A Network Tomography Approach to Imputing OD Traffic in an IP/MPLS Network#

Claudio Risso (Universidad de la República – FING), Carlos Martínez (LACNIC), Pablo Cuello (ANTEL)

• Why: Origin-destination traffic matrices drive capacity planning, what-if/failure analysis, anomaly detection, and interconnect optimization — and per-LSP statistics in IP/MPLS networks are unreliable precisely because they are not standardized.
• What: A network-tomography method that infers OD traffic by combining LSP statistics with trusted physical-interface counters. Production validation on ANTEL's international network (May 2025: 31 nodes, 930 tunnels, 128 directed links, 5-minute samples): 47,430 model parameters fit against 261,248 measurements (~5.5:1) via positivity-constrained least squares that exploit the daily traffic profile's temporal correlation. The top 25 LSPs (2.69% of tunnels) carried 60% of traffic; solving sequentially from the heaviest LSPs first stabilizes the initialization-sensitive iterative fit. The "blackbox" mode is technology-independent (native IP, MPLS, Segment Routing) and works even where LSP statistics are absent.
• How: Linear model Y = AX where A holds ECMP-aware routing fractions per link/tunnel, plus a piecewise-linear daily-profile constraint matrix B; constrained least squares; the toy example makes the underdetermination explicit (34 link measurements vs 132 OD unknowns).
• Where applicable: Validated on ANTEL's production network — org-specific data — but the method generalizes to any IP/MPLS backbone with interface counters and known topology. Transient failures/restorations are explicitly out of scope, and results are approximations of an inherently underdetermined problem.