Segment Routing Single Link Failure Congestion Optimization

Vitor Pereira

, Miguel Rocha

and Pedro Sousa

Centro Algoritmi, Department of Informatics, University of Minho, Portugal

Centre of Biological Engineering, Department of Informatics, University of Minho, Portugal

Keywords:

Segment Routing, Link Failure, Evolutionary Computation, Software Deﬁned Networking, Routing and Flow

Control.

Abstract:

Segment routing is an implementation of the source routing paradigm built over an Interior Gateway Proto-

col. It allows improving Trafﬁc Engineering in IP/MPLS networks by decomposing forwarding paths in lists

of smaller paths called segments. The ﬂexibility introduced by segment routing enables new responsive and

dynamic methods to react to network fault events, such as link failures. Typical responses to single link fail-

ures only aim to reestablish loop-free connectivity between affected routers and do not account for impacts

in network congestion. As Interior Gateway Protocols recompute shortest-paths after a link failure, SR paths

and links utilization are affected by the fault augmenting the overall network congestion. This paper addresses

this problem comparing post-convergence network congestion values obtained with industry and new proposal

Evolutionary Computation based recovery path computation methods for segment routing. Furthermore, by

comparing and analyzing each method’s advantages and disadvantages, we show that a multiplane optimiza-

tion procedure stands out with best results.

1 INTRODUCTION

Segment routing (SR) (Filsﬁls et al., 2013) is

a Software-Deﬁned Networking (SDN) (Feamster

et al., 2013) enabled technology proposed by the In-

ternet Engineering Task Force (IETF) in 2013. By

decomposing forwarding paths into segments, SR en-

ables new methods capable of optimizing network

performance and improving Trafﬁc Engineering (TE).

The ability to steer packets through a deﬁned list of

nodes provides the agility and ﬂexibility required for

enhanced tasks. With SR, network environments be-

come more responsive to topological events, such as

link failures, while enabling a broader range of re-

sponse options.

SR is built over already existing Interior Gate-

way Protocols (IGP), such as Open Shortest Path

First (OSPF) (Moy, 1997) and Intermediate System

to Intermediate System (IS-IS) (Gredler and Goralski,

2005). From the two main primary types of segments,

adjacency and node segments, the lasts are IGP de-

pendent and uniquely identiﬁed by an ID, a Segment

ID (SID). A Node SID pinpoints an address preﬁx

calculated by the IGP, a destination that should be

reached using an IGP shortest-path. An Adjacency

SID, on the other hand, represents a local interface to

a speciﬁc adjacent IGP node.

Any path can be represented by a combination of

Node SIDs and Adjacency SIDs. SR takes ad-

vantage of multiple IGP features, for example, it sup-

ports Equal Cost Multi-Path (ECMP) (Hopps, 2000)

by design. Paths identiﬁed by node segments are IGP

shortest-paths, and intrinsically include all the ECMP

paths to the destination node, which contributes to a

tremendous gain in network performance and scala-

bility. Another essential IGP feature on which SR

relies is the automatic rerouting of connections after

a link failure. Upon a link failure, the IGP protocol

recomputes all shortest-paths, and segments are au-

tomatically repaired without any additional interven-

tion.

The time required to detect a link failure, propa-

gate the fault and recompute the shortest-paths can be

excessively long and, therefore, recovery paths should

preferably be pre-computed and installed in the data

plane. Fast reroute (FRR) with loop-free alternates

(LFA) (Atlas and Zinin, 2008) follows this strategy

and provides sub-50msec loss of connectivity to IGPs.

However, LFA does not offer a complete network cov-

erage and is topology dependent. With SR, those lim-

itations ceased to exist. Topology-Independent LFA

(TI-LFA) (Bashandy et al., 2018) provides local pro-

Pereira, V., Rocha, M. and Sousa, P.

Segment Routing Single Link Failure Congestion Optimization.

DOI: 10.5220/0006856200760083

In Proceedings of the 15th International Joint Conference on e-Business and Telecommunications (ICETE 2018) - Volume 1: DCNET, ICE-B, OPTICS, SIGMAP and WINSYS, pages 76-83

ISBN: 978-989-758-319-3

tection for IGP SIDs in any topology. Backup paths

can be pre-computed on a per IGP SID basis along

the post-convergence path from the Points of Local

Repair (PLR) to all possible destinations. In the vast

majority of cases, a single segment is enough to en-

code the post-convergence path in a loop-free manner.

