Analysis of computer network performance models for structural management tasks throughout the life cycle

Abstract

The article addresses the problem of selecting models for calculating computer network characteristics in tasks related to managing network structure throughout the life cycle. It is proposed to distinguish between the system life-cycle stage and its structural and configuration state, which makes it possible to separately describe the network development context and specific structural changes. The study follows a review-and-methodological approach; the literature base was formed through a structured thematic search and analysis of fundamental and recent studies on network performance modeling, analytical approaches, neural networks, graph-based models, and generative methods. Model selection is proposed to be performed with regard to the nature of the structural and configuration change, the availability and structure of data, the required type of estimate or guarantee, and the allowable computational budget. The capabilities of queueing theory models, Network Calculus, classical neural networks, graph neural networks, and diffusion models are analyzed. Their functional division into the roles of generator, evaluator, verifier, and optimizer is proposed, together with an operational procedure for mapping task parameters to an appropriate functional role and model class. It is shown that the suitability of a model for verifying established Service Level Objectives (SLOs) is determined not only by prediction accuracy but also by the correspondence between the type of model output and the form of the requirement being verified. An illustrative case of upgrading an overloaded network link demonstrates the sequence of candidate configuration evaluation, SLO verification, and selection of an admissible structural solution. In addition, a limited retrospective assessment of the proposed procedure was performed using published scenarios involving GNNs, Network Calculus, and diffusion models; in the considered cases, the functional roles identified by the procedure were consistent with the actual use of the models in the corresponding studies. The feasibility of combining analytical and data-driven models within a unified process for supporting structural decisions is substantiated. The proposed approach can be used to select and combine models when evaluating alternative states of a computer network throughout its life cycle.