Virtual topology partitioning towards an efficient failure recovery of software defined networks

Abstract

Software Defined Networking is a new networking paradigm that has emerged recently as a promising solution for tackling the inflexibility of the classical IP networks. The centralized approach of SDN yields a broad area for intelligence to optimise the network at various levels. Fault tolerance is considered one of the most current research challenges that facing the SDN, hence, in this paper we introduce a new method that computes an alternative paths re-actively for centrally controlled networks like SDN. The proposed method aims to reduce the update operation cost that the SDN network controller would spend in order to recover from a single link failure. Through utilising the principle of community detection , we define a new network model for the sake of improving the network's fault tolerance capability. An experimental study is reported showing the performance of the proposed method. Based on the results, some further directions are suggested in the context of machine learning towards achieving further advances in this research area.

Full text

VIRTUAL TOPOLOGY PARTITIONING TOWARDS AN EFFICIENT FAILURE
RECOVERY OF SOFTWARE DEFINED NETWORKS
ALI MALIK1, BENJAMIN AZIZ1, CHIH-HENG KE2, HAN LIU1, MO ADDA1
1School of Computing University of Portsmouth, Portsmouth PO1 3HE, UK
2Department of Computer Science and Information Engineering, National Quemoy University, Taiwan
E-MAIL: {ali.al-bdairi, benjamin.aziz, han.liu, mo.adda}@port.ac.uk, smallko@gmail.com
Abstract:
Software Defined Networking is a new networking paradigm
that has emerged recently as a promising solution for tackling the
inflexibility of the classical IP networks. The centralized approach
of SDN yields a broad area for intelligence to optimise the network
at various levels. Fault tolerance is considered one of the most cur-
rent research challenges that facing the SDN, hence, in this paper
we introduce a new method that computes an alternative paths re-
actively for centrally controlled networks like SDN. The proposed
method aims to reduce the update operation cost that the SDN
network controller would spend in order to recover from a single
link failure. Through utilising the principle of community detec-
tion, we define a new network model for the sake of improving the
network’s fault tolerance capability. An experimental study is re-
ported showing the performance of the proposed method. Based
on the results, some further directions are suggested in the context
of machine learning towards achieving further advances in this
research area.
Keywords:
Network Topology, Community Detection, Graph Theory, Soft-
ware Defined Networking
1. Introduction
The Internet and networking system in general plays an es-
sential role in changing our life style through producing vari-
ous type of technologies that get involved in our daily activities
such as social media, economics and business. Networking de-
vices exchange data in a variety networks (e.g. wireless sensor
networks, the Internet-of-Things and Cloud networks) where
currently there are about 9 billion devices connected to the In-
ternet and this number is expected to more than double by 2020.
However, the today’s IP network infrastructures do not have
the ability to accommodate such a huge number of devices,
hence the Internet ossification is highly expected [1]. One ef-
fective solution to tackle the ossification issue is via replace
the complex and rigid networking systems by a programmable
networking instead.
Software-defined networking (SDN) was resulted from a
long history of efforts that have been exerted for the purpose
of simplifying the computer networks management and control
[2]. In SDN the control plane has been moved out from the
data plane and placed in a central location usually called the
network controller or the network operating system. Due to
this decoupling, the network forwarding elements (e.g. routers
and switches) became a dummy devices, which typically dic-
tated by the network controller. In reality, the SDN controller is
either has a direct connection to the data plane forwarding ele-
ments or indirect way through which the controller will be able
to instruct those devices. The OpenFlow [3] protocol is the
most commonly used to establish the connection between the
data plane forwarding elements and the controller, so the con-
troller will be able to send the forwarding rules to each device
that lies on its domain. So far, SDN has gained much attention
of both academia and industry community and adopted by some
of well known pioneering companies like Deutsche Telekom,
Google, Microsoft, Verizon, and Yahoo [4]. Although SDN has
brought a significant benefits to the concept of networks, some
new challenges accompanied this innovation such as faults,
configuration, security and performance [5]. Thus, our goal
in this paper is to eliminate the SDN drawbacks through miti-
gating the one of the most current challenges, which is the data
plane fault tolerance.
