An Online Algorithm for Optimizing Network Transmission Cost of Federated Learning in the Cloud

Abstract

Data privacy concerns and related regulations such as the General Data Protection Regulation in machine learning have fostered a boom in federated learning (FL). However, the costly infrastructure and time-consuming deployments pose significant barriers to the widespread adoption of FL in real-world scenarios. To increase the user-friendliness of federated learning while reducing deployment costs and improving its scalability, service providers are beginning to offer federated learning as a service (FLaaS) in the cloud. Due to the distributed nature of FL, communication overhead imposes significant network costs on FLaaS providers. In mainstream cloud platforms, there are two main types of billing methods for networking products, which are on-demand and reserved. How to optimally combine these two billing models to optimize communication cost in the face of time-varying demands of federated learning in the cloud poses a challenge to FLaaS providers. To address this problem, we propose OnlineNS, an online algorithm for optimally making networking product selection decisions without prior knowledge of future demand sequences. Our algorithm can achieve better cost performance compared to online algorithms that are widely used in practice. The theoretical analysis and simulations based on real-world traces as well as synthetic datasets validate the effectiveness of our online algorithm and demonstrate that it can achieve better cost performance compared to benchmarks with the same communication performance.