Shiny 2021
Shiny 2021
Computer Networks
journal homepage: www.elsevier.com/locate/comnet
Keywords: The sensor-generated data by Internet of Things are considered to be the most common source of big data.
Energy management A wide range of applications are relying on these data for analytics. While a considerable amount of data is
Software defined wireless sensor network sufficient for the application users to get valuable insights, sending vast amount of data to the cloud seems
(SDWSN)
inappropriate and it only increases the communication cost in the network. It is well-known that an increase
Wireless sensor networks (WSN)
in communication cost increases energy depletion in the network. Since sensor nodes have a restricted power
supply, it is necessary to harness the energy of nodes to prolong the network lifetime. In this paper, a solution
for energy management of sensor nodes is proposed by integrating the software defined framework with the
sensor network, software defined wireless sensor networks (SDWSN), that aids in processing the data inside the
network before transferring it to the sink node. To this context, a game model has been formulated for selecting
the appropriate nodes as reducers which will execute the reducer function. The software defined network
(SDN) controller, geographically placed outside of the wireless sensor network, is responsible for selecting the
reducers and dynamically load reducing function on them. Based on the selection, a routing protocol, routing
via respective reducer (RVRR), that forwards data packets via in-network processing path and control packets
via common path has been proposed. This remarkably reduces the communication cost, thereby prolonging
the lifetime of the deployed network. The RVRR algorithm is implemented in NS-3 simulator to evaluate the
performance of proposed work in SDWSN environment.
1. Introduction processed even before they are transferred to the sink node, the burden
imposed on the tiny sensor nodes will be diminished [4,5]. This leads
The Internet of Things (IoT) is altering everything on the earth and to a great reduction in overall communication cost in the network.
will continue to spark innovations in the future [1]. Today, social and This reduced cost, in terms of communication, in turn reduces network
commercial interactions among humans, between machines and people latency and, hence, paves way for better real-time performance.
produce a continuous flow of data for monitoring and analysis. While
IoT is now emerging as a mandate in every industry, including trans-
portation, healthcare, building, and so on, several barriers exist (the 1.1. Need for SDN
deployment of IPV6, agreement on standards, and power of sensors)
that slows down the growth of IoT [2]. A research conducted by Andy The software-defined networks (SDN) paradigm is fascinating the
Noronha [3] revealed that IoT is not about things — It is all about interest of both the academicians and industrial researchers. It is wor-
data. The sensor network, which is the heart of the IoT, produces an thy of attention as it remarkably simplifies network management and
unpredictable amount of data that becomes a major source to big data. control [6]. As shown in Fig. 1, SDN can help to alleviate many com-
As a result, the organizations are posed with significant challenges in plications of wireless sensor network (WSN). The elements of WSN are
building infrastructure that can handle big data. A report by Cisco
typically energy constrained by nature. Thus, running all the protocols
says that sensors are producing 5 Quintilian data bytes every day.
and functionalities within the tiny sensors makes them prone to early
It is expected to generate 79.4 zettabytes (ZB) of data by the year
failure. Using SDN, the energy-intensive functionalities are pulled out
2025 according to international data corporation forecast. Therefore,
reducing the volume of data before transmitting it to the cloud (in- from the physical sensor node to a logically centralized or distributed
network processing) is found to be essential. Also, when data are SDN controller. The SDN controller plays the role of an operating
∗ Corresponding author.
E-mail address: [email protected] (S.S.G. Shiny).
https://doi.org/10.1016/j.comnet.2021.108094
Received 22 December 2020; Received in revised form 8 March 2021; Accepted 5 April 2021
Available online 10 April 2021
1389-1286/© 2021 Elsevier B.V. All rights reserved.
S.S.G. Shiny et al. Computer Networks 193 (2021) 108094
2. Related works
The rest of this paper is organized as follows. Section 2 reviews the In [20], an energy aware cognitive based SDWSN prototype using
related works on the basis of energy consumption in SDWSN and game reinforced learning was proposed for monitoring systems. The authors
theory in WSN. Section 3 describes the network model along with the state that the proposed protocol decreases the energy consumption
energy consumption model. Reducer selection using game theory and of nodes with guaranteed QoS. Wang et al. [21] had developed an
2
S.S.G. Shiny et al. Computer Networks 193 (2021) 108094
Table 1
Notations.
