This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2959351, IEEE Access

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2017.DOI

Multiclass Queueing Network Modeling

and Traffic Flow Analysis for SDN
Enabled Mobile Core Networks with
Network Slicing
SANTHOSHA KAMATH1 , SANJAY SINGH1 .,(Senior Member, IEEE), and M.SATHISH KUMAR2
1
Department of Information and Communication Technology, Manipal Institute of Technology, MAHE, Manipal-576104, India
(e-mail:[email protected])
2
Department of Electronics and Communication Engineering, Manipal Institute of Technology, MAHE, Manipal-576104, India
(e-mail: [email protected])
Corresponding author: Sanjay Singh (e-mail: [email protected]).

ABSTRACT The back-haul networks of 5G are formed by heterogeneous links which need to handle
massive traffic. The service providers are not able to provide good QoS for their users. The technology like
Software Defined Networks(SDN) and Network Slicing helps a little for a service provider to providing
QoS for multiple links. The challenge with service providers in the efficient utilization of resources to
fulfill the QoS requirement of users to comply with the growth and thereby increasing the revenue. These
problems require an accurate traffic model to determine the steady-state of the system. The proposed
model uses an architecture that has the combination of two technologies: SDN and network slicing,
which empowers an administrator a flexible, programmable network, and the best management of network
resources. Heterogeneous application is well managed by creating multiple logical networks called slicing.
The slicing can be modeled using multi-class queuing networks. These technologies encourage service
providers to fulfill QoS and revenue growth. To leverage the benefits of these technologies in allocating
QoS is to identify the performance of the system, which requires a precise model of traffic to decide the
steady-state condition of the framework. In this paper, we focus on SDN and slicing in mobile networks
and quantify the performance measure considering an in-band OpenFlow architecture for a single node and
homogeneous traffic class, which is further extended to the multi-class heterogeneous class queuing model
and analyzed. The results obtained help a service provider to monitor the utilization of resources in every
node by every class of core network, which in turn helps to allocate the resources precisely to fulfill QoS
requirements.

INDEX TERMS Mobile networks, network modeling, network slicing, software defined network, traffic
flow analysis

I. INTRODUCTION situation demands that the service provider knows the user’s
Current mobile backhaul networks are built with a large QoS requirements and act accordingly. That enforces service
number of middle-boxes and forwarding devices. Each de- provider to enable Over The Top (OTT) [1] services on their
vice is embedded with some intelligence in it to perform network. Providing OTT on the mobile network requires to
some functions. However, despite up-gradation and limited solve many challenges. These OTT applications will have
resources of mobile network technology, mobile service different transmission characteristics. OTT on mobile net-
providers fall short of promising the customer requirement works follows the characteristics of heterogeneous Ultra-
with expanding data demand and traffic pattern. Customers Dense Network (UDN) of 5G backhaul networks. Heteroge-
are using data services more, which are heterogeneous and neous application is well managed by creating multiple log-
rarely use the services provided by the service providers like ical networks on the top of common physical infrastructure.
voice, and SMS. Making an informed decision in such a Therefore network slicing along with SDN will have a great

VOLUME 4, 2016 1

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2959351, IEEE Access

S Kamath et al.:Multiclass Queueing Network Modeling and Traffic Flow Analysis for SDN Enabled Mobile Core Networks with Network Slicing

impact on managing the network. mobile networks. The applications like mobile networks,
The transmission characteristic for a given UDN varies In-band architecture suits since it is deployed over a large
from user to user in mobile networks. In other words, policies geographical area.
from user to user or application to application changes. It
is worth noting that mobile traffic depends heavily on the
day-time and less in the night and more in urban than in
rural. All these make mobile traffic a non-ergodic pattern.
The main impact of this is significant and multiple. They
result in the migration of subscribers and hence, revenue loss
for the service providers. Thus it is necessary to limit and
manage the network traffic resources wisely and dynamically.
The analytical model that we propose can be viewed at
different levels: it can be used at the validation level to detect
congestion in the mobile network traffic and at monitoring
level, where the model analyzes the traffic flow and detect fu-
ture congestion. The proposed system needs to be analyzed, FIGURE 1: In-band SDN controller
which gives answers to the questions like
i) which application is used by the user most?
ii) which part of the system has more traffic? and
iii) which application is consuming what resources the
most?
Answers for these questions help the service providers to
deploy the policies and allocate the resources.
The above requirement demands an architecture for mobile
networks that is flexible to deploy policies and utilize the
limited resources to gain maximum benefits. One of the
architectures proposed in line with this is as shown in Fig. 3.
The proposed architecture uses two technologies: Software
Defined Networks (SDN), and Network Slicing. In the next
paragraph, we briefly discuss the SDN and network slicing.
SDN is a type of network system architecture that iso-
lates the network into the data plane and control plane. The
network devices that forward traffic from the control plane
and the software logic that decides how the traffic should
be processed through the network. The separation enables
a system administrator to control network behavior, from
a single high-level control program. OpenFlow [2] is used
to communicate between the data and the control plane.
Deployments of SDN helps to manage the variety of network
problem in real networks.
Unlike traditional networks, SDN based network man-
ages all the resources such as routing, firewall, encryption, FIGURE 2: Out-band SDN controller
gateways, Network Address Translation (NAT), Deep Packet
Inspection (DPI) by the controller. OpenFlow based switches Network Slicing is an emerging trend, which defines the
transmit the request to the controller, and the controller production network into logical slices [3]. It enables different
receives the request and applies the action on those packets. virtual systems to be made over a standard shared physical
The communication between the control and data plane framework. The virtual systems are then tweaked to meet the
in OpenFlow has two architectures: In-band and Out-band particular needs of applications, services, and devices to the
architecture, as shown in Fig. 1 and Fig. 2. In Out-band, there users or administrators.
is a separate and dedicated connection between every SDN Users might be able to control their network traffic in the
enabled device and the controller. This type of architecture is given slice. Slicing enforces strong isolation between each
followed when the distance between controller and devices slice, so that action in one slice does not affect another slice.
is small such as data-centers. Whereas in In-band, both data Therefore each slice controls its packet forwarding. There are
plane and southbound traffic are sent through the same chan- various reasons one might want to slice the network. One is
nel. This type of architecture can be used when the distance that the network might have multiple administrative groups,
between the controller and SDN devices is large such as such as different applications. The development of versatile
2 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2959351, IEEE Access
S Kamath et al.:Multiclass Queueing Network Modeling and Traffic Flow Analysis for SDN Enabled Mobile Core Networks with Network Slicing