But, even though TI-LFA solves the loss of connec-

tivity problem, it is insufﬁcient to ensure congestion

avoidance in the SR domain after the IGP has con-

verged.

The present work addresses the SR link failure

pos-convergence congestion problem. We start by

analyzing post-convergence congestion levels of net-

works that use TI-LFA and compare them with dis-

tinct possible approaches which resort to Evolution-

ary Computation (EC) algorithms. Network’s con-

ﬁguration on a normal state, that is, before a sin-

gle link failure, are optimized for congestion and use

the Single Adjacency Label Path Segment Routing

(SALP-SR) (Pereira et al., 2017) optimization model.

All proposed approaches extend SALP-SR, but some

may also be applied to any SR model. The remain-

ing of the article follows with a brief presentation

of the SALP-SR optimization method, the deﬁnition

of possible approaches to the addressed problem, the

presentation of results and their analysis and, ﬁnally,

some conclusions.

2 SALP-SR

Single Adjacency Label Path (SALP-SR) was pro-

posed by the authors as a technique to optimize

congestion for best-effort trafﬁc in Segment Routing

(Pereira et al., 2017). As stated by its name, this

method uses at the most a single adjacency segment

in the conﬁguration of each SR path and trafﬁc always

moves towards its destination. As a consequence,

each path is deﬁned using three or fewer segments

with the following alternatives of conﬁguration for-

mats:

• 1-Segment: The SR path conﬁguration only has a

single Node SID, the SID of the destination node

t. SR paths with such conﬁguration forward traf-

ﬁc across edge-to-edge IGP shortest-paths. The

majority of SR paths are 1-segment SR paths.

• 2-Segments: A SR path conﬁguration comprises

a Node SID and an Adjacency SID, with two pos-

sible formats: [(s, u);t] and [u;(u, t)]. In the ﬁrst,

an adjacency segment forwards trafﬁc to an adja-

cent node u, and from there on, trafﬁc is driven

across shortest-path. In the second case, trafﬁc

follows an IGP shortest-path to a destination ad-

jacent node and, lastly, an adjacent segment to the

destination.

• 3-Segments: A SR path conﬁguration comprises

two Node SIDs and an Adjacency SID formatted

as [u; (u, v);t], where (u, v) is the adjacency seg-

ment.

The optimization model assumes that all network

nodes are SR capable and that a Path Computation

Element (PCE) or SDN controller gathers topological

information and trafﬁc necessities required by the op-

timization algorithm. Figure 1 displays a conceptual

representation of SALP-SR. Evolutionary Algorithms

use information provided by the PCE to deliver solu-

tions which aim to minimize the network congestion.

An SR simulator evaluates each solution and deter-

mines how well it ﬁts the addressed problem.

An optimized solution includes three components:

1) a link weights conﬁguration for the link-state IGP

operation, 2) SR label paths and 3) trafﬁc splitting ra-

tios for provider edges (PE) operations. SALP-SR

conﬁgurations are preemptively computed consider-

ing foreseeable conditions and stored in a database

accessible by the PCE/SDN controller, so they can

be installed in the network when ever required. We

next describe the SALP-SR mathematical optimiza-

tion model.

The routing optimization model uses undirected

graphs G(N, A) to represent network topologies,

where N is a set of nodes and A a set of arcs. For

a given trafﬁc demand matrix and a routing conﬁgu-

ration, the congestion level of a network is evaluated

using the normalized sum of link cost Φ proposed in

(Fortz and Thorup, 2000). For every topology link a,

this convex function associates a congestion value Φ

to the link utilization ratio u(a). This work uses the Φ

function both as primary comparison metric and also

as primary optimization objective. The ﬁrst objective

of SALP-SR is to ﬁnd a conﬁguration W, a set of in-

teger values, which minimizes the normalized sum of

all links’ congestion, Φ

, i.e, minΦ =

∑

a∈A

(u(a)).

Other additional optimization objectives can also be

considered such as the network maximum link utiliza-

tion (MLU) or even delay restrictions.

A solution W is nothing more than a set of IGP’s

integer link weights w

u,v

, u, v ∈ N, but from which

the optimization algorithm also derives the other ad-

ditional settings, SR paths and load balancing ratios

between parallel paths.

