Skip to content

Topologies

XLRON ships with all 119 real-world optical network topologies from TopologyBench in addition to the historically standard research topologies (NSFNET, COST239, USNET, JPN48, German17, CONUS, …). Both directed (each fibre direction is a separate link) and undirected variants are pre-converted and cached, so you can run on any topology by just changing one flag:

--topology_name=coronet_directed
--topology_name=geant_undirected
--topology_name=germany50_undirected
--topology_name=usa100_directed     # 100 nodes, 342 directed links
--topology_name=tataind_directed    # 143 nodes, 362 directed links

To list all available TopologyBench topologies (or regenerate from source):

python xlron/data/topologies/topology_bench_to_xlron_conversion.py --list
python xlron/data/topologies/topology_bench_to_xlron_conversion.py --download  # re-download from upstream

If your topology JSON includes latitude and longitude per node, the render visualisation will use them to seed geographic node placement.

Bundled research topologies

Topology Nodes Directed links Avg degree Avg SP (hops) Avg SP (km)
COST239 11 52 4.73 1.56 1,810
NSFNET 14 44 3.14 2.12 2,054
German17 17 48 2.82 2.85 799
USNET 24 86 3.58 2.99 2,993
JPN48 48 164 3.42 5.21 1,201
CONUS 75 198 2.64 6.45 2,687
USA100 (TopologyBench) 100 342 1.71 6.52 2,684
TataInd (TopologyBench) 143 362 1.26 9.87 1,972

Why this matters

Most published RL-for-RMSA studies use a small handful of topologies (NSFNET, COST239, USNET, JPN48). Conclusions drawn on those tiny topologies do not always generalise — both because the action space is small and because the path diversity is limited. The largest topology any prior dynamic-RMSA RL study had attempted before the Graph Transformer paper was JPN48 (48 nodes); USA100 and TataInd are 2–3× larger and have substantially more candidate paths per node-pair.

XLRON makes those big topologies a one-flag change. The same Graph Transformer architecture and training script that is benchmarked on NSFNET also trains successfully on TataInd in 5 hours on a single H100 — see the reproduction guide.

Adding your own topology

Drop a JSON file into xlron/data/topologies/ describing the graph in node-link format with distance (km) on each edge, then refer to it via --topology_name=<filename without .json>. Optional fields:

  • latitude / longitude per node (used by render for geographic layouts).
  • The xlron/data/topologies/topology_bench_to_xlron_conversion.py script is a worked example of converting an external topology source to the XLRON format.

K-shortest paths are pre-computed and cached under xlron/data/topologies/ksp/ per (topology, k, sort_criteria).

Credit

The 119-topology bundle is a redistribution of TopologyBench: A Comprehensive Repository of Real-World Optical Network Topologies by Robin Matzner et al. If you use these topologies in published work, please cite TopologyBench:

@article{matzner_topology_2024,
  author  = {Matzner, Robin and others},
  title   = {{TopologyBench}: A Comprehensive Repository of Real-World Optical Network Topologies},
  year    = {2024}
}