Abstract
Recent years have witnessed a large amount of decentralized data in multiple (edge) devices of end-users, while the aggregation of the decentralized data remains difficult for machine learning jobs due to laws or regulations. Federated Learning (FL) emerges as an effective approach to handling decentralized data without sharing the sensitive raw data, while collaboratively training global machine learning models. The servers in FL need to select (and schedule) devices during the training process. However, the scheduling of devices for multiple jobs with FL remains a critical and open problem. In this paper, we propose a novel multi-job FL framework to enable the parallel training process of multiple jobs. The framework consists of a system model and two scheduling methods. In the system model, we propose a parallel training process of multiple jobs, and construct a cost model based on the training time and the data fairness of various devices during the training process of diverse jobs. We propose a reinforcement learning-based method and a Bayesian optimization-based method to schedule devices for multiple jobs while minimizing the cost. We conduct extensive experimentation with multiple jobs and datasets. The experimental results show that our proposed approaches significantly outperform baseline approaches in terms of training time (up to 8.67 times faster) and accuracy (up to 44.6% higher).
Introduction
Standard Federated Learning (FL) typically focuses on training a single global model at a time. However, in modern edge computing environments, servers often need to train multiple independent models simultaneously—such as image classifiers and speech recognizers. Sequential training in these scenarios is inefficient, leaving many devices idle and prolonging overall training time. Scheduling a shared pool of heterogeneous devices across multiple concurrent jobs is an NP-hard combinatorial optimization problem. Devices vary significantly in compute and communication speeds, and their local data is typically non-IID (not independent and identically distributed). Existing single-job schedulers often ignore device heterogeneity, leading to "straggler" issues, or ignore "data fairness," which results in models that fail to represent the diversity of the edge data, ultimately hurting accuracy.
The Multi-Job Federated Learning (MJ-FL) Framework
To address these challenges, we introduce the MJ-FL framework, designed to support the parallel training of multiple jobs. The core of this framework is a novel Cost Model that evaluates scheduling plans based on a linear combination of two critical factors:
- Time Cost: Determined by the slowest device (straggler) selected for a given training round.
- Data Fairness Cost: Measured by the variance in how frequently different devices are selected for a specific job. Low variance ensures that all devices' data is adequately represented in the global model.
To minimize this cost function, we propose two scheduling methods:
- BODS (Bayesian Optimization-based Device Scheduling): This method utilizes a Gaussian Process to model the cost function and an Expected Improvement (EI) acquisition function to efficiently sample and select optimal scheduling plans.
- RLDS (Reinforcement Learning-based Device Scheduling): This approach employs an LSTM-based policy network to learn complex device-sharing relationships. By processing device capabilities and historical fairness as inputs, it outputs the probability of assigning each device to specific jobs.
Experimental Evaluation and Results
We evaluated the MJ-FL framework using standard image classification datasets (CIFAR-10, EMNIST, Fashion-MNIST, and MNIST) under both IID and non-IID settings. Our proposed methods, BODS and RLDS, were compared against standard FL scheduling baselines, including Random, FedCS, Genetic, and Greedy algorithms, across various architectures like VGG-16, ResNet-18, and LeNet.
Key Performance Insights
- Significant Speed and Accuracy Gains: The proposed methods achieved single-job training speeds up to 8.67x faster and reduced total multi-job training time by up to 4.67x compared to baselines. Furthermore, MJ-FL achieved up to 44.6% higher accuracy than baseline methods, which often failed to reach target accuracies in non-IID scenarios.
- Fairness as a Catalyst: Prioritizing data fairness does more than just prevent bias; it accelerates convergence. By ensuring diverse device participation, models converged up to 9.35x faster in non-IID settings compared to fairness-agnostic methods.
- Algorithm Specialization: Our results suggest that RLDS is particularly effective for complex models (e.g., VGG) where the neural network can learn latent device interactions. Conversely, BODS is highly efficient for simpler models, providing statistical modeling benefits without the overhead of reinforcement learning.
Conclusion
The MJ-FL framework demonstrates that parallel training and intelligent device scheduling can drastically improve the efficiency and effectiveness of federated learning in multi-job environments. By balancing time costs with data fairness, we achieve substantial improvements in both convergence speed and final model accuracy. Future work will focus on dynamic hyperparameter tuning for the cost model weights and validating the framework on physical hardware in volatile real-world IoT networks.
Summary updated