Page 63 - ITUJournal Future and evolving technologies Volume 2 (2021), Issue 1

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

failure while considering network deployment costs and However, their approach suffers the same shortcoming as
satisfying switch–to–controller latency. In order to mimic that proposed by Rath et al. in that it does not determine
a production scenario, the authors take into account the the optimal controller placement. Furthermore, both
capacity of the controller and assume a varying switch these research works are limited to small–scale networks.
load. To maintain realism, they assessed their algorithms
on real tier–1 service provider topologies. The outcome
of their experiments demonstrated that controller Wendong et al. [21] study the trade‑off between
resiliency is topology dependent. The drawback of this reliability and latency using random placement, l–w
solution is that it is resource intensive and only ideal for greedy and simulated annealing. The results suggest that
simulated annealing yields the most optimal solution in
small and medium network instances. The algorithm
comparison with the other approaches. The outcome
used in this study is the capacitated k–center algorithm.
of the trade‑off analysis indicate a signi icant trade‑off
The research work of Yao et al. [14] proposes a between reliability and latency. Authors argue that the
heuristic algorithm for capacitated controller placement number of controllers must be chosen carefully. They
in consideration of the switch–to–controller latency demonstrate that using too few controllers has an adverse
and traf ic load of switches. The main objective effect on reliability while using too many controllers
of this work is to optimize controller load balancing can result in a broadcast storm on the east/westbound
under heterogeneous data‑plane load while minimizing interface.
switch–to–controller latency. Resiliency is handled by
deploying additional controllers in the network. The
main shortcoming of this solution is that it is less accurate Hock et al. [22] and Lange et al. [23] advocate for careful
in larger deployments and therefore applicable only for consideration of latency (controller–to–controller) and
small–scale networks. reliability (de ined as resiliency in the event of a node
or link failure and control‑plane load balancing) during
Jimenez et al. [15], also proposes a capacitated controller placement. This work proposes a resilient
controller placement solution to optimize load balancing. Pareto–based Optimal Controller placement framework
Contrary to Yao et al., this work is not limited to to achieve optimal controller placement. The authors
the size of the network and propose a divide and use load imbalance as the key metric, which is the
conquer philosophy to achieve scalability and robustness. difference between the controller having more switch
Moreover, authors assume homogeneous traf ic load on assignments and the controller having fewer switches
the data‑plane. The solutions proposed by Jimenez under its supervision. The results from this work
et al. and Yao et al. optimize controller placement indicate that the optimal solution is achieved when 20%
based on ixed traf ic observed initially, but do not of all network nodes are controllers. The downside
adapt to the changing traf ic load. This shortcoming of this solution is that, instead of partitioning the
is addressed by Bari et al. [16] and Jourjon et al. network into small administrative domains, the authors
[17] who propose a heuristic algorithm for dynamic treat the network as a whole with controllers working
controller placement i.e. controller placement based on
collaboratively. This means the controllers frequently
current data‑plane load. The metrics considered are
share their network state information with their peers
switch–to–controller latency and controller processing
to maintain an accurate global view. This increases
load. The solutions proposed rely on trial and error the probability of incurring a network broadcast storm
and are not as reliable. Sanner et al. [18] propose a which increases inter–controller latency. Therefore, this
genetic algorithm leveraging the Non‑dominated Sorting solution is restricted to small and medium–scale network
Genetic Algorithm (NSGA) II framework to optimize load instances. Furthermore, this solution ignores the average
balancing and inter–controller latency. Authors conclude switch–to–controller latency which is a critical parameter
that their solution consumes a lot of CPU resources and is in SDN.
only ideal for small and medium–sized networks.
Ksentini et al. [24] consider three objectives in optimizing
Rath et al. [19] propose a Non–Zero–Sum game controller placement: (i) switch–to–controller latency,
theory approach to optimize controllers’ utilization. (ii) inter–controller latency and (iii) control‑plane
In this solution, controllers can be added or removed load balancing simultaneously. The authors propose a
dynamically and can also go to sleep mode occasionally bargaining game–based algorithm to optimize controller
based on the traf ic load on the controllers. This solution placement. Authors claim that their results outperform
is intended to reduce network deployment costs (by other mono–objective–based controller placement
minimizing the number of controllers deployed) and results. However, their algorithm is only suitable for
operational costs (by optimizing energy consumption small–scale networks and is less accurate for larger
through on–demand controller deployment). This network instances.
solution ignores controller placement in the network.
Sallahi et al. [20] propose a mathematical formulation Last but certainly not least, He et al. [26] formulate a
to ind the optimum number of controllers to deploy. controller placement model to optimize low setup time,

© International Telecommunication Union, 2021 47