Modelling federated learning workflows for the cloud-to-edge continuum

Federated Learning (FL) is a key paradigm for distributed, data rich environments, leveraging devices across the Cloud-to-Edge continuum to minimise latency and bandwidth usage while preserving privacy. However, testing these systems (IoT, 5G/6G) with FL capabilities in physical testbeds is challenging due to high costs and complexity, making simulation tools essential. In this paper, we introduce a simulation framework for modelling FL workflows in Cloud-to-Edge scenarios. Implemented as an extension to the DISSECT-CF-Fog discrete-event simulator, our module reuses its workflow scheduling and execution functionalities to model FL processes with centralised orchestration, diverse network topologies, varying device capacities, and dynamic workloads. It abstracts low-level Machine Learning mechanics to focus on first-order system effects such as latency, network traffic, and energy use. We model a Holographic Type Communications (HTC) use case as a baseline scenario (S0) to orchestrate FL deployments and quantify trade-offs in round latency, traffic, and energy across four operational scenarios: scalability (S1), aggregation policies (S2), pacing strategies (S3), and privacy and compression (S4). Each scenario tracks a synthetic learning progress proxy (e.g., rising from ≈ 0.50 to > 0.9 over 20 rounds in S0) and per-round aggregator energy (e.g., ≈ 1 kJ for early aggregation to ≈ 10 kJ for timeout rounds in S0). Ultimately, the module exposes how operational variations shape system costs. For example, aggressive model compression (e.g., 0.2 × in S4) reduces per-round traffic proportionally, yet the corresponding energy saving is bounded by fixed per-round infrastructure costs that dominate when rounds close on a timeout rather than on transmission completion.