Segment Routing Use
CasesCisco Systems, Inc.BrusselsBEcfilsfil@cisco.comIMDEA NetworksLeganesESpifranco@cisco.comCisco Systems, Inc.Via Del Serafico, 200Rome00142Italysprevidi@cisco.comOrangeFRbruno.decraene@orange.comOrangeFRstephane.litkowski@orange.comDeutsche TelekomHammer Str. 216-226Muenster48153DEMartin.Horneffer@telekom.deTelekom SrbijaTakovska 2BelgradeRSigormilojevic@telekom.rsBritish TelecomLondonUKrob.shakir@bt.comTDC OyMechelininkatu 1aTDC00094FIsaku@ytti.fiAlcatel-LucentCopernicuslaan 50Antwerp2018BEwim.henderickx@alcatel-lucent.comEricsson300 Holger WaySan JoseCA95134USJeff.Tantsura@ericsson.comGoogle, Inc.1600 Amphitheatre ParkwayMountain ViewCA94043USedc@google.comNetwork Working GroupSegment Routing (SR) leverages the source routing and
tunneling paradigms. A node steers a packet through a controlled set of
instructions, called segments, by prepending the packet with an SR header.
A segment can represent any instruction, topological or service-based. SR
allows to enforce a flow through any topological path and service chain
while maintaining per-flow state only at the ingress node to the SR
domain.The Segment Routing architecture can be directly applied to the MPLS
dataplane with no change on the forwarding plane. It requires minor
extension to the existing link-state routing protocols. Segment Routing
can also be applied to IPv6 with a new type of routing extension
header.The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
"SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this
document are to be interpreted as described in RFC 2119.In this document, we document various SR use-cases. illustrates the ability to tunnel traffic
towards remote service points without any other protocol than the
IGP. reports various FRR use-cases leveraging the SR
functionality. and document
traffic-engineering use-cases, respectively without and with a notion of
bandwidth admission control. documents the co-existence and interworking
with MPLS Signaling protocols. illustrates OAM use-cases.The objective of this document is to illustrate the properties and
benefits of the SR architecture. The main reference for this document is the SR architecture defined
in .The SR instantiation in the MPLS dataplane is described in .IS-IS protocol extensions for Segment Routing are described in
.OSPF protocol extensions for Segment Routing are defined in .Fast-Reroute for Segment Routing is described in .The PCEP protocol extensions for Segment Routing are defined in
.The SR instantiation in the IPv6 dataplane will be described in a
future draft.A unique index is allocated to each IGP Prefix Segment. The related absolute
segment associated to an IGP Prefix SID is determined by summing the index and
the base of the SRSB. In the SR architecture, each node can be configured with
a different SRSB and hence the absolute SID associated to an IGP Prefix Segment
can change from node to node.We have described the first use-case of this document in the most generic
way, i.e. with different SRSB at each node in the SR IGP domain. We have
detailed the packet path highlighting that the SID of a Prefix Segment may
change hop by hop.For editorial simplification purpose, we will assume for all the other use
cases that the operator ensures a single consistent SRSB across all the nodes
in the SR IGP domain. In this specific context, we call the SRSB the SRGB and
we use global absolute SID's for the IGP Prefix SID's. Indeed, when all the
nodes have the same SRSB, all the nodes associate the same absolute SID with
the same index and hence one can use the absolute SID value instead of the
index.Several operators have indicated that they would deploy the SR technology in
this way: with a single consistent SRGB across all the nodes. They motivated
their choice based on operational simplicity (e.g. troubleshooting across
different nodes).While this document notes this operator feedback and we use this deployment
model to simplify the text, we highlight that the SR architecture is not
limited to this specific deployment use-case. Shared segments are always
advertised in the IGP extensions as indexed values.SR, applied to the MPLS dataplane, offers the ability to tunnel
services (VPN, VPLS, VPWS) from an ingress PE to an egress PE, without
any other protocol than ISIS or OSPF. LDP and RSVP-TE signaling
protocols are not required.The operator only needs to allocate one node segment per PE and the
SR IGP control-plane automatically builds the required MPLS forwarding
constructs from any PE to any PE.In above, the four nodes A, CE1, CE2
and Z are part of the same VPN. CE2 advertises to PE2 a route to Z. PE2
binds a local label LZ to that route and propagates the route and its
label via MPBGP to PE1 with nhop 192.168.0.2. PE1 installs the VPN prefix
Z in the appropriate VRF and resolves the next-hop onto the node segment
associated with PE2. Upon receiving a packet from A destined to Z, PE1 pushes two labels
onto the packet: the top label is the Prefix SID attached to
192.168.0.2/32, the bottom label is the VPN label LZ attached to the VPN
route Z.The Prefix-SID attached to prefix 192.168.0.2 is a shared segment within
the IGP domain, as such it is indexed.Let us assume that:- the operator allocated the index 2 to the prefix 192.168.0.2/32- the operator allocated SRSB [100, 199] at PE1- the operator allocated SRSB [200, 299] at P1- the operator allocated SRSB [300, 399] at P2- the operator allocated SRSB [400, 499] at P3- the operator allocated SRSB [500, 599] at P4- the operator allocated SRSB [600, 699] at PE2Thanks to this context, any SR-capable IGP node in the domain can
determine what is the segment associated with the Prefix-SID attached to
prefix 192.168.0.2/32:- PE1's SID is 100+2=102- P1's SID is 200+2=202- P2's SID is 300+2=302- P3's SID is 400+2=402- P4's SID is 500+2=502- PE2's SID is 600+2=602 Specifically to our example this means that PE1 load-balance the traffic
