Reproducing the ECOC 2026 capacity bounds paper

This page documents how to reproduce the figures and tables from:

Doherty, M.*, Deng, B.*, Beghelli, A., Toni, L., Bayvel, P. Comparison of Dynamic Elastic Optical Network Capacity Bound Estimation Methods. Submitted to ECOC 2026 (pending publication). *Equal contribution.

The paper compares three methods of estimating the capacity of dynamically-operated elastic optical networks across the 118 real-world topologies from TopologyBench:

• KSP — K-shortest-path heuristics (KSP-FF and FF-KSP, picking the better of the two per topology) as a realistic on-line baseline.
• CS — Cut-set bound: each bottleneck cut is treated as one aggregated resource of n_i fibres per slot, removing the per-link routing constraint within a cut while still enforcing spectrum continuity across cuts. Top-10 and top-100 most congested cuts are evaluated.
• RPD — Resource-prioritized defragmentation: an omniscient controller that reassigns all active connections on would-be-blocking events, prioritising by slot-hops.

The capacity metric is the offered load (Erlangs) at 0.1% blocking probability.

All scripts live under experimental/capacity_bounds_survey/.

---

Run the full sweep

The sweep runs in six phases. The master entrypoint is resume-safe (each phase skips topologies that already have output files), so it can be killed and restarted at any time.

uv run python experimental/capacity_bounds_survey/run_all.py

Or run individual phases as standalone scripts:

Phase	Script	What it does
1	01_discover_load_ranges.py	Probe each topology to find the load range that brackets 0.1 % blocking.
2	02_compare_heuristics.py	KSP-FF vs FF-KSP across the load range; pick the winner per topology. Output: results/heuristic_selection.json.
3	03_run_cutset_bounds.py	CS bound with the top-10 and top-100 most congested cuts. Calls xlron.bounds.cutsets_bounds.
3b	03b_run_cutset_bounds_top1pct.py	CS bound with the top-1 % of cuts (sensitivity check).
4	04_run_rr_bounds.py	RPD bound. Calls xlron/bounds/reconfigurable_routing_bounds.py.
5	05_run_k_sensitivity.py	KSP K-sensitivity sweep on the larger topologies (K = 25, 50, 75, 100).
6	06_analyze_results.py	Interpolate every method's bp-vs-load curve to find the load at 0.1 % blocking, write results/summary_table.csv, and generate the figures.

config.py defines the topology list and per-method defaults; warmup_ksp_cache.py precomputes the K-shortest-path cache so the bounds runs don't pay that cost.

Compute requirement. This is a heavy sweep — 118 topologies × 4 methods (KSP-FF, KSP-FF/FF-KSP winner, CS-top10, CS-top100, RPD) × multiple loads each, on top of the K-sensitivity sub-sweep. Expect tens of GPU-hours on an A100 even with the resume-safe restart logic. Single topologies are individually reproducible by editing config.get_topology_list().

---

Plot the figures

06_analyze_results.py writes plots to experimental/capacity_bounds_survey/results/figures/. The three figures used in the paper are:

Paper figure	Plot file
Heuristic-selection summary (which of KSP-FF / FF-KSP wins per topology)	heuristic_selection.png
Bound gap scatter (Δ % over KSP for CS-top10, CS-top100, RPD)	bounds_gap_scatter.png
All bounds normalised by edge count	all_bounds_normalized_by_edges.png

Copies of the rendered PNGs that ship with the manuscript live alongside the LaTeX source in the ECOC-2026-Capacity-Bounds sister repository.

---

Headline numbers

From summary_table.csv aggregated over 118 topologies:

Method	Median Δ over KSP at 0.1 % blocking
Cut-set, top-10 cuts	+9.4 %
Cut-set, top-100 cuts	+5.8 %
Resource-prioritised defragmentation	+17.0 %

The larger RPD gap implies that spectrum fragmentation is a bigger source of lost capacity than routing inefficiency — spectrum-allocation is the higher-leverage optimisation target. CS can underestimate (negative Δ) on a small number of large topologies for the mathematical reason set out in cutsets_explanation.tex (feasibility scales as \(\bar p^T\) across \(T\) traversed cuts vs \(\bar q^H\) across \(H\) physical hops; on large topologies \(T \gg H\)).

See Capacity Bound Estimation for the underlying XLRON estimator implementations (xlron.bounds.cutsets_bounds, xlron.bounds.reconfigurable_routing_bounds) and example single-topology commands.

---

Citation

Pending publication. Until the proceedings appear, please use:

@unpublished{doherty_deng_capacity_bounds_2026,
  author  = {Doherty, Michael and Deng, Bohua and Beghelli, Alejandra and Toni, Laura and Bayvel, Polina},
  title   = {Comparison of Dynamic Elastic Optical Network Capacity Bound Estimation Methods},
  note    = {Submitted to ECOC 2026},
  year    = {2026}
}