Performance Evaluation of Software Defined Network Controllers
Edna Dias Canedo
1 a
, F
´
abio L
´
ucio Lopes de Mendonc¸a
2
, Georges Daniel Amvame Nze
2
,
Bruno J. G. Praciano
2,3 b
, Gabriel P. M. Pinheiro
3
and Rafael T. de Sousa Jr.
2 c
1
Department of Computer Science, University of Bras
´
ılia (UnB), Bras
´
ılia-DF, Brazil
2
Cybersecurity INCT Unit 6, Decision Technologies Laboratory — LATITUDE, Electrical Engineering Department (ENE),
Technology College, University of Bras
´
ılia (UnB), Bras
´
ılia-DF, Brazil
3
Department of Mechanical Engineering, University of Bras
´
ılia (UnB), Bras
´
ılia-DF, Brazil
Keywords:
Software Defined Networks, Virtualization, Network Topology, Software Defined Networks Controller.
Abstract:
The increasing digitization of various industrial sectors, such as automotive, transportation, urban mobility,
telecommunications, among others, reveals the need for using point-to-multipoint communications services,
such as constantly updating software and reliably delivering messages users, as well as optimizing the use
of hardware and software resources. The implementation of Software Defined Networks (SDN) provides the
network with the flexibility and programming capabilities needed to accurately and reliably support point-to-
multipoint distribution services. This paper presents an account of the activities implemented to produce the
performance evaluation of the SDN controllers. We identified the metrics used in the literature to perform
performance evaluation of SDN controllers. The mechanisms adopted were used to describe the performance
evaluation environment and the respective operating processes of the controllers. Besides, a methodology is
proposed to perform the performance evaluation of the controllers.
1 INTRODUCTION
Some serious problems arise in the current net-
works, such as the limitations of standardized equip-
ment that runs proprietary software, the high demand
of routing tables, the complexity of various proto-
cols, the large number of applications executed by
users which create network bottlenecks. To mitigate
these problems, a new concept of network architec-
ture, called Software Defined Network (SDN) (Chen,
2019), (Karimzadeh et al., 2014) has emerged.
SDN is increasingly gaining ground in the context
of digital transformation and new technology capabil-
ities, because its resource provisioning is not much af-
fected by hardware constraints as SDN control system
is centralized allowing intelligent monitoring to adapt
the network automatically according to traffic condi-
tions (Hohlfeld et al., 2018). In short, it provides a
way to scan the network (Ochoa-Aday et al., 2019)
and to consequently adapt the network services.
Therefore, the SDN design is an architecture of
networking between computers, whose main differ-
a
https://orcid.org/0000-0002-2159-339X
b
https://orcid.org/0000-0002-7423-6695
c
https://orcid.org/0000-0003-1101-3029
ence to traditional models is that SDN allows central-
ized control of the network through software appli-
cations. In practice, this helps the operator to man-
age the entire network more consistently, regardless
of the underlying communications technology that is
used (Ochoa-Aday et al., 2019).
The SDN paradigm breaks vertical integration by
radically separating the packet forwarding and the
control plans, providing applications with a central-
ized and abstract view of network distribution (Jung
et al., 2019). This architecture allows software de-
velopers to build new network deployment methods
tailored to different user needs to optimize the per-
formance of their applications and consequently opti-
mize their tasks (Chen, 2019). Data and control plans
are interconnected by public interfaces and allow di-
rect programming of the control plan. This way, all
network policies can be implemented and managed
logically centrally on one controller, enabling net-
work management as a single, programmable entity
(Ai et al., 2019).
This paradigm has been gaining ground and great
attention from researchers and major computer net-
work industries (Alharbi and Portmann, 2019), (Cer-
roni et al., 2019), (Umair et al., 2019), (Rasool et al.,
Canedo, E., Lopes de Mendonça, F., Nze, G., Praciano, B., Pinheiro, G. and Sousa Jr., R.
Performance Evaluation of Software Defined Network Controllers.
DOI: 10.5220/0009414303630370
In Proceedings of the 10th International Conference on Cloud Computing and Services Science (CLOSER 2020), pages 363-370
ISBN: 978-989-758-424-4
Copyright
c
2020 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
363
2019), (Din et al., 2019).
