Topology Independent Fast Reroute using Segment Routing

DLFA: Remote LFA with Directed forwarding. FRR: Fast Re-route. IGP: Interior Gateway Protocol. LFA: Loop-Free Alternate. LSDB: Link State DataBase. PLR: Point of Local Repair. RL: Repair list. RLFA: Remote LFA. RSPT: Reverse Shortest Path Tree. SID: Segment Identifier. SLA: Service Level Agreement. SPF: Shortest Path First. SPT: Shortest Path Tree. SR: Segment Routing. SRGB: Segment Routing Global Block. SRLG: Shared Risk Link Group. TI-LFA: Topology Independant LFA.

Segment Routing aims at supporting services with tight SLA guarantees . By relying on SR this document provides a local repair mechanism for standard link-state IGP shortest path capable of restoring end-to-end connectivity in the case of a sudden directly connected failure of a network component. Non-SR mechanisms for local repair are beyond the scope of this document. Non-local failures are addressed in a separate document . The term topology independent (TI) refers to the ability to provide a loop free backup path irrespective of the topologies used in the network. This provides a major improvement compared to LFA and remote LFA which cannot provide a complete protection coverage in some topologies as described in . For each destination in the network, TI-LFA pre-installs a backup forwarding entry for each protected destination ready to be activated upon detection of the failure of a link used to reach the destination. TI-LFA provides protection in the event of any one of the following: single link failure, single node failure, or single SRLG failure. In link failure mode, the destination is protected assuming the failure of the link. In node protection mode, the destination is protected assuming that the neighbor connected to the primary link has failed. In SRLG protecting mode, the destination is protected assuming that a configured set of links sharing fate with the primary link has failed (e.g. a linecard or a set of links sharing a common transmission pipe). Protection techniques outlined in this document are limited to protecting links, nodes, and SRLGs that are within a link-state IGP area. Protecting domain exit routers and/or links attached to another routing domains are beyond the scope of this document Thanks to SR, TI-LFA does not require the establishment of TLDP sessions with remote nodes in order to take advantage of the applicability of remote LFAs (RLFA) or remote LFAs with directed forwarding (DLFA). All the Segment Identifiers (SIDs) are available in the link state database (LSDB) of the IGP. As a result, preferring LFAs over RLFAs or DLFAs, as well as minimizing the number of RLFA or DLFA repair nodes is not required anymore. Thanks to SR, there is no need to create state in the network in order to enforce an explicit FRR path. This relieves the nodes themselves from having to maintain extra state, and it relieves the operator from having to deploy an extra protocol or extra protocol sessions just to enhance the protection coverage. raised several operational considerations when using LFA or remote LFA. Section 3 presents a case where a high bandwidth link between two core routers is protected through a PE router connected with low bandwidth links. In such a case, congestion may happen when the FRR backup path is activated. introduces a local policy framework to let the operator tuning manually the best alternate election based on its own requirements. From a network capacity planning point of view, it is often assumed that if a link L fails on a particular node X, the bandwidth consumed on L will be spread over some of the remaining links of X. The remaining links to be used are determined by the IGP routing considering that the link L has failed (we assume that the traffic uses the post-convergence path starting from the node X). In , we consider a network with all metrics equal to 1 except the metrics on links used by PE1, PE2 and PE3 which are 1000. An easy network capacity planning method is to consider that if the link L (X-B) fails, the traffic actually flowing through L will be spread over the remaining links of X (X-H, X-D, X-A). Considering the IGP metrics, only X-H and X-D can be used in reality to carry the traffic flowing through the link L. As a consequence, the bandwidth of links X-H and X-D is sized according to this rule. We should observe that this capacity planning policy works, however it is not fully accurate. In , considering that the source of traffic is only from PE1 and PE4, when the link L fails, depending on the convergence speed of the nodes, X may reroute its forwarding entries to the remote PEs onto X-H or X-D; however in a similar timeframe, PE1 will also reroute a subset of its traffic (the subset destined to PE2) out of its nominal path reducing the quantity of traffic received by X. The capacity planning rule presented previously has the drawback of oversizing the network, however it allows to prevent any transient congestion (when for example X reroutes traffic before PE1 does).

