Evaluating communication frameworks for federated deep learning in simulated edge environments

Abstract

Federated Deep Learning (FDL) enables collaborative deep learning on distributed devices while preserving data privacy, but its practical performance can heavily depend on the chosen communication framework especially on the edge. This study evaluates three following protocols Hypertext Transfer Protocol (HTTP), gRPC, and MQ Telemetry Transport (MQTT) in a simulated edge environment using Docker and the CIFAR-10 dataset. We systematically compare their impact on overall training completion time, communication latency, data transfer overhead, and system resource utilization (CPU, memory, network I/O). Our reproducible Docker-based methodology revealed significant trade-offs: Contrary to expectations based solely on protocol design, HTTP achieved the fastest overall training time in our synchronous setup, despite higher data overhead from JSON serialization. gRPC offered superior bandwidth efficiency via binary Protobufs but suffered from increased processing time and server CPU load due to implementation-specific data handling overhead. MQTT proved resource-frugal, especially suitable for constrained devices and asynchronous communication, yet its broker-based architecture and JSON usage led to longer end-to-end training durations. The model accuracy remained consistent across frameworks. These findings highlight that framework selection requires balancing bandwidth efficiency (gRPC), overall speed in synchronous settings (HTTP), and resource constraints or asynchronicity (MQTT). This work provides crucial empirical evidence to guide the deployment of efficient federated deep learning systems on the edge.