← Interview Prep

Segment Routing — Theory

Source routing for modern IP/MPLS backbones: SIDs, SR-MPLS vs SRv6, TI-LFA, and Flex-Algo — the parts worth explaining at a whiteboard.

Segment Routing (SR) rethinks how a packet is steered through a network. Instead of the control plane signaling and holding per-flow / per-LSP state in every core node (LDP, RSVP-TE), the ingress node encodes the entire path as an ordered list of segments in the packet header. The core just executes instructions and forwards.

SR is source routing: the head-end decides the path and writes it into the packet as an ordered segment list. Each segment is an instruction ("go to node X via the shortest path", "cross this specific link", "run this function"). The core keeps no per-flow state — it only needs to know the globally-agreed meaning of each segment, which the IGP already floods.

Core idea — source routing, stateless core

A segment is an identifier called a SID (Segment Identifier). The ingress imposes a stack/list of SIDs; each hop pops or reads the active segment, forwards toward it, and the next segment becomes active. The path lives in the packet, not in the routers.

The practical payoff: a full-mesh TE topology that would need N² RSVP LSPs (and state in every hop) becomes a stateless SID list computed once at the edge.

SID types

Segments come in two broad flavors: global segments (meaningful network-wide, advertised by the IGP, everyone installs the same forwarding) and local segments (meaningful only on the node that allocated them).

The SRGB — Segment Routing Global Block

The SRGB is the range of label values (in SR-MPLS, e.g. 16000–23999) reserved on every node for global segments. A Prefix-SID is advertised as an index; each node computes the local label as SRGB_base + index.

• Keeping a consistent SRGB across the domain means the same prefix maps to the same label everywhere — this makes SIDs globally meaningful and troubleshooting sane. (SR permits per-node SRGBs, but a homogeneous SRGB is strongly recommended.)
• Because the ingress just references indices, the whole domain shares one small, predictable label plan instead of the per-node arbitrary labels LDP hands out.
• SRLB (Segment Routing Local Block) is the analogous local range, used for Adjacency-SIDs and other local functions.

Two data planes — SR-MPLS vs SRv6

SR is an architecture; it runs over two data planes. The control-plane concepts (SIDs, IGP flooding, TI-LFA, Flex-Algo) are shared — the difference is how the segment list is carried on the wire.

SR-MPLS — label stack = the path

The IGP (IS-IS or OSPF, with SR extensions) distributes prefix-SIDs, so LDP is no longer needed to build a transport LSP — one protocol (the IGP) both computes reachability and distributes labels. A TE path is just a stack of Node-SIDs (with Adjacency-SIDs where you must pin a link):

Ingress builds label stack to steer PE1 → PE2 via node R5:

   ┌──────────┐
   │  16005   │  ← Node-SID(R5)   (top: go to R5 shortest path)
   ├──────────┤
   │  16002   │  ← Node-SID(PE2)  (then to PE2)
   ├──────────┤
   │  payload │
   └──────────┘

Each core LSR: forward on shortest path to the top SID's prefix,
pop when it becomes the owner (PHP / next-segment), no per-LSP state.

SRv6 — IPv6 addresses as SIDs, "network programming"

In SRv6 each SID is a 128-bit IPv6 address structured as locator:function[:args]. The locator routes the packet to the node that owns the SID (it's just an IPv6 prefix in the IGP); the function is an instruction that node executes — this is SRv6 network programming.

• The segment list is carried in the SRH; the IPv6 destination address is set to the active SID and rewritten as the packet advances through the list (Segments Left counter decrements).
• Common behaviors (functions): End (plain shortest-path to the next SID), End.X (forward out a specific adjacency — the SRv6 Adjacency-SID), End.DT4/DT6 (decapsulate + lookup in a VRF — used for L3VPN), End.DX2 (L2 cross-connect).
• A big draw is service in the SID: VPN/TE encoded directly in IPv6 addresses, so the core needs no MPLS and no separate VPN label plane.
• uSID / micro-SID: full 128-bit SIDs are heavy — the SRH grows fast. uSID packs several short (e.g. 16-bit) micro-instructions into a single 128-bit address, so one IPv6 address can carry a chain of hops. This slashes header overhead and often removes the SRH entirely for short paths.

SRv6 SID layout (locator : function):

   2001:db8:  aaaa :  0100  :: 0
   └─ locator (routes to the node) ─┘ └ function (End, End.DT4, …) ┘

uSID carrier (micro-program in one 128-bit address):

   2001:db8:  0001 0002 0005 0002  ::   ← visit nodes 1→2→5→2 in order
              └── micro-SIDs shifted left as each is consumed ──┘

IGP extensions distribute SIDs

SR reuses the link-state IGP you already run to flood SID information — no new distribution protocol:

• IS-IS: SR sub-TLVs (Prefix-SID sub-TLV, Adjacency-SID sub-TLV, SR-Capabilities/SRGB in the Router Capability TLV). The most common transport-network choice.
• OSPF: SR carried in Opaque LSAs — the Extended Prefix / Extended Link Opaque LSAs (OSPFv3 uses its own extended-TLV LSAs).
• BGP-LS exports the full SR topology (nodes, links, SIDs, SRGB, Flex-Algo definitions) north to a controller / PCE, which computes explicit paths and installs them via PCEP or BGP-SR-TE.

Because reachability and label distribution collapse into one protocol, the failure/convergence story is the IGP's — one SPF, one set of timers — instead of coordinating IGP + LDP (and avoiding the classic LDP/IGP synchronization pitfalls).

TI-LFA — topology-independent fast reroute

TI-LFA (Topology-Independent Loop-Free Alternate) gives guaranteed, loop-free fast reroute with sub-50 ms protection against link, node, or SRLG failure — for essentially any topology. It leans on SR's key property: the ability to express an explicit repair path as a SID list.

• The head-end (PLR — Point of Local Repair) pre-computes, per destination, the SID list that reaches the P-space/Q-space connection point on the path the network will use after it reconverges — usually a Node-SID to a "P" node plus an Adjacency-SID for the final hop.
• Because the repair path is the same one the IGP will settle on, TI-LFA avoids the transient micro-loops that plague plain LFA.
• All in the forwarding plane and pre-installed, so activation on failure is a local, sub-50 ms swap — no signaling, no waiting for SPF.

Flex-Algo — custom IGP metrics

Flexible Algorithm (Flex-Algo) lets the IGP compute shortest paths against a custom optimization objective instead of only the default IGP metric — and binds that path to a prefix-SID. Steering onto a Flex-Algo path is then just "use that SID," no explicit SID list needed.

• An algorithm definition (a number, e.g. 128) specifies: a metric type (IGP metric, TE metric, or min-latency / delay), a calculation type (SPF), and constraints (include/exclude links by admin-group / affinity color).
• Each node advertises a Prefix-SID per algorithm. Sending traffic to a loopback's Algo-128 SID means "take the min-latency path to this node."
• Typical uses: a low-latency plane and a low-cost plane over the same physical topology; or a plane that avoids certain links (e.g. keep sovereign traffic off international links) via affinity constraints.
• Delay-based Flex-Algo pairs with IGP performance measurement (link-delay advertisement) so "min-latency" reflects measured one-way delay, not just static metric.

Benefits vs LDP / RSVP-TE

The headline talking points: no LDP, no per-LSP core state, ECMP-friendly, TE without a signaling protocol, and a simpler overall control plane — plus TI-LFA and Flex-Algo as capabilities that are hard or impossible in the legacy stack.