To compute SR paths, and later load balancing

ratios, the optimization algorithm forwards packets

from a ﬂow with source s and destination t by apply-

ing a penalization on longer paths. Trafﬁc at a node

u, in a path from s to t, is forwarded to a next-hop

v with a probability that decreases exponentially with

Segment Routing Single Link Failure Congestion Optimization

PCE/SDN Controller

Topology and

Traffic information

SR Configuration

IGP weights

Splitting Ratios

PE2

PE1

PE2

Destination

Prefix

SR Path Ratio

PE2

⁞

⁞ ⁞

4/5

1/5

R3 R3-R2 PE2

PE1 SR Recovery Paths

Failure

R1-PE2

simulator

Optimization

engine

Solution

Traffic Matrices

Topology

Destination

Prefix

SR Path Ratio

PE2

⁞ ⁞

⁞

PE1 SR Paths

PE2

new proposed module

Figure 1: SALP-SR and Recovery Paths Optimization Conceptual Architecture.

the extra length of the path to t, h

u,v

. This compu-

tation is depicted in Equations 1 and 2, proposed in

(Xu et al., 2007), where d

is the shortest path dis-

tance from u to the destination t. The portion of traf-

ﬁc at u with destination t to be routed to next-hop v is

computed by the proportion function Equation 3. This

trafﬁc splitting method enables the optimization algo-

rithm not only to take advantage of IGP’s Equal Cost

Multi-Path (ECMP) but also to use non-shortest-path

links to forward trafﬁc while ensuring that packets are

always routed towards the destination.

SR paths are obtained based on the previous for-

mulation and according to the three available conﬁg-

uration formats:

• 3-Segments: For each non-shortest-path link

(u, v) in a path from s to t, identiﬁed in the de-

scribed procedure, the algorithm produces a hop-

by-hop path such that paths from s to u and from

v to t are shortest-paths. Non-shortest-path links

(u, v) become adjacency segments, while remain-

ing path portions are converted to node segments

([u;(u, v);t]).

• 2-Segments: When u or v coincide with source

or destination nodes, s and t respectively, the

path becomes a 2-segment SR path ([(s, u);t] or

[u;(u, t)]).

• 1-Segment: A shortest-path between s and t is

converted into a 1-segment SR path, [t].

Following the deﬁnition of the SR paths, the algo-

rithm assigns to each parallel SR path a portion of the

trafﬁc necessities from s to t. The fractions of traf-

ﬁc F to be assigned to each parallel path are obtained

by applying Equations 4 and 5. The fraction of trafﬁc

assigned to a path from a source s to a destination t

that contains an adjacency segment (u, v) is the prod-

uct of the fraction of trafﬁc assigned to (u, v) at node

u, P



u,v



, with the fraction of trafﬁc from s which

arrives at u using shortest-paths. This formulation is

depicted in Equation 4 where P

s,u

represents the set

of all shortest paths from s to u, and A

s,t

is the set of

all adjacency segments on SR paths from s to t. The

optimization model assumes that individual ﬂows are

relatively small and that a hashing-based trafﬁc split-

ting scheme ensures that packets from the same ﬂow

are routed along the same SR path.

u,v

= d

+ w

u,v

− d

(1)



u,v



(

−

u,v

, if d

< d

0, otherwise

(2)



u,v





u,v



∑

(u,i)∈A



u,i



, (3)

F (s, t, u, v) =













u,v



, u = s



u,v



∑

p∈P

s,u

∏

(i, j)∈p



i, j



, u 6= s

(4)

F (s, t) = 1 −

∑

(u,v)∈A

s,t

F (s, t, u, v) (5)

SALP-SR has an important feature that allows dy-

namic trafﬁc load balancing corrections between par-

allel edge-to-edge paths. The exponential parameter

p in Equation 2 affects the splitting of trafﬁc between

outgoing links. By default, all values are set to 1 and

are only modiﬁed if circumstances impose alterations

on trafﬁc load balancing between parallel paths. Fig-

ure 2 presents an illustrative example of how differ-

ent p-node values affect trafﬁc splitting locally at a

node u. For a given set of weights (Figure 2a) and

considering a p-node value of 1 at node u, the pro-

cedure described in the previous section splits trafﬁc

for destination t as depicted in Figure 2b. To respond

DCNET 2018 - International Conference on Data Communication Networking

(a)