to VPN route Z between P1 and P4. The packets sent to P1 have a top label 202
while the packets sent to P4 have a top label 502. P1 swaps 202 for 302 and
forwards to P2. P2 pops 302 and forwards to PE2. The packets sent to P4
had label 502. P4 swaps 502 for 402 and forwards the packets to P3. P3
pops the top label and forwards the packets to PE2. Eventually all the
packets reached PE2 with one single lable: LZ, the VPN label attached to
VPN route Z.This scenario illustrates how supporting MPLS services (VPN, VPLS, VPWS) with SR has the following
benefits:- Simple operation: one single intra-domain protocol to operate:
the IGP. No need to support IGP synchronization extensions as
described in and .- Excellent scaling: one Node-SID per PE.SR leverages the technologies stemming from the IPFRR framework to
provide fast recovery of end-to-end connectivity upon failures. This
section assumes familiarity with Remote-LFA concepts described in
.Consider an arbitrary protected link S-E. In LFA FRR, if a path
from a neighbor N of S towards the destination does not cause packets
to loop back over the link S-E (i.e. N is a loop-free alternate
(LFA)), then S can forward packets to N and packets will be delivered
to the destination using the pre-failure LFA forwarding
information.If there is no such LFA neighbor, then S may be able to create a
virtual LFA by using a tunnel to carry the packet to a point in the
network which is not a direct neighbor of S and from which the packet
will be delivered to the destination without looping back to S. Remote
LFA (RLFA, ) calls such a
tunnel a repair tunnel. The tail-end of this tunnel is called a
"remote LFA" or a "PQ node". We refer to RLFA for the definitions of
the P and Q sets.In networks with symmetric IGP metrics (the metric of a link AB is
the same as the metric of the reverse link BA), we can prove that P
and the Q sets intersect or there is at least one P node that is
adjacent to a Q node.If the P and Q sets do not intersect (i.e. there is no RLFA PQ
node), we propose to use a Directed LFA (DLFA) repair tunnel from S to
a Q node that is adjacent to the P space (). The LFA repair tunnel only
requires two segments: a node segment to a P node which is adjacent to
the Q node and an adjacency segment from the P node to its adjacent Q
node.Thanks to the DLFA extension, we thus have a guaranteed LFA-based
FRR technique for any network with symmetric IGP metrics. Future
versions of the document will describe the solutions leveraging SR
capabilities to provided guaranteed FRR applicability in any IGP
topology.Resolving FRR with SR has the following benefits:Preservation of the simplicity properties of LFA FRR ().Preservation of the capacity planning properties (unlike SDH
and other FRR solutions, the repaired packet does not go back to
the next-hop or next-next-hop but uses shortest-path forwarding
from a much closer release point, ).Simplification of the RLFA operation: no
dynamically-established directed LDP sessions to the repair
nodes.No requirement for any extra computation on top of the one
required for RLFA.Guaranteed coverage for symmetric networks,The repair tunnel in symmetric network can be encoded
efficiently with only two segments.More details will be provided in a future version.Referring to the below figure, let us assume:A is identified by IP address 192.0.2.1/32 to which Node-SID
101 is attached.B is identified by IP address 192.0.2.2/32 to which Node-SID
102 is attachedA and B host the same set of services.Each service is identified by a local segment at each node:
i.e. node A allocates a local service segment 9001 to identify a
specific service S while the same service is identified by a local
service segment 9002 at B. Specifically, for the sake of this
illustration, let us assume that service S is a BGP-VPN service
where A announces a VPN route V with BGP nhop 192.0.2.1/32 and
local VPN label 9001 and B announces the same VPN route V with BGP
nhop 192.0.2.2/32 and local VPN label 9002.A generic mesh interconnects the three nodes M, Q and B.N prefers to use the service S offered by A and hence sends its
S-destined traffic with segment list {101, 9001}.Q is a node connected to A.Q has a method to detect the loss of node A within a few 10's
of msec.In that context, we would like to protect the traffic destined to
service S upon the failure of node A.The solution is built upon several components:B advertises a MIRROR sub-TLV in its IGP Link-State Router
Capability TLV with the values (TTT=000, MIRRORED_OBJECT=101,
CONTEXT_SEGMENT=10002),, and
for
more details in the encodings.Doing so, B advertises within the routing domain that it is
willing to backup any traffic originally sent to Node-SID 101
provided that this rerouted traffic gets to B with the context
segment 10002 directly preceding any local service segment
advertised by A. 10002 is a local context segment allocated by B to
identify traffic that was originally meant for A. This allows B to
match the subsequent service segment (e.g. 9001) correctly.We assume that B is able to discover all the local service
segments allocated by A (e.g. BGP route reflection and add-path). B
maps all the services advertised by A to its similar service
representations. For example, service 9001 advertised by A is mapped
to service 9002 advertised by B as both relate to the same service S
(the same VPN route V). For example, B applies the same service
treatment to a packet received with top segments {102, 10002, 9001}
or with top segments {102, 9002}. Basically, B treats {10002, 9001}
as a synonym of {9002}.In advance of any failure of A, Q (and any other node connected
to A) learns the identity of the IGP Mirroring node for each
Node-SID advertised by A (MIRROR_TLV advertised by B) and
