Packet Path/

Segment Routing vs MPLS Explained: Same Labels, Very Different Traffic Engineering

People talk about segment routing as if it replaced MPLS. It did not — at least not the part that forwards packets. SR-MPLS uses the exact same label switching as the MPLS networks built since the early 2000s. What segment routing replaces is the signalling: the protocols that decide which path a packet takes and tell the routers about it. LDP lets every router follow the shortest path on its own. RSVP-TE lets the first router choose a path, but then has to install and maintain that path in every router along it. Segment routing lets the first router choose the path and write it into the packet itself, as a short list of labels — so the routers in the middle just follow the list and remember nothing about the flow. This article puts the three side by side: where the state lives, what the headend decides, and one worked path taken three ways.

Routing · Intermediate · 11 min · November 14, 2026

Two label-switched paths across a dark network background: the upper path follows short straight hops between routers, the lower path detours through two extra routers guided by a glowing list of numbered labels held at the first router.

The short answer

What is the difference between segment routing and MPLS? MPLS is a forwarding method: routers swap short numbered labels instead of looking up long IP addresses, and it is not going away. Segment routing is a way of deciding and signalling the path those labels will take. In traditional MPLS, one of two older protocols hands out the labels: LDP, which simply follows the routing protocol's shortest path, or RSVP-TE, which lets the first router pin a tunnel to an explicit path but must install that tunnel's state in every router it crosses. In segment routing over MPLS (SR-MPLS), the first router — the headend — writes the path into the packet itself as a stack of labels, each label being a segment ID (SID) that names a node or a link. Middle routers execute the top instruction and forget the packet. So the honest comparison is not SR versus MPLS forwarding; it is SR versus LDP and RSVP-TE as the signalling system behind the labels.

Think of three ways to cross a city. LDP is following the road signs: every junction decides locally, you always end up on the standard shortest route, and nobody planned your journey as a whole. RSVP-TE is a booked convoy route: a planner chooses every street in advance, but then has to phone every junction, reserve the route at each one, and keep re-confirming every reservation for as long as the convoy runs. Segment routing is writing a short list of landmarks on the envelope before you set off — 'head for the ring road, then for the bridge, then for the depot' — and letting each junction just read the next landmark. The planner did the thinking once, at the start; the junctions hold no booking for you at all.

In plain EnglishLabel: a short number stuck on the front of a packet; routers read the number, swap it for the next hop's number, and send the packet on — faster and simpler than re-deciding the whole route at every hop. Headend: the first router on a path, the one that sticks the labels on. Midpoint: any router in the middle of the path. Segment / SID: an instruction with an ID number — 'go to node X' or 'leave by this specific link'. State: memory a router must keep about a path or flow. State is what runs out, what has to be refreshed, and what has to be rebuilt after a failure. Traffic engineering (TE): deliberately sending traffic by a route other than the shortest one — to dodge congestion, use idle links, or keep latency low.

The three signalling systems, side by side

QuestionMPLS with LDPMPLS with RSVP-TESR-MPLS (segment routing)
What forwards the packet?An MPLS label, swapped hop by hopAn MPLS label, swapped hop by hopAn MPLS label stack — the same forwarding plane, per RFC 8660
Who chooses the path?Nobody as a whole — each router follows the IGP shortest path to the destinationThe headend (or a controller talking to it) chooses an explicit pathThe headend (or a controller talking to it) chooses the path and encodes it as a segment list
How do routers learn the labels?LDP distributes a label binding for each prefix, hop by hopRSVP-TE signals each tunnel along its path and installs label state as it goesThe IGP (IS-IS or OSPF) advertises SIDs alongside the topology — no separate label protocol needed for SR paths
Where does per-path state live?No per-path state as such — label bindings are per prefix, and every flow to a prefix takes the same IGP pathEvery midpoint holds state for every tunnel crossing it, refreshed periodically for the tunnel's whole lifeOnly the headend holds the policy and its segment list; midpoints hold no per-flow or per-path state at all (RFC 8402, RFC 9256)
Can you force one specific link?No — not without changing the IGP metrics that move everyone else tooYes — the explicit route names the hopsYes — add that link's Adj-SID to the list; the rest of the list can stay loose
What does a new path cost the core?Nothing new — but you cannot have a custom pathSignalling plus state in every router on the path, for every tunnel, foreverNothing in the core — the headend pushes a different label stack
Scaling pressure pointThe IGP itself; paths are whatever it computesMidpoint memory and refresh load grow with the number of tunnels crossing the routerHeadend policy count and label-stack depth — how many labels the hardware can push and read
Bandwidth reservationNoneYes — reserving bandwidth along the path is a built-in part of the designNot by itself — SR steers paths; reservation, where needed, is handled by the controller or policy layer above it
One topology, two paths — the worked example below PE1 P1 P2 PE2 P3 P4 Shortest path (LDP): PE1 – P1 – P2 – PE2 — congested at peak Segment list [16004, 16005]: shortest to P4 (via P3), then on to PE2 The list is written once, at PE1. P3 and P4 keep no record of this flow.