42%

16 %

p = 1

(b)

34%

32 %

p = 10

(c)

Figure 2: p-node values example.

to some event, for example, to prevent congestion af-

ter the imposition of new trafﬁc necessities, a network

administrator may need to readjust the load balancing

at u such that trafﬁc becomes (almost) evenly divided

between the outgoing links. To achieve such result, it

is sufﬁcient to adjust to 10 the p-node value at node u,

as shown in Figure 2c, while preserving the remaining

functional conﬁgurations. The present work explores

this feature to improve the distribution of trafﬁc and

thereby reduce the network congestion after a single

link failure event.

3 LINK FAILURE

OPTIMIZATION APPROACHES

The traditional response to a link failure is to provide

a set of recovery paths to reestablish point-to-point

connectivity. However, this can be insufﬁcient. It is

also essential to take into consideration IGP shortest-

path recomputation and how it affects the network

congestion level. IGP shortest-paths recomputation

has an impact on node segments and, consequently,

on trafﬁc distribution over the available resources af-

ter a link failure. A simple approach, which provides

recovery paths from the Point of Local Repair (PLR)

on, may not be sufﬁcient to warrant adequate overall

network congestion levels. Other options might need

to be considered which include 1) the deﬁnition of

where SR segment ID stack should preferably be up-

dated, at the PLR or network ingress nodes; 2) which

portion of the recovery path should be updated, to the

destination or only to the next segment not affected

by the failure.

These questions deﬁne three end-to-end possibili-

ties: PLR to the destination (TI-LFA), PLR to the next

segment or edge-to-edge SR recovery paths (Figure

3). It might also be conceivable to implement cor-

rections on trafﬁc load balancing to improve trafﬁc

distribution, or even alter the SID stack of SR paths

that were not affected by the failure. In this con-

text, we devised distinct approaches to analyze sin-

gle link failure impacts of in SR network’s congestion

and simultaneously evaluate their responses. None of

the proposals consider SR path with service chaining,

and only paths that steer trafﬁc from source to des-

tination are considered. All approaches aim to min-

imize the network post-convergence congestion after

a single failure. Solutions for approaches that require

longer computation time are preemptively computed,

and stored in a PCE database as shown in Figure 1.

u v

1 2

3 4

Node-SID u

Adj-SID u-v

Adj-SID u-2

Node-SID C

Node-SID v

Initial SR Path Edge-to-Edge

PLR-to-Dst PLR-to-Next Seg.

Node-SID C

Node-SID u

Node-SID C

Node-SID u

Figure 3: SR recovery paths: Edge-to-Edge, PLR to desti-

nation, PLR to next segment.

3.1 Approach 1 - Simple Link

Protection

This approach is the most straightforward response to

a single link failure. Before any link failure, the net-

work is conﬁgured with an optimized SALP-SR con-

ﬁguration for congestion avoidance. Upon the fail-

ure of link (u, v), trafﬁc that previously traversed the

failed link is rerouted. SR paths that did not use the

failing link remain unchanged as are trafﬁc splitting

ratios between parallel paths.

Within this approach we consider two possibili-

ties to reroute trafﬁc that initially traversed the (u, v)

link: a) Edge-to-Edge rerouting and b) Point of Local

Repair rerouting.

1.a) Edge-to-Edge Rerouting: All SR paths that in-

cluded the failing link (u,v) are reconﬁgured with

a single Node Segment [node(t)], where t is the

provider’s edge destination. In practice, all trafﬁc

is rerouted according to the new computed shortest

paths. This approach has the disadvantage of not be-

ing responsive enough. It can only be implemented

after the fault is announced to all routers and the IGP

has converged, as edge routers need to become aware

Segment Routing Single Link Failure Congestion Optimization

of the fault. On the other hand, it does not require

any centralized control, and only edge nodes need to

recompute SR paths.

1.b) Point of Local Repair Rerouting: To enable an

under 50 msec response to a single link failure, IETF

proposed a Topology Independent Loop-Free Alter-

nate (TI-LFA) (Bashandy et al., 2018). The main

idea of TI-LFA is to provide loop-free recovery paths,

between the PLR and provider’s edge destinations,

which remain unchanged before and after the IGP

convergence.

