Page 77 - ITUJournal Future and evolving technologies Volume 2 (2021), Issue 1

ITU Journal on Future and Evolving Technologies, Volume 2 (2021), Issue 1

Fig. 15 – Impact of soft idle timeout on control‑plane overhead.

packet drop of 0.19% and 0.53% (respectively) before algorithms, namely Silhouette and Gap Statistics
switch‑to‑controller placement balancing, and 0.14% algorithms were applied to optimize the number of
and 0.07% (respectively) after switch‑to‑controller controllers to deploy in a given topology. Given the fact
placement balancing. that network operators are more concerned about the
cost associated with network deployment, this study also
Fig. 15 depicts the impact of tuning the soft idle timeout takes into consideration the trade‑off between the cost
and polling frequency on control‑plane overhead. The of installing a new SDN controller and performance. This
results indicate that, increasing the polling frequency and is necessary to facilitate decision making regarding the
soft idle timeout decreases the number of control packets number of controllers to deploy, based on performance
(synonymous with control‑plane overhead) generated requirements and cost constraints. To determine the
during reactive low instantiation. However, from 20 optimal locations to install the controllers, a classical
seconds forward, the number of control packets remains algorithm called PAM was used. The applied algorithms
constant. Therefore, we can conclude that con iguring are exhaustive making them ideal for static controller
the polling interval and idle timeout to 20 seconds placement with minimal to no time constraints. In
would be the suf icient choice to achieve an acceptable order to mimic a real SDN deployment and also to
control overhead and a potentially lower switch memory verify the outcome from our mathematical model,
utilization. Increasing the timeout beyond 20 seconds we use exhaustive search on an emulator to address
would not change the load on the control channel but will the controller placement problem. This approach
potentially lead to a higher memory utilization. also takes into account resiliency and control‑plane
overhead metrics. We use the ONOS SDN controller
8. SOURCE CODES due to its inherent self‑coordinating distributed core.
Our emulation results show that running a single
The source codes for the proposed solution controller yields high reaction times as some switches
have been made publicly available on Github,
are located too far away from the controller. Moreover,
a world’s leading code repository. The source
running a single controller is not enough to meet
codes can be downloaded from this link:
resiliency requirements. When the number of controllers
https://github.com/Lusani/SDN‑Controller‑Placement
was increased to two, the reaction time was reduced
considerably since the network was subdivided into two
9. CONCLUSION administrative domains. Moreover, the two controllers
worked collaboratively to alleviate control overhead
This study considers determining the number and
location of SDN controllers in a wide area network, and ensure resiliency in the network. Leveraging our
and associated performance and cost implications and controller placement results as well as balancing the
is intended to be used to address the SDN controller switch‑to‑controller placement, we also investigated
placement problem. The work is applied to a national the impact of soft idle timeout and polling frequency on
network from a developing country, SANReN. The work control‑plane overhead. Our indings suggested that a
includes mathematical modelling and a method for large soft idle timeout and polling frequency reduces
obtaining the results through emulation on a popular the overall control‑plane overhead. In reading the above
controller suitable for real world deployments. The conclusions, it should be noted that the solution to the
emulation con irmed the modelling and is also used controller placement problem is topology dependent
to derive important practical limits. The modelling and thus the results presented by this work only apply
included Silhouette, Gap and PAM approaches. Using to the SANReN topology studied. However, the proposed
graph modelling, two ”unsupervised” machine learning approach is “protocol” agnostic and can be adopted

© International Telecommunication Union, 2021 61