correspondences frameworks towards the Fifth Generation the performance of the switch factors like flow table size,
(5G) must meet the testing essentials of Mobile Broadband processing rate, and packet arrival rate. Bosakov and Rizk
(MBB) [4] use cases too and furthermore new various use [7] discussed the SDN switch performance using a network
cases related with clients of new market sections and vertical calculus-based approach without considering the feedback.
enterprises (e.g., e-wellbeing [5], IoT, MBB and so forth) as Naous [8] discussed the switch performance using the de-
shown in Fig. 3. terministic model. Naous [8] determined the performance of
These applications have altogether different transmission the SDN switch under the FPGA platform. Lazaris et al.,
qualities. For instance, UDN will commonly have a pervasive [9] showed the effect of processing time on a high packet
number of gadgets. However, every gadget may have low arrival rate. They demonstrated that the ability of the SDN
throughput. MBB has about the contrary properties since it switch negatively extends the service time of the switch,
will have a lot more modest number of gadgets, yet everyone which likewise influences the execution of the control plane.
will transmit or getting extremely high transfer bandwidth The researchers concluded that the location of rules has a
content. The objective of network slicing is to segment significant influence on a flow response time of the switch
the physical system at a start to finish level to permit the [10]. It is because of both the equipment and programming
ideal gathering of traffic, separation from different tenants. based switches having their processing abilities.
Therefore a slice type could be defined from a functional or Apart from these few works like [11], [12] is carried out by
behavioral perspective. It is therefore anticipated that mobile simulation or using some measurement tool. Mahmood et al.
network operator could deploy a single slice that satisfies [13] proposed a general M/M/1 based single node modeling
the need of multiple vertical, as well as multiple slices of for OpenFlow based SDN switch. In this, they considered
different types that are packaged as a single product targeted that processing time is only exponential distribution. They
on customers who has many and diverse requirements. For extended their study in [14] for multiple nodes using the
example, a vehicle may need simultaneously a high band- M/M/1 model. Sood et al., [15] proposed a M/Geo/1 based
width for entertainment as well as an ultra-reliable slice for mathematical model for SDN switch, where the author as-
assisted driving. Heterogeneous application is well managed sumes that the processing rate is geometric where it searches
by creating multiple logical networks on the top of common the policies in the table without prior knowledge and is
physical infrastructure. Therefore network slicing along with modeled as geometric distribution. It assumes that complete
SDN will have a great impact on managing the network. processing is geometric. Therefore it is not clear whether
In cellular networks such as 5G and Long Term Evolution the processing time should be considered as a geometric or
(LTE), the SDN concept is incorporated into different parts of exponential distribution.
the core and backhaul network, such as Serving Gateway (S- Several studies [6]–[11], [13]–[15] carried out for Out-
GW), Packet Gateway (P-GW), Mobility Management En- band OpenFlow networks. To the best of our knowledge,
tity (MME), Home Subscriber Service (HSS), etc. However, there is no work done in determining the performance of
most of the work related to SDN that we have come across SDN enabled network under slicing for in-band OpenFlow
was based solely on experiments and simulations. Once Architecture in the literature for the model that consists of
the packet enters into the network, its performance metrics both SDN and slicing. Therefore first, we need to clarify
depends on what are the different components on which the type of queue at a given SDN enabled switch under In-
application need to be processed. Such as S-GW, P-GW, band architecture, and we extend it to SDN enabled network
HSS, and MME and its population matrix. The population slicing for the architecture.
matrix can be calculated by the observation or by using In general, SDN is assumed to be adopted in the back-
some distribution. Similarly, many messages are exchanged haul/transport portion rather than the operation of a mobile
between a controller and a switch like Packetin, Flowmode, network. In this paper, we try to link different aspects, includ-
PortStatus to deploy the policies. These messages should ing in-band OpenFlow-based SDN, queuing theory, cellular
be treated with high priority to reduce the delay. Therefore networks, and network slicing aspects [5].
logically grouping these components based on applications We propose an analytical model for the mobile network
helps to manage the network easily. that takes advantage of both SDN and network slicing con-
A mathematical model can provide researchers with in- cepts like considered in [5], which helps the network op-
sights into how an OpenFlow architecture performs accord- erator to offer the required QoS and also helps to utilize
ing to specific parameters under given circumstances, thus the resources at great extent. The traffic monitoring is one
leading to propose efficient algorithms. Therefore, we pro- of the considerable challenges in the proposed architecture.
vide an analytical model based on results from the queuing The analytical tool can be used to monitor the traffic and
theory of an In-band OpenFlow-based mobile network. utilization of the resources at every node by every slice.
To the best of our knowledge, there is not much work done Our study is twofold; first, we want to find an adequate
in multi-class traffic problems for a mobile network in the analysis model to represent traffic characteristics at mobile
literature. However, there are a couple of experimental stud- networks. Second, we want to validate the analysis model by
ies have been done that investigate the switch performance performing a traffic simulation and determining performance
of SDN enabled networks. Ludwig et al., [6] highlighted measures of the mobile network. For the time being, we
VOLUME 4, 2016 3

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2959351, IEEE Access

S Kamath et al.:Multiclass Queueing Network Modeling and Traffic Flow Analysis for SDN Enabled Mobile Core Networks with Network Slicing

FIGURE 3: Example for network slicing

will consider the simulation model as a verification means to an analytical model for multi-class under slicing. Section IV
validate the analysis model. The analysis model is built with discusses the performance evaluation and application of our
a multiclass queuing theorem where each class represents a model to the network dimensioning and management, and
slice or application. The main contributions of this paper are Section V concludes this paper.
summarized as follows:
• This paper identifies the traffic characteristic at SDN II. TRAFFIC CHARACTERISTIC IN SDN ENABLED
enabled mobile networks under slicing. MOBILE CORE NETWORK USING NETWORK SLICING
• Design of an adequate analysis model of flow-based AND NETWORK MODEL
Internet traffic to represent the traffic characteristic at SDN enabled devices use OpenFlow protocol for communi-
mobile networks under In-band OpenFlow architecture cation [16] between a switch and a controller. An OpenFlow-
for single node homogeneous class. enabled switch essentially maintains a flow table. This table
• Applying an analytical model for multi-class hetero- holds a set of entries, each of which comprises of match
geneous applications queuing theory to measure the fields. The communication between a switch and a controller
performance matrices under network slicing for traffic is bidirectional with a certain type of standard messages [17],
monitoring at every node in an application. such as Packetin, Port-status, Flow-Removed, hello, echo,
The rest of the paper is organized as follows. Section II and Error messages. When a packet arrives at a switch,
gives the traffic characteristics in SDN enabled mobile net- policies for the packet in the flow table is checked and
work followed by a brief background of multi-class queuing applied. If not found, then Packetin message is initiated and
and the proof to show that the type of queue in OpenFlow forwarded to the controller to get the policies. Similarly, if
In-band architecture. Section III explains the application of a user moves from one place to another, then Port-status
4 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2959351, IEEE Access
S Kamath et al.:Multiclass Queueing Network Modeling and Traffic Flow Analysis for SDN Enabled Mobile Core Networks with Network Slicing

