Download as pdf or txt
Download as pdf or txt
You are on page 1of 7

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/340229620

Routing Optimization in SDN using Scalable Load Prediction

Conference Paper · December 2019


DOI: 10.1109/GIIS48668.2019.9044960

CITATIONS READS
4 172

3 authors:

Majdoub Manel Ali El Kamel


University of Sousse University of Sousse
6 PUBLICATIONS   21 CITATIONS    22 PUBLICATIONS   52 CITATIONS   

SEE PROFILE SEE PROFILE

Habib Youssef
University of Sousse
238 PUBLICATIONS   3,976 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Design and evaluation of an IP switch architecture optimized for switching packets flows presenting different qualities of services View project

SDN/NFV View project

All content following this page was uploaded by Majdoub Manel on 24 March 2021.

The user has requested enhancement of the downloaded file.


GIIS'19 1570605357
1

1  
2  
3  
4  
Routing Optimization in SDN using Scalable Load
5  
6  
Prediction
7  
Manel Majdoub ∗ , Ali El Kamel† and Habib Youssef‡
8  
9  
PRINCE Lab
10  
11   University of Sousse, Tunisia
12   Email: ∗ manelmaj 2007@hotmail.com, † ali.el.kamel@hotmail.com, ‡ habib.youssef@fsm.rnu.tn
13  
14  
15  
16   Abstract—With the exponential growth of data traffic and be unable to achieve those objectives. Therefore, it is necessary
17   the rapid development of smart devices, networks are becoming to introduce intelligence within routing processes.
18   more and more heterogeneous and complex. Therefore, managing Obviously, Deep Learning (DL) is a powerful technology
19   network resources with traditional routing features is no longer
able to add intelligence to various applications such as traffic
20   advised. More intelligence needs to be deployed.
However, deploying intelligence in traditional networks seems prediction, classification, recognition and, especially, routing
21   processes. Clearly, the architecture of SDN is very suitable for
to be hard to achieve since they are naturally distributed. The
22   emerging of Software Defined Networking (SDN) will ease the DL application since it is inherently centralized. Moreover,
23   introduction of intelligence in networks. In this vein, Deep the application of DL algorithms is straightforward since
24   Learning (DL) is considered the most promising concept for
programmability is a fundamental feature in SDN.
25   intelligence delivery.
In conclusion, the introduction of intelligence in SDN brings
26   In this paper, we combine Deep learning (DL) with SDN in or-
der to improve the performance of routing techniques efficiently. many advantages and leads to efficient real-time controller
27  
In this work, we analyse the algorithm denoted ”Predicting of decisions in the routing process and the path selection. In
28   Future load-based routing (PFLR)” and we prove that it is not the routing process, the prediction of time-dependent future
29   scalable in Very Large-Scale Networks(VLSN). Therefore, we
bandwidth is a key that helps to optimize resource allocation
30   suggest an enhancement of this algorithm to achieve scalability
31   by predicting available future bandwidth on path-basis in spite and to avoid bottlenecks. Besides, Artificial Neural Networks
32   of link-basis. Predicted values are obtained using a MultiLayer (ANN) are used to predict time series forecasting. Although
Perceptron (MLP) neural network and applied in the Dijkstra many ANN models are defined in the literature, MultiLayer
33  
algorithm to find the optimal path according to a reciprocal Perceptron (MLP) are the most commonly used neural net-
34   metric. The proposed approach is denoted Scalable Predicting
35   works in time series prediction, according to [2] .
of future load-based routing (SPFLR). Experiments show that
36   the proposed approach outperforms parallel ones by achieving In this paper, an intelligent routing algorithm is proposed. It
37   significant load balancing through the network. uses DL to achieve optimized resource allocation which helps
38   to provide best QoS for incoming flows. Mainly, the path
Index Terms—SDN , DL, load balancing, MLP.
39   selection is based on predicting future available bandwidth
40   through an MLP model. Initially, MLP is used to predict
41   I. I NTRODUCTION the future available bandwidth on each link. The predicted
42   values are used to optimize routing decisions. This is known
43  
44  
45  
I N the previous years, Software Defined Networking (SDN)
drew attention of service providers due to its ability of
offering flexibility, programmability and fine-grained control
as the ”Predicting of future load-based routing” algorithm
(PFLR) [3]. Although PFLR has provided promising results
in terms of rejection ratio, it falls short of scalability. Indeed,
46   over switches. SDN is an evolved logically centralized net- PFLR seems to be unsuitable to be applied in large-scale
47   work which consists of a controller with a global view of networks due to the large number of links to be controlled.
48   the network state and many switches which forward packets Moreover, the deployment of PFLR in traditional IP-networks
49   according to flow table rules. Communication between con- is associated with high complexity since those networks are
50   trollers and switches is achieved through efficient open and naturally distributed.
51   standard protocols such as OpenFlow [1]. To deal with the previous shortcomings, the Scalable Pre-
52   Furthermore, SDN consists of the separation between the dicting of Future load-based routing (SPFLR) algorithm is
53   control plane and the data plane. Thus, this separation makes introduced. SPFLR consists of predicting the available band-
54   the control of flows more flexible, maintenance management width on a selected subset paths instead of on each link. This
55   and network operation simpler and routing process more helps to reduce the complexity of links management and can
56   dynamic. With the tremendous amount of data resulting from ensure efficient QoS-aware packet forwarding.
57   the emerging of new applications, routing process has become The remaining of this paper is organized as follows : Section
60   a paramount issue that must be focused on in order to optimize 2 presents the related work. Section 3 details PFLR algorithm.
61   network resources utilization and perform routing process with FFNN architecture is described in section 4. the evaluation and
62   QoS support. However, traditional routing protocols seem to the experimental results of PFLR are presented in section 5.
63  
64  
65  
Authorized licensed use limited to: UNIVERSITE DE SOUSSE. Downloaded 1
on December 23,2020 at 09:08:52 UTC from IEEE Xplore. Restrictions apply.
2

