Performance Observations from Split Federated Learning on Heterogeneous Devices & Networks

Split Federated Learning (SFL) combines the scalability of Federated Learning (FL) with the computational efficiency of Split Learning (SL), making it a promising paradigm for distributed machine learning in heterogeneous environments. Existing studies have explored theoretical convergence and split layer selection, however, the practical implications of split layer selection with client heterogeneity and network variability remain under explored. This paper investigates the impact of heterogeneous client resources, network conditions, and split layer configurations on the performance of SFL. Using the CIFAR-10 dataset, we implement a SFL model with heterogeneous clients. Experiments include diverse configurations, such as uniform and varying split layers among clients. We analyze training time, idle and active times, and resource utilization to understand the effects of heterogeneity. Our observations show that strategic split layer configurations tailored to heterogeneous environments can improve training efficiency by 26.7%. Additionally, we observed the effects that heterogeneous networks have on SFL and how to compensate for it, offering valuable insights for deploying SFL in real-world computing systems.