performance-influence models such as P4, the proposed GPR surrogate approximates a non-linear response surface across the full configuration and workload space, yielding calibrated confidence intervals alongside point predictions. An experimental dataset of 420 samples was generated via Latin Hypercube Sampling. A complete ML pipeline was implemented, comparing three GPR kernels (RBF, Matérn 5/2, Rational Quadratic) and baseline models (XGBoost, Random Forest, additive linear). The Matérn 5/2 kernel achieved best results: R² = 0.9863, MAE = 34.4 ms, and 95% CI coverage = 0.937. Surrogate efficiency analysis determined that only 75 real measurements are required to achieve R² ≥ 0.90, versus 720 in a full grid an 89.6% reduction in measurement cost. SHAP analysis confirmed the dominant role of the engineered "load intensity" feature over primary configuration parameters. An active learning criterion based on GPR predictive variance was implemented for targeted dataset expansion. The application of the GPR surrogate to cybersecurity tasks is substantiated: predictive uncertainty σ(x, l) serves as a statistical detector of anomalous network load (DoS/DDoS attacks), and the optimised system configuration provides a minimal attack surface.
Keywords: Gaussian process regression; surrogate model; performance prediction; enterprise web systems; uncertainty quantification; Latin Hypercube Sampling; SHAP analysis; active learning; anomaly detection; infrastructure cybersecurity.