Load Balancing model based on Machine Learning

and Segment Routing in SDN

Dimitar Todorov Hristo Valchanov Veneta Aleksieva
Computer Science and Engineering Computer Science and Engineering Computer Science and Engineering
Department Department Department
Technical University of Varna Technical University of Varna Technical University of Varna
Varna, Bulgaria Varna, Bulgaria Varna, Bulgaria
[email protected] [email protected] [email protected]

Abstract—Because of increased number of network devices  Northbound application programming interfaces

in the world, as well as the need for the network environment to (APIs) – software interfaces between the software
provide dynamics of faults and adaptability to load changes, it modules and the SDN applications running on top of
is difficult and complicated network to be managed. By applying the network platform. They give universal network
an approach called Software Defined Networks (SDN), the abstraction models and functionality which can be
separation of the control and data planes is achieved. This allows used by network applications.
the creation and deployment of new network applications to be
easier, as well as provides simplification and flexibility in  East-west interfaces – allows communication between
network policy enforcement, facilitating network configuration multiple controllers if there is multi-controller
and management. The main problems in SDN are load architecture.
balancing and segment routing. This paper proposes a model
which aims to reduce not only the overall load on the network,  Data plane – represents the hardware for forwarding
but also to reduce the bandwidth and improve the routing in SDN architecture. Because the controller must
mechanism on the SDN networks. It combines segment routing communicate with the network infrastructure, it
algorithm and load balancing mechanisms based on machine requires to implement special protocols to control and
learning. manage different network equipment. Currently the
most popular and used protocol is OpenFlow [4].
Keywords—software defined networks, load balancing,
segment routing, machine learning

I. INTRODUCTION
With the large increase of the network devices around the
world, it becomes increasingly difficult and complicated,
networks to be managed. Each of the network devices require
to be configured separately with low-level and vendor-
specific. Also, the network environments must endure the
dynamics of faults and to be adaptive to load changes. Current
IP networks does not support these automatic reconfigurations
and response mechanisms. For these reasons, it is necessary to
apply another approach named Software Defined Networks
(SDN) [1].
Separation of the control and data planes is the main
feature of the SDN. It gives direct advantages, because
network programmability is achieved through the
centralization of the control plane in concurrence with the
availability of open APIs. With these capabilities, the creation
and deployment of new network applications are easier. SDN
also provides simplification and flexibility in network policy
enforcement, facilitating network configuration and
management [2].
There are four components [3] which are represented in
SDN architecture Fig.1:
 Control plane – represents an abstract view of the
complete network infrastructure. It enables the Fig. 1. SDN Architecture
administrator to apply custom policies or protocols
across the network hardware. With the help of SDN it becomes easy to control and scale
enterprise networks programmatically. It also provides the
ability to not affect the performance, reliability, or the user
The research, the results of which are presented in the present paper, experience. Simplifying the network management and
have been carried out under the TU-Varna Scientific Project “Research of bringing back virtualization to the network eliminates the
the possibilities of integrating machine learning and Blockchain technologies complexity of the infrastructure layer and adds visibility for
for the Internet of Things”, which is financed by the state budget. the applications and services.

