Parallel and distributed execution can support computational scalability by dividing work across processing units. However, partitioning is useful only when dependencies between model components do not create excessive cross-partition communication and synchronisation, workload imbalance, or repeated repartitioning. The thesis therefore examines how dynamic and stochastic interaction structures affect whether model components can be divided into groups with limited dependencies between them.

The study combines literature from agent-based economics, network theory, graph partitioning, and parallel and distributed simulation. It develops a dependency-based assessment framework covering dependency alignment, communication and synchronisation, and stability and adaptation, with workload balance as a cross-cutting constraint. The framework is examined through a controlled goods-market prototype and through firm input-purchase dependencies extracted from the Oxford/INET economic agent-based model.

The results show that partitioning depends on how dependencies are organised and how they develop over time. Local and community-based interactions can support partitioning when dependencies remain concentrated within identifiable groups, while hub-based structures have mixed implications and widely distributed random interactions are less favourable for static partitioning. Frequent changes in individual relationships do not automatically require repartitioning if the overall dependency structure remains stable. For the examined Oxford/INET input-purchase dependencies, country-based partitions aligned substantially better than the tested industry-based schemes, although the country groups were not isolated. The controlled prototype also showed that persistent structural change can alter which partitioning scheme is best aligned. The framework provides a structured way to compare candidate partitioning schemes based on the model’s economic organisation. It helps focus implementation and runtime testing on structurally promising candidates, while making clear that computational benefits still require direct workload and runtime evaluation.