message is initiated. These messages, that is, the messages or arrival link. For example, the capacity of the server is
exchanged between a controller and a switch are need to be shared between every packet equally or based on scheduling
considered as high priority messages. We logically divide the algorithms among the class. Frequently processor sharing
communication between the controller and the components refers to Egalitarian Processor Sharing (EPS).
as a slice. The logical division helps to manage and main- In SDN enabled network with slicing, we have different
tain the network and resources easily and effectively [5]. logical slices. Each slice has a packet belonging to different
This concept is well explained in the slicing of the mobile users or tenants. Each user or application carries packets
network, as shown in Fig 3. After deploying the policies, in its slice. Packets belonging to different tenants or users
depending upon the packets, it is necessary to process into a are considered as classes, and each tenant may use different
different component of core networks such as S-GW, P-GW, applications. It means user choice (of service plans) keeps on
or into EPC, etc. We consider all these components as a node changing from time to time; therefore, packet-level modeling
in our work, along with the simple forwarding devices. Sim- is suitable in this scenario. Similarly, the multi-tenants gives
ilarly, in the mobile network, each user has different plans the motivation for the multi-class model. Here, we use the
or different requirements. Therefore logically dividing the multiclass model to analyze the key factors: throughput,
application on a shared medium helps to deploy the policies delay, and utilization statistics.
that make ease if network slicing is used. We consider each There are two types of multiclass models [21]: i) Open
slice into a class. However, packet processing depends on class, in which packets arrive from the external world into
the table size [13] and the policy. It varies from packet to the system and exit out of the system, and ii) Closed class,
packet of every slice. A mobile network is deployed on a in which packets are generated inside the network and die
large geographical area, and it is difficult to have a separate within the network. One can have a mixed model where both
connection from every node to a controller. Therefore, In- open and closed classes are considered.
band architecture is more suitable in such a scenario. The appropriate analytical model for SDN enabled under
Packets that enter and exit a device may fall under packet- slicing is shown in Fig. 4. In In-Band, if the SDN device does
based, and flow or stream-based processing. Packet-based, not find the policies, it forwards a packet to the controller
or stateless processing treats packets discretely. Each packet using Packetin message as a new packet through other nodes.
is assessed individually for treatment. Packet level models Furthermore, the controller deploys the policies using Flow
for high-speed links are difficult to calibrate, because of the message. Therefore there is a new flow generated if policies
high level of multiplexing of many flows whose behavior are not found for the incoming packet; we model this as a
is strongly influenced by the transport protocol and by the change of class. First, we evaluate the homogeneous network
application. Besides, monitoring the traffic at the packet level that is a single class, and then we extend our work to
becomes critical at the optical carrier (OC-192) and above multiclass queuing of the In-band OpenFlow network.
link speeds. Packet-level models are useful in estimating pa- Therefore we begin with the arrival distribution and then
rameters such as queue sizes at buffers and throughput. They processing distribution at a node for SDN enabled networks
have been vigorously used to study the stability of congestion for a single class that is arrival and processing rate at every
control. Stream-based packet handling treats related packets node because of using an in-band architecture, and then we
in the same fashion. This model is used in the backbone extend it to multiple nodes with the multi-class network.
network [18]. Furthermore, the flow level traffic is easy to
monitor using tools like NetFlow [19]. In SDN, enabled
network policies for the packet belong to a slice are the same;
therefore, we consider flow-based processing.
Traffic Analysis (TA) of the above system can be mod-
eled using queuing models (QM) [20]. QM is reasonable
for analyzing frameworks where some constrained resources
are utilized to play out a specific task. The resources are
generally referred to as middle-boxes or nodes and the task
as packets. Every packet has a service prerequisite, and after
accepting it departs from the system. Kendall’s annotation
[20] is used to indicate the properties of a queue. Usually,
the entries (arrival) of packets are stochastic, and the ser-
vice prerequisites are accepted to pursue some likelihood
probability distribution, and thus, the queue procedure is
arbitrary (random) in nature. Queuing discipline decides how
the capacity is shared among the packets at a given time;
for instance, the First-Come-First-Serve (FCFS) discipline FIGURE 4: Multi-Class Model for In-band Openflow Archi-
serves the packets in their arrival order. Processor Sharing tecture
(PS) shares the processor capacity among multiple class
VOLUME 4, 2016 5

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2959351, IEEE Access

S Kamath et al.:Multiclass Queueing Network Modeling and Traffic Flow Analysis for SDN Enabled Mobile Core Networks with Network Slicing

A. DETERMINING QUEUE IN AN SDN ENABLED In our network, we assume p is the probability that a node
SWITCH FOR IN-BAND OPENFLOW ARCHITECTURE does not find the policy for the packet that it has received.
FOR HOMOGENEOUS NETWORK A packet that does not find the policy and the other traffic
The arrival and processing distribution at a node depends is forwarded according to its class to the next node. Besides,
on two possibilities: the availability of policies and the non- we assume that there are n nodes in the system between a
availability of policies. We discuss these possibilities and switch and a controller. The controller deploys the policies
conclude the type of queue that suits both the possibilities. for the request that comes from a node. The single-node open
network model for In-band is as shown in Fig. 5. We also
1) Arrival Distribution at In-Band OpenFlow Node and assume l as the probability that packets are dropped from the
Heterogeneous Network path, that is, packets are sent to trans-coders or firewall for
Let us assume that the packet arrives from outside network further processing. We consider it a leakage. If the event of
follows the Poisson distribution with arrival rate λ1 at node no policies is discovered in the table, the packet is fed back to
1, λ2 at node 2, and λn at node n respectively. It is done based the same node after processing at the controller. The external
on the following assumptions: arrival rate at the controller is λn . Switches and the controller
• Flow arrivals follow a homogeneous Poisson process of
processes each packet at a rate of µs and µc , respectively. Let
finite rate λ. Poisson might be the right model if we Λ be the arrival rate at any given node after the feedback,
consider recent findings by [22], [23] about the process which is the sum of external arrival and the packet that is
of flow arrivals in the backbone of the Internet, where fed back to the node by the controller. Let ρ = µ/λ is the
a large number of flows are multiplexed. It is shown in intensity of load on a node. Applying the traffic rate equation
[23] that a Weibull distribution very well approximates to Fig. 5, we get
the distribution of flow inter-arrival times with a shape
parameter smaller than 1, and that as the traffic intensity Λ1 = λ1 + pΛ1 (1)
increases, flow inter-arrival times become independent, At node 2
whereas the Weibull shape parameter gets close to 1. Λ2 = λ2 + pΛ1 − lΛ1 (2)
Thus, the flow arrival process tends to be in good agree-
ment with a Poisson process. The well-known results Expecting no packet is lost at the controller, intermediate
explain its limit on the superposition of marked point switches and also we assume that the only traffic that are sent
processes. The Poisson property is also known to apply to the controller for not finding policy will return back to the
to aggregates at the session-level. Note that since our intermediate device, then we can rewrite the equation (1) as
model does not depend on a particular definition of flow, λ1
one can group packets into sessions that have Poisson Λ1 = λ1 + pΛ1 = (3)
1−p
arrivals, and apply the model at the session-level.
For simplicity at all the node we assume arrival rate is same
• Flow rate functions are independent of each other and
that is λ, processing rate µ, and leakage l is same. Then we
identically distributed. The assumption on the indepen-
can rewrite the equation (3) and (2) as
dence of flow rate functions is based on the following
facts: λ
Λ1 = (4)
i) The link we consider is a backbone link kept under- 1−p
utilized by engineering rules. It does not, therefore,
experience congestion, and so it does not introduce Λ2 = λ + pΛ1 − lΛ1 (5)
dependency among the flow rate functions. Similarly for the nth node
ii) The flows sharing this link have a large number of dif-
ferent sources and destinations and use many differ- Λn = λ + pΛn−1 − lΛn−1 (6)
ent routes before being multiplexed on the backbone
From equation (3), one can make out that the total arrival
link. The assumption of identical distribution can be
input traffic is not Poisson, however bursty, that is, a packet
relaxed by introducing multiple classes (based on the
may pass at most two times through the queue depending on
transport protocol, flow size, or any other metric).
the p value. There exist no analytical results for such net-
works in which inter-arrival and service times are dependent.
The Kleinrock assumption [24] proposes that consolidating
traffic streams on a transmission line has the impact of re-
establishing the autonomy of arrival and service times. It
implies that the service time of a packet is autonomously
recovered each time the message is added back to the queue,
which in turn implies that the same message is considered to
have different service times each time it enters the queue.
FIGURE 5: Open network model for an in-band architecture It eliminates arrival or service time correlation. With this
6 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2959351, IEEE Access
S Kamath et al.:Multiclass Queueing Network Modeling and Traffic Flow Analysis for SDN Enabled Mobile Core Networks with Network Slicing

