KAFL: Achieving High Training Efficiency for Fast-K Asynchronous Federated Learning

Federated Averaging (FedAvg) and its variants are prevalent optimization algorithms adopted in Federated Learning (FL) as they show good model convergence. However, such optimization methods are mostly running in a synchronous flavor which is plagued by the straggler problem, especially in the real-world FL scenario. Federated learning involves a massive number of resource-weak edge devices connected to the intermittent networks, exhibiting a vastly heterogeneous training environment. The asynchronous setting is a plausible solution to fulfill the resources utilization. Yet, due to data and device heterogeneity, the training bias and model staleness dramatically downgrade the model performance. This paper presents KAFL, a fast-K Asynchronous Federated Learning framework, to improve the system and statistical efficiency. KAFL allows the global server to iteratively collect and aggregate (1) the parameters uploaded by the fastest K edge clients (K-FedAsync); or (2) the first M updated parameters sent from any clients (Mstep-FedAsync). Compared to the fully asynchronous setting, KAFL helps the server obtain a better direction toward the global optima as it collects the information from at least K clients or M parameters. To further improve the convergence speed of KAFL, we propose a new weighted aggregation method which dynamically adjusts the aggregation weights according to the weight deviation matrix and client contribution frequency. Experimental results show that KAFL achieves a significant time-to-target-accuracy speedup on both IID and Non-IID datasets. To achieve the same model accuracy, KAFL reduces more than 50% training time for five CNN and RNN models, demonstrating the high training efficiency of our proposed framework.