pre-installs the following new MIRROR_FRR entry:Upon detecting the loss of node A, Q intercepts any traffic
destined to Node-SID 101, pops the segment to A (101) and push a
repair tunnel {102, 10002}. Node-SID 102 steers the repaired traffic
to B while context segment 10002 allows B to process the following
service segment {9001} in the right context table.Upon the failure of A, all the neighbors of A will flood the loss
of their adjacency to A and eventually every node within the IGP
domain will delete 192.0.2.1/32 from their RIB.The RIB deletion of 192.0.2.1/32 at N is beneficial as it
triggers the BGP FRR Protection onto the precomputed backup next-hop
.The RIB deletion at node M, if it occurs before the RIB deletion
at N, would be disastrous as it would lead to the loss of the
traffic from N to A before Q is able to apply the Mirroring
protection.The solution consists in delaying the deletion of the SRDB entry
for 101 by 2 seconds while still deleting the IP RIB 192.0.2.1/32
entry immediately.The RIB deletion triggers the BGP FRR and BGP Convergence. This
is beneficial and must occur without delay.The deletion of the SRDB entry to Node-SID101 is delayed to
ensure that the traffic still in transit towards Node-SID 101 is not
dropped.The delay timer should be long enough to ensure that either the
BGP FRR or the BGP Convergence has taken place at N.In our reference figure, N sends its packets towards A with the
segment list {101, 9001}. The shortest-path from S to A transits via
M and Q.Within a few msec of the loss of A, Q activates its pre-installed
Mirror_FRR entry and reroutes the traffic to B with the following
segment list {102, 10002, 9001}.Within a few 100's of msec, any IGP node deletes its RIB entry to
A but keeps its SRDB entry to Node-SID 101 for an extra 2
seconds.Upon deleting its RIB entry to 192.0.2.1/32, N activates its BGP
FRR entry and reroutes its S destined traffic towards B with segment
list {102, 9002}.By the time any IGP node deletes the SRDB entry to Node-SID 101,
N no longer sends any traffic with Node-SID 101.The deletion of the SRDB entry to Node-SID101 is delayed to
ensure that the traffic still in transit towards Node-SID 101 is not
dropped.In conclusion, the traffic loss only depends on the ability of Q
to detect the node failure of its adjacent node A.This section describes traffic-engineering use-cases which do not
require bandwidth admission control.The first sub-section illustrates the use of anycast segments to
express macro policies. Two examples are provided: one involving a
disjointness enforcement within a so-called dual-plane network, and the
other involving CoS-based policies.The second sub-section illustrate how a head-end router can combine a
distributed CSPF computation with SR. Various examples are provided
where the CSPF constraint or objective is either a TE affinity, an SRLG
or a latency metric.The third sub-section illustrates how SR can help traffic-engineer
outbound traffic among different external peers, overriding the best
installed IP path at the egress border routers.The fourth sub-section describes how SR can be used to express
deterministic non-ECMP path. Several techniques to compress the related
segment lists are also introduced.The fifth sub-section describes a use-case where a node attaches an
Adj-SID to a set of its interfaces however not sharing the same
neighbor. The illustrated benefit relates to loadbalancing.The SR architecture defines an anycast segment as a segment
attached to an anycast IP prefix ().The anycast node segment is an interesting tool for traffic
engineering:Macro-policy support: anycast segments allow to express
policies such as “go via plane1 of a dual-plane
network” () or “go via
Region3” ().Implicit node resiliency: the traffic-engineering policy is not
anchored to a specific node whose failure could impact the
service. It is anchored to an anycast address/Anycast-SID and
hence the flow automatically reroutes on any ECMP-aware
shortest-path to any other router part of the anycast set.The two following sub-sections illustrate to traffic-engineering
use-cases leveraging Anycast-SID.Many networks are built according to the dual-plane
design:We assume a common design rule found in such deployments: the
inter-plane link costs (Cik-Cjk where i<>j) are set such that
the route to an edge destination from a given plane stays within the
plane unless the plane is partitioned.In the above network diagram, let us that the operator
configures:The four routers (C1A, C1B, C1C, C1Z) with an anycast
loopback address 192.0.2.1/32 and an Anycast-SID 101.The four routers (C2A, C2B, C2C, C2Z) with an anycast
loopback address 192.0.2.2/32 and an Anycast-SID 102.Edge router Z with Node-SID 109.A can then use the three following segment lists to control its
Z-destined traffic:{109}: the traffic is load-balanced across any ECMP path
through the network.{101, 109}: the traffic is load-balanced across any ECMP path
within the Plane1 of the network.{102, 109}: the traffic is load-balanced across any ECMP path
within the Plane2 of the network.Most of the data traffic to Z would use the first segment list,
such as to exploit the capacity efficiently. The operator would use
the two other segment lists for specific premium traffic that has
requested disjoint transport.For example, let us assume a bank or a government customer has
requested that the two flows F1 and F2 injected at A and destined to
Z should be transported across disjoint paths. The operator could
classify F1 (F2) at A and impose and SR header with the second
(third) segment list. Focusing on F1 for the sake of illustration, A
would route the packets based on the active segment, Anycast-SID
101, which steers the traffic along the ECMP-aware shortest-path to
the closest router part of the Anycast-SID 101, C1A is this example.