The remainder of this paper is organised as follows. In Sec-
tion 2 various techniques to solve the issue of SDN data plane
failures are presented. Section 3 and 4 illustrate our model and
the proposed framework. The experimental results of the eval-

uation are presented in section 5. Finally, the conclusion of this
paper and future work directions are provided in section 6.
2. Related work
The topic of SDN faults and recovery has been already stud-
ied, so we will discuss the relevant literature in this section.
Since the SDN has two separated planes (i.e. control and data),
hence each plane is susceptible to failure. Apart from control
plane failure, which is not the main focused in this work, the
recovery mechanism of data plane can be classified into pro-
tection and restoration techniques [6]. In protection method,
which also known as proactive, the flow entries of the backup
paths are determined in early stage of failure, hence the effected
data packets will convey through the backup path directly at the
moment of failure. On the other hand, in restoration mecha-
nism, which also known as reactive, the backup path will be de-
termined by the controller after the occurrence of failure, thus
extra time is required to set up the discovered alternative path.
Authors in [7] and [8] have shown how a fast data plane recov-
ery can be achieved within the protection approach. However,
the cost of proactive mechanisms is high as it consumes the ca-
pacity of the Ternary Content Addressable Memory (TCAM)
[6], where the forwarding rules to be stored. In addition, there
is no guarantee that the pre-planned paths will be in a good
health once the primary paths will fail.
In contrast, efforts have been exerted to investigate the effec-
tiveness of the reactive approach. In this context, the authors in
[9] and [10] have shown how a fast restoration can be accom-
plished, however in both works the experimental topologies
were small scale (i.e. 6 and 14 nodes respectively). Further-
more, the processing time for setting up the chosen path was
ignored, which is a requirement in SDNs in order to re-routing
from the affected primary path to the backup one.
Unlike the previous studies, the authors in [11] produced
a new method for fast restoration by reducing the processing
time, which typically paid by the controller, through finding
the alternative path (from end-to-end) that has a minimum op-
eration requirement. One main drawback with this work is that
it does not guarantee the sequentiality of the health nodes in the
effected path to be in similar order on the alternative one. Addi-
tionally, there is a lack of information regarding the simulation
tool that has been used. We also noted that there was a mistake
with the first network topology that used in their experiments,
since the authors mentioned that it contains 37 nodes and 57
links, while in reality their network contains a 56 links as it
missed the edge that connects between Hamburg and Frankfurt
according to the SNDlib library [12], hence their results could
be significantly deviant from the expected one.
All the above issues motivated us to put a further efforts in
order to investigate a more possible solutions for the problem of
SDN fault tolerance. The next section will present the proposed
system model of this work.
3. System model
The most frequently used notations throughout this paper are
listed in Table 1.
TABLE 1. List of notations
Symbol Description
CThe set of cliques
cThe failed clique
rsSource router
rdDestination router
ric Any arbitrary router in the failed clique
rjc Any arbitrary router in the failed clique
Pmin Dijkstra’s shortest path in terms of number of hops
DGDijkstra for finding the shortest path based on the graph G
DcDijkstra for finding the shortest path based on the clique c
We utilised the undirected graph theory as a basis to model
the computer network topology. Generally, each simple graph
G= (V, E )consists of a set of vertices (i.e. nodes), V, as well
as a set of edges (i.e. links), E, which connect the nodes to one
another. The set of all links in Gcan be defined as a 2-element
subset of nodes, E⊆V×V. We define a path P, from
source to destination, as a sequence of consecutive vertices
representing nodes or routers in the network1. The path starts
at the source router, rs, and ends with a destination router, rd,
with ric and rjc being any two adjacent routers along P:
P= (rs, . . . , ric, rj c, . . . , rd)
We define the set of all possible paths, Prs,rd, between any
source router rsand destination router rd, as the following set:
Prs,rd={P|(first(P) = rs)∧(last(P) = rd)}
and the definition of first and last is given as functions on any
general sequence (a1, . . . , an):
first((a1, . . . , an)) = a1
last((a1, . . . , an)) = an
1We use the terms router and node interchangeably.

