Federated Learning for Workload Forecasting for Network Service Management

Next generation networks are expected to connect and manage a vast number of heterogeneous devices stretching over diverse and distributed technological domains (e.g., radio, transport, core), by embracing a ubiquitous presence of AI and ML within their operations. This shift imposes a significant challenge for the management and orchestration (MANO) frameworks in order to efficiently handle that great demand for computational power, particularly in the edge of the networks. Therefore, in order to increase the effectiveness of the task scheduling of the next generation networks it is crucial to proactively detect both periodic and non-periodic patterns that could affect the network’s decision-making processes. Previous research has been primarily based on centralized predictive models gathering all the data from the edge devices on one central processing point. Such approaches impose substantial communication overhead on the network, impacting bandwidth to the point of channel congestion, and exposing data to security vulnerabilities. This paper showcases a computational load forecasting method that aims to feed the MANO frameworks with predicted insights of the infrastructure state. The forecasting is performed by employing machine learning techniques such as LSTMs and Bi-LSTMs, by utilizing the federated learning family of algorithms, FedOpt, facilitating the distributed training process. The proposed
method is evaluated with actual workload traces originating from a cluster of virtual machines. The experimental results show that the proposed federated learning approach provides results of comparable quality to the centralized learning methods, while minimizing the data exchanged between the edge nodes and the cloud.