Client scheduling and bandwidth slicing for multiple federated learning tasks over multiple passive optical networks

Abstract

Listen

Translate article

Federated Learning (FL) has attracted extensive attention in facilitating emerging edge intelligence applications for its inherent advantages of ensuring data security and privacy. Especially in the edge computing networks connected by Passive Optical Network (PON) system, FL is introduced to enable applications like autonomous driving, intelligent manufacturing, and precision medicine. However, in this system, the FL deployment over PON inevitably faces challenges induced by the conflict between a large volume of model data with the restrict latency limitation and the confined bandwidth resource of PON, especially for the multiple FL tasks. To address these issues, a novel scheme is proposed for tackling the client distribution and bandwidth allocation problems under the scenario of multiple simultaneous FL tasks supported by the multiple interconnected PON systems, which is consisted of the client scheduling and bandwidth slicing processes. To be specific, an easy-to-implement heuristic algorithm is first performed to assign the client numbers to PONs based on iterative method, with which certain operations are repetitively executed to achieve optimal solutions. And then, the serial bandwidth slicing which adapts the traditional policy, i.e., one-task-per-cycle, to the situation with multiple FL tasks, and parallel slicing with the multi-task-per-cycle, are designed for the investigated system. Furthermore, the simulation system is constructed to verify our method. The corresponding results exhibit that, the largest 43.2 % round time reduction is achieved by client scheduling compared to benchmark without the scheduling. Compared to the benchmark with serial slicing, our scheme can achieve a maximum 50.11 % of the round time decrease. It's also validated that, our proposed client scheduling and parallel bandwidth slicing method can improve the learning efficiency by reducing communication delay, especially for the situation with less client number and smaller FL threshold.