In practice, and considering SALP-SR conﬁgu-

rations, SR paths deﬁned by a single Node SID are

kept unchanged as they are automatically repaired by

the IGP after its convergence. SR paths which ex-

plicitly use the (u, v) link are repaired at u. For ex-

ample, packets reaching u with a segment list [node-

SID u, adj-SID u-v, node-SID t] would leave u with

a segment list matching [RP,node-SID t], where RP

is a routing path to the Q node, a node segment to a

PQ node or a direct neighbor of node u (Bashandy

et al., 2018). The repair path conﬁgures the post-

convergence shortest-path from the PLR to the desti-

nation. In Figure 3 this case corresponds to the header

list [adj-SID u-2, node-SID C]. The adjacent segment

(adj-SID u-2) would not be included in the segment

header as node 2 is a neighbor of the PLR, and is only

included to ease comprehension.

3.2 Approach 2 - Link Protection with

SALP-SR Paths Recomputation

As stated before, a SALP-SR conﬁguration encom-

passes an IGP link weights conﬁguration, edge-to-

edge SR path deﬁnitions and load balancing splitting

ratios between parallel paths. All conﬁgurations are

derived from a set of integers, the IGP link weights,

and a set of real values assigned to each node (p-node

values). When a topology change is announced to

all network nodes (or a PCE), such as a link failure,

a previously computed new conﬁguration can be in-

stalled and deployed in a short amount of time. Ap-

proach 2 takes advantage of this SALP-SR feature.

Upon the failure o link (u,v), and considering that

link weights conﬁguration remains unchanged, two

recomputations are performed which exclude the fail-

ing link: IGP shortest-paths and edge-to-edge SALP-

SR paths. This is equivalent to applying the optimiza-

tion process described in Section 2 with the link fail-

ure topology alteration. A disadvantage of this pro-

cedure is that a small percentage of SR paths not af-

fected by the fault may need to be altered to reduce

congestion. Furthermore, the fault needs to propagate

before this recovery procedure can be applied. Addi-

tionally to IGP and SALP-SR recomputations, by ad-

justing p-node values, new trafﬁc splitting ratios be-

tween parallel SR paths may be installed, which leads

to the following two cases:

2.a) Default P-node Values: All p-node values are

kept unchanged with the default value of 1.

2.b) Optimized P-node Values: When the installed

conﬁguration becomes inadequate, due to trafﬁc

changes or link failures, different p-node values con-

ﬁgurations alter trafﬁc load balancing between par-

allel SR paths and may improve network operational

conditions. Hence, this approach in addition to steps 1

and 2, also optimizes the p-node values conﬁguration.

The p-node values can be preemptively optimized for

each link failure and stored in a database to be later

applied if necessary.

All recovery paths are edge-to-edge SR paths that

can be preemptively computed and stored (Figure 1).

3.3 Approach 3 - Multi-objective

Optimization

Although similar to the previous, this approach con-

siders two objectives instead of a single objective for

the initial SALP-SR optimized conﬁguration. Pre-

vious research by the authors in congestion multi-

objective optimization for link-state routing proto-

cols demonstrated promising results for the single

link failure congestion problem (Pereira et al., 2015).

However, and contrary to the aims of the present

work, link weights were only optimized for conges-

tion avoidance after the failure of a single or a reduced

set of speciﬁc links. Approach 3 aims to improve net-

works performance by minimizing both congestion

values simultaneously, before and after the failure of

any single link.

In this approach, the initial network optimization

is performed considering simultaneously: First Ob-

jective - minimize the congestion of the network on

a fully functional state, Φ; and Second Objective -

minimize the maximum congestion after a single link

failure. The formulation of this second objective is

Min



Max



(n−1,a)



, where (n − 1, a) denotes the

failure of each individual link a.

This scenario contemplates the recomputation of

only SR paths conﬁgurations (3.a), as well as the ad-

ditional installation of optimized trafﬁc splitting ra-

tios (3.b). It is expected that, by adding the second

optimization objective, the congestion levels of the

network, after a link failure, improve when compared

with the previous scenario.

DCNET 2018 - International Conference on Data Communication Networking

3.4 Approach 4 - Multiplane Recovery

Path Optimization

One of the attributes of SALP-SR is that trafﬁc always

moves towards the destination when the distance is

the shortest-path lengths. Although this characteristic

is a positive SALP-SR property, it nonetheless nar-

rows the number of possible recovery path solutions.

