What question did this study set out to answer?

This work addresses the orchestration of federated learning systems to enhance efficiency and adaptability.

May 7, 2026Information and Software Technology0 citationsOpen Access

FedJoint: A software architecture for adaptive orchestration in federated learning systems

MUMd Raihan UddinSoftware Engineering Institute PKPrabhat KumarSoftware Engineering Institute NINajmul IslamSoftware Engineering Institute

Key Points

This work addresses the orchestration of federated learning systems to enhance efficiency and adaptability.
Introduced FedJoint as a modular orchestration framework for federated learning.
Utilized a Selection Manager for client participation and an Aggregation Manager for updates.
Implemented a Deep Reinforcement Learning Controller to adapt decisions based on runtime feedback.
Achieved final accuracy between 63.62% and 97.82% on CIFAR-10 and MNIST datasets.
Demonstrated 5.2–14.0 × speedup over synchronous and semi-asynchronous baselines.
Achieved up to 12.2 × lower wall-clock cost compared to RL-based approaches.

Abstract

Federated Learning (FL) operates as a distributed software system in which a central coordinator orchestrates training across heterogeneous and intermittently available clients. In practice, client selection and aggregation policies are configured as static and independent parameters, which become brittle under fluctuating computation capacity, network latency, and client reliability. Consequently, FL systems often suffer from slow convergence, unstable training, and limited adaptability under dynamic execution conditions. This work formulates FL orchestration as a software architecture problem and introduces FedJoint, a modular framework that jointly coordinates client selection and aggregation timing. Unlike prior approaches that optimize these mechanisms independently, FedJoint treats them as coupled runtime control decisions within a unified architecture. The objective is to enable adaptive co-management that improves efficiency, robustness, and operational adaptability under heterogeneous and non-stationary conditions. FedJoint is realized as a modular orchestration architecture composed of three decoupled components: a Selection Manager controlling client participation, an Aggregation Manager governing update integration, and a Deep Reinforcement Learning (DRL) Controller that adapts orchestration decisions based on runtime feedback. The components interact through well-defined interfaces that support configurability, substitution, and integration with existing FL platforms. Within this architecture, a DRL–based policy adapts selection and aggregation parameters to balance accuracy, latency, client dropout, and communication overhead under dynamic execution conditions. Evaluation on CIFAR-10 and MNIST under two heterogeneity levels (Dirichlet α ∈ 0. 1, 0. 5) and non-stationary conditions shows final accuracy of 63. 62%–97. 82%, with 5. 2–14. 0 × speedup over synchronous and semi-asynchronous baselines. Compared to RL-based baselines, FedJoint achieves up to 12. 2 × lower wall-clock cost at higher final accuracy. Ablation confirms that gains stem from joint coordination, with single-component variants showing accuracy gaps up to 31. 84 percentage points. Treating FL orchestration as a coupled architectural concern enables more robust and manageable distributed learning systems and offers concrete guidance for adaptive FL platform design.

Ask AI

Helpful

Bookmark

View Full Paper