978-1-7281-9308-3/20/$31.00 ©2020 IEEE

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO CEARA. Downloaded on December 19,2024 at 12:24:05 UTC from IEEE Xplore. Restrictions apply.
 Dividing the workload onto multiple resources in order to and collecting information for the link usage, then it calculates
avoid overload on any of the resources is a technique called the less loaded path. This algorithm applies simple evaluation
Load balancing. criteria for collecting network path information, which is its
main disadvantage.
One of the main goals of load balancing is to minimize the
response time, optimize the traffic and minimize throughput Another algorithm of this type is QoS-aware Adaptive
[5, 6]. Conventional networks do not have view of the global Routing (QAR) for providing Quality of Service routing
network compared to SDN networks. Thus, load balancing mechanism. It uses Markov Decision Process (MDP)
methods in SDN are more accurate and have higher combined with reward function for modelling the system [11,
performance. Because of industry concerns, the load 12]. This approach contains of multi-layer hierarchical control
balancing issue is one of the most important one plane architecture. The bottleneck of this mechanism is failure
of the super controller which will affect the entire network.
Nowadays machine learning (ML) algorithms are more
and more widespread in various information spheres. Their These algorithms have one common disadvantage – high
application allows solution of more complex problems to be network communication between the controller and network
dynamically, as well as to improve without human devices, which reflects to the network bandwidth. PASR is
intervention. A lot of research is being done on including ML segment routing mechanism which encapsulates routing
algorithms in the field of network architectures. Some of them information in packet header [13]. It minimizes the traffic
are successfully used in DDoS attach prevention mechanisms. exchanged for flow table packets and increases the bandwidth.
Unfortunately, it does not have load balancing mechanism and
This paper analyzes different static and dynamic load the traffic is calculated based on Shortest Path First or ECMP.
balancing mechanisms in SDN using different techniques and
improvements for Quality of Service. Finally, it proposes a III. ARCHITECTURE MODEL
network architectural model for dynamic network load
balancing which takes advantages of machine learning and This paper proposes a model which aims to reduce not
segment routing. only the overall load on the network, but also to reduce the
bandwidth and improve the routing mechanism on the SDN
II. RELATED WORKS networks. It combines segment routing algorithm and load
balancing mechanisms based on machine learning. The main
Not so many researches have been proposed for network purpose of this model is to investigate the most compatible
load balance in SDN architecture. Most of proposed solutions machine learning model for network load balancing and to
are using the controller to analyze the information from minimize the network traffic between the controller and
OpenFlow switches and to modify the flow-tables by specific network devices.
load balance strategies. These strategies are classified as static
load balance methods. They cannot make dynamic routing The model is developed as a cross platform SDN
decisions according to real-time network load conditions. controller using C++ programming language. It implements
the OpenFlow protocol and follows OS specific system calls
Another approach is to use dynamic load balance for higher performance.
algorithms, which can make decisions based on network load.
One of this kind algorithms is Dynamic Load Balance A. System Modules
Algorithm (DLB) [7, 10]. It uses greedy selection strategy The system contains four main modules: SDN controller
which transmits least data load to pick the next-hop link. This module, Prediction module, Path compute module and Path
approach has clear disadvantage, because the routing strategy encoding module (Fig. 2).
does not use the global network view and calculates the link
load on every next hop. This algorithm also does not minimize
the traffic from the controller to the network devices when
updating the flow tables. Processed status

Prediction module
The Fuzzy comprehensive evaluation [8, 10] is another
algorithm which collects hops of each path in every switch
port together with the number of transmission packets, key Network
status
switch number and the transmission rate. Then the algorithm Path Compute module

combines these parameters to compute using fuzzy

comprehensive evaluation the load condition on each path.
The disadvantage of this algorithm is that it cannot accurately SDN Controller module
Flow path

predict the real-time load condition. Control decision

All these algorithms provide load balancing strategy based Information Flow table Path Encoding module
on collected data for the status of the network, but they cannot collection installation
predict if the load will be increased or decreased. Another
approach is to use Machine Learning algorithms, which can
compute the change of the traffic based on different network
parameters and to send flow table updates before the traffic
change occurs.
One of suggested approaches is to use Ant Colony Data plane

Optimization based algorithm which makes classification of

the best path and best servers merged with the global network
Fig. 2. Proposed model modules
view [9, 10]. It is heuristic method which works by monitoring