FIGURE 1. Community detection
Community detection (or sometimes known as cliques iden-
tification) has been productively proposed as a solution to re-
solve various kinds of network-related problems including the
problem of network path optimisation. In this context, we will
use the concept of non-overlapping cliques as an approach to
optimise SDN restoration through accelerating the process of
failure recovery. By dividing the network’s graph Ginto a cer-
tain number of cliques, we are assuming that when the link
failure occurs then only one clique will suffer from the failure.
Meantime, the other cliques should be working fine. It is highly
likely that most of the path’s links will be distributed over var-
ious cliques, hence at the link failure moment, the only clique
c, which includes the effected link, will be treated rather than
dealing with the whole graph Gin looking for an alternative
path from end-to-end.
Informally, Figure 1 depicts how the virtual cliques can be
extracted from the original network topology graph through
applying any community detection algorithm, where the
number of the extracted cliques is vary and typically depends
on the network topology and the used algorithm itself. In our
model, we define the set of cliques dividing the network graph
Gas C⊆G, where individual cliques c∈Care defined as
follows:
c= (V0, E0)|V0⊆V∧E0⊆E
The definition of cliques is mutually exclusive, in other words:
∀c1, c2∈C|c1= (V1, E1)∧c2= (V2, E2)⇒(V1∩V2=
∅)∧(E1∩E2=∅)
We define a failed link (i.e. a 2-router sub-path), F, in a path
Pas follows:
F= (ric, rj c)| ∃ c:c= (V0, E 0)∧F∈E0
This definition of Fassumes only those cases where failed
links are always of an intra-clique type, i.e. it will never be the
case that the failed link connects two cliques, i.e.:
@F= (ric, rj c), c1= (V1, E1), c2= (V2, E2)|
c16=c2⇒ric ∈V1∧rjc ∈V2
Inter-clique failures will be the focus of future work.
We use the term longest-shortest path as the path that has
the maximum number of hops amongst the set of solutions
returned by applying Dijkstra’s algorithm [13] to our network
topology for finding the shortest path between every possible
two nodes in that network. To find this longest-shortest path,
we define the special function LS as follows:
LS(PDset) = x, such that x∈ PDset and
∀y∈ PDset :len(y)≤len(x)
If there are more than one longest-shortest paths in PDset,
we pick one randomly. PDset itself represents the set of all
Dijkstra-based solutions for some network topology (V, E ):
PDset ={P| ∀ rs, rd∈V:P=D(Prs,rd)}
4. The Proposed Framework
From a high level view, Figure 2 illustrates the main com-
ponents of our proposed framework where the Fault Tolerance
Enhancer component is the primary contribution of this frame-
work. We discuss next in more detail the developed component
as well as the tools that we used in this framework.
FIGURE 2. Proposed framework components
4.1 SDN controller
SDN controller representing the network’s brain and the
most vital part, which is the place where the intelligence and

decision making resides. Presently, there are more than 30 dif-
ferent SDN controller offering from both academia (e.g. for
research purposes) and industry (e.g. for commercial use). Our
framework currently supports the POX controller [14], which
is a python-based open source SDN controller. We selected
the POX as it is more suitable for research purposes and also
more convenient for fast prototyping than the other available
ones [15]. The OpenFlow [3] protocol is used to communicate
between the control and data planes in order to gather statistics
from the data plane and carry the controller instructions to pro-
gram the data plane elements, whereas the set of POX APIs can
be utilised for developing various network applications.
4.2 Fault tolerance enhancer component
Currently, there are three main parts that consisting this com-
ponent as follows:
A.Topology parser: will be responsible for fetching the under-
lying network topology characteristics and build a topological
view with the aid of the POX openflow.discovery, which is an
already developed component. In order to represent the gained
network topology as a graph G, we utilised the NetworkX [16]
tool, which is a pure python package with a set of functions
that can be used to manipulate and simplify the network graphs.