This approach forsakes this restriction, i.e., recovery

paths may include segments which locally drive traf-

ﬁc away from the destination. This goal is achieved

using additional network planes during the optimiza-

tion process, where each plane uses a different SR

conﬁguration. A conceptual representation of this ap-

proach is presented in Figure 4.

Plane 1

Plane 0

Plane 2

Node-SID B

Node-SID A

(a) Trafﬁc necessities identiﬁcation

Plane 1

Plane 0

Plane 2

Node-SID B

Adj-SID C-2

Node-SID A

Node-SID 4

Adj-SID 4-v

Node-SID B

1 2

(b) Recovery paths computation

Figure 4: Multiplane recovery path optimization.

The computation of recovery paths for the fail-

ure of each link (u, v) is divided into two main

steps. First, the optimization procedure identities traf-

ﬁc which, before failure, travels over (u, v). From

this analysis, the algorithm produces two trafﬁc de-

mand matrices, one for each of the failing link en-

try ports, that represent trafﬁc necessities which need

to be rerouted after the link failure, Figure 4a. New

hop-by-hop paths are then optimized in two separate

planes (plane 1 and 2), one for each trafﬁc matrix,

and such that the overall network congestion is min-

imized. Finally, hop-by-hop paths are translated into

SR paths for the installed SR conﬁguration (plane 0),

Figure 4b. Planes 1 and 2 only compute unique re-

covery paths between each source/destination pair to

ensure a functional mapping of paths’ translation and

also of SR paths into recovery paths.

Multi-objective optimizations (approach 3) use

NSGA-II (Non-dominated Sorting Genetic Algorithm

II) (Deb et al., 2002) as optimization engine. The au-

thors experimentally showed in (Pereira et al., 2015)

that NSGA-II has good behavior in IGP link weights

multi-objective optimization problems. The remain-

ing optimizations are performed using a Single Ob-

jective Evolutionary Algorithm (SOEA).

4 EXPERIMENTS AND RESULTS

4.1 Experiments Setup

The evaluation of each approach was performed using

a publicly available network optimization framework,

NetOPT, developed by the authors. We considered a

set of distinct synthetic network topologies, summa-

rized in Table 1, varying in size (30 and 50 nodes) and

minimum in/out node degree (2,3 and 4). Trafﬁc de-

mand matrices were randomly generated such that 1)

requirements between two nodes are inversely propor-

tional to their Euclidean distance, and 2) the expected

average link utilization is one-third of their capacity.

Trafﬁc matrices, for topologies with the same num-

ber of nodes but with more directed links, are conse-

quently more demanding.

Table 1: Network Topologies used in the experiments.

Topology Type Nodes Links Min. degree

Rand30

Synthetic 30 110 2

Rand30

Synthetic 30 190 3

Rand30

Synthetic 30 220 4

Rand50

Synthetic 50 194 2

Rand50

Synthetic 50 380 4

Solutions take integers values from the range [1;

20] for link weights, while p-node values are real

numbers in the range [0.01; 10]. The optimization ob-

jectives are the minimization of the normalized sum

of links’ congestion cost Φ. It is of notice that when

this normalized cost equals 1, all loads are below 1/3

of the link capacity, and when all arcs are exactly full

the congestion value is 10 2/3.

4.2 Results

Results presented in Table 2 are average post-

convergence congestion values from 10 runs of each

experiment. They are divided into two main groups,

before (normal state) and after a single link failure.

In the last, organized by ”Approach”, values are the

mean of network congestion after the failure of each

link, one at a time. Optimized congestion values for

networks in their normal state, that is, fully functional,

Segment Routing Single Link Failure Congestion Optimization

are operational benchmarks with which the remaining

values contrast.

The minimum node in/out-degree of a network

topology signiﬁcantly inﬂuences the quality of recov-

ery paths. Topologies with higher minimum node de-

gree have more edge-to-edge recovery paths available

after a single link failure. In this context, it is un-

derstandable that topologies with a higher minimum

node in/out-degree present globally better results.