Once the packets have reached C1A, the second segment becomes
active, Node-SID 109, which steers the traffic on the ECMP-aware
shortest-path to Z. C1A load-balances the traffic between C1B-C1Z
and C1C-C1Z and then C1Z forwards to Z.This SR use-case has the following benefits:Zero per-service state and signaling on midpoint and tail-end
routers.Only two additional node segments (one Anycast-SID per
plane).ECMP-awareness.Node resiliency property: the traffic-engineering policy is
not anchored to a specific core node whose failure could impact
the service.Frequently, different classes of service need different path
characteristics.In the example below, a single-area international network with
presence in four different regions of the world has lots of cheap
network capacity from Region4 to Region1 via Region2 and some scarce
expensive capacity via Region3.In such case, the IGP metrics would be tuned to have a
shortest-path from A to Z via Region2.This would provide efficient capacity planning usage while
fulfilling the requirements of most of the traffic demands. However,
it may not suite the latency requirements of the voice traffic
between the two cities.Let us illustrate how this can be solved with Segment
Routing.The operator would configure: With this in mind, the operator would instruct A to apply the
following policy for its Z-destined traffic:This SR use-case has the following benefits:Zero per-service state and signaling at midpoint and tailend
nodes.One additional anycast segment per region.ECMP-awareness.Node resiliency property: the traffic-engineering policy is
not anchored to a specific core node whose failure could impact
the service.In this section, we illustrate how a head-end router can map the
result of its distributed CSPF computation into an SR segment
list.Let us assume that in the above network diagram:The operator configures a policy on A such that its Z-destined
traffic must avoid SRLG1.The operator configures SRLG1 on the link BC (or is learned
dynamically from the IP/Optical interaction with the DWDM
network).The SRLG’s are flooded in the link-state IGP.The operator respectively configures the Node-SIDs 101, 102,
103, 104, 105 and 109 at nodes A, B, C, D, E and Z.In that context, A can apply the following CSPF behavior: The same use-case can be derived from any other C-SPF objective or
constraint (TE affinity, TE latency, SRLG, etc.) as defined in and . Note that the
bandwidth case is specific and hence is treated in .Let us assume that:C in AS1 learns about destination Z of AS 4 via two BGP paths
(AS2, AS4) and (AS3, AS4).C sets next-hop-self before propagating the paths within
AS1.C propagates all the paths to Z within AS1 (add-path).C only installs the path via AS2 in its RIB.In that context, the operator of AS1 cannot apply the following
traffic-engineering policy:Steer 60% of the Z-destined traffic received at A via AS2 and
40% via AS3.Steer 80% of the Z-destined traffic received at B via AS2 and
20% via AS3.This traffic-engineering policy can be supported thanks to the
following SR configuration.The operator configures: C with a loopback 192.0.2.1/32 and attach the Node-SID 101 to
it.C to bind an external adjacency segment () to each of its
peering interface.For the sake of this illustration, let us assume that the external
adjacency segments bound by C for its peering interfaces to (D, AS2)
and (E, AS3) are respectively 9001 and 9002.These external adjacencies (and their attached segments) are
flooded within the IGP domain of AS1 .As a result, the following information is available within
AS1:The operator of AS1 can thus meet its traffic-engineering objective
by enforcing the following policies:A should apply the segment list {101, 9001} to 60% of the
Z-destined traffic and the segment list {101, 9002} to the
rest.B should apply the segment list {101, 9001} to 80% of the
Z-destined traffic and the segment list {101, 9002} to the
rest.Node segment 101 steers the traffic to C.External adjacency segment 9001 forces the traffic from C to (D,
AS2), without any IP lookup at C.External adjacency segment 9002 forces the traffic from C to (E,
AS3), without any IP lookup at C.A and B can also use the described segments to assess the liveness
of the remote peering links, see OAM section.The previous sections have illustrated the ability to steer traffic
along ECMP-aware shortest-paths. SR is also able to express
deterministic non-ECMP path: i.e. as a list of adjacency segments. We
illustrate such an use-case in this section.In the above figure, it is assumed all nodes are SR capable and
only the following SIDs are advertised:The operator can steer the traffic from A to Z via a specific
non-ECMP path ABCDEFGHZ by imposing the segment list {9001, 9002,
9003, 9004, 9001, 9002, 9003, 9004}.The following sub-sections illustrate how the segment list can be
compressed.Clearly the same exact path can be expressed with a two-entry
segment list {101, 109}.This example illustrates that a Node Segment can also be used to
express deterministic non-ECMP path.The operator can configure Node B to create a
forwarding-adjacency to node H along an explicit path BCDEFGH. The
following behaviors can then be automated by B:B attaches an Adj-SID (e.g. 9007) to that forwarding
adjacency together with an ERO sub-sub-TLV which describes the
explicit path BCDEFGH.B installs in its Segment Routing Database the following
entry:Active segment: 9007.Operation: NEXT and PUSH {9002, 9003, 9004, 9001, 9002,
9003}As a result, the operator can configure node A with the following
compressed segment list {9001, 9007, 9004}.A given node may assign the same Adj-SID to multiple of its
adjacencies, even if these ones lead to different neighbors. This may
be useful to support traffic engineering policies.In the above example, let us assume that the operator:Requires PE1 to load-balance its PE2-destined traffic between
the ABCDE and ABFE paths.Configures B with Node-SID 102 and E with Node-SID 202.Configures B to advertise an individual Adj-SID per adjacency
(e.g. 9001 for BC and 9002 for BF) and, in addition, an Adj-SID
for the adjacency set (BC, BF) (e.g. 9003).With this context in mind, the operator achieves its objective by
configuring the following traffic-engineering policy at PE1 for the
PE2-destined traffic: {102, 9003, 202}: Node-SID 102 steers the traffic to B.Adj-SID 9003 load-balances the traffic to C or F.From either C or F, Node-SID 202 steers the traffic to PE2.In conclusion, the traffic is load-balanced between the ABCDE
and ABFE paths, as desired.The implementation of bandwidth admission control within a network
(and its possible routing consequence which consists in routing along
explicit paths where the bandwidth is available) requires a capacity
planning process.The spreading of load among ECMP paths is a key attribute of the
capacity planning processes applied to packet-based networks.The first sub-section details the capacity planning process and the
role of ECMP load-balancing. We highlight the relevance of SR in that
context.The next two sub-sections document two use-cases of SR-based traffic
engineering with bandwidth admission control.The second sub-section documents a concrete SR applicability
involving centralized-based admission control. This is often referred to
as the “SDN/SR use-case”.The third sub-section introduces a future research topic involving
the notion of residual bandwidth introduced in .Capacity Planning anticipates the routing of the traffic matrix
onto the network topology, for a set of expected traffic and topology
variations. The heart of the process consists in simulating the
placement of the traffic along ECMP-aware shortest-paths and
accounting for the resulting bandwidth usage.The bandwidth accounting of a demand along its shortest-path is a
basic capability of any planning tool or PCE server.For example, in the network topology described below, and assuming
a default IGP metric of 1 and IGP metric of 2 for link GF, a 1600Mbps
A-to-Z flow is accounted as consuming 1600Mbps on links AB and FZ,
800Mbps on links BC, BG and GF, and 400Mbps on links CD, DF, CE and
EF.ECMP is extremely frequent in SP, Enterprise and DC architectures
and it is not rare to see as much as 128 different ECMP paths between
a source and a destination within a single network domain. It is a key
efficiency objective to spread the traffic among as many ECMP paths as
possible.This is illustrated in the below network diagram which consists of
a subset of a network where already 5 ECMP paths are observed from A
to M.Segment Routing offers a simple support for such ECMP-based
shortest- path placement: a node segment. A single node segment
enumerates all the ECMP paths along the shortest-path.When the capacity planning process detects that a traffic growth
scenario and topology variation would lead to congestion, a capacity
increase is triggered and if it cannot be deployed in due time, a
traffic engineering solution is activated within the network.A basic traffic engineering objective consists of finding the
smallest set of demands that need to be routed off their shortest path
to eliminate the congestion, then to compute an explicit path for each
of them and instantiating these traffic-engineered policies in the
network.Segment Routing offers a simple support for explicit path policy.
