Optimizing Federated Learning via SDN Telemetry

HybridFLow achieves up to 40% faster convergence in cross-silo federated learning by moving client orchestration decisions from the application layer to the network layer. Traditionally, federated learning systems treat client participation as a simple application-level scheduling problem. This ignores the reality of wide-area network latency, where slow clients become stragglers that force the entire system to wait. HybridFLow integrates an SDN controller to monitor network state, allowing the system to see bottlenecks that are invisible to individual clients.

Moving Beyond the Application Layer

In standard synchronous federated learning, the entire training round is gated by the slowest participant. Asynchronous approaches avoid this gate but introduce model staleness that degrades accuracy. HybridFLow creates a middle path by using network telemetry to predict completion times. It uses protocols such as gNMI to stream real-time metrics—including interface utilization and buffer occupancy—from switches across the topology. The SDN controller aggregates this data into a performance map, calculating per-link latency. Because the controller understands the path from each client to the aggregator, it can estimate the communication overhead for every participant.

For engineers building distributed training pipelines, the shift here is that the network ceases to be an opaque pipe. A local client only knows that a packet is slow; it cannot distinguish between a congested link, a circuitous routing path, or a saturated aggregator port. By feeding link-cost telemetry into the scheduling logic, the system identifies the specific hop causing congestion. This link-cost awareness allows the scheduler to proactively relegate clients on constrained paths to the asynchronous group before the synchronous round begins, ensuring the primary training loop remains fast while still incorporating data from slower participants.

The Engineering Trade-off

This approach works because it treats link saturation as a manageable input rather than a random error. However, the system is constrained by the overhead of collecting and processing real-time telemetry. While the current framework provides a significant speedup, it remains limited to a centralized control architecture. Scaling this logic to massive, thousand-node networks is currently blocked by the latency of the control plane itself. Distributing this orchestration logic across secondary SDN controllers to maintain performance at scale remains a known, unsolved hurdle for practitioners.