The widely used and consolidated standard for
SDN is OpenFlow (Poncea et al., 2019). In SDN, all
intelligence is logically centered on software-based
controllers (control plane) and network devices be-
come simple packet forwarding elements (data plan)
that can be programmed through an open interface
provided by OpenFlow (Poncea et al., 2019).
This paper presents a proposal for a methodology
for performing the performance evaluation of SDN
controllers. The proposed methodology will be im-
plemented and used by the projects developed in the
Latitude research laboratory and will allow the per-
formance monitoring of the resources used.
The remainder of this paper is organized as fol-
lows. Section 2 presents the background. Section 3
describes the proposed model for the comparison and
performance of SDN controllers. Section 5 concludes
the paper.
2 BACKGROUND
SDN arose due to the need to automate, scale and
optimize networks, in order to better deal with ap-
plications coming from the public cloud, from pri-
vate storage services and also from databases, that
is, it was developed to follow the changes by which
service providers and operators have faced in recent
years (Alonso et al., 2019).
The SDN keep traffic flowing and assist with near-
instant problem resolution, offering cost savings as
costs are similar to consumption, generating long-
term savings (Gal
´
an-Jim
´
enez et al., 2018).
SDN also help improve the quality of services
and/or products offered by organizations to end con-
sumers by enabling centralized cloud monitoring,
configuration, management, service delivery, control,
and automation (Schwabe, 2018).
SDN controller performs network-intensive oper-
ations, and performance is a factor that deserves at-
tention in the network role virtualization scenario, es-
pecially considering the virtualization overheads that
are already widely investigated.
In the field of SDN controllers, the two main per-
formance metrics widely considered in various inves-
tigations are throughput and latency(Lai et al., 2019;
Li et al., 2019). Throughput refers to the performance
of the network function, namely, the amount of Open-
Flow streams the controller is capable of processing
per second, while latency measures the controller re-
sponse time to a PACKET IN message received from
a switch (Lai et al., 2019)(Li et al., 2019).
Some other secondary metrics may also be con-
sidered, such as:
CPU Performance A set of network functions
are sensitive to throughput and latency because
they essentially operate with packet send/receive.
However, computing is also a critical factor for the
performance of much of the virtualization of net-
work functions, so it is necessary to analyze the
percentage of system CPU utilization, particularly
in virtual environments that can impose consider-
able processing degradation and limit the overall
performance of the network function.
Memory Performance Managing RAM utiliza-
tion efficiently is a critical factor in the perfor-
mance of any computer system, especially in sce-
narios where resources are limited. Memory al-
location aspects need to be considered in per-
formance evaluation of virtualization of network
functions because overall system performance can
be negatively affected if poorly managed RAM
results in secondary (slower) memory accesses.
This situation gets even worse in virtualized sce-
narios because disk I/O is still a big performance
bottleneck for virtualized systems.
Software Defined Networks is a paradigm that
breaks vertical integration with the radical separation
of packet forwarding and control plans, providing ap-
plications with an abstract centralized view of the dis-
tributed state of the network (Kantor et al., 2019).
Data and control plans are interconnected by pub-
lic interfaces and allow direct programming of the
control plan. This way, all network policies can be
implemented and managed logically centrally as a
whole on the controller, enabling network manage-
ment as a single, programmable entity. Figure 1
presents the architecture of a SDN (Bozakov, 2016).
The controller is the main element of software-
defined networks because it centralizes all control and
management of the network. Thus, performance is
a critical factor in SDN and has been the subject of
much research (Bholebawa and Dalal, 2018).
In production networks, the controller manages an
OpenFlow network of real switches. However, in ex-
perimental scenarios, when implementation of a real
structure is not possible, tools that emulate switches
are generally used, such as the Cbench tool, much ex-
plored in OpenFlow controller performance investiga-
tion (Bholebawa and Dalal, 2018).
2.1 Related Works
Since the development of SDN, many comparisons
between SDN controllers have been made, mainly ad-
CLOSER 2020 - 10th International Conference on Cloud Computing and Services Science
364
Figure 1: Architecture SDN (Bozakov, 2016).
dressing a comparative analysis of Throughput and la-
tency using the Cbench (Laissaoui et al., 2015) (Khat-
tak et al., 2014).
Works typically compare various open-source
controllers, providing a quantitative analysis of
throughput and latency. Scenarios often include vary-
ing the number of switches and threads to assess con-
troller scalability (Khattak et al., 2014).
