Večciljno ogrodje za razvrščanje bremen v okolju Kubernetes z nevronskim algoritmičnim sklepanjem

Modern cloud native applications operate in increasingly complex, dynamic and distributed environments, where efficient workload scheduling is essential to ensure performance, cost efficiency, and compliance with operational constraints. This thesis presents a multi-objective scheduling framework for Kubernetes that combines a graph-based modeling and representation approach with neural algorithmic reasoning (NAR). By representing workloads and infrastructure as graphs, the framework captures dependencies, resource requirements, and constraints while enabling multi-layer observability. To design and evaluate the framework, existing Kubernetes simulation tools were analyzed, and a tailored environment was developed to test diverse scheduling strategies on semi-realistic datasets. Multiple neural models were implemented and trained, including variants with structured numerical inputs and text-conditioned encoder-decoder architectures, to assess the impact of multimodal supervision and modular training. Experimental results show that NAR-based models consistently outperform traditional rule-based schedulers, achieving significant reductions in constraint violations and scheduling costs while maintaining scalability with growing problem size. Text-conditioned models further improve efficiency and adherence to constraints, albeit with a considerable trade-off in assignment completeness. In general, the research demonstrates the effectiveness of integrating symbolic graph modeling with neural reasoning for adaptive and scalable workload scheduling in complex computing environments. The graph-based representation of infrastructure additionally improves transparency, offering a promising foundation for more intelligent orchestration of cloud and edge workloads.