DRL-Based OSPF Convergence Optimization: From Simulation to Network Emulation
International Journal of Information and Communication Technology Trends · 19.10.2026 · 6 (2) · 393-405
DOI: 10.66525/ijictt142 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Information and Communication Technology Trends, online erschienen am 19.10.2026. Verfasst von Jonah Desai und 2 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Deep reinforcement learning (DRL) has been proposed to tune OSPF convergence timers online, but such agents are trained and evaluated in simulators whose fidelity to real routing stacks is rarely checked. We take DRL-based OSPF convergence optimization from simulation to network emulation. An FRRouting testbed running production ospfd/zebra daemons in Linux namespaces is paired with a protocol-level simulator at three fidelity levels (idealized, FRR-faithful, trace-calibrated) that share scenarios, observations, actions and metrics. Calibration identifies 4.4 ms carrier-loss detection, 2.2 ms per-hop flooding and about 1.0 s of configuration-push skew. On real ISP topologies and traffic matrices, the idealized simulator misestimates the per-episode cost by 62.7% on average and the restoration time of hard failures by a factor of 66, whereas the calibrated simulator stays within 12.6%. We re-implement the DRL-Adapt agent, use its Markov decision process as the common control formulation, and extend it with a feasibility-aware action-branching design. Deployed on the testbed, the re-implemented DRL-Adapt agent lowers the cost by 31.6% (trained in the idealized simulator) and 32.9% (trained in the calibrated simulator) relative to RFC default timers, a gain comparable in magnitude to the convergence-time reduction originally reported for DRL-Adapt. The agent nevertheless converges to a near-static operating point, and a static configuration found by searching the calibrated simulator reduces the cost by 32.7%, within the scenario-to-scenario variability of both DRL-Adapt variants. Our results indicate that DRL-Adapt’s improvement over default timers transfers from simulation to a production routing stack, and that, under realistic protocol constraints, calibrated configuration search captures most of the benefit of adaptive timer control on the evaluated topologies.
Zitieren
APA
Desai, J., Phillips, E., & Miller, S. (2026). DRL-Based OSPF Convergence Optimization: From Simulation to Network Emulation. International Journal of Information and Communication Technology Trends, 6(2), 393-405. https://doi.org/10.66525/ijictt142
Vancouver
Desai J, Phillips E, Miller S. DRL-Based OSPF Convergence Optimization: From Simulation to Network Emulation. International Journal of Information and Communication Technology Trends. 2026;6(2):393-405. doi:10.66525/ijictt142
BibTeX
@article{Desai2026,
author = {Desai, Jonah and Phillips, Elijah and Miller, Seojun},
title = {DRL-Based OSPF Convergence Optimization: From Simulation to Network Emulation},
journal = {International Journal of Information and Communication Technology Trends},
year = {2026},
volume = {6},
number = {2},
pages = {393--405},
doi = {10.66525/ijictt142}
}
EUVIMED