Tootoonchian et al. (Tootoonchian et al., 2012)
have pioneered comparative studies of SDN con-
trollers, the authors conducted a comparative analysis
of the effectiveness of open source controllers.
Zhao et al. (Zhao et al., 2015) conducted a com-
prehensive performance evaluation of open source
SDN controllers, also assessing how system config-
urations can interfere with controller performance.
Salman et al. (Salman et al., 2016) performed an
achievement appraisal of the latest controllers such as
ONOS and Libfluid-based (raw, base) controllers us-
ing Cbench (Jawaharan et al., 2018) as a testing tool
and concluded that the choice of the ideal driver de-
pends on the user’s criteria. Some work goes beyond
comparing controller performance to focus on bench-
mark tools.
Erickson (Erickson, 2013) addresses the perfor-
mance evaluation of controllers and their relationship
to the programming language used to develop them.
Alencar et al. (Alencar et al., 2014) inscribe
a comparative study between two Java-based con-
trollers (Floodlight and Beacon) from a different per-
spective: software aging. In this study, controllers
are compared primarily from a memory leak stand-
point, with Beacon outperforming Floodlight with
less memory consumption.
Rashma and Poornima (Rashma and Poornima,
2019) proposed a rapid SDN prototyping of Single
Controller (SC) and Multi Controller (MC) architec-
ture on mininet, a programming flexible simulator.
The emulation displays a Multi Controller (MC) SDN
architecture which improves efficiency and exhibits
competence in handling scalable network on contrary
to Single-Controller (SC) architecture.
The proposed work demonstrates establishing
user-defined controllers on inbuilt Open Flow con-
trollers, which also brings intra-cluster and inter-
cluster communication in a hierarchical network.
They empirically ratify the scalability of the network
by increasing the number of host nodes (Rashma and
Poornima, 2019).
de Jesus et al. (de Jesus et al., 2014) analyze re-
cent security proposals derived from the use of SDN
and verify if its usage can help the increase of con-
fidence, security and privacy in cloud computing en-
vironments. Furthermore, the authors approach con-
cerns regarding security introduced by the SDN archi-
tecture and how they may compromise cloud services.
Performance Evaluation of Software Defined Network Controllers
365
3 SDN CONTROLLER
PERFORMANCE
COMPARISON MODEL
This section presents the proposed environment for
the performance evaluation experiment developed
with an SDN controller implemented as a virtual net-
work function. We use the SDN Cbench controller
benchmark (Jawaharan et al., 2018), (Laissaoui et al.,
2015) to gauge the performance of the network func-
tion deployed directly on the host and on KVM and
XEN virtual machines to measure performance degra-
dation caused by the virtualization layer.
Xen is an x86 virtual machine monitor based on
a virtualization technique called paravirtualization,
which has been introduced to avoid the drawbacks of
full virtualization by presenting a virtual machine ab-
straction that is similar but not identical to the under-
lying hardware. Xen does not require changes to the
application binary interface, and hence no modifica-
tions are required to guest applications (Armstrong.
and Djemame., 2011). The XEN approach is very
similar to that of KVM as is based considerably on the
works of the XEN developers. XEN supports block
devices through a hypercall that makes use of an al-
tered version of the operating system.
The VNF metrics collected in the experiments
were throughput, response time, CPU consumption
and RAM. Table 1 presents a summary of the perfor-
mance evaluation criteria of the SDN.
Table 1: SDN Performance Metrics (Lai et al., 2019; Li
et al., 2019).
ID Metric
01 Throughput
02 Latency
03 CPU performance
04 Memory Performance
To simulate a real SDN scenario, the experiments
were performed in a peer-to-peer network topol-
ogy such that a traffic generator transmits OpenFlow
streams to a controller hosted on a dedicated server,
as presented in Figure 2.
The client and server are interconnected by a 1
Gbps network link. Floodlight has been set to default
according to your guidelines. Regarding the execu-
tion environment, we have used the Java language.
The client machine has two (2) processors with 12
GB of RAM and 500 GB of the hard disk. The vir-
tual machines have been configured to be identical to
the base operating system (without virtualization) that
housed the physical network function for a better bal-
ance in comparing network functions. For each ex-
periment, several tests were performed.
Figure 2 shows the SDN controller deployment
environment for performing Proposed Model Valida-
tion.