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO CEARA. Downloaded on December 19,2024 at 12:24:05 UTC from IEEE Xplore. Restrictions apply.
 The SDN controller module is simple SDN controller, C. Simple System Workflow
which have a Global Network View database in which it stores Fig. 3 represents simple workflow chart of proposed
the network architecture and the connections between network model.
devices. Also, it keeps information for the network latency,
network speed and bandwidth of the devices. The switches The connection between the switch and the controller is
connect to the controller via exchange of handshake messages established via exchanging a handshake packets. These
and starts two-way communication to track flow table packets contain the maximum OpenFlow version supported
information, packet information or other kind of statistics. by the controller or the network device. If the versions are
different, they will pick the maximum version, which is
The Prediction module is used to compute the data supported by both devices.
collected from the controller and to make a decision for
potential flow overload. It integrates four different machine Once the connection is established, the controller sends a
learning algorithms which helps to forecast path overload. status check packet to get the network information of the
Implemented machine learning algorithms are: Q-learning device.
[14], Genetic algorithm [15], Particle swarm optimization
algorithm [16] and Hidden Markov model [17, 18]. They work
independently one from another and can be specified at Start
controller start time.
Listening
The Path compute module is used to compute the optimal for end
End
flow path based on network resources and Quality of Service process
Waiting for
new requests
signal
requirements. The computation is achieved using the
No
parameters such as bandwidth, network speed connection and
packets sent per second computed by the Prediction module. Accept Success
connection
The Path encoding module receives the flow path
Yes
information calculated by the Path compute module and
combines it with Path Segment Database to produce the No Handshake Parsing the
optimal path aggregation scheme for needed path. Then the request No
optimal path segment scheme is encoded in the packet header,
which will be installed and shared between network devices.
Sucssess Compute
Yes Is Echo
B. Prediction process bandwidth

Collected network parameters from the system are used to Yes

No
compute the optimal path based on machine learning
algorithms. The parameters are reduced to a single coefficient, Status check
request
which is then used for training and prediction purpose. Is Packet In

Using a Q-Learning algorithm, the process is split in two Yes

Success
flows. First there are no data on which to make the prediction.
To fill that data, the module starts to learn and receives a Yes
Predict
flow
reward for every successful conjunction. Once the model is increase
trained, the module can predict any flow change. Store status
check response
Compute and
For Genetics algorithm, first the system generates all Encode packet
possible paths which are called Population. The computed flow table
Global
coefficient represents the Genes and they are joined to form a Network
flow path solution (Chromosome). Once we have all the View
Install flow
needed data, a fitness function is called to find the individual’s table
ability to compute with each other. At last we make a selection
of fittest population, which represents the chosen path. Fig. 3. Simple workflow chart

The Particle swarm optimization algorithm purpose is to When receiving the information, the controller stores it in
determine the variable which minimizes in a vector. The a Network Global View database. After that, the controller and
model contains swarm with P particle, which represents the the switch start to exchange echo packets to track the
network devices. Each particle contains a vector and the connectivity between them. These packets are used to track
values stored in the vector are coefficients computed by the the bandwidth of the device.
controller. After the computation process finishes, the result
contains the minimum loaded path. When a host sends a packet to another destination, the first
switch, which receives the packet, will check its flow table for
Hidden Markov model represents a Markov process. It has further forwarding. If the flow table of the switch contains the
two states – loaded and free. The states represent if the destination address, it will pass the packet to the next network
network path is not loaded or it is overloaded. The used device. If the flow table does not contain such one, it will send
observations presented from the Hidden Markov model are – packet-in message to the controller to get information what to
start and stop. Start stands for low used path and stop do with it. Then the controller checks the Global Network
represents overloaded path. The probability is computed based Database and sends back flow table installation packet to the
on the passed from the controller coefficient. switch.

