HybridFLow: SDN-Orchestrated Client Partitioning for Hybrid Federated Learning

In cross-silo Federated Learning (FL), communication delays across wide-area networks often create a straggler effect, where the entire training round is gated by the slowest participant. While hybrid FL approaches attempt to mitigate this by allowing some clients to participate asynchronously, they often lack the network-level visibility required to make optimal partitioning decisions. This paper asks: can integrating global SDN-based network intelligence into the hybrid FL orchestration loop improve training efficiency without sacrificing model convergence?

HybridFLow introduces a closed-loop orchestration framework that bridges the SDN control plane with the FL server. The SDN controller maintains a global view of network topology, link utilization, and path contention. Before each training round, the controller generates calibrated per-client communication-time estimates. These estimates are fed to a Hybrid Partitioner, which uses a greedy sweep algorithm to categorize clients into synchronous and asynchronous groups, balancing the trade-off between round latency and the staleness risk of asynchronous updates. After each round, the system uses actual measured transfer times to update correction factors, continuously refining future predictions.

HybridFLow demonstrates that network-aware orchestration significantly outperforms static or compute-only partitioning strategies. By identifying and reclassifying potential stragglers as asynchronous participants before they can delay the synchronous phase, the system reduces average round duration by 30-40 seconds across various network topologies. Crucially, this management strategy preserves model utility; HybridFLow reaches 80% target accuracy 33-40% faster than the SmartFLow baseline. Furthermore, the study shows that while pure asynchronous methods like FedAsync fail to converge under non-IID data distributions, HybridFLow maintains convergence quality by carefully managing the staleness of asynchronous contributions.