approach, we can approximate that the arrival distribution is From Little’s law [27] the length of the queue Lq is given
Poisson with a mean of Λ. Note that it is valid only if one by
uses Kleinrock’s approach. ρ2t
Lq = (12)
1 − ρt
2) Processing Distribution at Single Node Equation (10) reveals that as the number of nodes in-
To determine the type of queue packet processing distribution creases the nth node will have more load which is sum
at a node is also one of the vital features. Processing the of arrival rate from its previous node with small portion
packet follows two steps: i) searching the policy and ii) of leakage at every node. On using equation (11) in
applying the policy. Searching policy follows geometric dis- (12) it is straightforward to verify that the overall queue
tribution as mentioned in [25] (denoted by variable A), and length increases exponentially when the load on any
applying that policy follows an exponential distribution, as node in the path approaches 1. And as the number of
mentioned in [13] (denoted by variable B), and the combined nodes increases in the path, the waiting time increases.
effect of this is discussed in [26]. According to Santhosh et Similarly if the probability of availability of policies
al., [26], geometric distribution effect can be seen only when p is less then the model behaves as M/H/m that is
the load is low; otherwise, it follows an exponential distribu- hyper-exponential distribution which means processing
tion. Similarly, geometric progression leads to exponential at two times at m switches and once at the controller else
infinity. That means when the number of policies increases, the model behaves as M/M/m.
the searching time leads to exponential. Therefore, there are
two independent exponential distributions, one after another. III. ANALYTICAL MODEL FOR MULTICLASS
Overall it gives an exponential response. The challenge here In multiple class queuing models, we assume that there are C
is that in [26], the authors considered Out-band architecture, classes of requests, where C = {ci |i = 1, 2, . . . , n}. In our
but here the request needs to go through the n number of case we assume three classes. Our first slice is the messages
nodes; therefore, the total delay needs to be explored. that are exchanged between controller and the switches like
From Fig. 5 we infer that the total delay depends on the Packetin, Flow-mode. The remaining two classes are the
load on the other nodes also. If there are n nodes in a path, messages that are entered into the core network from outside
and we consider ρk = λk /µk as the load on k th node, where of given applications. Each request from class c spends on
k ∈ n, we explore the processing time at a node for two cases average time Sck at node k. For open models, we use λck
that is with leakage and without leakage. to denote the arrival rates, where λck is the external arrival
i) Without Leakage at a Node: Total load in a path can be rate of class c requests at center k, which may be a node or
written as switch. The transition probability matrix, P for multiple class
ρt = ρ1 + ρ2 + . . . + ρn (7) networks is a C × K × C × K matrix with Pij = Picl jcm
where i, j ∈ K and l, m ∈ C. The Pij is the probability that
The equation (7) can be rewritten in terms of arrival rate a packet of the cth th
l class at the i node is transferred to the
and processing rate as follows. th th
cm class at the j node (routing probability) [28].
Λ1 Λ2 Λn The traffic equation that is the set of the expected number
ρt = + + ... + (8) of visits eck of class c for node k is given by [29]
µ1 µ2 µn
C K
Substituting (2), (3) and assumption made in (4) we get X X
eck = λck + λck Picl jcm (13)
λ1 λ1 + λ2 λ1 + λ2 + λ3 c=1 k=1
ρt = + + + ... l,m=1 i,j=1
µ1 µ2 µ3
λ1 + λ2 + λ3 · · · + λn For the class 1 Picl jcm ∈ {0, 1} which means if the policies
... + (9) available then it is either 0 or 1. For other class it depends on
µn
processing required at other nodes like S-GW, P-GW, MME
If we assume same processing time for all the nodes that etc.
is µ1 = µ2 = . . . µn = µ, then equation (9) can be In equation (13) C corresponds to different slice cate-
written as gories, λck is an arrival rate of a class-c packet at node k
nλ1 + (n − 1)λ2 + · · · + λn in the system, which describes an exogenous arrival rate and
ρt = (10) the arrivals directed from different nodes. It is equivalent to
µ
the departure rate of class c packets from switch k if there
ii) With Leakage of l at a Node: If we consider l leakage,
should be stability.
that is, the packets exit that node after processing.
Let ρck be the traffic intensity of the class c at node k.
nλ1 + (n − 1)λ2 + · · · + λn The performance parameter can be calculated for both PS and
ρt = FCFS systems using following equations [30]
µ
nλ1 + (n − 1)λ2 + · · · + λn C
 
−l (11)
X eck λck
µ ρk = ρck , whereρck = , (14)
c=1
µck
VOLUME 4, 2016 7

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2959351, IEEE Access

S Kamath et al.:Multiclass Queueing Network Modeling and Traffic Flow Analysis for SDN Enabled Mobile Core Networks with Network Slicing

where ρk is the traffic intensity at node k, and µck is the or delay-sensitive applications. Therefore we consider multi-
processing rate for packet of class c at node k. Utilization class based modeling for In-Band architecture.
Uck of k th node by cth class is given by
Uck = λc Dck (15) B. EVALUATION OF MULTI-CLASS NETWORK MODEL
For the multi-class processing we consider result obtained
where Dck is the demand for class c at node k, and defined in the section IV-A for modeling. From the equation (13)
as to equation (20) one can make out that to calculate the
Dck = eck .Sck . (16) performance parameters one need arrival distribution, pro-
The utilization Uk of node k is given by cessing distribution, and the transition probability matrix P .
In the following section we explain how these parameters are
K
X adopted in our model.
Uk = Uck . (17)
k=1