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO CEARA. Downloaded on December 19,2024 at 12:24:05 UTC from IEEE Xplore. Restrictions apply.
 The prediction process monitors possible changes of the [13] Z. Li and Y. Hu, “PASR: An Efficient Flow Forwarding Scheme based
network load. If it detects such ones, it sends a signal to the on Segment Routing in Software-Defined Networking”, DOI
10.1109/ACCESS.2020.2964800, IEEE Access
Path Compute Module, to update the Flow Tables with needed
[14] S. Nanda, F. Zafari, C. DeCusatis, E. Wedaa and B. Yang, “Predicting
routes to make the network load balanced. Once the optimal network attack patterns in SDN using machine learning approach”, In
paths are computed, they are sent back and installed to the Proc of IEEE Conference on Network Function Virtualization and
switches. The controller also notifies the Prediction process of Software Defined Networks (NFV-SDN), Nov. 2016, pp. 167-172.
dropped network devices, which will automatically trigger the [15] L. Chou, Y. Yang, M. Hong, K. Hu and B. Jean, “A genetic-based load
Flow table changes. balancing algorithm in openflow network”, Advanced Technologies,
Embedded and Multimedia for Human-centric Computinc, 2014, pp.
IV. CONCLUSION 411-417.
[16] M. Awad, M. El-Shafei, T. Dimitriou, Y. Rafique, M. Baidas and A.
In this paper is considered one of the main problems in Alhusaini, “Power-efficient routing for SDN with discrete link rates
SDN networks, namely network load balancing. There are and size-limited flow tables: A tree-based particle swarm optimization
different load balancing mechanisms in SDN, which uses two approach”, International Journal of Network Management, 2017, pp.
main approaches – static and dynamic load balancing. The e1972.
shortcomings and problems with approaches and algorithms [17] C. Zhang, X. Wang, F. Li, Q. He and M. Huang, “Deep learning-based
in related works are considered. network application classification for SDN”, Transactions on
Emerging Telecommunications Technologies, 2018, pp. 1-18.
At last the paper proposes an architecture model, which [18] Z. Fan, Y. Xiao, A. Nayak and C. Tan, “An improved network security
combines machine learning algorithms with segment routing situation assessment approach in software defined networks”, Peer-to-
Peer Networking and Applications, 2019, pp. 295-309.
for achieving better performance and network load balancing.
The main advantage of the model is that it improves the QoS
and provides ability to predict overload of network routes.
The aim of future work is to study the reduction of
messages in flows and to make a comparison with similar
proposed approaches. Another goal is to study the benefits of
the proposed model on different operating systems.
REFERENCES
[1] B. Nunes, "A survey of software-defined networking: Past, present, and
future of programmable networks" Communications Surveys &
Tutorials, IEEE 16.3, 2014, pp. 1617-1634.
[2] S. Michel, K. Bonfim, L. Dias, and S. Fernandes, “Integrated
NFV/SDN Architectures: A Systematic Literature Review”, ACM
Comput. Survey, 2019, DOI: https://doi.org/10.1145/3172866.
[3] M. Jammal, T. Singh, A. Shami, R. Asal, and Y. Li, “Software defined
networking: State of the art and research challenges”, Computer
Networks, 2014, DOI: 10.1016/j.comnet.2014.07.004.
[4] Open Networking Foundation, “OpenFlow Switch Specification”,
March 2015, https://www.opennetworking.org/wp-
content/uploads/2014/10/openflow-switch-v1.5.1.pdf
[5] W. Lin and L. Zhang, “The load balancing research of SDN based on
ant colony algorithm with job classification”, in Proc. 2ndWorkshop
Adv. Res. Technol. Ind. Appl., 2016, pp. 472-476.
[6] S. K. Askar, “Adaptive load balancing scheme for data center networks
using software defined network,” Sci. J. Univ. Zakho, vol. 4, no. 2,
2016, pp. 275-286.
[7] Y. Pan, “OpenFlow based load balancing for Fat Tree networks with
multipath support”, in Proc. 12th IEEE International Conference on
Communications (ICC’13), Budapest, Hungary, 2013, pp. 1-5.
[8] J. Chang, Y. Ren, Z. Zhang, and G. Wang, “An Effective Path Load
Balancing Mechanism Based on SDN”, Trust, Security and Priva cy in
Computing and Communications (TrustCom), 13th International
Conference on IEEE, 2014, pp. 527-533.
[9] A. Koushik and S. Selvi, “Load balancing Using Software Defined
Networking in cloud environment”, Recent Trends in Information
Technology (ICRTIT), 2014 International Conference on. IEEE, 2014,
pp. 1-8.
[10] Cui Chen-xiao and Xu Ya-bin, “Research on Load Balance Method in
SDN”, International Journal of Grid and Distributed Computing, vol.
9, No. 1, 2016, pp.25-36.
[11] M. Latah and L. Toker, “Artificial Intelligence Enabled Software
Defined Networking: A Comprehensive Overview”, IET Networks
8(2), 2018, DOI: 10.1049/iet-net.2018.5082.
[12] S. Lin, I. Akyildiz, P. Wang, and M. Luo, “QoS-aware adaptive routing
in multilayer hierarchical software defined networks: a reinforcement
learning approach”, in Proc. of IEEE International Conference on
Services Computing (SCC), San Francisco, USA, June 2016, pp. 25-
33.

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO CEARA. Downloaded on December 19,2024 at 12:24:05 UTC from IEEE Xplore. Restrictions apply.