In section 6, SPFLR approach is described and its validation Where α denotes the weight prediction.
through experimentation is evaluated and compared to PFLR. The prediction of future available bandwidth Bw(li , t+1) is
Finally, section 7 draws a general conclusion. computed each period of X seconds, recognized as the window
size (WS), using an MLP model which has shown a high
II. R ELATED WORK accuracy with nonlinear time-series forecasting.
Many studies have focused on introducing Deep Learning In PFLR algorithm, the prediction is made in decentralized
(DL) in different network architectures and various domain way. This means that in each section of the network, a specific
applications. node is designated to achieve prediction. This node must have
In legacy IP networks, author in [3] proposes the predicting a global view of the section for which it is the responsible.
of future load-based routing (PFLR) algorithm which is based Moreover, it must be so efficient to run DL algorithms.
on DL. In order to reduce network congestion and enhance Clearly, the application of PFLR in classical networks seems
resource utilization, this algorithm combines the predicted load to be complex and its feasibility must be deeply discussed.
with the actual load of each link. PFLR is divided into two Obviously, SDN is well-designed to support this kind of
phases. In the first phase, the future available bandwidth on algorithms. An SDN controller is natively suitable for running
each link is predicted. In the second phase, the link weight DL algorithms so that it is responsible for the prediction
is updated basing on a reciprocol formula which combines process. Its implication helps to enhance performance of QoS-
the predicted future bandwidth with the current residual band- aware routing algorithms. Especially by obtaining the global
width. Thus, the best path is selected using classical Dijkstra status of network, the logically centralized node will be able to
algorithm basing on already computed link weights. In this take intelligent decisions on routing and path selection which
work, the author uses Feed Forward Neural Networks (FFNN) may significantly improve QoS provisioning and delivery.
to predict future available bandwidth.
In SDN networks, DL is used in various domain applica- Algorithm 1: predicting of future load-based routing
tions such as transportation sector, wireless network and IOT. (PFLR)
Indeed, DL and SDN are used in transportation sector and Input :
especially in Vehicular Cyber-Physical Systems (VCPS) to The network topology and all residual link
control end-to-end traffic [4]. Moreover, Convolutional Neural capacities.
Networks (CNN) are proposed basing on time-dependent The path requests between different
network state (available bandwidth, capacity, resource utiliza- ingress-egress pairs.
tion...) in order to achieve an optimal path selection for an Output : Routed paths through the network.
efficient packet delivery. Algorithm:
IN [5], DL is introduced to allocate channels intelligently on 1) Predict the future available BW in all links in the
each link in the network which can avoid congestion and network after a specified WS using ANN.
improve the quality of transmission. 2) Repeat the following steps until the time of WS has
The work in [6] deploys a deep reinforcement learning mecha- elapsed.
nism using an agent and three signals (state, action and reward)
a) Compute the reciprocal of available BW using
to optimize routing.
the following equation:
Finally, [7] defines a set of features which are built to forecast
in advance the results of a re-routing operation, basing on the (1 − α) α
RBW (li , t) = +
quantities returned by OpenFlow. bw(li , t) bw(li , t + 1)

