Bandwidth constrained client selection and scheduling for federated learning over SD‐WAN

Abstract

Cross-silo federated learning (FL) has been proposed and applied in many domains such as financial risk prediction, pharmaceutical discovery, electronic health records mining. In this paradigm, multiple clients can collaboratively train machine learning models with higher accuracy while protecting their privacy under the coordination of a central server. To interconnect these geo-distributed clients and the central server, the architecture of FL is proposed over software-defined wide area network (SD-WAN). The total time to achieve the required accuracy for FL depends on both the time of each iteration and the number of iterations. Because the bandwidth for each path from client to central server is constrained, different client selection schemes and different path routing mechanisms may lead to different times of iteration. The authors propose FedMT algorithm to minimize the total time through well-designed client selection and routing, which greatly reduces the upload time of each iteration while slightly affecting the number of iterations. Extensive experiments are conducted and demonstrated the time to achieve the same accuracy for FedMT is less than half of that for FedAvg and FedProx.