Graduation Semester and Year
Spring 2026
Language
English
Document Type
Dissertation
Degree Name
Doctor of Philosophy in Computer Engineering
Department
Computer Science and Engineering
First Advisor
Hao Che
Second Advisor
Engin Arslan
Third Advisor
Md Arifuzzaman
Fourth Advisor
Faysal Shezan
Abstract
Modern wide-area networks increasingly adopt hybrid architectures that combine high-capacity wired backbones with flexible wireless links to extend connectivity to remote and underserved locations. However, the bandwidth variability inherent in wireless segments creates routing challenges that traditional protocols, designed for static link capacities, cannot adequately address. Simultaneously, Federated Learning (FL) has emerged as a privacy-preserving distributed machine learning paradigm in which geographically dispersed clients collaboratively train shared models without exchanging raw data. When deployed over wide-area networks, FL training is severely bottlenecked by communication overhead, particularly in cross-silo settings where model payloads reach hundreds of megabytes and synchronous aggregation protocols leave all participants idle while awaiting a single straggler.
This dissertation addresses these challenges through five progressive contributions. First, BACAR introduces an SDN-based routing system that infers wireless capacity changes from lightweight delay measurements and reroutes traffic within seconds, improving aggregate utilization by 15-20% over static baselines. Second, RENET extends this approach with QoS-aware, dual-telemetry path selection that sustains approximately 93% per-flow bandwidth satisfaction under severe weather-induced degradation.
Third, SmartFlow translates this networking intelligence to FL by dynamically assigning communication paths to clients based on real-time network state, reducing synchronous training time by up to 47% without modifying the learning algorithms. Fourth, HybridFlow integrates SDN network estimates into a hybrid synchronous-asynchronous FL partitioning framework; its greedy partitioner minimizes a joint cost function balancing round latency against staleness risk, achieving 33-40% faster convergence compared to synchronous FL. Finally, FLEET provides an open-source FL emulation testbed that fuses production-grade FL orchestration with realistic SDN-controlled network emulation, quantifying communication overhead at up to 59% of total round time.
Together, these contributions demonstrate that the programmable network is a powerful, orthogonal lever for accelerating distributed machine learning. By bridging adaptive traffic engineering with federated optimization, this dissertation shows that cross-layer co-design yields performance gains unattainable by either domain in isolation.
Keywords
Software-Defined Networking, Federated Learning, Wide-Area Networks, Traffic Engineering, Communication Optimization
Disciplines
Computer and Systems Architecture | Computer Engineering | Digital Communications and Networking
License
This work is licensed under a Creative Commons Attribution 4.0 International License.
Recommended Citation
Abu Hamdan, Osama, "ADAPTIVE ROUTING AND COMMUNICATION OPTIMIZATION IN SOFTWARE-DEFINED NETWORKS: FROM HYBRID WIDE-AREA NETWORKS TO FEDERATED LEARNING" (2026). Computer Science and Engineering Dissertations-Archive. 433.
https://mavmatrix.uta.edu/cse_dissertations/433
Included in
Computer and Systems Architecture Commons, Digital Communications and Networking Commons