Abstract
Federated learning (FL) enables multiple data owners (a.k.a. FL clients) to collaboratively train machine learning models without disclosing sensitive private data. Existing FL research mostly focuses on the monopoly scenario in which a single FL server selects a subset of FL clients to update their local models in each round of training. In practice, there can be multiple FL servers simultaneously trying to select clients from the same pool. In this paper, we propose a first-of-its-kind Fairness-aware Federated Job Scheduling (FairFedJS) approach to bridge this gap. Based on Lyapunov optimization, it ensures fair allocation of high-demand FL client datasets to FL jobs in need of them, by jointly considering the current demand and the job payment bids, in order to prevent prolonged waiting. Extensive experiments comparing FairFedJS against four state-of-the-art approaches on two datasets demonstrate its significant advantages. It outperforms the best baseline by 31.9% and 1.0% on average in terms of scheduling fairness and convergence time, respectively, while achieving comparable test accuracy.
Introduction
In the standard Federated Learning (FL) narrative, a single server coordinates a group of clients to train a model. However, the real world is rarely a monopoly. Imagine a marketplace where multiple companies (servers) are simultaneously trying to train different models (jobs) using the same pool of mobile devices or edge sensors (clients). This creates a "multi-job" bottleneck. Clients have finite resources and can usually only participate in one training job at a time. If several jobs all need a specific, rare data type, some jobs might face "prolonged waiting," leading to slow convergence or total starvation.
Current methods often focus on minimizing costs for the servers but ignore the fairness of how these jobs are actually scheduled. This paper introduces FairFedJS (Fairness-aware Federated Job Scheduling) to bridge this gap by ensuring that high-demand client datasets are allocated fairly across competing jobs, preventing resource starvation for lower-priority or lower-budget tasks.
The FairFedJS Framework
FairFedJS treats the scheduling problem as a dual-objective optimization: maximizing system revenue while minimizing scheduling unfairness. The system maintains Virtual Queues for each data type; if the demand for a specific data type exceeds the supply of available clients, the queue grows, signaling a higher priority for jobs needing that data.
To stabilize these queues, the framework employs Lyapunov Optimization. By minimizing the "drift" of these queues, the system naturally prioritizes jobs that have been waiting the longest for specific resources. Decisions are made in real-time using a Job Scheduling Index (JSI), which ranks jobs based on queue length, payment bids, and expected client reliability.
As shown in Algorithm 1, the process involves generating payments, computing the JSI (denoted as $\Psi_k(t)$), and ordering jobs accordingly. Once a job is scheduled, it selects clients based on a Beta Reputation System (reliability) and Data Fairness (preventing client over-sampling).
Experimental Evaluation
The approach was evaluated using two classic computer vision benchmarks adapted for a multi-job environment: Fashion-MNIST and CIFAR-10. The experimental setup involved 50 FL clients: 20 dedicated to each dataset and 10 "versatile" clients possessing both. The authors tested both IID (Independent and Identically Distributed) and Non-IID data distributions to simulate real-world heterogeneity.
Key Results
FairFedJS was compared against four baselines: Random scheduling, Alternating (round-robin), Utility-Based (UB), and MJ-FL (a cost-minimization approach).
- Scheduling Fairness (SF): FairFedJS outperformed the best baseline (MJ-FL) by 31.9% on average. As seen in Table II, FairFedJS consistently achieves the lowest SF scores (where lower is better), indicating a more equitable distribution of client resources.
- Convergence Speed: The system achieved 1.0% faster convergence on average. While seemingly small, this represents significant savings in time and compute energy in large-scale FL deployments.
- Model Accuracy: Despite the focus on fairness, the final test accuracy remained comparable to (and sometimes better than) the baselines, proving that system fairness does not necessitate a trade-off in model quality.
Conclusion and Future Work
FairFedJS demonstrates that scheduling fairness is not just a moral goal but a functional one that stabilizes the training ecosystem. By treating client availability as a managed resource, the framework prevents job starvation and ensures predictable convergence.
Future research will expand this model to Vertical FL, where jobs may require combinations of different data types from multiple clients simultaneously, and Federated Transfer Learning, where the relationships between concurrent jobs are more complex. Additionally, the authors aim to test the scalability of the JSI ranking in environments with thousands of concurrent jobs.
Summary updated