III. P REDICTING OF F UTURE L OAD - BASED ROUTING : where α = (prediction weight).


PFLR b) Update the link weight formula with RBW
c) Compute the best path using the normal routing
The PFLR [alg1] consists of collecting a time-series value of algorithm without changing anything.
the available bandwidth on each link which is used to predict
3) Repeat the previous steps and so on.
the future available bandwidth. A network is defined as a graph
G=(N,L), where N is the set of nodes and L is the set of links.
For each li ∈ L, we denote Bw(li , t) the current available Upon receiving the first packet of a new flow, the ingress
bandwidth in link li at time t. Previous t − n values of the switch checks for a flow table entry that matches the flow at-
available bandwidth in this link, collected each period of m tributes. If no entry is found, the ingress switch communicates
seconds, must be memorized to be used to predict Bw(li , t + a Packet-In message to the controller. As a result, the controller
1). This predicted value is used to update link weights in order selects the suitable path using an appropriate routing process
to find the best path that maximizes the available bandwidth and an up-to-date network state database and replies using a
in the future. In PFLR, a new weight is introduced as the Packet-Out message toward the ingress switch. Generally, the
reciprocal of available bandwidth (RBW). It is computed for routing process is based on two features: i) a cost value on
each link using the equation 1. The best path is the path with each link from the network and ii) an optimization algorithm
the minimum weight. that minimizes the resulting cost on a given path. Dijkstra al-
(1 − α) α gorithm is recognized as the most used algorithm to minimize
RBW = + (1) end-to-end path costs. Furthermore, costs are based on current
Bw(li , t) Bw(li , t + 1)

Authorized licensed use limited to: UNIVERSITE DE SOUSSE. Downloaded 2


on December 23,2020 at 09:08:52 UTC from IEEE Xplore. Restrictions apply.
3

status of links and do not consider future changes in that status. n), · · · , bw(lm , t)), each perceptron n1i from the first hidden
As a result, future network fluctuations which are not expected layer is associated with an output h1i which is computed using
may lead to routing process failure. The introduction of Deep the equation 2:
Learning has overcome this shortcoming. It is possible now Xn
to predict the future link state through time-series forecasting h1i = sigmoid( wji bw(lm , t − j)) (2)
using DL techniques. j=1
To deal with the feasibility of the PFLR algorithm, we
where wji denotes the link weight between the input j and the
suggest an extension to the basic architecture of the RYU
perceptron i.
controller. This extension consists of adding 3 modules which
In the same way, given a set of p outputs H 1 = (h11 , · · · , h1p )
are: data collection module, prediction module and routing
from the first hidden layer, each perceptron from the second
algorithm module (Figure 1).
hidden layer n2i is associated with an output h2i computed
using the equation3:
Xp
h2i = sigmoid( wji h1j ) (3)
j=1

where wji denotes the link weight between the perceptron n1j
and the perceptron n2i . Finally, the output bw(li , t + 1) is
computed as follows (equation 4):
Xk
bw(lm , t + 1) = sigmoid( wj h2j ) (4)
j=1

where wj denotes the link weight between the perceptron n2j


and the output perceptron. This value is then used to compute
the link weight RBW (lm , t) needed for the routing process.