For each experiment, 30 tests were performed
with duration of 500 seconds each, using 100 MAC
addresses per switch emulated so that both the aver-
age performance for each second and the average of
the generated files were stored in text files. This sce-
nario is based on the experiments performed by Zhao
et al (Zhao et al., 2015).
The validation of the proposed model for the ex-
periments involves a set C = {1, 2, 3, 4, 5, 6} of sce-
narios, a set A of test environments that correspond to
the network functions, a set M of variations of each
environment that corresponds to the Cbench (factor)
modes and the S set of switches. Considering the sets
A = {physical function, VNF-KVM, VNF-XEN}, M
= {throughput, latency} and S = {1, 2, 4, 8, 16, 32},
the number of experiments performed to validate the
proposed model must be equal to 36.
In environment 1: physical network function, the
SDN controller must be installed directly over the op-
erating system, simulating the same architecture as a
dedicated hardware device, thus constituting a non-
virtualized architecture that we may call a native sce-
nario (Klosowski and Fiorese, 2019), as can be seen
in Figure 3 (a).
In this scenario, its necessary to remove 8 GB of
RAM from the server for a fair comparison to the vir-
tual network function in the VNF-KVM and VNF-
XEN scenarios, since the virtual machines have been
configured with 8 GB of RAM. All processor cores
were used in this scenario.
In environment 2: VNF-KVM, the network role
must be virtualized on the standard KVM, which uses
full hardware-assisted virtualization, as shown in Fig-
ure 3 (b). The emulator used must be QEMU 2.1.2.
In carrying out the experiments in this scenario, it
was necessary to replace the RAM, totaling 12 GB
in the physical machine because the VNF must have
the same amount of memory as the physical network
function (8 GB), allowing the same configuration of
functions Network.
In environment 3: VNF-XEN, the network func-
tion was virtualized on the XEN 4.4 amd64 hy-
pervisor in the default configuration, namely using
paravirtualization. XEN disables automatically the
KVM hypervisor when configuring XEN at system
boot, preventing KVM from consuming resources and
causing inconsistencies in performance evaluation.
Figure 3 (c) shows the VNF-XEN deployment envi-
ronment.
CLOSER 2020 - 10th International Conference on Cloud Computing and Services Science
366
Figure 2: Controller Deployment Environment.
Figure 3: VNF-XEN Deployment Environment.
3.1 SDN Performance Assessment
Process Mapping
During the performance evaluation of SDN, Cbench
was configured to emulate a specific number of
switches capable of sending streams to the controller
installed on the server for a certain period. At the end
of each performance validation, the data generated by
Cbench was collected.
Simultaneously with the execution of Cbench, the
CPU and RAM data of the physical and virtual net-
work functions were recorded by the server. It is im-
portant to note that when initiating a benchmarking,
it is important to run a script to send commands to
the server, instructing it in its network function (SDN
Controller) host CPU and memory monitoring activ-
ities. This script has stored all the data configured to
evaluate SDN performance.
The developed script should be responsible for
synchronizing the distributed system components of
the SDN benchmarking process and for emptying the
RAM, system buffers, and memory buffers at each
start of the performance monitoring process.
The script should also have the function of writ-
ing controller performance data collected by Cbench
to monitoring files on the server. Files generated dur-
ing the monitoring process should store monitoring
data on the network (controller) performance and per-
formance of physical and virtual network functions
related to system CPU and RAM utilization.
The proposed SDN performance evaluation pro-
cess for the experiments should follow the following
steps, as shown in Figure 4.
1. Define the number of switches: The performance
evaluation must be planned and the switches that
will participate in the experiment must be defined
and configured, according to the proposed sce-
nario and need of the evaluation project;
2. Configure hardware and software features: All
features that will participate in the experiment
must be configured as planned in activity 1 of the
Performance Evaluation of Software Defined Network Controllers
367
process;
3. Define the tools to be used during monitoring, on
both client and server: The tools should be defined
according to plan and criteria adopted during the
experiment;
4. Define performance monitoring metrics: As pre-
sented in Table 1, in the literature there are pro-
posed metrics and they should be adopted and up-
dated, according to the investigated scenario;
5. Define the data collection script: In benchmark-
ing, during controller execution, a data collection
script must be designed and installed on the client
to trigger experiments and data collection.
The script authorized the start of Cbench and the
data monitoring tool that will be adopted in the
experiment. The script must be installed on the
server to send SSH commands to the server in-
structing it in its network function (SDN Con-
troller) CPU and memory monitoring activities.