B.Cliques producer: is responsible for virtually partitioning
the network topology graph Ginto C(i.e. sub-graphs) by
incorporating the well known community detection algorithm
Girvan and Newman [17] to produce the possible cliques
(with any size) on the basis of the network graph that acquired
from the topology parser. The densely connection between
the resulted clique’s vertices is the main feature of Girvan
and Newman that interesting to us, in other words, the strong
connection among the nodes in each clique could provide a
multiple alternative paths that would utilised at the failure
events. To do so, we have incorporated the igraph tool [18]
to our framework, which is a python open source library for
analysing and manipulating the graphs where the Girvan and
Newman algorithm is already implemented.
C.Route finder: is used to identify the paths for the network’s
flows on the basis of the global view of the underlay topology,
which can be gained from the topology parser. in fact, the route
finder will be called on two occasions; the first one when a new
packet arrives to the network, while the second one when the
failure occurs so a new path should be computed. For this pur-
pose, two algorithms have been developed to obtain the shortest
path based on the well known Dijkstra’s algorithm. We adopt
the hop count as an additive metric on which the Dijkstra’s al-
gorithm will be able to form the shortest path for any incoming
requests. The first algorithm reflects the default action that per-
formed by the SDN controller at failure moments, which is to
erase the flow entries of the effected path and then install the
rules of the backup one from the source router rsto destination
rd, the pseudo code is demonstrated in Algorithm 1.
Algorithm 1: First algorithm to find the shortest path with
Dijkstra from End-to-End based on the Graph G
On Normal: Set Primary Path as Pmin ∈ P rs,rd
On Failure : Do the following procedure
1Prs,rd:= Prs,rd− {Pmin}
2Pmin := DG(Prs,rd)
On the other hand, the second algorithm will tackle the fail-
ure based on the effected clique c∈Crather than searching
from end-to-end based on the whole graph G, the pseudo code
of this algorithm is illustrated in Algorithm 2. Currently, the
Algorithm 2 is capable to manage the clique’s intra-link failure
and for the future we will extend the framework to include the
inter-link failure among the cliques.
Algorithm 2: Second algorithm to find the shortest path
with Dijkstra on the basis of either Gor effected c
On Normal: Set Primary Path as Pmin ∈ P rs,rd
On Failure : Do the following procedure
1if F= (ric, rj c)then
2Pmin := Dc(Pric,rj c )
3end
The complexity of the two above algorithms are similar to
the Dijkstra, which is O(|V|+|E|log |V|). It also worth not-
ing here that both algorithms have the same strategy to obtain
the shortest path for the newly arriving packets, which is by
running the Dijkstra based on the network graph G. However,
they differ in respect of finding the alternative path to recover
from the failure, in which the Algorithm 2 will turn to apply the
Dijkstra on the crather than G.
5. Simulations and experimental results
In order to test the performance of the proposed framework,
we first build it on top of POX controller as illustrated in Figure
2, then we used the popular SDN emulator Mininet [19] to eval-
uate the prototype of our proposed framework. As mentioned
earlier that SDN has splitted the network architecture into con-
trol plane, the brain, and data plane, the muscles. So, we have

FIGURE 3. Cliques-based network topology of Germany50, [12]
modelled the Germany50 from SNDlib [12] as an instance of
a real world network topology to act our data plane structure,
which contains 50 nodes and 88 edges. The black links in
Figure 3 represent the original Germany50 topology and we
added the colored layer (i.e. from c1 to c6) to demonstrate the
cliques that achieved after applying the Girvan and Newman
algorithm. According to the Germany50 topology, the longest-
shortest path lies between 1 and 13 through the route:(1, 5, 27,
22, 32, 42, 48, 50, 26, 13), there might be several longest-
shortest paths in the network topology and in such a case we
choose randomly. We attached two virtual hosts H1and H2
to the rsand rdof the longest-shortest path respectively in or-
der to simulate the scenario of packet injection. For now let’s
assume that the link (27, 22) fails. Consequently, Algorithm
1 will response to the failure by removing the failed-primary
path and install (1, 25, 47, 3, 39, 7, 41, 17, 40, 13) as an alter-
native shortest path to mask the failure. Since our selected path
passes through 4 cliques, which are c1,c2,c3 and c6, this
means only one clique (e.g. sub-graph) will be involved at the
moment of the intra link failure in the scenario of Algorithm
2, which will react first by detecting the effected clique c(i.e.