Fig. 1: The structure of Ryu C. Routing Decision Algorithm


controller The suitable path computation is based on the Dijkstra
algorithm using the RBW link weight which depends on both
the residual bandwidth and the predicted available bandwidth.
A. Data Collection Module Given a link li ∈ L, the bandwidth bw(li , t) and bw(li , t + 1)
Periodically, the controller probes all switches in the net- are used to compute the metric RBW (li , t) using the equation
work by sending an OFPPortStatsRequest message to query 1.
statistics on local ports. Hence, switches communicate re- Let bw(li , t + 1)1 and bw(li , t + 1)2 be two predicted values
quested values through an OFPortStatsReply message. There- of the available bandwidth on link li where bw(li , t + 1)1 >
after, the controller computes available bandwidth for each link bw(li , t + 1)2 . The new metric has reciprocal changes re-
in the network. This process is repeated every WS seconds. garding the bandwidth which means that RBW (li , t)1 <
All information are saved with the history link state module. RBW (li , t)2 . Thus, the best path is defined as the path
that achieves RBW (P, t) = min(RBW (li , t)). This path is
li ∈P
B. Future Available Bandwidth Prediction Module associated with the maximum available bandwidth.
This module consists of two processes: a training process
and a validation process. For each link, an MLP model is IV. D EEP L EARNING BASED ON FFNN
defined. It is expressed as a quadruple Mli = (X, H1 , H2 , O), The learning model of MLP is expressed through a
where X is the set of inputs, H1 and H2 are two hidden quadruple (X, H1 , H2 , O) where X is an input layer, H1 and
layers and O is the output layer. The controller runs training H2 are hidden layers and O is the output layer (Fig. 2). The
process for all links in a parallel way. The training process input layer contains 5 neurons. The two hidden layers contain
consists of the application of the previously defined model 12 and 10 neurons respectively. The output layer contains one
on a subset of the collected data. The resulting model after neuron. The set of collected data is divided into forks. Each
training must be validated using another subset of data. The fork contains 5 inputs plus one output. F=48 forks are used
result of the validation process must endorse or disapproveof in the training process while 12 forks are used to validate the
the effectiveness of the proposed learning model. If the model model. The training process is finalized through n epochs. At
is approved, the controller starts using it to predict the future the end of each epoch, the prediction accuracy is computed
available bandwidth on the link. Predicted bandwidth will be through the Mean Squared Error (Equation 5):
used to adjust the link metric defined as RBW, then, the metric n
is used in Dijkstra algorithm to establish suitable paths. 1X
MSE = (yi − ỹi )2 (5)
Given a set of n inputs X(lm , t) = (bw(lm , t − n i=1

Authorized licensed use limited to: UNIVERSITE DE SOUSSE. Downloaded 3


on December 23,2020 at 09:08:52 UTC from IEEE Xplore. Restrictions apply.
4

Number of epoch 100


Activation function sigmoid
Nbre neuron in input layer 5
Nbre neuron in hidden layer1 10
Nbre neuron in hidden layer2 8
Nbre neuron in out layer1 1
Nbre WS 5 seconds

TABLE II: MLP parameters

connected to the Ryu controller through the OpenFlow1.3. H1


is assumed to be the source while H4 is the destination. A
path needs to be established between H1 and H4 to support
bandwidth-sensitive flows. Besides, all link capacities are fixed
at 10Mb/s. In the controller, the MLP model is implemented
over the TensorFlow plateform [10]. It is an open-source
library for machine learning developed by the Google brain
Fig. 2: The structure of MLP team.
In order to evaluate the impact of the PFLR algorithm on
MLP Model X H1 H2 MSE the network performance, we adopt the root-mean-square
1 5 12 - 1.98
2 5 8 8 1.95
deviation (RMSD) metric as follows:
3 5 10 8 1.87 s
1X
4 5 12 8 1.87 σ= (bw(li ) − bw(lj ))2 (6)
n
TABLE I: Comparative Study of various MLP models i6=j

where n is the number of links in the network and bw(li )