H --- I --- J | | \ PE4 | | PE3 \ | (L) | / A --- X --- B --- G / | | \ PE1 | | PE2 \ | | / C --- D --- E --- F Based on this assumption, in order to facilitate the operation of FRR, and limit the implementation of local FRR policies, it looks interesting to steer the traffic onto the post-convergence path from the PLR point of view during the FRR phase. In our example, when link L fails, X switches the traffic destined to PE3 and PE2 on the post-convergence paths. This is perfectly inline with the capacity planning rule that was presented before and also inline with the fact X may converge before PE1 (or any other upstream router) and may spread the X-B traffic onto the post-convergence paths rooted at X. It should be noted, that some networks may have a different capacity planning rule, leading to an allocation of less bandwidth on X-H and X-D links. In such a case, using the post-convergence paths rooted at X during FRR may introduce some congestion on X-H and X-D links. However it is important to note, that a transient congestion may possibly happen, even without FRR activated, for instance when X converges before the upstream routers. Operators are still free to use the policy framework defined in if the usage of the post-convergence paths rooted at the PLR is not suitable. Readers should be aware that FRR protection is pre-computing a backup path to protect against a particular type of failure (link, node, SRLG). When using the post-convergence path as FRR backup path, the computed post-convergence path is the one considering the failure we are protecting against. This means that FRR is using an expected post-convergence path, and this expected post-convergence path may be actually different from the post-convergence path used if the failure that happened is different from the failure FRR was protecting against. As an example, if the operator has implemented a protection against a node failure, the expected post-convergence path used during FRR will be the one considering that the node has failed. However, even if a single link is failing or a set of links is failing (instead of the full node), the node-protecting post-convergence path will be used. The consequence is that the path used during FRR is not optimal with respect to the failure that has actually occurred. Another consideration to take into account is: while using the expected post-convergence path for SR traffic using node segments only (for instance, PE to PE traffic using shortest path) has some advantages, these advantages reduce when SR policies () are involved. A segment-list used in an SR policy is computed to obey a set of path constraints defined locally at the head-end or centrally in a controller. TI-LFA cannot be aware of such path constraints and there is no reason to expect the TI-LFA backup path protecting one the segments in that segment list to obey those constraints. When SR policies are used and the operator wants to have a backup path which still follows the policy requirements, this backup path should be computed as part of the SR policy in the ingress node (or central controller) and the SR policy should not rely on local protection. Another option could be to use FlexAlgo () to express the set of constraints and use a single node segment associated with a FlexAlgo to reach the destination. When using a node segment associated with a FlexAlgo, TI-LFA keeps providing an optimal backup by applying the appropriate set of constraints. The relationship between TI-LFA and the SR-algorithm is detailed in . Thanks to SR and the combination of Adjacency segments and Node segments, the expression of the expected post-convergence path rooted at the PLR is facilitated and does not create any additional state on intermediate nodes. The easiest way to express the expected post-convergence path in a loop-free manner is to encode it as a list of adjacency segments. However, this may create a long SID list that some hardware may not be able to push. One of the challenges of TI-LFA is to encode the expected post-convergence path by combining adjacency segments and node segments. Each implementation will be free to have its own SID list compression optimization algorithm. This document details the basic concepts that could be used to build the SR backup path as well as the associated dataplane procedures.

We use to illustrate the TI-LFA approach. The Point of Local Repair (PLR), S, needs to find a node Q (a repair node) that is capable of safely forwarding the traffic to a destination D affected by the failure of the protected link L, a set of links including L (SRLG), or the node F itself. The PLR also needs to find a way to reach Q without being affected by the convergence state of the nodes over the paths it wants to use to reach Q: the PLR needs a loop-free path to reach Q. defines the main notations used in the document. They are in line with . suggests to compute the P-Space and Q-Space properties defined in , for the specific case of nodes lying over the post-convergence paths towards the protected destinations. Using the properties defined in , describes how to compute protection lists that encode a loop-free post-convergence path towards the destination. defines the segment operations to be applied by the PLR to ensure consistency with the forwarding state of the repair node. By applying the algorithms specified in this document to actual service providers and large enterprise networks, we provide real life measurements for the number of SIDs used by repair paths. summarizes these measurements.

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 when, and only when, they appear in all capitals, as shown here.

We define the main notations used in this document as the following. We refer to "old" and "new" topologies as the LSDB state before and after the considered failure. SPT_old(R) is the Shortest Path Tree rooted at node R in the initial state of the network. SPT_new(R, X) is the Shortest Path Tree rooted at node R in the state of the network after the resource X has failed. PLR stands for "Point of Local Repair". It is the router that applies fast traffic restoration after detecting failure in a directly attached link, set of links, and/or node. Similar to , we use the concept of P-Space and Q-Space for TI-LFA. The P-space P(R,X) of a router R w.r.t. a resource X (e.g. a link S-F, a node F, or a SRLG) is the set of routers reachable from R using the pre-convergence shortest paths without any of those paths (including equal-cost path splits) transiting through X. Consider the set of neighbors of a router R and a resource X. Exclude from that set of neighbors that are reachable from R using X. The Extended P-Space P'(R,X) of a node R w.r.t. a resource X is the union of the P-spaces of the neighbors in that reduced set of neighbors w.r.t. the resource X. The Q-space Q(R,X) of a router R w.r.t. a resource X is the set of routers from which R can be reached without any path (including equal-cost path splits) transiting through X. A symmetric network is a network such that the IGP metric of each link is the same in both directions of the link.

