Traffic engineering: steering traffic when the shortest path is the wrong one
The shortest path ignores load and capacity; steering traffic around it is not free.

A routing protocol picks the shortest path from an additive metric. It knows the topology, not the load or the free capacity of each link. When a short link saturates while a longer parallel path sits idle, the shortest-path answer is the wrong one. Traffic engineering (TE) is the set of mechanisms that steer traffic onto paths a pure shortest-path computation would not choose. The field’s overview defines it as “that aspect of Internet network engineering dealing with the issue of performance evaluation and performance optimization of operational IP networks” (rfc3272.txt lines 177–179), and names routing control as its subject: “the control and optimization of the routing function, to steer traffic through the network in the most effective way” (lines 203–205).
The problem the shortest path creates
Traditional shortest-path-first (SPF) interior gateway protocols (IGPs) “have limited control capabilities for traffic engineering” (rfc3272.txt lines 2548–2549), and the limit is in the metric: “since IGPs always use the shortest paths (based on administratively assigned link metrics) to forward traffic, load sharing cannot be accomplished among paths of different costs” (lines 2554–2557). Two failure shapes follow. A source whose traffic exceeds a link on its shortest path congests that link “while a longer path between these two nodes may be under-utilized” (lines 2559–2562); and the shortest paths of several sources can overlap on the same link. Routing is “topology driven, so bandwidth availability and traffic characteristics are not factors considered in routing decisions” (rfc2702.txt lines 329–331). The congestion that results is a mapping problem, not a capacity shortage, and it appears even with spare capacity elsewhere: “These scenarios manifest even when feasible alternate paths with excess capacity exist” (rfc2702.txt lines 350–351).
A loop, not a setting
Traffic engineering is not a setting but an iterative loop: the process model is “iterative” and its phases are “repeated continually” (rfc3272.txt lines 1183–1184): state the policies that govern the network, acquire measurement data, analyse the network state and characterise the traffic, then optimise. Measurement comes first for a reason: “Measurement is crucial to the traffic engineering function. The operational state of a network can be conclusively determined only through measurement” (lines 1254–1256). The traffic matrix — demand between origin and destination nodes — “may be constructed as part of the analysis process” (line 1218).
the shortest path: chosen from an
additive metric, blind to load
|
v (a wrong answer, measured)
+----------------------------------------+
| 1 policies -> 2 measure -> |
| ^ 3 analyse -> |
| | 4 optimise |
| +-------- repeat -------< |
+----------------------------------------+
|
v
a path chosen for its constraints,
not for its cost in hops
What engineering means: constraints on the path
A path can be decided against stated requirements, not only computed. “Constraint-based routing computes routes to fulfill requirements subject to constraints” (rfc3272.txt lines 2601–2602), and the result is deliberate: “Routes computed through constraint-based routing are not necessarily the shortest paths” (lines 2607–2608). The constraints are “bandwidth, hop count, delay, and administrative policy instruments such as resource class attributes” (lines 2602–2604), where a resource class is an “administratively specified property of the link” (line 2630) — the basis of what operators call affinity.
The levers, in order of cost
The mechanisms differ in who chooses the path and how much state they add. Routing can be “administratively controlled at various levels of abstraction including the manipulation of BGP attributes and manipulation of IGP metrics” (rfc3272.txt lines 970–972), and the path of an explicit label switched path (LSP) “can be computed and established in various ways including: (1) manually, (2) automatically online using constraint-based routing processes implemented on label switching routers, and (3) automatically offline” (lines 976–981).
- The IGP metric. The cheapest lever is the metric itself, the administratively assigned weight on a link: raising it on a congested link moves traffic — “increasing the OSPF or IS-IS metric of a congested link until enough traffic has been diverted from that link” (
rfc3272.txtlines 3321–3322). But “Modifying IGP metrics to control traffic routing tends to have network-wide effect” (lines 2589–2590), so the change can move traffic no one meant to move. - Explicit routing in MPLS. Multiprotocol Label Switching (MPLS) assigns a packet to a forwarding class once, at the edge, and a label then stands for the route: “a label can be used to represent the route” (
rfc3031.txtlines 264–265). Route selection is hop by hop or explicit, and “In an explicitly routed LSP, each LSR does not independently choose the next hop; rather, a single LSR, generally the LSP ingress or the LSP egress, specifies several (or all) of the LSRs in the LSP” (lines 1261–1263). - Signalling with RSVP-TE. RSVP-TE — the Resource Reservation Protocol with traffic-engineering extensions — turns explicit paths into tunnels: “A key application of LSP tunnels is traffic engineering with MPLS as specified in RFC 2702” (
rfc3209.txtlines 41–43), and bandwidth “can be allocated to an LSP tunnel using standard RSVP” (line 294). The tunnel is movable — you can “dynamically reroute the session by simply changing the EXPLICIT_ROUTE object” (lines 437–439) — and “make-before-break” (line 646) builds the replacement before the old path is torn down. - Centralised path computation (PCE). When the head-end router cannot see enough of the network, a Path Computation Element computes the path: a PCE “is an entity that is capable of computing a network path or route based on a network graph, and of applying computational constraints during the computation” (
rfc4655.txtlines 191–193).
