Do link-weight optimizations deliver at packet level? A packet-level ns-3 validation of OSPF weight-setting traffic engineering on a real backbone
Journal of Al-Farabi for Engineering Sciences · 01.12.2025 · 4 (2) · 428-442
DOI: 10.59746/fzcekr91 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Al-Farabi for Engineering Sciences, online erschienen am 01.12.2025. Verfasst von Mohammed Ibrahim Salman.
Abstract (Originalsprache)
Setting OSPF/IS-IS link weights to balance load is a 25-year-old optimization problem, and every solution — from Fortz and Thorup’s local search to today’s learned traffic-engineering models — has been evaluated with a fluid model whose score is the maximum link utilization (MLU). We ask whether a weight vector that lowers fluid MLU also lowers the delay, loss and flow completion time that packets experience. We couple a Python optimizer (Fortz–Thorup-style local search, simulated annealing, a CVaR-robust variant) with ns-3.41 running OSPF-like routing over the real Abilene backbone, its real 5-minute traffic matrices and the operator’s real 2003 weights, and we add per-flow hashed ECMP with per-router seeds to ns-3. Across 1,000+ simulations we find: (i) near saturation the fluid model ranks loss and tail latency almost perfectly (Spearman ρ ≥ 0.94) and optimized weights deliver large packet-level gains — at three times the 2004 peak demand, p99 UDP delay 26 ms versus 44 ms (inverse-capacity weights) and 76 ms (operator weights), TCP p99 flow completion 0.9 s versus 3.4 s and 5.5 s, and at four times the peak 0.04 % loss versus 9–11 %; (ii) below saturation the same metric ranks delay backwards, because the operator’s distance-based weights minimize propagation delay — the crossover lies near twice the peak demand; (iii) the gains presuppose real equal-cost splitting: without ECMP they largely vanish, and a shared hash seed across routers (hash polarization) silently removes one of every two splits; (iv) the ordering survives single-link failures (ρ = 0.99 against the fluid model on the failed graph), buffer sizes from 10 to 100 ms and FqCoDel — but not a week of real traffic: weights optimized for the peak hour, whose traffic is dominated by one origin–destination pair, are worse than the vendor default in 79 of 85 other hours once each hour is stressed equally, whereas CVaR-robust weights trained on the twelve busiest hours match the default across the week (0.1 % vs 0 % loss) while keeping most of the peak-hour gain. We also document three fidelity traps — constant-rate sources, polarized hashing and survivorship bias in completion times — and a reproducible pipeline with two small ns-3 patches.
Zitieren
APA
Salman, M. I. (2025). Do link-weight optimizations deliver at packet level? A packet-level ns-3 validation of OSPF weight-setting traffic engineering on a real backbone. Journal of Al-Farabi for Engineering Sciences, 4(2), 428-442. https://doi.org/10.59746/fzcekr91
Vancouver
Salman MI. Do link-weight optimizations deliver at packet level? A packet-level ns-3 validation of OSPF weight-setting traffic engineering on a real backbone. Journal of Al-Farabi for Engineering Sciences. 2025;4(2):428-442. doi:10.59746/fzcekr91
BibTeX
@article{Salman2025,
author = {Salman, Mohammed Ibrahim},
title = {Do link-weight optimizations deliver at packet level? A packet-level ns-3 validation of OSPF weight-setting traffic engineering on a real backbone},
journal = {Journal of Al-Farabi for Engineering Sciences},
year = {2025},
volume = {4},
number = {2},
pages = {428--442},
doi = {10.59746/fzcekr91}
}
EUVIMED