FL-Task-aware Routing and Resource Reservation over Satellite Networks

Abstract

Earth observation satellites using asynchronous ground-assisted federated learning (FL) can avoid transmitting massive raw image data to ground. However, current FL approach uses only satellite-to-ground-station links, causing long delay for model parameter transfer. A promising direction to reduce delay is to use inter-satellite links. We identify that ground-assisted asynchronous FL requires a satellite to send all data of its model parameter to ground before ground station can start to update the model. This new feature prevents current routing algorithms (e.g., CGR) from being applicable. Therefore, we propose an FL task-aware routing and resource reservation (FLRRS) scheme to optimize the delay of FL model parameter transfer. First, we formulate the problem as an integer linear programming (ILP) problem, which is non-convex and intractable. Thus, we enhance the storage time-aggregated graph to model computing, storage and transmission resources of satellite network, and propose a graph-based routing and resource reservation algorithm. The numerical simulation based on a real-world satellite network shows that FLRRS runs much faster than CVXPY solver. Besides, FLRRS also significantly improves average delay and number of completed tasks, as compared to current routing algorithms.