HANOI UNIVERSITY OF SCIENCE AND TECHNOLOGY

SCHOOL OF ELECTRONICS AND TELECOMUNNICATIONS

----------

REPORT TOPIC

ARTIFICIAL INTELLIGENCE ENABLED

SOFTWARE DEFINED: A COMPREHENSIVE OVERVIEW

Students: Vũ Hoàng Anh MSSV 20172402

Hanoi - 5/2020

Abstract

1
 --Software defined networking (SDN) represents a promising networking
architecture that combines central management and network programmabllity.
SDN separates the control plane from the data plane and moves the network
management to a central point, called the controller, that can be programmed and
used as the brain of the network. Recently, the research community has showed an
increased tendency to benefit from the recent advancements in the artifical
intelligence ( AI) field to provide learning abilities and better decision making in
SDN. In this study, we provide a detailed overview of the recent efforts to include AI
in SDN. Our study showed that the research efforts focused on the field of AI
namely: machine learning. According, in this work we investigate their potential
use, as well as the improvements achieved by including AI-base techniques in the
SDN paradigm.

2
Table of Contents
1. Introduction.........................................................................................................4
2. Software Defined Networking............................................................................4
3. OpenFlow protocol..............................................................................................5
4. Conclusion...........................................................................................................5
5. References...........................................................................................................5

3
1. Introduction
In the last years, the tranditional internet traffic has been changed to be more
complex, especially in the era of big data, today’s datacenters require more
flexibility and scalability. Beside of the raising of advanced network applications
requires more scalable architecture which should be able to provide reliable and
sufficient services base on a specific traffic type. Furthermore because the
existence of various types of devices even in the same area, thousands of end point
devices could share and exchange different patterns of network traffic. Also the
traditional network deal with distributed management and processing of network
decision making. Therefore, we are dealing with a huge number of network nodes
and that could be very expensive in some cases like Virtual Machine (VM)
migration. Adding Quanility of Service (QoS), Quality of Experience (QoE) and
security policies to every end point in complex network architectures shows
another disadvantage in the legacy network approach . This is present an overview
on these information and modeling techniques in order to illustrate the advances
and achievements in this field. Hence, each of the next coming chapters will focus
on one of the techniques. Section 2 will focus on definition SDN. Section 3
disscusses mainly about the OpenFlow protocol. Then discussions on artifical
intelligence in SDN in section 4.

2. Software Defined Networking

Software promotes innovation by introducing the concept of centralized
programmable control of the data plane, which facilitates the development of new
network services and protocols. The SDN architecture is designed based on the
idea of the separation between control and data planes.
The applications of SDN reside in the application plane of SDN architecture
where the northbound application programming interface (API) provides the
commutation between the application and control planes [1], which allows
implementing a set of network services such as traffic engineering, intrusion
detection, quality of service, firewall and monitoring applications[2]. Northbound
API allows developers to write their own application without the need for a
detailed knowledge of the controller functions or understanding how the data plane
works. It is worth mentioning that several SDN controller provide their own
northbound APIs[3].

4
 The communication between control and data planes is provided using a
southbound API such as forwarding and control element separation, open vSwitch
database, protocol oblivious forwarding (POF) [4], Openstate [5], OpenFlow (OF)
[7] and OpFlex [5], which enables exchanging control messages with forwarding
elements. E.g OF-enabled switch adopts a flow-based SDN controller, which is
responsible for preparing the forwarding tables of each switch [1]. A typical OF
-enabled switch has a pipeline of flow tables, which consis of flow entries, each of
which has three parts: (1) matching rules, which are used for matching incoming
packets(2) counters that maintain statistics of matched flows and (3) actions or
instructions, which can be configured proactively or reactively to be executed upon
a match. The forwarding elements(i.e. OF-enabled switches) can be implemented
in either software or hardware. Some software switches such as OpenvSwitch have
a great potential for providing a solution for data centres and virtual networks [4].
On the other hand, other APIs are proposed for a specific purpose (e.g, VOIP
applications and inter-domain routing.

