Abstract
Traditional RSVP-TE MPLS deployments have served WAN traffic engineering well for over two decades, but the operational burden of maintaining distributed signaling state across thousands of LSPs has pushed large-scale operators toward Segment Routing MPLS (SR-MPLS). By encoding forwarding intent directly in the packet header as an ordered stack of SIDs (Segment Identifiers), SR-MPLS eliminates per-LSP state from transit nodes entirely, shrinking the operational surface dramatically while retaining explicit path control. This piece examines where SR-MPLS delivers on that promise and where the tradeoffs bite.
How SR-MPLS Eliminates Signaling State
In an RSVP-TE network, every transit router along an LSP holds soft state - refreshed periodically, synchronized across failures. At scale, this becomes a liability: convergence after a link failure requires distributed state teardown and re-signaling, which can stretch into seconds depending on refresh timers and LSP count. SR-MPLS replaces this with a source-routing model. The ingress node (or a PCE, Path Computation Element) pushes a label stack encoding the entire path. Transit nodes simply swap or pop labels without any per-path knowledge. Cisco IOS-XR and Junos both support SR-MPLS with IS-IS or OSPF extensions (RFC 8665, RFC 8667) advertising node and adjacency SIDs, and PCE integration via PCEP (RFC 8281) enables centralized path computation across the topology.
Traffic Engineering Policies and Flexible Algorithms
SR-MPLS introduces the concept of Flexible Algorithms (Flex-Algo, RFC 9350), which allow operators to define custom topology subsets - for instance, a low-latency graph excluding high-delay satellite links, or a path constrained to equipment within a specific regulatory boundary. Each Flex-Algo gets its own SID space, so packets destined for the low-latency graph simply carry the corresponding node SID computed under that algorithm. This eliminates the need for separate RSVP-TE tunnels for each service class. Operators at major carriers including NTT and Deutsche Telekom have reported using Flex-Algo to separate latency-sensitive voice traffic from bulk data flows without maintaining parallel tunnel infrastructure.
Operational Tradeoffs and Label Stack Depth
SR-MPLS is not without constraints. The label stack grows linearly with path segment count, and hardware label stack depth limits on older ASICs - often 8-12 labels on merchant silicon like Broadcom Trident series - cap the number of explicit hops. This is generally manageable on intra-domain deployments but becomes a real constraint for inter-domain SR paths. TI-LFA (Topology-Independent Loop-Free Alternates) provides fast reroute without pre-computing backup tunnels, restoring traffic in under 50ms for link failures, but its coverage depends on topology. Operators should validate TI-LFA coverage using simulation before decommissioning legacy RSVP-TE FRR paths on high-value circuits.
Migration Path from RSVP-TE
A gradual migration is straightforward: IS-IS or OSPF SR extensions can be enabled alongside RSVP-TE, with new services provisioned on SR-MPLS while legacy LSPs remain until their circuits or contracts end. Segment Routing Path Monitoring (SRPM) and the SR Policy model (draft-ietf-spring-segment-routing-policy) provide operational tooling comparable to existing RSVP-TE management workflows, and integration with existing Kafka-based telemetry pipelines via gRPC/gNMI closes the observability gap.