Optimized Multi-Service Tasks Offloading for Federated Learning in Edge Virtualization

Abstract

Edge federated learning (EFL) utilizes edge computing (EC) to alleviate direct round communications of multi-dimensional model updates between local participants and the central parameter server. With an elastic edge-assisted aggregation procedure, the burden of backbone congestion can be handled for multi-service model communications. To provide an adaptive virtual resource orchestration, the integration of software-defined networking (SDN), network functions virtualization (NFV), EC, and double deep q-networks is emerged, which accurately gathers the state observation from local and edge nodes in data plane and NFV infrastructure (NFVI) for applying actions of allocated resource predetermination and recommended offloading policy. The reward formulation considers the minimization of completion time, energy consumption, and round communications between local participants and the selected NFV-enabled EC (NFVeEC) node until the global model is constructed with satisfying accuracy. In this paper, we proposed a decentralized SDN controller as an agent to enable a softwarization mechanism and interact with the virtualization-based EFL environment. The architecture adoption, agent controller, and orchestrator allow the centralized view for virtual network functions placement with efficient virtual machine mapping to perform edge update procedures and create the forwarding path effectively. The recommended NFVeEC for local-EFL updates is based on reward output and proposed agent valuation.