Totoro+ decentralizes federated learning, scales to millions of edge nodes

A research team from multiple institutions (including authors from Georgia Tech and other universities) has introduced Totoro+, a federated learning system that fully decentralizes the traditional client-server architecture. The key innovation is using a distributed hash table (DHT) to create a peer-to-peer network where any edge node can act as coordinator, aggregator, or worker for any application—eliminating the single-point bottleneck of a centralized parameter server.

Totoro+ introduces three core innovations: a locality-aware P2P multi-ring structure for efficient node organization, a publish/subscribe-based forest abstraction for scalable model and gradient routing, and a game-theoretic path planning model that guarantees an ε-approximate Nash equilibrium for optimal client selection. In real-world experiments across 500 Amazon EC2 servers, Totoro+ scaled gracefully with the number of FL applications. Training sped up by 1.2x to 14x, model dissemination required only O(log N) hops even with millions of nodes, and the system handled practical edge network churns efficiently. The paper has been accepted to IEEE Transactions on Parallel and Distributed Systems (TPDS).