What each lever costs
Every lever buys control with state and complexity, and the higher levers also demand the skills to run them. An explicit path is a path someone must know when to move. With a single centralised computation, paths derived from the same link-state information “compete for limited resources within the network” (rfc4655.txt lines 540–544), and the model does not reach the open Internet: “PCE is not considered to be a solution that is applicable to the entire Internet” (lines 532–533). Feedback that chases load also risks oscillation, and the guidance is to prefer stability: “the tradeoff should be made in favor of stability” (rfc3272.txt lines 2480–2483).
When the shortest path is the right answer
Traffic engineering earns its cost only against a specific failure. Its own motivation names two congestion situations, and only one belongs to it: where resources are inadequate, the remedy is “(i) expansion of capacity, or (ii) application of classical congestion control techniques, or (iii) both” (rfc2702.txt lines 251–253). If a link is too small, more capacity is cheaper and simpler than steering around it. Where steering is warranted, the cheapest honest form is a few paths established “manually” (line 978), not an online computation; it pays where capacity cannot be expanded, where a latency or contractual requirement fixes the path, or where a few large demands can be moved onto explicit paths without redesigning the routing domain. The batch’s sources hold the principles and the architecture; they carry no data on any real traffic profile, so this sheet says what the mechanisms are for and what they cost, not which one a given network should run.
Level and prerequisites
L4 — architecture and mechanism selection. It assumes the L1 and L2 forwarding vocabulary (routing table, next hop, default route) and the L3 dynamic-routing foundation, and stays above configuration: no vendor syntax, command or platform default is given.
Where to go next
References
- RFC 3272 — Overview and Principles of Internet Traffic Engineering (Awduche, Chiu, Elwalid, Widjaja, Xiao; Informational; May 2002) — the definition of Internet traffic engineering (lines 177–179); steering traffic as the distinctive function of routing control (lines 203–205); the iterative process model (lines 1183–1184); measurement as the only conclusive source of network state (lines 1254–1256); the traffic matrix (line 1218); the limits of SPF IGPs (lines 2548–2549, 2554–2557, 2559–2562); the network-wide effect of IGP-metric changes (lines 2589–2590); constraint-based routing and its constraint set (lines 2601–2604, 2607–2608, 2630); the administrative control of routing and the manual/online/offline explicit-path options (lines 970–981); the reactive IGP-metric lever (lines 3321–3322); the preference for stability (lines 2480–2483).
- RFC 2702 — Requirements for Traffic Engineering Over MPLS (Awduche, Malcolm, Agogbua, O’Dell, McManus; Informational; September 1999) — the two congestion situations and the remedies (lines 251–253); topology-driven routing that ignores bandwidth and traffic characteristics (lines 329–331); congestion even where spare capacity exists (lines 350–351).
- RFC 3031 — Multiprotocol Label Switching Architecture (Rosen, Viswanathan, Callon; Standards Track; January 2001) — a label standing for the explicit route (lines 264–265); the two route-selection options and the explicit mode (lines 1261–1263).
- RFC 3209 — RSVP-TE: Extensions to RSVP for LSP Tunnels (Awduche, Berger, Gan, Li, Srinivasan, Swallow; Standards Track; December 2001) — LSP tunnels as the traffic-engineering application of MPLS (lines 41–43); bandwidth reserved over RSVP (line 294); dynamic rerouting via the explicit-route object (lines 437–439); make-before-break (line 646).
- RFC 4655 — A Path Computation Element (PCE)-Based Architecture (Farrel, Vasseur, Ash; Informational; August 2006) — the definition of a PCE (lines 191–193); competing demands for limited resources (lines 540–544); the limit of applicability (lines 532–533).