Residence time of the packet of class c at node k is given by ·104

Dck n=2
Rck = . (18) 2
n=3

Average Waiting Time (µsec)

1 − Uk
The average queue length of class c at node k is given by
1.5
Uck
Qck = . (19)
1 − Uk
The system response time for class c is given by 1
K
X
Rc = Rck . (20) 0.5
k=1

These equations are calculated based on assumption that

packet arrives from outside network follows the Poisson 0
distribution with arrival rate λck . 0 0.2 0.4 0.6 0.8
Arrival Rate
IV. PERFORMANCE EVALUATION AND DISCUSSION
A. EVALUATION OF SINGLE NODE SINGLE CLASS
FIGURE 6: Relationship between average waiting time with
QUEUE
the number of nodes in series
We have tested the effect of using In-band communication
and multi-class using MATLAB. For In-band, we have con-
nected M/M/m nodes in series and increased the arrival
rate at every node from 0.1 to 0.9 using Poisson variate.
4,000 L = 0.1
The processing time is used by generating an exponential
L = 0.2
Average Waiting Time (µsec)

variate with a mean of 9.8 µs. According to [31] the average

processing time of a 1500 byte packet takes 9.8µs at the
3,000
switch because Maximum Transferable Unit (MTU) value
in LTE core network is 1428 bytes [32]. Later we have
increased the number of the node to three to test the effect 2,000
of waiting time with the number of nodes. The simulation
results are shown in Fig. 6 and 7 prove that waiting time
increases exponentially at a shallow arrival rate as the number 1,000
of nodes in the path increases; the result validates equation
(11). Similarly, without loss of generality, we conclude that
nodes in series form a M/M/m model with different mean. 0
The new arrival depends on arrival at other nodes as well as
0.1 0.2 0.3 0.4 0.5 0.6 0.7
leakage at every node.
It is necessary to treat the policy with high priority to Arrival Rate
reduce the delay for the controller request packet. One more
FIGURE 7: Relationship between average waiting time with
observation is that if the controller is close to the node, delay
the leakage factor in series
in loading the policies into the node reduces. Therefore the
hierarchical architecture for the placement of the controller
will be a good choice in the use cases like mobile networks
8 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2959351, IEEE Access
S Kamath et al.:Multiclass Queueing Network Modeling and Traffic Flow Analysis for SDN Enabled Mobile Core Networks with Network Slicing

1) Arrival rate at multi-class networks values of p. The class switching probability matrix C for
The arrival rate at multi-class defines the number of packets p = 0.9 is given in equation (23), and q is the probability
arriving at the node k of class c. We use λck , that is, the of availability of class 2 policies in the switch.
arriving rate, which is according to the Poisson arrival. We c1 c2 c3
increase the arrival rate from 1 to N . We generate a Poisson c1
"
p 0 1−p
#
variate depicting the Poisson arrival with a mean sample 1−q
taken between 1 to N . C = c2 0 q (22)
c3 1−p 1−q 1−p+1−q
2) Processing rate at multi-class networks
The service rate matrix S, where Sck is the time taken to
process the packet of class c at node k. We assume the service c1 c2 c3
c1 0.9 0 0.1
" #
rate of every single server is state-independent. The flow
C = c2 0 0.9 0.1 (23)
is Poisson arrival, and processing is FIFO, as discussed in
section IV-A. We also assume that there is a single queue at c3 0.1 0.1 0.8
the node instead of assuming a separate queue per line card.
We consider average processing time at a switch for 1500 Traffic matrix
bytes of the packet is 9.8 µs, and processing time at controller
The traffic matrix is the population exchanges between every
µc is 240 µs [31] This value is close to the LAN. We generate
pair of network nodes. In SDN enabled slicing we assume
exponential variate with µk =9.8 µs to depict the exponential
that data transmitted between nodes follows a uniform dis-
service time. The output matrix S for the switch will have
tribution, except to the same node that is Mi,j = 0 when
the values in µs, as shown in equation (21). Where S is the
i = j, where Mi,j is the amount of traffic movement from ith
class-node wise service matrix, and rows represent the given
node to the j th node. For example, once the packet enters into
class, and column represents nodes.
the core network, S-GSN has to decide whether to send the
packet to EPC and to which S-GW it should forward, which
k1 k2 k3 k4 P-GW it should use to connect to the Internet. The sample of
c1 " 2.33 19.10 7.70 18.90# traffic matrix M is as shown in equation (24).
S = c2 11.13 5.30 12.50 9.10 (21)
c3 9.80 9.00 13.40 12.40
k1 k2 k3 k4
k1 0.0000 0.2837 0.4540 0.2627
3) Transition probability matrix in multi-class Networks M = k2 0.3414 0.0000 0.3073 0.3513 (24)
k3 0.2617 0.2808 0.0000 0.4573
 
Routing probability matrix P is a C × K × C × K matrix.
Transition probability depends on two main concepts they k4 0.2068 0.4489 0.34501 0.0000
are i) Class switching probability and ii) Traffic matrix. The
following section explains how they are computed in our By multiplying traffic matrix and class switching probabil-
model. ity one can calculate P for the equation 13. Using equations
(14) to (20) and matrices C, M , we can calculate the perfor-
Class Switching Probability mance matrices of SDN enabled network with slicing. The
Class switching probability is a matrix of size C × C, as results obtained are shown in Fig. 8 to 14.
shown below. It is the probability of switching a class after The graphs are shown in Fig.8 to 14 is the effect of avail-
processing at a node. In our model (see Fig. 4), we consider ability of the policies on utilization, response time, through-
that class c3 carries the controller packets, that is, if a node put and the mean queue length of both the node and class at
does not find its policies it changes its class from any class to different probability p of class switching and arrival of class1
c3 like Packetin, Port-status, Flow-removed etc. Switching by keeping class 2 constant. From figure 9 the utilization of
from other classes is not considered in this simulation; how- nodes increases linearly as per the equation (15). It is because
ever, one can consider this for priority-based scheduling in the demand for a given class in a node remains fixed, and
reality. The c1 , c2 are application slice classes where packets the arrival rate increases linearly. The utilization, particularly
arrives at the controller. The change of the classes is done at node 1 reaches maximum at around λck of 80-90 arrivals
according to the probability say p, where p is the probability per second for p = 0.5 and it takes around 170-180 arrivals
that a switch finds policies in its table. Therefore change of per second for p = 0.9. This observation indicates that the
class from c1 → c2 and c2 → c1 is 0. Wamser [33] has shown availability of policies in the node supports more and more
that the probability of a new flow being observed is 0.16 at the arrival of the packet for a given processing rate and proba-
switch. However, if the traffic is using a virtualized network bility matrix. In Fig. 8, the utilization of class 1 is less when
like slicing, then the probability varies from p = 0.2 to a node depends more on the controller for the policies. That
p = 0.5. Similarly, p = 1 indicates that the controller handles is because a given node need not process the packet again.
all the traffic. We calculate the traffic parameters for different Moreover, it increases its utilization when there are policies
VOLUME 4, 2016 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2959351, IEEE Access