Let us provide two examples based on .First example: let us assume that the process has selected the flow
AM for traffic-engineering away from its ECMP-enabled shortest path
and flow AM must avoid consuming resources on the LM and the FG
links.The solution is straightforward: A sends its M-destined traffic
towards the nhop F with a two-label stack where the top label is the
adjacent segment FI and the next label is the node segment to M.
Alternatively, a three-label stack with adjacency segments FI, IK and
KM could have been used.Second example: let us assume that AM is still the selected flow
but the constraint is relaxed to only avoid using resources from the
LM link.The solution is straightforward: A sends its M-destined traffic
towards the nhop F with a one-label stack where the label is the node
segment to M. Note that while the AM flow has been traffic-engineered
away from its natural shortest-path (ECMP across three paths), the
traffic-engineered path is still ECMP-aware and leverages two of the
three initial paths. This is accomplished with a single-label stack
and without the enumeration of one tunnel per path.Under the light of these examples, Segment Routing offers an
interesting solution for Capacity Planning because:One node segment represents the set of ECMP-aware shortest
paths.Adjacency segments allow to express any explicit path.The combination of node and adjacency segment allows to express
any path without having to enumerate all the ECMP options.The capacity planning process ensures that the majority of the
traffic rides on node segments (ECMP-based shortest path), while a
minority of the traffic is routed off its shortest-path.The explicitly-engineered traffic (which is a minority) still
benefits from the ECMP-awareness of the node segments within their
segment list.Only the head-end of a traffic-engineering policy maintains
state. The midpoints and tail-ends do not maintain any state.The heart of the application of SR to the SDN use-case lies in the
SDN controller, also called Stateful PCE ().The SDN controller is responsible to control the evolution of the
traffic matrix and topology. It accepts or denies the addition of new
traffic into the network. It decides how to route the accepted
traffic. It monitors the topology and upon failure, determines the
minimum traffic that should be rerouted on an alternate path to
alleviate a bandwidth congestion issue.The algorithms supporting this behavior are a local matter of the
SDN controller and are outside the scope of this document.The means of collecting traffic and topology information are the
same as what would be used with other SDN-based traffic-engineering
solutions (e.g. and .The means of instantiating policy information at a
traffic-engineering head-end are the same as what would be used with
other SDN-based traffic-engineering solutions (e.g.: , and ).Let us assume that in the above network diagram:An SDN Controller (SC) is connected to the network and is
able to retrieve the topology and traffic information, as well
as set traffic-engineering policies on the network nodes.The operator (likely via the SDN Controller) as provisioned
the Node-SIDs 101, 102, 103, 104, 105, 106, 107, 201, 202 and
203 respectively at nodes A, B, C, D, E, F, G, V, Y and Z.All the links have the same BW (e.g. 10G) and IGP cost (e.g.
10) except the links BG and GD which have IGP cost 50.Each described node connectivity is formed as a bundle of two
links, except (B, G) and (G, D) which are formed by a single
link each.Flow FV is traveling from A to destinations behind V.Flow FY is traveling from A to destinations behind Y.Flow FZ is traveling from A to destinations behind Z.The SDN Controller has admitted all these flows and has let A
apply the default SR policy: “map a flow onto its
ECMP-aware shortest-path”. In this example, this means that A respectively maps the
flows FV onto segment list {201}, FY onto segment list {202}
and FZ onto segment list {203}.In this example, the reader should note that the SDN
Controller knows what A would do and hence knows and
controls that none of these flows are mapped through G.Let us describe what happens upon the failure of one of the two
links E-D.The SDN Controller monitors the link-state database and detects a
congestion risk due to the reduced capacity between E and D.
Specifically, SC updates its simulation of the traffic according to
the policies he instructed the network to use and discovers that too
much traffic is mapped on the remaining link E-D.The SDN Controller then computes the minimum number of flows that
should be deviated from their existing path. For example, let us
assume that the flow FZ is selected.The SDN controller then computes an explicit path for this flow.