c2 in our case) and thence find a loop-free shortest path be-
tween the routers on both sides of the failed link, i.e. between
the two nodes of F. According to the given example, the sub-
path (27, 45, 22) will be returned by the second algorithm as
a quick solution based on c2 without taking into consideration
the other clique’s nodes as they will remain on the same settings
(untouched). Hence, it can be clearly seen that the second algo-
rithm does not guarantee the shortest path from end-to-end as
it works over a clique and not a whole graph. We are expecting
a significant reduction in the duration of the recovery process
for two reasons, (I) because the proposed algorithm will con-
sider only one clique but not the whole graph and (II) due to
minimising the number of rules modification. Figure 4 shows
the behaviour of the two algorithms, which is based on the se-
lected path (i.e. from 1 to 13), in both failure and non-failure
scenarios. We note that the two algorithms have nearly similar
FIGURE 4. Before and after failure measurements
durations for setting up the path (i.e. before failure scenario).
However unlike the Algorithm 1, Algorithm 2 has an extremely
positive impact in terms of enhancing the time cost of fault tol-
erance. The reduction rate can be calculated through compar-
ing the consumed time before and after the failure, which can
be arrived through the following formula:
R=Time BF −Time AF
Time BF ×100 (1)
Where Rrepresents the reduction rate, Time BF is the time
of setting up the primary path before the failure and Time AF
is the time of setting up the alternative path after the failure.
Based on (1) and according to the result of figure 4, the Rof
Algorithm 2 is 84.333%. Let us assume that len(rs, ..., rd) =
n, then the utilisation rate of Algorithm 2 can be measured
through (n−2
n)×100%, hence, the pre-installed rules utilisa-
tion percentage (based on our example case) will be 80%, this
ratio interprets the high reduction rate that gained by the algo-
rithm. As a result, the larger the value of n, the more utilisation
we obtain out of the selected primary path.
6. Conclusions
This paper provides a new approach for ameliorating the
fault tolerance mechanism of software-defined networks. We

have defined a new network model based on the set theory and
graph theory to represent the theoretical part. We also have
developed a new framework on the basis of the created model
to represent the practical side of this work. The experimen-
tal results, which have been carried out through a well-known
tools and emulation, demonstrate how the proposed method has
led to improving the performance of reactive failure recovery
through segmenting the network topology into a certain num-
ber of non-overlapping cliques. The proposed approach has not
been explored so far, which indicates that this paper first time
suggests the partitioning as a technique towards enhancing the
restoration mode of the SDN’s fault tolerance.
In future, we will position the study in the setting of ma-
chine learning, towards achieving that the routing solution can
be dynamically adjusted according to the update of the statis-
tical details of topological data. In other words, we aim to in-
volve learning strategies in routing to achieve globally optimal
search towards finding the shortest path.
Acknowledgements
This work is supported by the College of Computer Science
and Information Technology at the University of Al-Qadisiyah
under the scholarship program of the Iraqi Ministry of Higher
Education and Scientific Research.
References
[1] Lin, P., Bi, J., Hu, H., Feng, T., & Jiang, X. (2011, Novem-
ber). A quick survey on selected approaches for preparing pro-
grammable networks. In Proceedings of the 7th Asian Internet
Engineering Conference (pp. 160-163). ACM.