Read the state row twice — it is the whole argument. Everything else in the table is a consequence of it. RSVP-TE's power and its cost are the same fact (the path is real, installed state in the network), and SR's elegance and its limit are the same fact too (the path exists only as instructions carried in the packet, so the packet can only carry so many instructions, and nothing in the core has promised it any bandwidth).

Traditional MPLS: signs, or reservations

LDP (Label Distribution Protocol) is the simple end of the trade. Routers already know the shortest path to every destination from the IGP — IS-IS or OSPF. LDP simply attaches a label to each destination prefix and tells the neighbours, so a packet can be label-switched along that shortest path instead of IP-routed. It is robust, automatic, and almost free to run. Its limit is built into the deal: the path is the shortest path, full stop. If the shortest path to a site is melting at peak hour while a longer path sits idle, LDP has no opinion about that — and neither can you, short of bending IGP metrics and moving every other flow with them.

RSVP-TE is the engineered end. The headend computes an explicit path — avoiding a congested link, honouring a latency budget, reserving bandwidth — and RSVP signals that path hop by hop, installing label state in every router it crosses and refreshing that state for the tunnel's entire life. This genuinely works and carried the traffic engineering of large backbones for two decades. The bill arrives as scale: a busy midpoint can hold state for thousands of tunnels it neither starts nor ends, every one of them signalled, refreshed, and torn down through it. Adding a tunnel is a network-wide event, and a midpoint failure means rebuilding other people's paths, not just its own.

Segment routing: the path rides in the packet

Segment routing keeps the MPLS forwarding plane exactly as it is and replaces the signalling with two moves. First, the IGP itself advertises the instructions: every node announces a Prefix-SID — for a router's own loopback, a Node-SID — meaning 'the shortest path to me', and an Adj-SID for each of its links, meaning 'out this exact interface'. In SR-MPLS a SID becomes a label by a shared rule: each node reserves a label range, the Segment Routing Global Block (SRGB), and a global SID's label is the SRGB base plus the SID's index — so if the SRGB starts at 16000 and a node's index is 4, its Node-SID is label 16004 on every router in the domain. No per-tunnel negotiation is needed to agree that number; the IGP carried it.

Second, the headend does the steering. To send a flow by a chosen route, it pushes a short ordered stack of SIDs onto the packet — an SR Policy's segment list, in the language of RFC 9256. Each midpoint reads only the top label: a node segment means 'forward along the shortest path toward that node, spreading over equal-cost links'; an adjacency segment means 'send it out this one link, whatever the shortest path says'. When a segment is complete its label comes off the stack and the next instruction takes over. A list of two or three well-chosen segments can pin a path across a whole backbone, because each node segment covers the entire shortest-path stretch to that node for free — you only spend a segment where you need to bend the route.

One worked path, taken three ways

Use the diagram's topology. PE1 sends a latency-sensitive flow to PE2. The IGP shortest path is PE1 – P1 – P2 – PE2, and the P1 – P2 link is congested at peak. The bottom route — PE1 – P3 – P4 – PE2 — is one hop longer and empty. Every node uses an SRGB starting at 16000; P4 has index 4 (Node-SID label 16004) and PE2 has index 5 (Node-SID label 16005).

With LDP: the flow goes over P1 and P2 and queues with everything else. There is no configuration on PE1 that sends just this flow the other way; the labels only exist to follow the IGP.

With RSVP-TE: PE1 signals an explicit tunnel PE1 – P3 – P4 – PE2. P3 and P4 each install and keep refreshing state for this tunnel, alongside every other tunnel crossing them, for as long as it exists. The path is exactly as ordered — and exactly as stateful.