The script should be responsible for synchroniz-
ing between the distributed system components of
the experiment and for emptying the RAM, sys-
tem buffers, and memory buffers at the start of a
new experiment;
6. Develop SDNs performance reports: At the end
of each experiment, reports should be generated
from the results of the data collection scripts.
3.2 Performance Indicators for SDN
Controllers
While network applications are generally sensitive to
processing, network functions perform intensive net-
work input/output operations. For example, a web
server that receives an HTTP request can send a huge
amount of CPU cycles in processing the request be-
fore sending the response to the client, while a router
basically operates with a large number of forwarding
decisions, using only few CPU cycles.
The SDN controller performs intensive network
operations, with performance being a factor that de-
serves enough attention in the VNF scenario, es-
pecially if we consider the virtualization overheads
that have already been extensively investigated. In
the area of SDN controllers, the two main perfor-
mance metrics widely considered in several scientific
researches are throughput and latency. Throughput
refers to the performance of the network function, that
is, the number of OpenFlow flows that the controller
is able to process per second. Additionally, the la-
tency measures the response time of the controller to
a PACKET IN message received from a switch.
4 DISCUSSION AND THREATS
TO VALIDITY
If any performance metric of an SDN is less than 1, it
indicates that a considerable degradation of the SDN
controller’s performance occurred due to overloading
imposed by the virtualized environment. With para-
virtualization, a better result was achieved, demon-
strating the high cost of emulating network devices in
the host’s user space used during execution.
The function of a physical network engages nearly
all CPU time with the execution of the controller
and the flow processing, with a reasonable number of
executions, explaining the higher performance when
compared to VNFs. The VNF-XEN utilized most
of the processing for the controller, but the process-
ing power engaged in the XEN environment, spe-
cially in privileged domain interactions, is consider-
able and results in performance degradation. When
the controller is too overloaded, the physical func-
tions as well as the virtual machines generate a great
number of interruptions. Then, the processor spends
much time idle while the VNF-XEN consumes more
processing time with the system, but employed CPU
time with flow processing is longer than the remain-
ing functions of SDN processes.
Latency tests usually indicate a considerable
degradation of performance, the the additional soft-
ware layer responsible for virtualization intro-
duced between the hardware’s host and the host’s sys-
tem imply in higher network and data processing
complexities. This results in higher transmission de-
lay between the VNFs and its clients. The physical
function and the virtual machines consume less pro-
cessing time, as they produce more frequent interrupts
per second and leave the CPU idle for a longer inter-
val. Therefore, much of CPU time is used for virtual-
ization and system operations.
VNF-XEN works along with the privileged do-
main and strives to obtain the best possible perfor-
mance, including generating considerably less inter-
ruptions and shorter processor idle time. However,
inter-domain operations diminish CPU performance,
since XEN does not perform direct CPU scheduling
for the virtual machines, once it is not a hypervisor
integrated to the Kernel.
The hypervisor has the task to manage physical
devices of the computer, allowing for multiple in-
stances of virtual computers to be executed simultane-
ously through hardware sharing, thus optimizing re-
source usage. Virtual machine scheduling for CPU
usage is controlled by XEN instead of by the system
Performing a swapping operations during the ex-
periment could lead to lower controller performance
CLOSER 2020 - 10th International Conference on Cloud Computing and Services Science
368
Figure 4: Performance Evaluation Process SDN.
on recording flow information and the state of accom-
plished connections on the hard drive, once disk I/O
operations represent a critical bottleneck in perfor-
mance of virtualized systems, posing a major chal-
lenge for current virtualization attempts.
As with any experiment, the configuration of the
simulation environment may affect the obtained re-
sults, as it is known that, in some situations, the con-
figurations and used resources are prone to unforeseen
events. A way of mitigating this default is to run the
experiment in a number of environments, with differ-
ent configurations.
5 CONCLUSION
This paper presents an experiment using an SDN con-
troller implemented as virtual functions in the KVM
and XEN environments to compare the most critical
performance parameters in a native scenario to verify
the performance of network functions and the innova-
tive processes linked to the performance evaluation of
SDN. A simulation environment was built to test and
possible SDN configurations.
The experiments showed that the virtualization of
an SDN controller in the KVM and XEN environ-
ments did not overload the resources of the proposed
scenario. From the proposed scenario, we defined the
mapping of the performance evaluation process to be
adopted in the experiments.
