Joint Communication Scheduling and Resource Allocation for Distributed Edge Learning: Seamless Integration in Next-Generation Wireless Networks

Technical Summary: Joint Communication Scheduling and Resource Allocation for Distributed Edge Learning

Problem Statement
The paper addresses the challenge of integrating Distributed Edge Learning (DL), specifically synchronous Federated Learning (FL), into next-generation (6G) wireless networks that must coexist with existing high-bandwidth (HB) traffic services, such as enhanced Mobile Broadband (eMBB). Current literature predominantly relies on "rigid" resource allocation schemes where resource blocks (RBs) and transmission parameters are fixed for the entire duration of a communication round (CR). The authors argue that this approach is inefficient and inaccurate, particularly when CR durations exceed channel coherence times or when client heterogeneity (in data size, computational capacity, and channel conditions) causes clients to be ready for uplink transmission at staggered times.

The core problem is formulated as a Joint Communication Scheduling and Resource Allocation (JCSRA) problem. Unlike prior works that assume single-server systems (where only one client transmits at a time) or rigid allocation, this work targets a multi-server environment under Orthogonal Frequency Division Multiple Access (OFDMA), where multiple User Equipments (UEs) can transmit simultaneously. The objective is to minimize the total latency of a FL CR subject to strict energy budgets for edge devices and minimum rate requirements for coexisting HB traffic.

Methodology
The authors propose a framework that moves from a time-slot-wise optimization to a tractable session-based formulation:

1. System Modeling: The system considers a single cell with OFDMA, supporting both FL clients and HB traffic UEs. The model accounts for downlink broadcasting, local computation (using Dynamic Voltage and Frequency Scaling), and uplink transmission. Crucially, it incorporates the non-linear relationship between uplink rate, allocated RBs, and transmit power, which distinguishes it from linear-rate assumptions in prior literature.
2. Session-Based Reformulation: To handle the intractability of the time-slot-wise problem (which involves complex time-dependencies and non-convex constraints), the authors assume the CR occurs within a large-scale channel coherence time. They introduce "sessions" defined by the completion of downlink reception or the readiness of local training. This allows the optimization to focus on the duration of these sessions and the allocation of resources within them, rather than every individual time slot.
3. Optimization Framework:
   • Problem Formulation: The problem is cast as minimizing the CR duration (sum of downlink, idle, and uplink session durations) subject to energy constraints, completion constraints, and HB traffic rate guarantees.
   • Theoretical Properties: The authors prove that the optimal solution generally does not follow a single-server structure (sequential transmission) due to the non-linear rate function. They also establish that optimal scheduling is non-preemptive and non-idle.
   • Algorithm Design: To solve the resulting non-convex and non-block-separable problem, an iterative algorithm is developed. It employs:
     • Quadratic Transform: To handle product terms in rate constraints (fractional programming).
     • Majorization-Minimization (MM): To approximate the product of power and duration in energy constraints with a convex upper bound.
     • Heuristic Ordering: Since evaluating all possible uplink transmission orderings is NP-hard, a heuristic based on the results of a rigid resource allocation baseline is proposed to determine the session ordering.

Key Contributions

• First Time-Slot-Wise Formulation: This work presents the first formulation of a time-slot-wise resource allocation and computational speed optimization problem for vanilla FL coexisting with HB traffic, explicitly minimizing end-to-end latency under energy constraints.
• Multi-Server JCSRA Framework: The paper generalizes JCSRA to multi-server (simultaneous transmission) scenarios under OFDMA, contrasting with existing single-server literature. It establishes that multi-server transmission is often superior to sequential transmission.
• Theoretical Guarantees: Necessary and sufficient feasibility conditions are derived. The paper proves that optimal scheduling is non-preemptive and non-idle in the multi-server case.
• Iterative Algorithm: A convergence-guaranteed iterative algorithm is proposed to solve the non-convex optimization problem, utilizing convex sub-problems at each step.
• Real-Time Integration: The study demonstrates how the time-averaged JCSRA solutions can be integrated into real-time algorithms handling fast fading and integer RB constraints with negligible performance gaps.

Results
Simulation results, conducted with 10 FL clients and 20 HB clients in a resource-constrained environment, validate the proposed approach:

• Latency Reduction: The proposed JCSRA method significantly outperforms rigid resource allocation. In scenarios with moderate energy constraints (e.g., 50 J), latency improvements of approximately 40% were observed compared to rigid baselines.
• Multi-Server vs. Single-Server: While single-server JCSRA performs well in heterogeneous or energy-constrained systems, the multi-server approach yields substantial gains in homogeneous systems with sufficient energy, where simultaneous transmission is more efficient.
• Robustness of Heuristic: The rigid-based heuristic for determining uplink ordering was found to be highly effective, with relative gaps to the optimal ordering (found via exhaustive search for smaller systems) remaining below 1% in nearly all tested realizations.
• Real-Time Feasibility: When integrated into a real-time algorithm with fast fading, the performance gap between the predicted time-average solution and the actual real-time execution was negligible (consistently below 1%).

Significance and Claims
The paper claims that rigid resource allocation is an inaccurate representation of system behavior in realistic, heterogeneous DL scenarios. By introducing a time-dependent, session-based design, the work demonstrates that significant efficiency gains are possible without altering the fundamental FL update rules or convergence behavior.

The authors emphasize that while single-server JCSRA may suffice for highly constrained or heterogeneous systems, multi-server JCSRA is crucial for maximizing efficiency in resource-sufficient and homogeneous environments. The work highlights the necessity of considering time-dependent optimization within each CR to accurately estimate and minimize latency under energy constraints in future wireless networks. The paper concludes that its proposed framework provides a foundational step toward the seamless integration of DL services in 6G networks, acknowledging that future work will need to address asynchronous DL frameworks and non-stationary channel conditions.