Glint: Decentralized Federated Graph Learning with Traffic Throttling and Flow Scheduling

Federated learning has been proposed as a promising distributed machine learning paradigm with strong privacy protection on training data. Existing work mainly focuses on training convolutional neural network (CNN) models good at learning on image/voice data. However, many applications generate graph data and graph learning cannot be efficiently supported by existing federated learning techniques. In this paper, we study federated graph learning (FGL) under the cross-silo setting where several servers are connected by a wide-area network, with the objective of improving the Quality-of-Service (QoS) of graph learning tasks. We find that communication becomes the main system bottleneck because of frequent information exchanges among federated severs and limited network bandwidth. To conquer this challenge, we design Glint, a decentralized federated graph learning system with two novel designs: network traffic throttling and priority-based flows scheduling. To evaluate the effectiveness of Glint, we conduct both experiments on a testbed and trace-driven simulations. The results show that Glint can significantly outperform existing federated learning solutions.