For example, let us assume that the chosen path is ABGDFZ.The SDN controller then maps the chosen path into an SR-based
policy. In our example, the path ABGDFZ is translated into a segment
list {107, 203}. Node-SID steers the traffic along ABG and then
Node-SID 203 steers the traffic along GDFZ.The SDN controller then applies the following traffic-engineering
policy at A: “map any packet of the classified flow FZ onto
segment-list {107, 203}". The SDN Controller uses PCEP extensions to
instantiate that policy at A ().As soon as A receives the PCEP message, it enforces the policy
and the traffic classified as FZ is immediately mapped onto segment
list {107, 203}.This immediately eliminate the congestion risk. Flows FV and FY
were untouched and keep using the ECMP-aware shortest-path. The
minimum amount of traffic was rerouted (FZ). No signaling hop-by-hop
through the network from A to Z is required. No admission control
hop-by-hop is required. No state needs to be maintained by B, G, D,
F or Z. The only maintained state is within the SDN controller and
the head-end node (A).In the context of Centralized-Based Optimization and the SDN
use-case, here are the benefits provided by the SR
architecture:Explicit routing capability with or without
ECMP-awareness.No signaling hop-by-hop through the network.State is only maintained at the policy head-end. No state is
maintained at mid-points and tail-ends.Automated guaranteed FRR for any topology (.Optimum virtualization: the policy state is in the packet
header and not in the intermediate node along the policy. The
policy is completely virtualized away from midpoints and
tail-ends.Highly responsive to change: the SDN Controller only needs to
apply a policy change at the head-end. No delay is lost
programming the midpoints and tail-end along the policy.A future version of this document will report some analysis of
the application of the SDN/SR use-case to real operator data
sets.A first, incomplete, report is available here below.The first data-set consists in a full-mesh of 12000
explicitly-routed tunnels observed on a real network. These
tunnels resulted from distributed headend-based CSPF
computation.We measured that only 65% of the traffic is riding on its
shortest path.Three well-known defects are illustrated in this data set:
The lack of ECMP support in explicitly—routed
tunnels: ATM-alike traffic-steering mechanisms steer the
traffic along a non-ECMP path.The increase of the number of explicitly-routed non-ECMP
tunnels to enumerate all the ECMP options.The inefficiency of distributed optimization: too much
traffic is riding off its shortest path.We applied the SDN/SR use-case to this dataset. This means
that:The distributed CSPF computation is replaced by centralized
optimization and BW admission control, supported by the SDN
Controller.As part of the optimization, we also optimized the
IGP-metrics such as to get a maximum of traffic
load-spread among ECMP-paths by default.The traffic-engineering policies are supported by SR
segment-lists.As a result, we measured that 98% of the traffic would be kept
on its normal policy (ride shortest-path) and only 2% of the
traffic requires a path away from the shortest-path.Let us highlight a few benefits:98% of the traffic-engineering head-end policies are
eliminated.Indeed, by default, an SR-capable ingress edge node
maps the traffic on a single Node-ID to the egress edge
node. No configuration or policy needs to be maintained at
the ingress edge node to realize this.100% of the states at mid/tail nodes are eliminated.The notion of Residual Bandwidth (RBW) is introduced by .A future version of this document will describe the SR/RBW research
opportunity.The first section describes the co-existence of SR with other MPLS
Control Plane. The second section documents a method to migrate from LDP
to SR-based MPLS tunneling. The third section documents the interworking
of LDP and SR in the case of non-homogenous deployment. The fourth
use-case describes how a partial SR deployment can be used to provide SR
benefits to LDP-based traffic. The fifth section describes a possible
application of SR in the context of inter-domain MPLS use-cases.We call “MPLS Control Plane Client (MCC)” any
control-plane protocol installing forwarding entries in the MPLS
dataplane. SR, LDP, RSVP-TE, BGP 3107, VPNv4, etc. are examples of
MCCs.An MCC, operating at node N, must ensure that the incoming label it
installs in the MPLS dataplane of Node N has been uniquely allocated
to himself.Thanks to the defined segment allocation rule and specifically the
notion of the SRGB, SR can co-exist with any other MCC.This is clearly the case for the adjacency segment: it is a local
label allocated by the label manager, as for any MCC.This is clearly the case for the prefix segment: the label manager
allocates the SRGB set of labels to the SR MCC client and the operator
ensures the unique allocation of each global prefix segment/label
within the allocated SRGB set.Note that this static label allocation capability of the label
manager has been existing for many years across several vendors and
hence is not new. Furthermore, note that the label-manager ability to
statically allocate a range of labels to a specific application is not
new either. This is required for MPLS-TP operation. In this case, the
range is reserved by the label manager and it is the MPLS-TP NMS
(acting as an MCC) that ensures the unique allocation of any label
within the allocated range and the creation of the related MPLS
forwarding entry.Let us illustrate an example of ship-in-the-night (SIN)
coexistence.The EVEN VPN service is supported by PE2 and PE4 while the ODD VPN
service is supported by PE1 and PE3. The operator wants to tunnel the
ODD service via LDP and the EVEN service via SR.This can be achieved in the following manner:The operator configures PE1, PE2, PE3, PE4 with respective
loopbacks 192.0.2.201/32, 192.0.2.202/32, 192.0.2.203/32,
192.0.2.204/32. These PE’s advertised their VPN routes with
next-hop set on their respective loopback address.The operator configures A, B, C with respective loopbacks
192.0.2.1/32, 192.0.2.2/32, 192.0.2.3/32.The operator configures PE2, A, B, C and PE4 with SRGB {100,
300}.The operator attaches the respective Node-SIDs 202, 101, 102,
103 and 204 to the loopbacks of nodes PE2, A, B, C and PE4. The
Node-SID’s are configured to request
penultimate-hop-popping.PE1, A, B, C and PE3 are LDP capable.PE1 and PE3 are not SR capable.PE3 sends an ODD VPN route to PE1 with next-hop 192.0.2.203 and VPN
label 10001.From an LDP viewpoint: PE1 received an LDP label binding (1037) for
FEC 192.0.2.203/32 from its nhop A. A received an LDP label binding
(2048) for that FEC from its nhop B. B received an LDP label binding
(3059) for that FEC from its nhop C. C received implicit-null LDP
binding from its next-hop PE3.As a result, PE1 sends its traffic to the ODD service route
advertised by PE3 to next-hop A with two labels: the top label is 1037
and the bottom label is 10001. A swaps 1037 with 2048 and forwards to
B. B swaps 2048 with 3059 and forwards to C. C pops 3059 and forwards
to PE3.PE4 sends an EVEN VPN route to PE2 with next-hop 192.0.2.204 and
VPN label 10002.From an SR viewpoint: PE1 maps the IGP route 192.0.2.204/32 onto
Node-SID 204; A swaps 204 with 204 and forwards to B; B swaps 204 with
204 and forwards to C; C pops 204 and forwards to PE4.As a result, PE2 sends its traffic to the VPN service route
advertised by PE4 to next-hop A with two labels: the top label is 204
and the bottom label is 10002. A swaps 204 with 204 and forwards to B.