S Kamath et al.:Multiclass Queueing Network Modeling and Traffic Flow Analysis for SDN Enabled Mobile Core Networks with Network Slicing

0.6 Class1 1 N ode1

Class2 N ode2
Class3 N ode3
Calss-wise utilization

0.8

Nodewise Utilization
N ode4
0.4
0.6

0.4
0.2
0.2

0 0

0 20 40 60 80 0 20 40 60 80
Arrival rate of class 1 Arrival Rate of Class 1

(a) At p = 0.5 (a) At p = 0.5

Class1 1 N ode1
1
Class2 N ode2
Class3 N ode3
0.8
Nodewise Utilization
Class-wise utilization

0.8 N ode4

0.6 0.6

0.4 0.4

0.2 0.2

0 0

0 50 100 150 200 0 50 100 150 200

Arrival rate of class 1 Arrival Rate of Class 1

(b) At p = 0.9 (b) At p = 0.9

FIGURE 8: Average class-wise utilization in the network at FIGURE 9: Average node-wise utilization
different values of p

deployed before the packet arrives at a node. Similarly, from From Fig.12 and 13 we can infer that node and class
Fig. 8, it is clear that the availability of the policies increases throughput is more when the probability of having policies
utilization of class 1. Also, at p = 0.5 utilization of class 3 is in the node are more. Therefore, deploying the policies in ad-
more, this is because of class switching from class 1 to class vance helps the SP to utilize and plan the network resources.
3. Since mobility is involved in the cellular network deploying
the policies in advance is a complex operation, lot of pack-
From Fig. 11 we infer that the response time for the given
ets like Packetin, Port-status, Flow-removed messages are
class is less if the policies are available. Therefore it is a good
exchanged between controller and switches, which increases
choice to use learning-based policy deployment in the nodes
the traffic in the network hence leading to more resource
to reduce the response time of the application. From Fig. 10
consumption. Such an increase in resource consumption can
we can conclude that in both the cases that are at p = 0.5
be reduced using learning based techniques. Figure 14 gives
and p = 0.9, the response time increases exponentially
the statistics of the queue in every node for different classes.
when the node reaches its utilization close to 1. The average
It helps the SP to plan the resource allocation for a different
queue length increases when the load on a node reaches its
slice.
maximum capacity, and the throughput of the class increases
with more increase in probability.
10 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2959351, IEEE Access
S Kamath et al.:Multiclass Queueing Network Modeling and Traffic Flow Analysis for SDN Enabled Mobile Core Networks with Network Slicing

1 Class1 N ode1
Class1 N ode2
0.8
Class3 N ode3

Response Time (msec)

Response Time (msec)

0.8
N ode4
0.6
0.6
0.4
0.4

0.2
0.2

0
0
0 20 40 60 80 0 20 40 60 80 100
Arrival Rate of Class 1 Arrival Rate of Class 1

(a) At p = 0.5 (a) At p = 0.5

Class1 N ode1
2.5
Class2 N ode2
Class3 Response Time (msec) N ode3
Response Time (msec)

2 N ode4
2
1.5

1
1

0.5

0 0

0 50 100 150 0 50 100 150

Arrival Rate of Class 1 Arrival Rate of Class 1

(b) At p = 0.9 (b) At p = 0.9

FIGURE 10: Class-wise response time FIGURE 11: Node-wise response time

traffic on the node of the backbone, by only monitoring the

C. APPLICATION OF THE MODEL TO THE NETWORK edges.
DIMENSIONING AND MANAGEMENT The detailed information provided by our model on the
In this section, we discuss some applications of our model traffic helps to dimension backbone nodes. Given the char-
to network dimensioning and management. Suppose that a acteristics of flows composing the traffic, the nodes of the
Service provider (SP) collects statistics on flow delay and backbone network can be dimensioned to avoid congestion.
flow arrivals using tools such as Cisco NetFlow. With this Note that for highly variable traffic, dimensioning the nodes
sole information, the SP can compute the moments of the of the backbone based only on the average utilization is not
total rate. This way, the SP would have more detailed infor- enough to avoid congestion. Rate variation also needs to be
mation than that provided by Simple Network Management considered.
(SNMP) (one of the problems of SNMP is that it does not Computing the traffic in the backbone using the infor-
capture traffic variation at short time scales). The information mation on flows is not the only application of our model
on flows can be collected on the node or the class at a node to network dimensioning and management. A fundamental
we want to monitor. It can also be collected at the edges of problem the operator faces is the planning and up-gradation
the backbone. Combined with the routing information in the of the backbone nodes or links in order to maintain the
edge routers, this will give information on flows on each class absence of congestion. One may ask the following questions:
in the backbone. Our model can then be used to compute the i) What is the impact on the node/class utilization caused
VOLUME 4, 2016 11

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2959351, IEEE Access

S Kamath et al.:Multiclass Queueing Network Modeling and Traffic Flow Analysis for SDN Enabled Mobile Core Networks with Network Slicing

120 N ode1
Class1
MeanThroughput in Packets/sec

N ode2
Class1
100 150 N ode3
Class3
N ode4
80

Throughput
60 100

40
50
20

0 20 40 60 80 0 20 40 60 80

Arrival Rate of Class 1 Arrival Rate of Class 1

(a) At p=0.5
(a) At p=0.5

Class1 250 N ode1

MeanThroughput in Packets/sec

200 Class2 N ode2

Class3 200 N ode3
N ode4
150
Throughput

150

100
100

50
50

0 0
0 50 100 150 200 0 50 100 150
Arrival Rate of Class 1 Arrival Rate of Class 1
(b) At p = 0.9 (b) At p = 0.9

