Page 64 - ITU Journal Future and evolving technologies Volume 2 (2021), Issue 1

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

where low setup time is the total amount of time taken by 3.2 Contribution
the controller to install a low instruction on the switch’s
low table. The authors argue that dynamic controller From the state‑of‑the‑art review, it is apparent that most
placement is necessary to help reduce low setup time. studies (with the exception of the work by Sallahi et al.
The results from this work reveal that, for low low [20]) assume the number of controllers to be known in
densities, dynamic controller placement can reduce the advance. However, the model proposed by Sallahi et
low setup time by up to 50% in comparison with static al. is ideal to plan a small–scale SDN and runs out of
controller placement. However, for high low densities, memory when solving larger instances. Moreover, most
static controller placement produced better results. studies relied on heuristic algorithms to reduce algorithm
runtime. However, this is achieved at the expense
of solution accuracy. To the best of our knowledge,
the only research studies that implement exhaustive
As demonstrated by Heller et al. [8], Hock et al. [22] and algorithms are by Heller et al. [8] and Tanha et al.
Wendong et al. [21], there exists a signi icant trade‑off [9]. Both Heller et al. and Tanha et al. propose
between load balancing, reliability (also known as the use of k‑center to solve the controller placement
resiliency) and latency. Therefore it is almost impossible problem. However, k‑center is sensitive to outliers and
to optimize one objective without sacri icing the other. does not always consistently yield accurate results [27].
This study attempts to address the controller placement Perhaps more importantly, there is currently no analysis
problem in consideration of switch‑to‑controller latency of the controller placement problem purely using an
metric. This metric has emerged as an important emulation platform to mimic a real SDN deployment.
QoS determinant in SDN. This is primarily because the Most studies relied on mathematical modelling to address
communication between the controller and data‑plane the controller placement problem, making it dif icult to
has to be seamless to ensure an accurate view of the verify validity and reliability of the results.
network state and prompt data‑plane low installations.
Controller placement is a network planning problem,
and is normally not time sensitive. Consequently,
Table 1 provides a summary of the state of the art in this study proposes exhaustive algorithms to optimize
research pertaining to SDN controller placement. solution accuracy. In order to ind the best locations

Table 1 – Classi ication of existing controller placement solutions

Network
Solution Topology(s) Scale of Network Environment Algorithm(s) Placement Metric(s)
Partitioning
average switch–to–controller latency
Heller et al. [8] Internet2 OS3E Large–scale Static k–center No
worst–case latency
Small and
Hu et al.[11] Internet2 OS3E Static l–w greedy Reliability No
medium‑sized
Sprint
ATT NA Capacitated switch–to–controller latency
Tanha et al. [9] Large‑scale Static No
PSINET k‑center Reliability
UUNET
Linear switch–to–controller latency
Yao et al. [14] Internet Zoo Large–scale Dynamic No
relaxation Load balancing
Sparse
Jimenez et al. [15] Medium Large–scale Dynamic k–critical Load balancing Yes
Dense
switch–to–controller latency
Bari et al. [16] RF‑I Large–scale Dynamic DCP‑GK Yes
Load balancing
switch–to–controller latency
Jourjon et al. [17] Not discussed Large–scale Dynamic LiDy+ Yes
Load balancing
inter–controller latency
Sanner et al. [18] Internet2 OS3E Large–scale Dynamic NSGA Yes
load balancing
Random network Non–zero–
Rath et al. [19] small‑scale Dynamic Load balancing No
with 28 switches Sum Game
Random network
Sallahi et al. [20] with 10, 20, 30, 40, 50, small‑scale Dynamic CPLEX Load balancing No
75, 100, 150 switches
switch–to–controller latency
Wendong et al. [21] Internet2 OS3E Large–scale Static l–w greedy No
Reliability
switch–to–controller latency
Hock et al. [22] Internet2 OS3E Small and medium–sized Static POCO Reliability No
Load balancing
switch–to–controller latency
Internet2 OS3E Simulated
Lange et al. [23] Large–scale Dynamic Reliability No
Internet Zoo Annealing
Load balancing
switch–to–controller latency
Ring No speci ic
Ksentini et al.[24] Large–scale Static Inter–controller latency Yes
Binary Tree name
Load balancing
Partition Around Medoids (PAM) average switch–to–controller latency
Gap Statistics worst–case latency
Mamushiane et al.[25] SANReN Small‑scale Static Silhouette Analysis switch‑to‑controller balancing Yes
Johnson’s Algorithm propagation +queuing + processing latency
Emulation signalling overhead
48 © International Telecommunication Union, 2021