[2] Feamster, N., Rexford, J., & Zegura, E. (2014). The road to
SDN: an intellectual history of programmable networks. ACM
SIGCOMM Computer Communication Review, 44(2), 87-98.
[3] McKeown, N., Anderson, T., Balakrishnan, H., Parulkar, G., Pe-
terson, L., Rexford, J., ... & Turner, J. (2008). OpenFlow: en-
abling innovation in campus networks. ACM SIGCOMM Com-
puter Communication Review, 38(2), 69-74.
[4] Kreutz, D., Ramos, F. M., Esteves Verissimo, P., Esteve Rothen-
berg, C., Azodolmolky, S., & Uhlig, S. (2015). Software-defined
networking: A comprehensive survey. Proceedings of the IEEE,
103 (1), 14-76.
[5] Wickboldt, J. A., De Jesus, W. P., Isolani, P. H., Both, C. B., Ro-
chol, J., & Granville, L. Z. (2015). Software-defined network-
ing: management requirements and challenges. IEEE Commu-
nications Magazine, 53(1), 278-285.
[6] Akyildiz, I. F., Lee, A., Wang, P., Luo, M., & Chou, W. (2014).
A roadmap for traffic engineering in SDN-OpenFlow networks.
Computer Networks, 71, 1-30.
[7] Kempf, J., Bellagamba, E., Kern, A., Jocha, D., Takcs, A., &
Skldstrm, P. (2012, June). Scalable fault management for Open-
Flow. In Communications (ICC), 2012 IEEE International Con-
ference on (pp. 6606-6610). IEEE.
[8] Sgambelluri, A., Giorgetti, A., Cugini, F., Paolucci, F., & Cas-
toldi, P. (2013). OpenFlow-based segment protection in Ethernet
networks. Journal of Optical Communications and Networking,
5(9), 1066-1075.
[9] Sharma, S., Staessens, D., Colle, D., Pickavet, M., & De-
meester, P. (2011, October). Enabling fast failure recovery in
OpenFlow networks. In Design of Reliable Communication Net-
works (DRCN), 2011 8th International Workshop on the (pp.
164-171). IEEE.
[10] Staessens, D., Sharma, S., Colle, D., Pickavet, M., & Demeester,
P. (2011, October). Software defined networking: Meeting car-
rier grade requirements. In Local & Metropolitan Area Networks
(LANMAN), 2011 18th IEEE Workshop on (pp. 1-6). IEEE.
[11] Astaneh,S.A.,& Heydari,S. S. (2016).Optimization of SDN flow
operations in multi-failure restoration scenarios.IEEE Transac-
tions on Network and Service Management,13(3),421-432.
[12] SNDlib library, http://sndlib.zib.de.
[13] Dijkstra, E. W. (1959). A note on two problems in connexion
with graphs. Numerische mathematik, 1(1), 269-271.
[14] POX , https://openflow.stanford.edu/display/ONL/POX+Wiki
[15] Shalimov, A., Zuikov, D., Zimarina, D., Pashkov, V., & Smelian-
sky, R. (2013, October). Advanced study of SDN/OpenFlow
controllers. In Proceedings of the 9th central & eastern euro-
pean software engineering conference in russia (p. 1). ACM.
[16] Schult, D. A., & Swart, P. (2008, August). Exploring network
structure, dynamics, and function using NetworkX. In Proceed-
ings of the 7th Python in Science Conferences (SciPy 2008) (Vol.
2008, pp. 11-16).
[17] Girvan, M., & Newman, M. E. (2002). Community structure
in social and biological networks. Proceedings of the national
academy of sciences, 99(12), 7821-7826.
[18] Csardi, G., & Nepusz, T. (2006). The igraph software package
for complex network research. InterJournal, Complex Systems,
1695(5), 1-9.
[19] Lantz, B., Heller, B., & McKeown, N. (2010, October). A net-
work in a laptop: rapid prototyping for software-defined net-
works. In Proceedings of the 9th ACM SIGCOMM Workshop on
Hot Topics in Networks (p. 19). ACM.