FIGURE 12: Class-wise throughput FIGURE 13: Node-wise throughput

by a change in the distribution of policies, for example, can be done quickly using SDN. The availability of the
to the arrival of a new application where policies are not policies reduces the number of messages exchanged between
deployed? the node and a controller, which means high priority class
ii) What is the impact on the node/class utilization caused packets are reduced, and application packets are propagated
by a change in policies/slices, for example, to an in- quickly out of the network, which gives high utilization.
crease in the number of users in the congested access If policies are not there, then packets need to wait till the
networks? policies are deployed which in turn gives more delay. It is
We illustrate this application by the following two examples. because we are considering In-band OpenFlow architecture,
The first example shows the impact of a change in the flow in which packet needs to travel through other nodes to reach
arrival rate λ on the traffic, and hence on the dimensioning of the controller, and while deploying the policies also it should
the core network. The second example shown is the impact follow the same path in the reverse direction.
of the availability of the policies in the nodes.
When the arrival rate of flows increases, the bandwidth V. CONCLUSION
of the backbone links requirement and nodes with high pro- In this paper, we have proposed an analytical model to
cessing has to be increased as well. Similarly, load balancing determine the type of queue in an In-band Open-Flow archi-
should be used to reduce the burden on a given node, which tecture for uncongested backbone link, and we have extended
12 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2959351, IEEE Access
S Kamath et al.:Multiclass Queueing Network Modeling and Traffic Flow Analysis for SDN Enabled Mobile Core Networks with Network Slicing

of every slice can be evaluated using this method, which

N ode1 helps the management of the resources efficiently among all
50
N ode2 the slices. The experimental results shows that it is advisable
N ode3 to keep the policies in advance in the switches before the
40 N ode4 flow arrives at the system. We will be using a scheduling
Queue Length

algorithm that reduces the response time of the high priority

30
class packets in our future work.

20
REFERENCES
[1] A. Asava, “Ott is not just for consumers! how
10 businesses can benefit,” https://www.kandy.io/blog/
ott-not-just-consumers-how-businesses-can-benefit, April 2015.
[2] N. McKeown, T. Anderson, H. Balakrishnan, G. Parulkar, L. Peterson,
0 J. Rexford, S. Shenker, and J. Turner, “Openflow: Enabling innovation in
campus networks,” SIGCOMM Comput. Commun. Rev., vol. 38, no. 2,
0 20 40 60 80 pp. 69–74, Mar. 2008. [Online]. Available: http://doi.acm.org/10.1145/
Arrival Rate of Class 1 1355734.1355746
[3] F. Malik, “How to configure 5g wireless technology and future
mobile communication?” https://www.loginworks.com/blogs/
(a) At p=0.5
configure-5g-wireless-technology-future-mobile-communication/,
Jan 2018.
[4] O. Sallent, J. Perez-Romero, R. Ferrus, and R. Agusti, “On radio access
N ode1
250 network slicing from a radio resource management perspective,” IEEE
N ode2 Wireless Communications, vol. 24, no. 5, pp. 166–174, October 2017.
N ode3 [5] J. Ordonez-Lucena, P. Ameigeiras, D. Lopez, J. J. Ramos-Munoz, J. Lorca,
200 N ode4 and J. Folgueira, “Network slicing for 5g with sdn/nfv: Concepts, architec-
tures, and challenges,” IEEE Communications Magazine, vol. 55, no. 5,
Queue Length

pp. 80–87, May 2017.

150 [6] A. Ludwig, S. Dudycz, M. Rost, and S. Schmid, “Transiently secure
network updates,” SIGMETRICS Perform. Eval. Rev., vol. 44, no. 1,
100 pp. 273–284, Jun. 2016. [Online]. Available: http://doi.acm.org/10.1145/
2964791.2901476
[7] Z. Bozakov and A. Rizk, “Taming sdn controllers in heterogeneous hard-
50 ware environments,” in 2013 Second European Workshop on Software
Defined Networks, Oct 2013, pp. 50–55.
[8] J. Naous, D. Erickson, G. A. Covington, G. Appenzeller, and
0 N. McKeown, “Implementing an openflow switch on the netfpga
platform,” in Proceedings of the 4th ACM/IEEE Symposium on
0 50 100 150 200 Architectures for Networking and Communications Systems, ser. ANCS
’08. New York, NY, USA: ACM, 2008, pp. 1–9. [Online]. Available:
Arrival Rate of Class 1 http://doi.acm.org/10.1145/1477942.1477944
[9] A. Lazaris, D. Tahara, X. Huang, E. Li, A. Voellmy, Y. R. Yang, and
(b) At p = 0.9 M. Yu, “Tango: Simplifying sdn control with automatic switch property
inference, abstraction, and optimization,” in Proceedings of the 10th ACM
FIGURE 14: Mean queue length International on Conference on Emerging Networking Experiments and
Technologies, ser. CoNEXT ’14. New York, NY, USA: ACM, 2014,
pp. 199–212. [Online]. Available: http://doi.acm.org/10.1145/2674005.
2675011
this work to a multiclass network system using multiclass [10] N. Katta, O. Alipourfard, J. Rexford, and D. Walker, “Cacheflow:
queuing theory. The proposed queuing system assumes that Dependency-aware rule-caching for software-defined networks,” in
Proceedings of the Symposium on SDN Research, ser. SOSR ’16.
the flows are heterogeneous: requests come from a variety New York, NY, USA: ACM, 2016, pp. 6:1–6:12. [Online]. Available:
of sources, from customers to OpenFlow switches and from http://doi.acm.org/10.1145/2890955.2890969
OpenFlow switches to the controller, with different intensity. [11] A. Khan and N. Dave, “Enabling hardware exploration in software-defined
networking: A flexible, portable openflow switch,” in Proceedings of the
The objective is to analyze non-ergodic traffic patterns and 2013 IEEE 21st Annual International Symposium on Field-Programmable
load-dependent network architecture like mobile networks Custom Computing Machines, ser. FCCM ’13. Washington, DC,
under network slicing. The applicability of the proposed USA: IEEE Computer Society, 2013, pp. 145–148. [Online]. Available:
http://dx.doi.org/10.1109/FCCM.2013.15
model is determined by the parameters such as the average [12] R. Sherwood, G. Gibb, K.-K. Yap, G. Appenzeller, M. Casado, N. McKe-
waiting time for multiclass equipment both OpenFlow switch own, and G. Parulkar, “Flowvisor: A network virtualization layer,” Open-
and controller, average queue length on controller and Open- Flow Switch Consortium, Tech. Rep, vol. 1, p. 132, 2009.
Flow switch, and switch processor utilization for each slice. [13] M. Jarschel, S. Oechsner, D. Schlosser, R. Pries, S. Goll, and P. Tran-Gia,
“Modeling and performance evaluation of an openflow architecture,” in
For each traffic class, routing probabilities were randomly Proceedings of the 23rd International Teletraffic Congress, ser. ITC ’11.
defined so that a particular choice of distribution does not International Teletraffic Congress, 2011, pp. 1–7. [Online]. Available:
bias the simulation. Given the routing matrix, the statistics http://dl.acm.org/citation.cfm?id=2043468.2043470
[14] K. Mahmood, A. Chilwan, O. Østerbø, and M. Jarschel, “Modelling of
of the network provides the knowledge of valuable resource openflow-based software-defined networks: the multiple node case,” IET
consumptions for every class at every node. The performance Networks, vol. 4, no. 5, pp. 278–284, 2015.