is the available bandwidth in the link li .
where: yi is predicted value and ỹi is actual value. The root-mean-square deviation reflects the achievement of the
The choice of our model is not arbitrary. The table (I) depicts load balancing among a network. Defined as the square of the
a comparative study between different models. Clearly, the variance, this metric provides information on the bandwidth
introduction of a second hidden layer helps to reduce the MSE. oscillation among the overall links of the network.
Moreover, increasing the number of neurons in the first hidden
layer has a significant impact on the resulting MSE. However, A. Results
beyond 10 neurons in the first hidden layer seems to not have
further impact on the MSE. Therefore, we fix the number of This experimental test aims to evaluate the impact of intro-
neurons in our model as follows: (X = 5, H1 = 10, H2 = ducing DL in routing process. The simulation test is repeated
8, O = 1). Accordingly, the MLP parameters are defined in 10 times. As shown in Fig. 4, the proposed solution performs
the table II. better than the classical routing algorithm in terms of resource
utilization. Indeed, the root-mean-square deviation of resource
V. PFLR ANALYSIS AND CRITICS utilization reaches over 16 Mbps for non intelligent routing.
It goes around 10 Mbps when prediction is used. Obtained
results (Fig 4) ratify results of parallel researches. Obviously,
DL improves the overall network performance. Indeed, the
resource utilization is optimized over the time.
However, as the number of switches increases, the man-
agement of the prediction process becomes more and more
laborious which probably burdens the controller. Moreover,
computation time becomes harder due to the growth of the
number of links. The scalability issue must be focused in
the PFLR algorithm to make it suitable for application in
large-scale networks such as SD-WAN. In the next section,
a scalable release of the PFLR is presented.

VI. S CALABLE P REDICTING OF F UTURE L OAD - BASED


Fig. 3: Network topology.
ROUTING : SPFLR
In this section, we discuss the performance of the PFLR A. General concept
algorithm following the results of experimental tests. Tests The basic idea of SPFLR algorithm (scalable Predicting
are deployed on Ryu controller [8] and Mininet [9]. The of Future Load-based Routing) is to forecast the available
network topology is shown in Fig. 3 and consists of 5 switches bandwidth of end-to-end paths instead of predicting available

Authorized licensed use limited to: UNIVERSITE DE SOUSSE. Downloaded 4


on December 23,2020 at 09:08:52 UTC from IEEE Xplore. Restrictions apply.
5

Algorithm 2: Scalable Predicting of Future Load-based


Routing (SPFLR)
Input :
The network topology and all residual link
capacities.
The path requests between different
ingress-egress pairs.
Output : Routed paths through the network.
Algorithm:
1) select a set of paths P that connects a source to a
destination
2) For each path pi ∈ P
Fig. 4: Optimization of the
a) compute bandwidth utilization Bw(pi ):
resource utilization (RU) in
PFLR Bw(pi ) = min ( bwij ) ∀ eji ∈ pi
3) select a subset of paths P̂ ⊆ P that has available bw
more than 40%
4) for each path pi ∈ P̂ :
a) predict future bandwidth utilization Bw(pi , t + 1)
of pi
b) Compute the reciprocal of available BW using
the following equation:
(1 − α) α
RBW (pi ) = +
Bw(pi , t) Bw(pi , t + 1)
where α = (prediction weight).
c) reroute flows through the path
p∗ = min RBW (pi )
Fig. 5: Optimization of the pi ∈P̂
resource utilization (RU) in 5) Repeat previous steps each WS
SPFLR

