(POSTER) Network Testbed for Experimenting With Decentralized Federated Learning

Federated Learning (FL) has seen widespread adoption in different domains - healthcare, IoT, edge-computing, etc. While the predominant FL approach is centralized (CFL), more recent approaches are being investigated where the global model is aggregated in a decentralized fashion using innovative client topologies, as in Decentralized FL (DFL), addressing issues of single point of failure, trust etc. Various factors like resource heterogeneity, network performance, data distribution, data volumes, and FL configuration impact the optimization and performance of FL applications, with challenges exacerbating for DFL due to additional factors like client topologies, communication patterns, dynamic updates, etc. It is often non-trivial for FL application developers to provision and configure the resources to experiment with FL applications. This paper makes a case for using the FABRIC testbed and tools to efficiently provision, configure, deploy and manage FL architectures on a real distributed network testbed, by (a) providing a sandbox to emulate a range of infrastructure behaviors to test and optimize FL applications, and (b) enabling a rich set of network topologies beyond the client-server model, e.g. mesh, ring, etc. with support for high volume data movement, dynamic network functions and realistic communication patterns, especially relevant for DFL. We present our architecture and example DFL deployments.