With SR-MPLS: PE1 pushes the two-label stack [16004, 16005]. The top label, 16004, says 'shortest path to P4' — from PE1 that path runs through P3, which forwards toward P4 knowing nothing about this flow's final destination or policy. At P4 the first segment is done, 16005 is on top, and it says 'shortest path to PE2' — the direct link. P3 and P4 hold no state for the flow; if PE1 wants a different route tomorrow, it pushes a different stack and the core notices nothing. That is the whole mechanism — and if you want to feel how segment choice bends a path, the free lab linked at the end of this article lets you build lists like this one and watch where the packets actually go.

What segment routing does not do, and the honest limits

It does not remove the IGP — it leans on it. SIDs ride in IS-IS or OSPF; a broken IGP is a broken SR network. It does not reserve bandwidth. An SR Policy steers; guaranteeing capacity along the path is a separate job, usually given to the controller that computes the policies. The stack is finite. Hardware can only push and process so many labels (its maximum SID depth), so segment lists are kept short by design — loose node segments doing the long stretches, as in the worked example. The SRGB must be planned. The base-plus-index trick only produces one agreed label per node if the domain's ranges and indexes are assigned consistently. Migration is coexistence, not a flag day: SR and LDP interworking is standardised (RFC 8661) precisely because real networks run both for years. And SRv6 is a different data plane — the same source-routing idea with IPv6 addresses as SIDs in an IPv6 routing header, not an MPLS label stack; choosing between SR-MPLS and SRv6 is a greenfield-versus-installed-base decision, separate from the LDP and RSVP-TE comparison above. For the mechanics of SIDs, TI-LFA fast reroute and Flex-Algo in depth, read the companion guide to segment routing explained — this article is the comparison; that one is the machine room.

The decision, in three questions, and the mistakes

1. Do you need custom paths at all? If shortest-path transport serves the network, LDP's zero-engineering simplicity is a feature, and adding TE machinery buys nothing but new failure modes.

2. If you need custom paths, what should scale — midpoint state, or headend policy? Established RSVP-TE estates with working bandwidth reservation have a real asset, not just inertia. New traffic engineering, especially controller-driven, usually picks the model where a new path costs the core nothing: SR.

3. What computes the paths? Both RSVP-TE and SR let a headend or a controller decide — the difference is what that decision costs to install. A controller steering SR headends changes paths by changing lists at the edge; the same controller working through RSVP-TE is asking every midpoint to re-signal.

The mistakes to avoid. Saying 'we replaced MPLS with segment routing' — you replaced the signalling; the labels still forward every packet, and troubleshooting skills transfer. Writing segment lists hop by hop, Adj-SID by Adj-SID, when two node segments would do — long stacks hit hardware depth limits and brittle paths. Assuming an SR path implies reserved bandwidth — nothing was reserved. And forgetting the SRGB plan during a merger or a new vendor's arrival, so the same index quietly means different nodes in different parts of the domain.

Fact-check sourcesSegment routing as source routing with per-flow state only at the ingress or headend, segments encoded as an MPLS label stack in SR-MPLS with no change to the forwarding plane, Prefix-SID as the ECMP-aware shortest path to a prefix or node and Adj-SID as one specific adjacency — RFC 8402, Segment Routing Architecture. SR-MPLS forwarding behaviour, the Segment Routing Global Block, and mapping a SID index to a label as SRGB base plus index — RFC 8660. SR Policy as a headend-instantiated ordered segment list identified by headend, colour and endpoint, with intermediate per-path state eliminated — RFC 9256. That SR paths in MPLS networks do not require LDP or RSVP-TE signalling, while being able to coexist and interwork with LDP — the IGP extension RFCs for SR-MPLS and RFC 8661. The SRGB starting at 16000 and the node indexes in the worked example are this article's own worked numbers, chosen so the base-plus-index arithmetic can be checked by hand; a production domain uses its own assigned range and indexes.

Try it yourself

Build the worked example with your own hands in the free segment routing traffic engineering lab: pick a headend and a destination, write a segment list, and watch the flow follow your list instead of the shortest path — then delete a segment and see which stretch falls back to IGP behaviour. When the lab's paths make sense, come back to the table above and answer the three decision questions for a network you know: does it need custom paths, and if it does, where should the state live?

Key takeaways

Keep reading