systematically decreased and therefore the controller is less


resources on all links from the network. First, all feasible paths
overloaded.
from a source to a destination are selected. Then, only those
As for example, Given a network G=(V,E) where kV k=5
with at least 40% of available resources are considered in
and kEk=6 denote the set of nodes and the set of links,
remaining steps of the algorithm. The available bandwidth on
respectively, the algorithm of PFLR uses 6 MLP(one for each
each link belonging to a path from the selected set is requested.
link) to predict the future available bandwidth. The complexity
Received data is not saved but used to compute the available
of PFLR is equal to O(kEk + W S(kEk + O(Dij))), where
bandwidth of the corresponding path using the equation 7:
O(Dij) is the complexity of the Dijkstra algorithm and it
Bw(P, t) = min bw(li , t) (7) is equal to (O(kEk+kV k log kV k). However, SPFLR has a
li ∈P
complexity of O(kEk2 log(kEk)) and uses only 4 MLP since
Where P is a path, li is a link belonging to P and bw(li , t) is there are only 4 paths between source s1 and the destination s5
the available bandwidth on link li at t. The set of bw(P, t − i), (1-2-3-5;1-2-4-5;1-4-5;1-4-2-3-5). Furthermore, the paths from
∀0 ≤ i ≤ n are saved than used to predict bw(P, t + 1), as it previous subset with less than 40% of bandwidth availability
is done with PFLR. are excluded from the remaining process. This solution can
make our algorithm scalable in Very large-Scale Networks
B. Results (VLSN) such as SD-WAN.
As shown in Fig 5, SPFLR algorithm performs better than
PFLR algorithm. Indeed, all paths in the network are estab- VII. C ONCLUSION
lished with respect to bandwidth requirements while resources Obviously, SDN is the most marking evolution in network
reservation is done in a fair way. This can contribute to the architectures. It improves efficiently network performance
efficiency of the network resource management. Moreover, the through the introduction of programmability, flexibility and
number of threats used to predict the available bandwidth is fine-grained network control. This efficiency can be ratified
reduced significantly since only a subset of paths between by the combination of intelligence with SDN capabilities.
a source and a destination is considered during the predic- In this way, this paper proposes a scalable intelligent routing
tion process. As a result, the number of monitored links is algorithm which improves QoS-aware path establishment in

Authorized licensed use limited to: UNIVERSITE DE SOUSSE. Downloaded 5


on December 23,2020 at 09:08:52 UTC from IEEE Xplore. Restrictions apply.
6

SDN networks. Mainly, it predicts the future available band-


width in paths scope instead of links scope. Future bandwidth
values in each path are predicted through the application of a
per-path MLP model. SPFLR uses immediate results to reroute
flows through the best path without need to run the Dijkstra
algorithm.
The integration of DL in routing algorithm performs better
than classical routing in terms of the resource utilization as
shown in this paper. The most problem of DL is the complexity
of its algorithms which can be avoided particularly in SDN
using a powerful controller or a distributed architecture of
collaborating controllers. This issue will be pointed out in
future work.

R EFERENCES
[1] Open networking foundation. https://www.opennetworking.org/.
[2] Mehdi KHASHEI, Mehdi et BIJARI. A novel hybridization of artificial
neural networks and arima models for time series forecasting. Applied
Soft Computing, 11(2):2664–2675, 2011.
[3] Andreas TURKY, Abutaleb Abdelmohdi et MITSCHELE-THIEL. Use
of load prediction mechanism for dynamic routing optimization. IEEE
Symposium on Computers and Communications, pages 782–786, 2009.
[4] AUJLA Gagangeet Singh-KUMAR Neeraj et al JINDAL, Anish. Seda-
tive: Sdn-enabled deep learning architecture for network traffic control
in vehicular cyber-physical systems. IEEE Network, 32(6):66–73, 2018.
[5] FADLULLAH Zubair Md-MAO Bomin et al TANG, Fengxiao. An
intelligent traffic load prediction-based adaptive channel assignment
algorithm in sdn-iot: A deep learning approach. IEEE Internet of Things
Journal, 5(6):5141–5154, 2018.
[6] LAN Julong GUO-Zehua et al. YU, Changhe. Drom: Optimizing the
routing in software-defined networks with deep reinforcement learning.
IEEE Access, 6:64533–64539, 2018.
[7] MARCHESE Mario et MONGELLI Maurizio. CELLO, Marco. On the
qos estimation in an openflow network: The packet loss case. IEEE
Communications Letters, 20(3):554–557, 2016.
[8] Ryu sdn framework. https://osrg.github.io/ryu/.
[9] Mininet. http://mininet.org/.
[10] Tensorflow. https://www.tensorflow.org/.

Authorized licensed use limited to: UNIVERSITE DE SOUSSE. Downloaded 6


View publication stats on December 23,2020 at 09:08:52 UTC from IEEE Xplore. Restrictions apply.

You might also like