3. OpenFlow protocol
OpenFlow (OF) is considered the most commonly used southbound API in SDN,
which is being continuosly developed and standardized by open networking
foundation (ONF) [7]. OF provides an abstracstion layer that enables the SDN
controller to securely communicate with OF- enabled forwarding elements [7].
OpenFlow has become the de-facto standard for southbound APIs used in SDNs
[7] and therefore, in this study, we mainly focus on OF-based SDNs. OF-based
forwarding devices have been developed to coexist together with conventional
Ethener devices [4]. Hybrid switches, on the other hand, reveal new possibilities by
including both OF and non-OF ports [7]. As we mentioned previously, a set of
control messages can be sent by the controller to prepare and update a paticular
switch’s flow tables. A typical OF-enabled switch handles new coming packets
based on it’s flow table. A table-miss occurs when a new packet does not match
any of the flow table entries. In this case, the switch may either drop the packet or
forward it to the corresponding controller using OF protocol [7].
Recently, different versions of OF protocol have been introduced to add
more flexible and reliable capabilities by including multiple flow tables, enhanced
matching/action abilities, optical ports, group tables, meter tables and synchronized
tables. More details concerning various verisons of OF specification can be found
in [30]. In addition, there are many availables OF controllers, such POX, Beacon,
OpenDayLight, Floodlight..

5
4. Artificial Intelligence in SDN
Recently, the soft computing and artificial intelligence methods have started to
play a significant role in most of modern systems such as intelligent transportation.
That gives us the chance to improve the performance of the current computer
networks. The integration between the abstraction concept in SDN paradigm and
AI techniques can lead to more adaptive behavior of network elements. Also it will
introduce new mechanism for dealing with both traditional network issues and
SDN related new ones. In this section the recent efforts in this context will be
discussed.
4.1 Load Balance and Flow Routing
Load balance function is a requirement for minimizing the latency and maximizing
the throughput in computer networks that support multiple routing approaches.
Load balancing also considered as a defense technique against some types of
networks attacks such as DdoS attack .
The abstraction in SDN approach provides an important advantage which is the
global view and discovering of topology of the network. In a Back Propagation
Neural Network used for achieving real time dynamic load balance and latency has
been decreased by 19.3% compared to DLB and static Round Robin methods. The
input vector for the neural network contains path information which are : 1.
Bandwidth utilization ratio 2. Packet loss rate 3. Trasmission latency and 4.
Transmission hops[8]. Authors in also proposed a BPNN based approach for load
balancing in data centers. The BPNN applied internally inside the Open vSwitch in
a way that reduce the time consumed for sending routing decision from the
controller to the Open vSwitch. The input vector consists of :1. Available
bandwidth and 2. Packet loss. And from, which proposes a genetic algorthm in
SDN based clident sever architecture. The fitness function defined by the
Formula(1):

6
Where K represents the severs and each one has X set of workload. The
performance in has been compared with random and round robin methods and
show better performance. In also a genetic algorithm for flow routing optimization
in SDN based audio over IP network has been introduced. The network described
as a connected graph. The problem is to show that the graph meet the demand
which is bandwidth and latency requirements of the source and destination. The
fitness function given by the formula shown in Equation (2).

Due to time consumption issues the authors did not implement the crossover
operation. The population size and non allocation probability. Also because the
gentic algorithm has been implemented in python, the time efficiency was 10x time
less than a mixed-integer linear programming algorithm implemented in C++. The
advantage to use the genetic algorithm approach is to get a paritcal solution of the
problem through the solving stage while this is not possible in linear program; this
partial solution helps in evaluating other algorithms. In another context, an Ant
Colony Optimization (ACO) approach for QoE-aware flow routing is presented
in[9]. ACO a swarm intelligence method that uses metaheuristic optimization. In
computer networks the Quality of Experience (QoE) indicates requirements for the
customer’s perspective. In the SDN applications deliver user session parameters to
the controller which in turn runs the ACO algorithm on a weighted graph, where
the weights between vertices are delay and loss rate for each network device. The
fitness function depends on the flow type and estimated value of corresponds QoE