The basic algorithm to compute the repair path is to pre-compute SPT_new(R,X) and for each destination, encode the repair path as a loop-free SID list. One way to provide a loop-free SID list is to use adjacency SIDs only. However, this approach may create very long SID lists that hardware may not be able to handle due to MSD (Maximum SID Depth) limitations. An implementation is free to use any local optimization to provide smaller SID lists by combining Node SIDs and Adjacency SIDs. In addition, the usage of Node-SIDs allow to maximize ECMPs over the backup path. These optimizations are out of scope of this document, however the subsequent sections provide some guidance on how to leverage P-Spaces and Q-Spaces to optimize the size of the SID list.

One of the challenges of defining an SR path following the expected post-convergence path is to reduce the size of the segment list. In order to reduce this segment list, an implementation MAY determine the P-Space/Extended P-Space and Q-Space properties (defined in ) of the nodes along the expected post-convergence path from the PLR to the protected destination and compute an SR-based explicit path from P to Q when they are not adjacent. Such properties will be used in to compute the TI-LFA repair list.

We want to determine which nodes on the post-convergence path from the PLR R to the destination D are in the extended P-space of R w.r.t. resource X (X can be a link or a set of links adjacent to the PLR, or a neighbor node of the PLR). This can be found by intersecting the set of nodes belonging to the post-convergence path from R to D, assuming the failure of X, with P'(R, X).

We want to determine which nodes on the post-convergence path from the PLR R to the destination D are in the Q-Space of destination D w.r.t. resource X (X can be a link or a set of links adjacent to the PLR, or a neighbor node of the PLR). This can be found by intersecting the set of nodes belonging to the post-convergence path from R to D, assuming the failure of X, with Q(D, X).

raises scaling concerns about computing a Q-Space per destination. Similar concerns may affect TI-LFA computation if an implementation tries to compute a reverse Shortest Path Tree () for every destination in the network to determine the Q-Space. It will be up to each implementation to determine the good tradeoff between scaling and accuracy of the optimization.

The TI-LFA repair path (RP) consists of an outgoing interface and a list of segments (repair list (RL)) to insert on the SR header. The repair list encodes the explicit post-convergence path to the destination, which avoids the protected resource X and, at the same time, is guaranteed to be loop-free irrespective of the state of FIBs along the nodes belonging to the explicit path. Thus, there is no need for any co-ordination or message exchange between the PLR and any other router in the network. The TI-LFA repair path is found by intersecting P(S,X) and Q(D,X) with the post-convergence path to D and computing the explicit SR- based path EP(P, Q) from P to Q when these nodes are not adjacent along the post convergence path. The TI-LFA repair list is expressed generally as (Node_SID(P), EP(P, Q)). Most often, the TI-LFA repair list has a simpler form, as described in the following sections. provides statistics for the number of SIDs in the explicit path to protect against various failures.

When a direct neighbor is in P(S,X) and Q(D,x) and on the post-convergence path, the outgoing interface is set to that neighbor and the repair segment list SHOULD be empty. This is comparable to a post-convergence LFA FRR repair.

When a remote node R is in P(S,X) and Q(D,x) and on the post-convergence path, the repair list MUST be made of a single node segment to R and the outgoing interface SHOULD be set to the outgoing interface used to reach R. This is comparable to a post-convergence RLFA repair tunnel.

When a node P is in P(S,X) and a node Q is in Q(D,x) and both are on the post-convergence path and both are adjacent to each other, the repair list SHOULD be made of two segments: A node segment to P (to be processed first), followed by an adjacency segment from P to Q. This is comparable to a post-convergence DLFA (LFA with directed forwarding) repair tunnel.

In some cases, there is no adjacent P and Q node along the post- convergence path. As mentioned in , a list of adjacency SIDs can be used to encode the path between P and Q. However, the PLR can perform additional computations to compute a list of segments that represent a loop-free path from P to Q. How these computations are done is out of scope of this document and is left to implementation.