VOLUME 4, 2016 13

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2959351, IEEE Access

S Kamath et al.:Multiclass Queueing Network Modeling and Traffic Flow Analysis for SDN Enabled Mobile Core Networks with Network Slicing

[15] K. Sood, S. Yu, and Y. Xiang, “Performance analysis of software-defined SANJAY SINGH (M’2005, SM’2013) graduated
network switch using m/geo/1 model,” IEEE Communications Letters, from Institution of Electronics and Telecommu-
vol. 20, no. 12, pp. 2522–2525, Dec 2016. nications Engineers, New Delhi, India in 2001,
[16] W. Stallings, “Software-defined networks and OpenFlow,” The Internet and received his M.Tech and PhD degrees from
Protocol Journal, vol. 16, no. 1, pp. 2–14, 2013. Manipal Institute of Technology, Manipal, India in
[17] S. Contributors, “Openflow switch specification version 1.3.1,” Open 2003 and 2010, respectively. In 2004, he joined the
Networking Foundation, Tech. Rep., 01 2012, tech.report. Department of Information and Communication
[18] S. B. Fred, T. Bonald, A. Proutiere, G. Régnié, and J. W. Roberts,
Technology at Manipal Institute of Technology,
“Statistical bandwidth sharing: A study of congestion at flow level,”
MAHE, Manipal, where currently he is working
SIGCOMM Comput. Commun. Rev., vol. 31, no. 4, pp. 111–122, Aug.
2001. [Online]. Available: http://doi.acm.org/10.1145/964723.383068 as a Professor.
[19] R. Sommer and A. Feldmann, “Netflow: Information loss or win?” He is also heading Centre for Artificial and Machine Intelligence (CAMI)
in Proceedings of the 2Nd ACM SIGCOMM Workshop on Internet at Manipal Institute of Technology,Manipal.His work spans various areas,
Measurment, ser. IMW ’02. New York, NY, USA: ACM, 2002, pp. including wireless communications, artificial intelligence, machine learning,
173–174. [Online]. Available: http://doi.acm.org/10.1145/637201.637226 neural networks, fuzzy logic, and information retrieval. He is a member of
[20] U. N. Bhat, An Introduction to Queueing Theory: Modeling and Analysis ACM, and senior member of IEEE.
in Applications, 2nd ed. Birkhäuser Basel, 2015.
[21] S. Balsamo and A. Marin, Queueing Networks. Berlin, Heidelberg:
Springer Berlin Heidelberg, 2007, pp. 34–82. [Online]. Available:
https://doi.org/10.1007/978-3-540-72522-0_2
[22] S. Ata, M. Murata, and H. Miyahara, “Analysis of network traffic and
its application to design of high-speed routers,” IEICE Transactions on
Information and Systems, vol. E83-D, 05 2000.
[23] J. Cao, W. S. Cleveland, D. Lin, and D. X. Sun, “On the
nonstationarity of internet traffic,” SIGMETRICS Perform. Eval. Rev.,
vol. 29, no. 1, pp. 102–112, Jun. 2001. [Online]. Available: http:
//doi.acm.org/10.1145/384268.378440
[24] L. Kleinrock, Communication Nets; Stochastic Message Flow and Delay.
New York, NY, USA: Dover Publications, Inc., 1972.
[25] K. Sood, S. Yu, and Y. Xiang, “Performance analysis of software-defined
network switch using m/geo/1 model,” IEEE Communications Letters,
vol. 20, no. 12, pp. 2522–2525, Dec 2016.
[26] S. Kamath, S. Singh, and M. S. Kumar, “Performance analysis of out-band
openflow switch architecture: The single node case,” in Proceedings of SATHISH KUMAR M graduated from the Institu-
IEEE International Conference on Advanced Networks and Telecommu- tion of Engineers (India) in Electronics and Com-
nications Systems (ANTS) (IEEE ANTS 2018), Indore, India, Dec 2018, munication Engineering and obtained his Master’s
pp. 1–6. in Engineering specializing in Optical Communi-
[27] D. Bertsimas and D. Nakazato, “The distributional little’s law and its
cation from College of Engineering, Guindy, Anna
applications,” Oper. Res., vol. 43, no. 2, pp. 298–310, Apr. 1995. [Online].
University. He was awarded the PhD degree by
Available: http://dx.doi.org/10.1287/opre.43.2.298
[28] E. D. Lazowska, J. Zahorjan, G. S. Graham, and K. C. Sevcik, Quanti- the National Institute of Technology Karnataka
tative System Performance: Computer System Analysis Using Queueing (NITK), Surathkal for his thesis on Performance
Network Models. Upper Saddle River, NJ, USA: Prentice-Hall, Inc., Enhancement of Optical CDMA Networks and
1984. Hybrid Multiplexed Optical Communication Sys-
[29] S. Balsamo and A. Marin, “Queueing networks,” in International School tems. He started his academic career as a lecturer in Electronics and
on Formal Methods for the Design of Computer, Communication and Communication Engineering at Dr. T. M .A. Pai Polytechnic, Manipal, in
Software Systems. Springer, 2007, pp. 34–82. the year 1992.
[30] X. Chao and M. Pinedo, “Networks of queues with batch services, In 1998 he joined the Electronics and Communication Engineering De-
signals and product form solutions,” Operations Research Letters, partment of MIT, Manipal, as a lecturer. He continued till 2008 as a reader
vol. 17, no. 5, pp. 237 – 242, 1995. [Online]. Available: http: in the same department. He was the recipient of the International Scholar
//www.sciencedirect.com/science/article/pii/016763779500026G Exchange Fellowship of the Korea Foundation for Advanced Studies for
[31] R. Sherwood, “Cbench (controller benchmark),” Available:https://github. the year 2006-2007. From 2008 to 2010 he was a Research Assistant
com/mininet/oflops/tree/master/cbench, Jan 2013. Professor in the school of Electrical Engineering and Computer Science,
[32] USAT, “What are the main mtu and mss design considerations?”
Seoul National University, South Korea. He joined back MIT, Manipal in
http://usatcorp.com/faqs/main-mtumss-design-considerations/.
2010 as a Professor in the Electronics and Communication Engineering
[33] F. Wamser, R. Pries, D. Staehle, K. Heck, and P. Tran-Gia, “Traffic
characterization of a residential wireless internet access,” Telecommun. Department. Currently, he is heading the department.
Syst., vol. 48, no. 1-2, pp. 5–17, Oct. 2011. Dr.Sathish Kumar works in the area of wireless communications, optical
communications and nanophotonics.

SANTHOSH KAMATH received his B.E. from

Vishweshwarayya University Belgaum, Kar-
nataka, India in 2004. He received M.Tech in
Network Engineering from Manipal University
in 2009. He is pursuing his Ph.D. in the area of
Software Defined Networking at Manipal Institute
of Technology, MAHE, Manipal, India. His re-
search interest includes performance analysis and
enhancement of the mobile core network for OTT
application using SDN.

14 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.