As future work, the proposed process will be
tested in various scenarios, with experiments to col-
lect metrics to verify if the adopted process is ade-
quate to evaluate the performance of SDN controllers.
These experiments will be important to validate and
evolve the proposed process.
ACKNOWLEDGMENTS
This work was supported by the Brazilian Coor-
dination for the Improvement of Higher Education
Personnel (CAPES), Grants 23038.007604/2014-69
FORTE and 88887.144009/2017-00 PROBRAL; the
Brazilian National Council for Scientific and Techno-
logical Development (CNPq), Grants 303343/2017-6,
312180/2019-5 PQ-2, BRICS2017-591 LargEWiN,
and 465741/2014-2 INCT on Cybersecurity; The
Brasilian Federal District Research Support Founda-
tion (FAP-DF), Grants 0193.001366/2016 UIoT and
0193.001365/2016 SSDDC; the LATITUDE/UnB
Laboratory (Grant 23106.099441/2016-43 SDN); the
Ministry of Economy (TEDs DIPLA 005/2016 and
ENAP 083/2016); the Institutional Security Office
of the Presidency of the Republic (TED ABIN
002/2017); the Administrative Council for Eco-
nomic Defense (TED CADE 08700.000047/2019-
14); and the Federal Attorney General (TED AGU
697.935/2019).
REFERENCES
Ai, J., Guo, Z., Chen, H., and Cheng, G. (2019). Improv-
ing the routing security in software-defined networks.
IEEE Communications Letters, 23(5):838–841.
Alencar, F., Santos, M. A., Santana, M., and Fernandes, S.
(2014). How software aging affects SDN: A view on
the controllers. In GIIS, pages 1–6. IEEE.
Alharbi, T. and Portmann, M. (2019). The (in)security of
virtualization in software defined networks. IEEE Ac-
cess, 7:66584–66594.
Alonso, R. S., Sitt
´
on-Candanedo, I., Rodr
´
ıguez-Gonz
´
alez,
S., Garc
´
ıa,
´
O., and Prieto, J. (2019). A survey on
software-defined networks and edge computing over
iot. In PAAMS (Workshops), volume 1047 of Commu-
Performance Evaluation of Software Defined Network Controllers
369
nications in Computer and Information Science, pages
289–301. Springer.
Armstrong., D. and Djemame., K. (2011). Cultivating cloud
computing - a performance evaluation of virtual image
propagation & i/o paravirtualization. In Proceedings
of the 1st International Conference on Cloud Comput-
ing and Services Science - Volume 1: CLOSER,, pages
539–550. INSTICC, SciTePress.
Bholebawa, I. Z. and Dalal, U. D. (2018). Perfor-
mance analysis of sdn/openflow controllers: POX ver-
sus floodlight. Wireless Personal Communications,
98(2):1679–1699.
Bozakov, Z. (2016). Architectures for virtualization and
performance evaluation in software defined networks.
PhD thesis, University of Hanover, Hannover, Ger-
many.
Cerroni, W., Galis, A., Shiomoto, K., and Zhani, M. F.
(2019). Telecom software, network virtualization, and
software defined networks. IEEE Communications
Magazine, 57(5):88.
Chen, L. (2019). Resource Allocation in Multi-Domain
Wireless Software-Defined Networks. (Allocation de
ressources dans des r
´
eseaux d
´
efinis par logiciels sans-
fil multi-domanes). PhD thesis, INSA Toulouse,
France.
de Jesus, W. P., da Silva, D. A., de Sousa J
´
unior, R. T., and
da Frota, F. V. L. (2014). Analysis of SDN contri-
butions for cloud computing security. In UCC, pages
922–927. IEEE Computer Society.
Din, S., Paul, A., and Rehman, A. (2019). 5g-enabled hi-
erarchical architecture for software-defined intelligent
transportation system. Computer Networks, 150:81–
89.
Erickson, D. (2013). The beacon openflow controller. In
Proceedings of the second ACM SIGCOMM workshop
on Hot topics in software defined networking, pages
13–18. ACM.
Gal
´
an-Jim
´
enez, J., Polverini, M., and Cianfrani, A. (2018).
Reducing the reconfiguration cost of flow tables in
energy-efficient software-defined networks. Com-
puter Communications, 128:95–105.