B swaps 204 with 204 and forwards to C. C pops 204 and forwards to
PE4.The two modes of MPLS tunneling co-exist.The ODD service is tunneled from PE1 to PE3 through a
continuous LDP LSP traversing A, B and C.The EVEN service is tunneled from PE2 to PE4 through a
continuous SR node segment traversing A, B and C.We want to highlight that several MPLS2MPLS entries can be
installed in the dataplane for the same prefix.Let us examine A’s MPLS forwarding table as an
example:Incoming label: 1037Incoming label: 203These two entries can co-exist because their incoming label is
unique. The uniqueness is guaranteed by the label manager allocation
rules.The same applies for the MPLS2IP forwarding entries.By default, we propose that if both LDP and SR propose an IP2MPLS
entry for the same IP prefix, then the LDP route is selected.A local policy on a router MUST allow to prefer the SR-provided
IP2MPLS entry.Several migration techniques are possible. We describe one
technique inspired by the commonly used method to migrate from one IGP
to another.T0: all the routers run LDP. Any service is tunneled from an
ingress PE to an egress PE over a continuous LDP LSP.T1: all the routers are upgraded to SR. They are configured with
the SRGB range (100, 200). PE1, PE2, PE3, PE4, P5, P6 and P7 are
respectively configured with the node segments 101, 102, 103, 104,
105, 106 and 107 (attached to their service-recursing loopback). At this time, the service traffic is still tunneled over LDP
LSP. For example, PE1 has an SR node segment to PE3 and an LDP LSP
to PE3 but by default, as seen earlier, the LDP IP2MPLS
encapsulation is preferred.T2: the operator enables the local policy at PE1 to prefer SR
IP2MPLS encapsulation over LDP IP2MPLS.The service from PE1 to any other PE is now riding over SR. All
other service traffic is still transported over LDP LSP.T3: gradually, the operator enables the preference for SR IP2MPLS
encapsulation across all the edge routers.All the service traffic is now transported over SR. LDP is
still operational and services could be reverted to LDP.T4: LDP is unconfigured from all routers.In this section, we analyze a use-case where SR is available in one
part of the network and LDP is available in another part. We describe
how a continuous MPLS tunnel can be built throughout the network.Let us analyze the following example:P6, P7, P8, PE4 and PE3 are LDP capable.PE1, PE2, P5 and P6 are SR capable. PE1, PE2, P5 and P6 are
configured with SRGB (100, 200) and respectively with node
segments 101, 102, 105 and 106.A service flow must be tunneled from PE1 to PE3 over a
continuous MPLS tunnel encapsulation. We need SR and LDP to
interwork.In this section, we analyze a right-to-left traffic flow.PE3 has learned a service route whose nhop is PE1. PE3 has an LDP
label binding from the nhop P8 for the FEC “PE1”. Hence
PE3 sends its service packet to P8 as per classic LDP behavior.P8 has an LDP label binding from its nhop P7 for the FEC
“PE1” and hence P8 forwards to P7 as per classic LDP
behavior.P7 has an LDP label binding from its nhop P6 for the FEC
“PE1” and hence P7 forwards to P6 as per classic LDP
behavior.P6 does not have an LDP binding from its nhop P5 for the FEC
“PE1”. However P6 has an SR node segment to the IGP
route “PE1”. Hence, P6 forwards the packet to P5 and
swaps its local LDP-label for FEC “PE1” by the
equivalent node segment (i.e. 101).P5 pops 101 (assuming PE1 advertised its node segment 101 with
the penultimate-pop flag set) and forwards to PE1.PE1 receives the tunneled packet and processes the service
label.The end-to-end MPLS tunnel is built from an LDP LSP from PE3 to
P6 and the related node segment from P6 to PE1.In this section, we analyze the left-to-right traffic flow.We assume that the operator configures P5 to act as a Segment
Routing Mapping Server (SRMS) and advertise the following mappings:
(P7, 107), (P8, 108), (PE3, 103) and (PE4, 104).The mappings advertised by an SR mapping server result from local
policy information configured by the operator. IF PE3 had been SR
capable, the operator would have configured PE3 with node segment
103. Instead, as PE3 is not SR capable, the operator configures that
policy at the SRMS and it is the latter which advertises the
mapping. Multiple SRMS servers can be provisioned in a network for
redundancy.The mapping server advertisements are only understood by the SR
capable routers. The SR capable routers install the related node
segments in the MPLS dataplane exactly like if the node segments had
been advertised by the nodes themselves.For example, PE1 installs the node segment 103 with nhop P5
exactly as if PE3 had advertised node segment 103.PE1 has a service route whose nhop is PE3. PE1 has a node segment
for that IGP route: 103 with nhop P5. Hence PE1 sends its service
packet to P5 with two labels: the bottom label is the service label
and the top label is 103.P5 swaps 103 for 103 and forwards to P6.P6’s next-hop for the IGP route “PE3” is not SR
capable (P7 does not advertise the SR capability). However, P6 has
an LDP label binding from that next-hop for the same FEC (e.g. LDP
label 1037). Hence, P6 swaps 103 for 1037 and forwards to P7.P7 swaps this label with the LDP-label received from P8 and
forwards to P8.P8 pops the LDP label and forwards to PE3.PE3 receives the tunneled packet and processes the service
label.The end-to-end MPLS tunnel is built from an SR node segment from
PE1 to P6 and an LDP LSP from P6 to PE3.SR can be deployed such as to enhance LDP transport. The SR
deployment can be limited to the network region where the SR benefits
are most desired.In , let us assume:All link costs are 10 except FG which is 30.All routers are LDP capable.X, Y and Z are PE’s participating to an important service
S.The operator requires 50msec link-based FRR for service S.A, B, C, D, E, F and G are SR capable.X, Y, Z are not SR capable, e.g. as part of a staged migration
from LDP to SR, the operator deploys SR first in a sub-part of the
network and then everywhere.The operator would like to resolve the following issues:To protect the link BA along the shortest-path of the important
flow XY, B requires an RLFA repair tunnel to D and hence a
directed LDP session from B to D. The operator does not like these
