Contents I. Introduction
Federated learning (FL) is a popular distributed learning paradigm that has captured the interest of researchers globally. In a typical FL approach, the central server and the partic-ipating client devices are assumed to be fully synchronous [1]. The server waits for all clients to complete their local training before aggregating the weights. The availability of each client varies considerably over time due to system heterogeneities and internet connectivity. To address these challenges, researchers explored the potential of asynchronous updates, resulting in the emergence of asynchronous federated learning (AsynchFL) [2]. AsynchFL introduce flexibility and improve scalability to the FL framework by enabling clients to independently and asynchronously update their local models. The transition from synchronous to asynchronous FL mark a significant advancement in improving the performance and capabilities of federated learning. However, this transition also introduced new challenges, such as managing asynchrony, addressing stale gradients, convergence properties,and ensuring consistency in the global model. In a fully AsynchFL method [2], [3], each client update leads to an update in the server model. However, this approach introduces a challenge where faster clients update the global model more frequently. Whereas, slower clients may have their model updates based on an earlier round of the global model, leading to potential staleness that affect the model convergence [4]. In [2], [5] and [6], the authors introduce a staleness function that assigns lower weights to model updates from stale clients during server aggregation. In another work, [7], the authors propose a semi-asynchronous approach named SAFA. SAFA incorporates a lag tolerance hyperparameter to quantify client staleness and disregards local training results from clients considered too stale based on the lag tolerance threshold. Additionally, FedBuff [8] suggests an approach where the server waits for updates from a minimum number of clients before performing aggregation and stores the weights in a buffer. Both SAFA and FedBuff operate as hybrids between fully asynchronous and synchronous modes of operation. The TimelyFL approach, presented in [9], introduces a heterogeneity-aware semi-AsynchFL method with adaptive partial training. They address the inclusion of more available devices in global aggregation without introducing staleness by implementing partial model training for clients with lower capacity.
