Summary

  • BMWG revision 09, submitted on 22 September, describes laboratory benchmarking for SR-MPLS and SRv6 forwarding; it is an active Internet-Draft, not an approved RFC or device result.
  • Its policy-scaling procedure checks traffic preservation and ECMP/wECMP function as segment lists, candidate paths and policies grow, while explicitly excluding ECMP/wECMP performance benchmarking.
  • A result is interpretable only with the test's SID depth, policy counts, traffic mix, oversubscription and hash inputs; none of these laboratory fields alone establishes live-network capacity.

Suppose a procurement table says a router “passed ECMP at scale.” The phrase can describe at least two very different experiments. One may measure throughput as traffic is spread across paths. The other may merely verify that packets still arrive, without drops, while increasingly many paths and policies are configured. The distinction cannot be recovered from a headline pass mark.

The BMWG's 22 September revision of Benchmarking Methodology for Segment Routing Forwarding makes that distinction unusually visible. Section 5.7 stages a forwarding-scale exercise at an SR Policy headend. It starts with one policy and candidate path, then tests maximum SID depth, more segment lists in a candidate path, more candidate paths in a policy, and more policies on the device. The procedure checks that policies are installed and active where expected and that traffic continues to flow without loss. ECMP and weighted ECMP are exercised as functions at scale. The text also says that benchmarking their performance is outside its scope, as is measuring control-plane performance.

That qualification matters because a functional check is valuable but not interchangeable with a multipath speed curve. A policy count is a configuration dimension, not packets per second. A no-drop observation has an offered load and duration. A successful weighted distribution depends on the tested weights and hash fields. Moving any one of these details can change what a second laboratory is able to repeat.

The proposed reporting format therefore asks for more than throughput, latency and frame loss. It calls for interface numbers and oversubscription, the upstream/downstream traffic split, SID-list and policy counts, compression details, SRv6 behaviors, payload types, policy construction and steering methods, and the ECMP hash algorithm, input fields and balancing interfaces. These are properties of a particular test record, not specifications already satisfied by a named product.

Revision 09 also explains its relationship to earlier benchmarking RFCs. It lists SR-specific differences such as TCP as well as UDP traffic, trial duration adjusted around protocol hold times, and ECMP/wECMP in scaling checks. The distinction from RFC 5695 is important: that earlier MPLS method excluded ECMP to maintain deterministic forwarding during its tests. The new text does not simply rename an old benchmark; its additional test dimensions need their own interpretation.

There are firm status and location limits. The document remains an active, intended-Informational Internet-Draft in IESG AD Followup, with two DISCUSS positions still recorded at research time. Its methodology is for an isolated lab, not deployed Internet paths. It reports no vendor scores, production traffic, service-level outcomes or security incident. A claimed field result would need independent evidence outside this draft.

Sources