Simple Link Protection

Results comparison for the ﬁrst group of link fail-

ure experiments, simple link protections (Approach

1), evidence no signiﬁcant differences in the conges-

tion values between edge-to-edge (1.a) and point of

local repair rerouting (1.b). They both globally dis-

play values below (but near) the operational thresh-

old of the network (10 2/3). These approaches, 1.a

and 1.b, only provide recovery shortest-paths for the

affected trafﬁc. In particular, the TI-LFA simula-

tions (1.b), although capable of shortening connec-

tivity lost to under 50 msec, show high congestion

values. The TI-LFA path recovery strategy does not

account for impacts on congestion resulting from IGP

post-convergence shortest-path changes. PLR recov-

ery paths may also present additional issues. After

the IGP convergence, the PLR might no longer be on

the path to the destination. The IGP shortest-path re-

computation would repair the SR path. Another issue

may also arise if the PLR is explicitly deﬁned in the

SR path. Due to network design, trafﬁc may need to

be routed back to already traversed nodes.

Link Protection with SALP-SR Paths

Recomputation

Approaches 2.a and 2.b enhance the simple edge-to-

edge recovery approach (1.a) by taking advantage of

SALP-SR functionalities, and as expected, enabling

network congestion to diminish. It is important to

reemphasize that none of the approaches 2.a and 2.b

alters IGP weights. In 2.a, only SR paths are reconﬁg-

ured by the SALP-SR algorithm, reﬂecting the newly

computed IGP shortest-paths. If additionally load bal-

ancing ratios are adjusted (2.b), network congestion

values drop to half, on average, of those observed with

a simple rerouting (approach 1).

Multi-objective Optimization

Results for this approach show that, after a SALP-SR

paths recomputation (3.a vs. 2.a), networks perform

slightly better with multi-objective than with single

objective optimizations. Nonetheless, these differ-

ences fade with the adjustment of load balancing ra-

tios (3.b vs. 2.b). A multi-objective optimization

establishes a compromise between the optimization

goals, a trade-off, by relaxing conﬁguration ﬁtness on

both objectives. However, an increase in SR conﬁg-

urations ﬂexibility, with a penalization on fully func-

tional network congestion, is insufﬁcient to improve

on all results obtained with single objective optimiza-

tion. Additional measures need to be installed to im-

prove the already obtained results. Approach 4, mul-

tiplane recovery path optimization, gives the improve-

ment that is needed.

Multiplane Recovery Path Optimization

Multiplane recovery path optimization adds to SALP-

SR the ability to have more ﬂexible SR paths. Al-

though recovery SR paths may not always locally for-

ward trafﬁc towards the destination, they permit a re-

duction in post-failure congestion signiﬁcantly. Re-

sults show that in most experiments this approach fa-

cilitates post-failure congestion values almost equiv-

alent to those observed in fully functional networks.

The higher minimum network’s node in/out degree,

the lower are post-convergence congestion values.

Moreover, this approach only requires edge nodes

to re-conﬁgure SR paths affected by the link failure.

SR recovery paths are pre-emptively computed and

stored in a PCE/SDN controller or locally at each

router. Therefore, replacement paths can be imme-

diately installed after the fault propagation. If traf-

ﬁc necessities, which are used to compute recovery

paths, change signiﬁcantly and undermine the quality

of the recovery SR paths, a PCE can quickly compute

new load balancing ratios between parallel paths and

improve the overall network congestion.

5 CONCLUSIONS

Single link failures have two main impacts on net-

work’s operations, they undermine connectivity and

impact the overall network congestion negatively.

Segment Routing, a recent routing technology, en-

ables the deployment of more complete and effective

response to the problem of preserving network’s post-

failure congestion levels. We derive two main conclu-

sions from the possible approaches to this problem ex-

plored in this work. First, although TI-LFA is an ex-

cellent solution to reestablish networks connectivity

in a brief amount of time, it is insufﬁcient to provide

functional levels of congestion after a link failure.

Multiplane preemptive path recovery computation, on

DCNET 2018 - International Conference on Data Communication Networking

Table 2: Average congestion before and after single link failure.