The following sections describe how to build the repair lists using the terminology defined in . The procedures described in are equally applicable to both SR-MPLS and SRv6 dataplane, while the dataplane-specific considerations are described in .

In this section, we explain how a protecting router S processes the active segment of a packet upon the failure of its primary outgoing interface for the packet, S-F.

The active segment MUST be kept on the SR header unchanged and the repair list MUST be added. The active segment becomes the first segment of the repair list. The way the repair list is added depends on the dataplane used (see ).

We define hereafter the FRR behavior applied by S for any packet received with an active adjacency segment S-F for which protection was enabled. As protection has been enabled for the segment S-F and signaled in the IGP (for instance using protocol extensions from and ), any SR policy using this segment knows that it may be transiently rerouted out of S-F in case of S-F failure. The simplest approach for link protection of an adjacency segment S-F is to create a repair list that will carry the traffic to F. To do so, one or more “PUSH” operations are performed. If the repair list, while avoiding S-F, terminates on F, S only pushes segments of the repair list. Otherwise, S pushes a node segment of F, followed by the segments of the repair list. For details on the "NEXT" and "PUSH" operations, refer to . This method which merges back the traffic at the remote end of the adjacency segment has the advantage of keeping as much as possible the traffic on the pre-failure path. As stated in , when SR policies are involved and a strict compliance of the policy is required, an end-to-end protection should be preferred over a local repair mechanism. However, this method may not provide the expected post-convergence path to the final destination as the expected post-convergence path may not go through F. Another method requires to look to the next segment in the segment list. We distinguish the case where this active segment is followed by another adjacency segment from the case where it is followed by a node segment.

If the next segment in the list is an Adjacency segment, then the packet has to be conveyed to F. To do so, S MUST apply a "NEXT" operation on Adj(S-F) and then one or more “PUSH” operations. If the repair list, while avoiding S-F, terminates on F, S only pushes the segments of the repair list. Otherwise, S pushes a node segment of F, followed by the segments of the repair list. For details on the "NEXT" and "PUSH" operations, refer to . Upon failure of S-F, a packet reaching S with a segment list matching [adj(S-F),adj(F-M),...] will thus leave S with a segment list matching [RL(F),node(F),adj(F-M),...], where RL(F) is the repair path for destination F.

If the next segment in the stack is a node segment, say for node T, the segment list on the packet matches [adj(S-F),node(T),...]. In this case, S MUST apply a "NEXT" operation on the Adjacency segment related to S-F, followed by a "PUSH" of a repair list redirecting the traffic to a node Q, whose path to node segment T is not affected by the failure. Upon failure of S-F, packets reaching S with a segment list matching [adj(S-F), node(T), ...], would leave S with a segment list matching [RL(Q),node(T), ...].

MPLS dataplane for Segment Routing is described in . The following dataplane behaviors apply when creating a repair list using an MPLS dataplane: If the active segment is a node segment that has been signaled with penultimate hop popping and the repair list ends with an adjacency segment terminating on the tail-end of the active segment, then the active segment MUST be popped before pushing the repair list. If the active segment is a node segment but the other conditions in 1. are not met, the active segment MUST be popped then pushed again with a label value computed according to the SRGB of Q, where Q is the endpoint of the repair list. Finally, the repair list MUST be pushed.

SRv6 dataplane and programming instructions are described respectively in and . The TI-LFA path computation algorithm is the same as in the SR-MPLS dataplane. Note however that the Adjacency SIDs are typically globally routed. In such case, there is no need for a preceding Prefix SID and the resulting repair list is likely shorter. If the traffic is protected at a Transit Node, then an SRv6 SID list is added on the packet to apply the repair list. The addition of the repair list follows the headend behaviors as specified in section 5 of . If the traffic is protected at an SR Segment Endpoint Node, first the Segment Endpoint packet processing is executed. Then the packet is protected as if its were a transit packet.

SR allows an operator to bind an algorithm to a prefix SID (as defined in . The algorithm value dictates how the path to the prefix is computed. The SR default algorithm is known has the "Shortest Path" algorithm. The SR default algorithm allows an operator to override the IGP shortest path by using local policies. When TI-LFA uses Node-SIDs associated with the default algorithm, there is no guarantee that the path will be loop-free as a local policy may have overriden the expected IGP path. As the local policies are defined by the operator, it becomes the responsibility of this operator to ensure that the deployed policies do not affect the TI-LFA deployment. It should be noted that such situation can already happen today with existing mechanisms as remote LFA. defines a flexible algorithm (FlexAlgo) framework to be associated with Prefix SIDs. FlexAlgo allows a user to associate a constrained path to a Prefix SID rather than using the regular IGP shortest path. An implementation MAY support TI-LFA to protect Node-SIDs associated to a FlexAlgo. In such a case, rather than computing the expected post-convergence path based on the regular SPF, an implementation SHOULD use the constrained SPF algorithm bound to the FlexAlgo (using the Flex Algo Definition) instead of the regular Dijkstra in all the SPF/rSPF computations that are occurring during the TI-LFA computation. This includes the computation of the P-Space and Q-Space as well as the post-convergence path. An implementation MUST only use Node-SIDs bound to the FlexAlgo and/or Adj-SIDs that are unprotected to build the repair list.