dynamically established multi-hop LDP sessions and would seek to
eliminate them.There is no LFA/RLFA solution to protect the link BE along the
shortest path of the important flow XZ. The operator wants a
guaranteed link-based FRR solution.The operator can meet these objectives by deploying SR only on A,
B, C, D, E and F:The operator configures A, B, C, D, E, F and G with SRGB (100,
200) and respective node segments 101, 102, 103, 104, 105, 106 and
107.The operator configures D as an SR Mapping Server with the
following policy mapping: (X, 201), (Y, 202), (Z, 203}.Each SR node automatically advertises local adjacency segment
for its IGP adjacencies. Specifically, F advertises adjacency
segment 9001 for its adjacency FG.A, B, C, D, E, F and G keep their LDP capability and hence the
flows XY and XZ are transported over end-to-end LDP LSP’s.For example, LDP at B installs the following MPLS dataplane
entries:The novelty comes from how the backup chains are computed for these
LDP-based entries. While LDP labels are used for the primary nhop and
outgoing labels, SR information is used for the FRR construction. In
steady state, the traffic is transported over LDP LSP. In transient
FRR state, the traffic is backup thanks to the SR enhanced
capabilities.This helps meet the requirements of the operator:Eliminate directed LDP session.Guaranteed FRR coverage.Keep the traffic over LDP LSP in steady state.Partial SR deployment only where needed.B’s MPLS entry to Y becomes:In steady-state, X sends its Y-destined traffic to B with a top
label which is the LDP label bound by B for FEC Y. B swaps that top
label for the LDP label bound by A for FEC Y and forwards to A. A
pops the LDP label and forwards to Y.Upon failure of the link BA, B swaps the incoming top-label with
the node segment for Y (202) and sends the packet onto a repair
tunnel to D (node segment 104). Thus, B sends the packet to C with
the label stack {104, 202}. C pops the node segment 104 and forwards
to D. D swaps 202 for 202 and forwards to A. A’s nhop to Y is
not SR capable and hence A swaps the incoming node segment 202 to
the LDP label announced by its next-hop (in this case, implicit
null).After IGP convergence, B’s MPLS entry to Y will
become:And the traffic XY travels again over the LDP LSP.Conclusion: the operator has eliminated its first problem:
directed LDP sessions are no longer required and the steady-state
traffic is still transported over LDP. The SR deployment is confined
to the area where these benefits are required.B’s MPLS entry to Z becomes:In steady-state, X sends its Z-destined traffic to B with a top
label which is the LDP label bound by B for FEC Z. B swaps that top
label for the LDP label bound by E for FEC Z and forwards to E. E
pops the LDP label and forwards to Z.Upon failure of the link BE, B swaps the incoming top-label with
the node segment for Z (203) and sends the packet onto a repair
tunnel to G (node segment 106 followed by adjacency segment 9001).
Thus, B sends the packet to C with the label stack {106, 9001, 203}.
C pops the node segment 106 and forwards to F. F pops the adjacency
segment 9001 and forwards to G. G swaps 203 for 203 and forwards to
E. E’s nhop to Z is not SR capable and hence E swaps the
incoming node segment 203 for the LDP label announced by its
next-hop (in this case, implicit null).After IGP convergence, B’s MPLS entry to Z will
become:And the traffic XZ travels again over the LDP LSP.Conclusion: the operator has eliminated its second problem:
guaranteed FRR coverage is provided. The steady-state traffic is
still transported over LDP. The SR deployment is confined to the
area where these benefits are required.In Inter-AS Option C , B2 advertises to B1
a BGP3107 route for PE2 and B1 reflects it to its internal peers, such
as PE1. PE1 learns from a service route reflector a service route
whose nhop is PE2. PE1 resolves that service route on the BGP3107
route to PE2. That BGP3107 route to PE2 is itself resolved on the AS1
IGP route to B1.If AS1 operates SR, then the tunnel from PE1 to B1 is provided by
the node segment from PE1 to B1.PE1 sends a service packet with three labels: the top one is the
node segment to B1, the next-one is the BGP3107 label provided by B1
for the route “PE2” and the bottom one is the service
label allocated by PE2.The same straightforward SR applicability is derived for CsC and
Seamless MPLS ().This section documents a few representative SR/OAM use-cases.
In the above figure, a monitoring system (MS) needs to assess the
dataplane availability of all the links within a remote bundle
connected to routers R1 and R2.The monitoring system retrieves the segment information from the
IGP LSDB and appends the following segment list: {72, 662, 992, 664}
on its IP probe (whose source and destination addresses are the
address of AA).MS sends the probe to its connected router. If the connected router
is not SR compliant, a tunneling technique can be used to tunnel the
SR-based probe to the first SR router. The SR domain forwards the
probe to R2 (72 is the node segment of R2). R2 forwards the probe to
R1 over link L1 (adjacency segment 662). R1 forwards the probe to R2
over link L2 (adjacency segment 992). R2 forwards the probe to R1 over
link L3 (adjacency segment 664). R1 then forwards the IP probe to AA
as per classic IP forwarding.In , node A can monitor the dataplane
liveness of the unidirectional peering link from C to D of AS2 by
sending an IP probe with destination address A and segment list {101,
9001}. Node-SID 101 steers the probe to C and External Adj-SID 9001
steers the probe from C over the desired peering link to D of AS2. The
SR header is removed by C and D receives a plain IP packet with
destination address A. D returns the probe to A through classic IP
forwarding. BFD Echo mode () would support
such liveliness unidirectional link probing application.TBDTBDTBDWe would like to thank Dave Ward, Dan Frost, Stewart Bryant, Thomas
Telkamp, Ruediger Geib and Les Ginsberg for their contribution to the
content of this document.Segment Routing ArchitectureIS-IS Segment Routing ExtensionsOSPF Segment Routing ExtensionsPCEP Extensions for Segment RoutingBGP Prefix Independent Convergence