Hohlfeld, O., Kempf, J., Reisslein, M., Schmid, S., and
Shah, N. (2018). Guest editorial scalability is-
sues and solutions for software defined networks.
IEEE Journal on Selected Areas in Communications,
36(12):2595–2602.
Jawaharan, R., Mohan, P. M., Das, T., and Gurusamy, M.
(2018). Empirical evaluation of SDN controllers using
mininet/wireshark and comparison with cbench. In
ICCCN, pages 1–2. IEEE.
Jung, O., Smith, P., Magin, J., and Reuter, L. (2019).
Anomaly detection in smart grids based on software
defined networks. In Proceedings of the 8th Interna-
tional Conference on Smart Cities and Green ICT Sys-
tems - SMARTGREENS,, pages 157–164. INSTICC,
SciTePress.
Kantor, M., Biernacka, E., Borylo, P., Domzal, J., Ju-
rkiewicz, P., Stypinski, M., and W
´
ojcik, R. (2019). A
survey on multi-layer IP and optical software-defined
networks. Computer Networks, 162.
Karimzadeh, M., Valtulina, L., and Karagiannis, G. (2014).
Applying sdn/openflow in virtualized LTE to sup-
port distributed mobility management (DMM). In
CLOSER, pages 639–644. SciTePress.
Khattak, Z. K., Awais, M., and Iqbal, A. (2014). Perfor-
mance evaluation of opendaylight SDN controller. In
ICPADS, pages 671–676. IEEE Computer Society.
Klosowski, E. A. and Fiorese, A. (2019). Freecontroller:
A framework for relative efficiency evaluation of
software-defined networking controllers. In ICEIS (1),
pages 349–360. SciTePress.
Lai, Y., Ali, A., Hossain, M. S., and Lin, Y. (2019). Perfor-
mance modeling and analysis of TCP and UDP flows
over software defined networks. J. Network and Com-
puter Applications, 130:76–88.
Laissaoui, C., Idboufker, N., Elassali, R., and Baamrani,
K. E. (2015). A measurement of the response times
of various openflow/sdn controllers with cbench. In
AICCSA, pages 1–2. IEEE Computer Society.
Li, F., Xu, X., Yao, H., Wang, J., Jiang, C., and Guo,
S. (2019). Multi-controller resource management for
software-defined wireless networks. IEEE Communi-
cations Letters, 23(3):506–509.
Ochoa-Aday, L., Cervello-Pastor, C., and Fern
´
andez-
Fern
´
andez, A. (2019). etdp: Enhanced topology dis-
covery protocol for software-defined networks. IEEE
Access, 7:23471–23487.
Poncea, O. M., Pistirica, S. A., Moldoveanu, F., and Asavei,
V. (2019). Design and implementation of an openflow
SDN controller in NS-3 discrete-event network simu-
lator. IJHPCN, 14(1):17–29.
Rashma, B. and Poornima, G. (2019). Performance eval-
uation of multi controller software defined network
architecture on mininet. In International Conference
on Remote Engineering and Virtual Instrumentation,
pages 442–455. Springer.
Rasool, R. U., Ashraf, U., Ahmed, K., Wang, H., Rafique,
W., and Anwar, Z. (2019). Cyberpulse: A machine
learning based link flooding attack mitigation sys-
tem for software defined networks. IEEE Access,
7:34885–34899.
Salman, O., Elhajj, I. H., Kayssi, A., and Chehab, A.
(2016). Sdn controllers: A comparative study. In
2016 18th Mediterranean Electrotechnical Confer-
ence (MELECON), pages 1–6. IEEE.
Schwabe, A. (2018). Data-centre traffic optimisation using
software-defined networks. PhD thesis, University of
Paderborn, Germany.
Tootoonchian, A., Gorbunov, S., Ganjali, Y., Casado, M.,
and Sherwood, R. (2012). On controller performance
in software-defined networks. In Hot-ICE. USENIX
Association.
Umair, Z., Qureshi, U. M., Cheng, T. Y., and Jia, X. (2019).
An efficient wireless control plane for software de-
fined networking in data center networks. IEEE Ac-
cess, 7:58158–58167.
Zhao, Y., Iannone, L., and Riguidel, M. (2015). On the
performance of SDN controllers: A reality check. In
NFV-SDN, pages 79–85. IEEE.
CLOSER 2020 - 10th International Conference on Cloud Computing and Services Science
370