Flotilla: A scalable, modular and resilient federated learning framework for heterogeneous resources

2025-07-03

AI Summary

Existing federated learning (FL) frameworks predominantly rely on pseudo-distributed simulation, lacking real-world edge hardware deployment and thus failing to systematically evaluate core FL properties; moreover, most lack support for asynchronous aggregation and exhibit weak fault tolerance. To address these limitations, we propose a highly elastic FL framework tailored for heterogeneous edge resources. Our approach introduces a novel “client-first” modular architecture, featuring stateless clients and decoupled session-state management—enabling both synchronous and asynchronous aggregation modes alongside sub-second fault recovery. The state-separation server design supports periodic and incremental checkpointing, ensures compatibility with mainstream DNN models, and achieves efficient execution on resource-constrained edge devices (e.g., Raspberry Pi, Jetson). Experiments demonstrate robust master–worker fault tolerance with 200+ clients and significantly superior scalability over Flower, OpenFL, and FedML at 1,000+ nodes—while incurring lower resource overhead.

Technology Category

• Machine Learning: Distributed Machine Learning & Federated Learning
• Multiagent Systems: Multiagent Learning
• Search and Optimization: Distributed Search

Application Category

• Systems and Infrastructure for Web, Mobile and WoT: Federated Web and WoT systems, including distributed, federated and edge-based data processing
• Responsible Web: Machine-in-the-loop, human agency and autonomy
• User Modeling, Personalization and Recommendation: Federated recommendation systems and personalization

Abstract

With the recent improvements in mobile and edge computing and rising concerns of data privacy, Federated Learning(FL) has rapidly gained popularity as a privacy-preserving, distributed machine learning methodology. Several FL frameworks have been built for testing novel FL strategies. However, most focus on validating the learning aspects of FL through pseudo-distributed simulation but not for deploying on real edge hardware in a distributed manner to meaningfully evaluate the federated aspects from a systems perspective. Current frameworks are also inherently not designed to support asynchronous aggregation, which is gaining popularity, and have limited resilience to client and server failures. We introduce Flotilla, a scalable and lightweight FL framework. It adopts a ``user-first'' modular design to help rapidly compose various synchronous and asynchronous FL strategies while being agnostic to the DNN architecture. It uses stateless clients and a server design that separates out the session state, which are periodically or incrementally checkpointed. We demonstrate the modularity of Flotilla by evaluating five different FL strategies for training five DNN models. We also evaluate the client and server-side fault tolerance on 200+ clients, and showcase its ability to rapidly failover within seconds. Finally, we show that Flotilla's resource usage on Raspberry Pis and Nvidia Jetson edge accelerators are comparable to or better than three state-of-the-art FL frameworks, Flower, OpenFL and FedML. It also scales significantly better compared to Flower for 1000+ clients. This positions Flotilla as a competitive candidate to build novel FL strategies on, compare them uniformly, rapidly deploy them, and perform systems research and optimizations.

Problem

• Lack of FL frameworks for real edge hardware deployment
• Insufficient support for asynchronous FL aggregation
• Limited resilience to client and server failures

Innovation

• Modular design for diverse FL strategies
• Stateless clients with resilient server architecture
• Scalable performance on edge devices