7
model (e.g. audio, video or data). ACO has achieved 24.1% increasing for the
maximal QoE value obtained by the shortes path routing approach.
4.2 Intelligent network application
The integration between SDN and AI field opens the door for building more
intelligent network applications. As authors in51 proposed a reinforcement
learning approach for adaptive video streaming in SDN paradigm. The controller
represents a periodically decision maker that determines the time of selecting a
new path and when the server needs to change the quality of the video. Markov
decision process used for modelling the actions of the decision making. The Q-
learning technique used in the case of unknown rewards for moving between the
current and next state. The percentage of packet losses and the number of quality
changes represent the most significant parameters to define the reward. The Q-
values are updated by the function shown in Equation (3) and stored in Q-table
where γ and η represent the discount factor and learning factor respectively.

And the softmax function shown in Equation (4) represents the probability of
selecting an action in state s time t

Where T represents a random move already used in simulated annealing

method to escape from the local optima problem. The controller can change the
current path and/or adaptively extract/add the selected layers based on the available

8
bandwidth to increase the QoE of the video streaming service. The mentioned
approach compared with the shortest path routing and greedybased approaches has
shown a decreasing of the frame loss rate by 89% and 70% respectively.

5. Conclusion
In this paper, we provide an overview of the integration between SDN paradigm
and AI techniques. We described the basic SDN architecture and the significant
role of OpenFlow protocol in it. Then we summarized the recent research
contributions to provide more intelligent network behavior in the SDN approach.
Neural networks have been applied in different scopes such as load balancing. Also
the results for the adaptive video streaming system have showed decreasing of the
frame loss rate. As a result we see that the research in this area is growing rapidly
and hybrid intelligent approaches can also bring more improvements to the field of
SDN-based networks.

9
6. Conferences

[1] Jain, R., Paul, S.: ‘Network virtualization and software defined networking for
cloud computing: a survey', IEEE Communication Magazine, 2013, 51,(11),
pp. 24-31.
[2] “Akyildiz, I.F., Lee, A., Wang, P., Luo, M., Chou, W.: ‘A roadmap for traffic
engineering in SDN-OpenFlow networks’, Computer Networks, 2014, 71, pp.
1-30”.
[3] “Bianchi, G., Bonola, M., Capone, A., Cascone, C.: ‘OpenState: programming
platform-independent stateful openflow applications inside the switch’, ACM
SIG COMM Computer Communication Review, 2014,44,(2), pp. 44-51.”.
[4] “Kreutz, D., Ramos, F.M., Verissimo, P.E., Rothenberg, C.E., Azodolmolky,
S., Uhlig, S.: ‘Software-defined networking: A comprehensive survey’,
Proceedings of the IEEE, 2015,103,(1),pp. 14-76”.
[5] “McKeown, N., Anderson, T., Balakrishnan, H., et al.: ’OpenFlow: enabling
innovation in campus networks’, ACM SIGCOMM Computer Communication
Review, 2008,38,(2), pp. 69-74.”.
[6] “Nunes, B.A.A., Mendonca, M., Nguyen, X.N., Obraczka, K., Turletti, T.: ‘A
survey of software-defined networking: Past, present, and future of
programmable networks’, IEEE Communications Surveys & Tutorials, 2014,
16,(3), pp. 1617-1634.”.
[7] “Song, H.: ‘Protocol-oblivious forwarding: Unleash the power of SDN through
a future-proof forwarding plane' In: Proc. 2nd ACM SIGCOMM Workshop
Hot Topics Softw. Defined Netw, Hong Kong, China, August 2013, pp. 127-
132”.
[8] “Chen-Xiao C, Ya-Bin X. Research on load balance method in SDN.
International Journal of Grid and Distributed Computing. 2016; 9(1):25–36.”.
[9] “. Dobrijevic O, Santl M, Matijasevic M. Ant colony optimization for QoE-
centric flow routing in software-defined networks. 2015 11th International

10
Conference on Network and Service Management (CNSM); Barcelona. 2015
Nov.p. 274–8.”.

11