The repair list of segments computed by TI-LFA may contain one or more adjacency segments. An adjacency segment may be protected or not protected.

S --- R2 --- R3 --- R4 --- R5 --- D \ | \ / R7 -- R8 | | R9 -- R10 In , all the metrics are equal to 1 except R2-R7,R7-R8,R8-R4,R7-R9 which have a metric of 1000. Considering R2 as a PLR to protect against the failure of node R3 for the traffic S->D, the repair list computed by R2 will be [adj(R7-R8),adj(R8-R4)] and the outgoing interface will be to R7. If R3 fails, R2 pushes the repair list onto the incoming packet to D. During the FRR, if R7-R8 fails and if TI-LFA has picked a protected adjacency segment for adj(R7-R8), R7 will push an additional repair list onto the packet following the procedures defined in . To avoid the possibility of this double FRR activation, an implementation of TI-LFA MAY pick only non protected adjacency segments when building the repair list. However, this is important to note that FRR in general is intended to protect for a single pre-planned failure. If the failure that happens is worse than expected or multiple failures happen, FRR is not guaranteed to work. In such a case, fast IGP convergence remains important to restore traffic as quickly as possible.

This section presents analysis performed on real service provider and large enterprise network topologies. The objective of the analysis is to assess the number of SIDs required in an explicit path when the mechanisms described in this document are used to protect against the failure scenarios within the scope of this document. The number of segments described in this section are applicable to instantiating segment routing over the MPLS forwarding plane. The measurement below indicate that for link and local SRLG protection, a 1 SID repair path delivers more than 99% coverage. For node protection a 2 SIDs repair path yields 99% coverage. Table 1 below lists the characteristics of the networks used in our measurements. The number of links refers to the number of "bidirectional" links (not directed edges of the graph). The measurements are carried out as follows: For each network, the algorithms described in this document are applied to protect all prefixes against link, node, and local SRLG failure For each prefix, the number of SIDs used by the repair path is recored The percentage of number of SIDs are listed in Tables 2A/B, 3A/B, and 4A/B The measurements listed in the tables indicate that for link and local SRLG protection, 1 SID repair paths are sufficient to protect more than 99% of the prefix in almost all cases. For node protection 2 SIDs repair paths yield 99% coverage.

The rest of this section presents the measurements done on the actual topologies. The convention that we use is as follows 0 SIDs: the calculated repair path starts with a directly connected neighbor that is also a loop free alternate, in which case there is no need to explicitly route the traffic using additional SIDs. This scenario is described in . 1 SIDs: the repair node is a PQ node, in which case only 1 SID is needed to guarantee loop-freeness. This scenario is covered in . 2 or more SIDs: The repair path consists of 2 or more SIDs as described in and . We do not cover the case for 2 SIDs () separately because there was no granularity in the result. Also we treat the node-SID+adj-SID and node-SID + node-SID the same because they do not differ from the data plane point of view. Table 2A and 2B below summarize the measurements on the number of SIDs needed for link protection

The techniques described in this document are internal functionalities to a router that result in the ability to guarantee an upper bound on the time taken to restore traffic flow upon the failure of a directly connected link or node. As these techniques steer traffic to the post-convergence path as quickly as possible, this serves to minimize the disruption associated with a local failure which can be seen as a modest security enhancement. The protection mechanisms does not protect external destinations, but rather provides quick restoration for destination that are internal to a routing domain. Security considerations described in and apply to this document. Similarly, as the solution described in the document is based on Segment Routing technology, reader should be aware of the security considerations related to this technology () and its dataplane instantiations (, and ). However, this document does not introduce additional security concern.

No requirements for IANA

In addition to the authors listed on the front page, the following co-authors have also contributed to this document: Francois Clad, Cisco Systems Pablo Camarillo, Cisco Systems

We would like to thank Les Ginsberg, Stewart Bryant, Alexander Vainsthein, Chris Bowers, Shraddha Hedge for their valuable comments.