What is it about?
Federated Learning (FL) enables collaborative machine learning across distributed local devices without requiring raw private data to be centrally shared. However, standard paradigms such as FedAvg rely on strict synchronous execution: the global server must pause and wait until every selected client finishes its local training before updating the global model. In real-world edge and IoT environments, slower hardware or unstable connections (stragglers) force faster clients into unproductive idle states and stall overall training pace. To solve this challenge, this paper introduces FedSaSync, the first Semi-Asynchronous Federated Learning (SAFL) extension implemented within the Flower framework. Instead of waiting for all participants, FedSaSync triggers server aggregation as soon as a predefined threshold of $M$ clients complete their local workloads. The participating clients are immediately released to rejoin training, decoupling system progress from the slowest stragglers while maintaining convergence stability.
Featured Image
Why is it important?
* First native SAFL extension in Flower: While Flower has become a standard in federated learning research, it previously lacked built-in semi-asynchronous coordination mechanisms. * Breaking the isolation barrier: Prior semi-asynchronous studies relied largely on custom, standalone simulation scripts, making comparative benchmarking and reproducibility difficult. * Sweet spot between sync and async: Pure asynchronous FL often causes severe staleness and convergence drift, while strict synchrony wastes massive wall-clock time. FedSaSync bridges both paradigms to optimize throughput and convergence. * Minimal system overhead: The built-in polling loop introduces negligible communication and runtime overhead (consistently measured between 0.4% and 0.6%).
Perspectives
Device heterogeneity and slow stragglers severely bottleneck traditional synchronous Federated Learning, while fully asynchronous approaches often harm model convergence. Semi-asynchronous FL (SAFL) offers the ideal balance, yet researchers have long been forced to rely on isolated, non-reproducible simulation scripts because standard frameworks like Flower lacked native SAFL support. With FedSaSync, we developed the first open-source SAFL middleware integrated directly into Flower. By cleanly decoupling concurrency orchestration from algorithmic logic, researchers can easily plug in custom aggregation heuristics, staleness policies, or derivative strategies such as FedProx. We released the full codebase on GitHub under the Apache-2.0 license to give the community a reliable, standardized baseline for reproducible experimentation.
Dra Carmen Carrión
Universidad de Castilla-La Mancha
Read the Original
This page is a summary of: Semi-asynchronous Federated Learning in Flower: Framework Extension and Performance Assessment, January 2026, Elsevier,
DOI: 10.2139/ssrn.7067742.
You can read the full text:
Contributors
The following have contributed to this page