Topology

Normal State Single Link Failure State

SALP-SR Simple Reroute SAPL-SR SO SAPL-SR MO Multiplane

SO MO 1.a 1.b 2.a 2.b 3.a 3.b 4

Rand30

1.40 1.42 10.40 10.45 8.49 6.09 8.19 6.26 5.06

Rand30

1.69 1.76 4.07 4.09 3.74 2.41 2.87 1.88 1.74

Rand30

1.96 2.35 9.05 9.11 6.75 3.46 5.14 3.47 2.16

Rand50

1.79 2.07 9.41 9.53 8.17 6.10 7.01 5.98 4.86

Rand50

3.74 4.91 9.48 9.57 8.98 5.31 9.14 5.87 3.73

the other hand, delivers good post-convergence con-

gestion levels but requires more time to be deployed.

Although we defend that changes on SR paths should,

when possible, be implemented at edge nodes instead

of at points of local repair, a combination of both

approaches presents itself as a good compromise to

achieve both goals, and shorten reaction time and low

network congestion.

After a link failure, connectivity can be reestab-

lished using TI-LFA, and as soon as the IGP con-

verges, optimized SR paths can be installed at edge

nodes, achieving this way both desired goals. Re-

covery paths are preemptively computed, considering

foreseeable trafﬁc necessities, and can be stored lo-

cally at edge router or centrally by a PCE. By provid-

ing a better distribution of trafﬁc among available re-

sources, resilience to trafﬁc variations also increases.

Nonetheless, changes in trafﬁc necessities may also

be addressed by correcting trafﬁc load balancing be-

tween parallel paths, using the p-node values opti-

mization feature. This last can be implemented in

less than a 100 seconds in topologies with less than

50 nodes. The NetOpt framework is available at

http://darwin.di.uminho.pt/netopt.

ACKNOWLEDGMENTS

This work has been supported by COMPETE: POCI-

01-0145-FEDER-007043 and FCT Fundac¸

ao para

a Ci

encia e Tecnologia within the Project Scope:

UID/CEC/00319/2013.

REFERENCES

Atlas, A. and Zinin, A. (2008). Basic Speciﬁcation for IP

Fast Reroute: Loop-Free Alternates. RFC 5286.

Bashandy, A., Filsﬁls, C., Decraene, B., Litkowski, S.,

and Francois, P. (2018). Topology Independent Fast

Reroute using Segment Routing. Internet-Draft draft-

bashandy-rtgwg-segment-routing-ti-lfa-02. Work in

Progress.

Deb, K., Pratap, A., Agarwal, S., and Meyarivan, T. (2002).

A fast and elitist multiobjective genetic algorithm:

Nsga-ii. IEEE Transactions on Evolutionary Compu-

tation, 6(2):182–197.

Feamster, N., Rexford, J., and Zegura, E. (2013). The road

to sdn. Queue, 11(12):20:20–20:40.

Filsﬁls, C., Previdi, S., Bashandy, A., Decraene, B.,

Litkowski, S., Horneffer, M., Milojevic, I., Shakir, R.,

Ytti, S., Henderickx, W., Tantsura, J., and Crabbe,

E. (2013). Segment Routing Architecture. Internet-

Draft draft-ﬁlsﬁls-rtgwg-segment-routing-01, Internet

Engineering Task Force. Work in Progress.

Fortz, B. and Thorup, M. (2000). Internet trafﬁc engineer-

ing by optimizing ospf weights. In INFOCOM, pages

519–528. IEEE.

Gredler, H. and Goralski, W. (2005). The Complete IS-IS

Routing Protocol. Springer.

Hopps, C. (2000). Analysis of an Equal-Cost Multi-Path

Algorithm. RFC 2992 (Informational).

Moy, J. (1997). OSPF Version 2. RFC 2178 (Draft Stan-

dard). Obsoleted by RFC 2328.

Pereira, V., Rocha, M., and Sousa, P. (2017). Optimiz-

ing segment routing using evolutionary computation.

In FNC/MobiSPC, volume 110 of Procedia Computer

Science, pages 312–319. Elsevier.

Pereira, V., Sousa, P., Cortez, P., Rio, M., and Rocha, M.

(2015). Comparison of single and multi-objective evo-

lutionary algorithms for robust link-state routing. In

EMO, volume 9019 of Lecture Notes in Computer Sci-

ence, pages 573–587. Springer.

Xu, D., Chiang, M., and Rexford, J. (2007). Deft: Dis-

tributed exponentially-weighted ﬂow splitting. In

IEEE INFOCOM 2007 - 26th IEEE International

Conference on Computer Communications, pages 71–

79.

Segment Routing Single Link Failure Congestion Optimization