Notation Description
𝑁 Set of sensor nodes
𝑚 Number of sensor nodes
𝑠𝑖 Identifier of the type of sensor
𝑝𝑖𝑠 Probability that node 𝑛𝑖 has sensor type 𝑠𝑖
𝑟𝑗𝑠 Binary variable to represent whether node 𝑛𝑗 is reducer for 𝑠𝑖
𝜏 Time slot
𝑇 Total number of periods in the game
𝛿𝑖 Strategy chosen by node 𝑛𝑖
𝜁 Strategy profile
𝜔𝑗 Total expected load on node 𝑛𝑗
𝑝𝑖𝑗 Binary variable to represent whether node 𝑛𝑖 is generating value
for 𝑛𝑗
𝑋̃ 𝑛(𝑠𝑖) (𝑡) Load on sensor node 𝑛𝑖 of sensor type 𝑠𝑖 at time 𝑡
𝑖
|𝑋𝑛(𝑠𝑖) | Number of data packets from sensor node 𝑛𝑖
𝑖
Fig. 2. An illustration of multi-tasking in SDWSN. 𝐷𝑖𝑗 Distance between node 𝑛𝑖 and reducer 𝑛𝑗
𝐷𝑗𝑠𝑛 Distance from reducer 𝑛𝑗 to sink node 𝑠𝑛
𝜃 Data packets generated by reducers
𝜙𝑗 Number of packets transmitted from reducers after executing
energy-efficient trust management and a routing mechanism for SD- reducer function
WSN that efficiently gathered the topology information of the network 𝐸𝑑𝑖𝑠 Energy dissipated when the transceiver operates
𝐸𝑎𝑔𝑔 Energy required to perform aggregation
and achieved organized network management in the event of malicious 𝑘 Total number of tasks
forwarding threats. 𝑟 Total number of reducers
Xiang et al. [22], proposed a routing algorithm for multi-task sens- 𝐸𝑓 𝑠 Energy required for amplification in 𝑑 2 power loss model
ing in SDWSNs that mainly focused on energy efficiency. Control nodes 𝐸𝑚𝑝 Energy required for amplification in 𝑑 4 power loss model
were selected in each cluster using PSO algorithm which assigns dif-
ferent sensing tasks to the cluster members dynamically. Each control
node has a direct link to the controller. The common nodes send data
𝑛𝑖+1 are immediate neighbors. It is assumed that, at any instance, the
to the control node as stated by the TDMA scheduler and the control
nodes 𝑛𝑖 and 𝑛𝑖+1 will not be assigned with the same task 𝑠𝑖 (both
node aggregates the data which later is sent to the controller. sensor type and sensor task are used interchangeably throughout the
In [23], a fork and join adaptive particle swarm optimization al- paper and are represented by the notation 𝑠𝑖 ) [25]. The probability
gorithm was presented to increase the network lifetime of SDWSN by that the sensor node 𝑛𝑖 has sensor 𝑠𝑖 is denoted with 𝑝𝑖𝑠 . The data
selecting the optimum number of control nodes that assign tasks to generated by the sensors are transformed into key–value pairs, ⟨𝑠, 𝑣⟩
common nodes in clusters and also route packets to sink nodes. How- prior to forwarding. There will be only one reducer for each sensor
ever, with billions of connected devices interacting with one another, ∑
type, i.e., 𝑚
𝑗=1 𝑟𝑗𝑠 = 1 ∀𝑠. 𝑟𝑗𝑠 is a binary variable that takes the value
routing would become more complex and security would still be a 1 when the node 𝑛𝑗 act as reducer for sensor type 𝑠𝑖 and 0 otherwise.
major concern. This can be represented as follows:
The authors of [24] formulated the minimal energy sensor activa- {
1, if node 𝑛𝑗 reduces sensor type 𝑠𝑖 ,
tion problem here as mixed integer quadratic programming problem 𝑟𝑗𝑠 = (1)
0, otherwise.
both by considering sensor activation and task mapping. They also
proposed an adaptive online algorithm to comply with dynamic events Therefore, nodes that generate data for a particular task, 𝑠𝑖 must
during SDWSN run time. forward the data to their respective reducer. To ensure the precision of
Ever since the birth of WSN, it is considered to be energy constraint sensing for a specific task, minimum sensing rate must be guaranteed.
by nature. It is evident from the above studies that there was indeed a i.e., 𝑛𝑠 ⊂ 𝑁 sensor nodes must run task 𝑠𝑖 at the time slot 𝜏. Each
substantial decrease in the energy consumption issue in WSN with the node 𝑛 ∉ 𝑛𝑠 is treated as relay node for transmitting data packets
emergence of SDWSN. This work also aims to minimize total energy to appropriate reducer. On considering the storage capacity of sensor
consumption by enabling the SDN controller to employ reducers in data nodes, the controller imposes the reducing function on the reducers
plane that reduce the amount of data transferred locally. as soon as they are selected. Thus, computation and communication
can happen simultaneously on sensor nodes as supported by various
3. System model platforms.
Selection of reducers is a key issue that needs special consideration
in SDWSN where the topology changes dynamically. Also, the sensor
3.1. Network model
type, 𝑠𝑖 is activated at a time instance 𝜏 in various nodes, dispersed
throughout the network. Thus, the problem of selecting the reducers
A typical SDWSN architecture which consists of one or more logi-
for each sensor type 𝑠 ∈ 𝑆 is formulated cautiously in Section 4. Some
cally centralized controller(s) and a set of sensor nodes, 𝑁 =
of the notations used in this paper are given in Table 1.
{𝑛1 , 𝑛2 , … , 𝑛𝑚 }, where 𝑚 is the number of sensor nodes is considered.
As per SDN paradigm, the data layer and control layer are distinctly 4. Reducer selection using repeated game theory
separated. More definitely, the data layer consists of sensor nodes
and sink node(s), that are responsible for packet forwarding and in- In this section, the problem of reducer selection is formulated as
network processing. Whereas, the control layer which holds the SDN a repeated game model. This provides a solution to help in-network
controller, in principle, is outside the WSN. The SDN controller does processing in SDWSN. The key–value pairs are forwarded to the respec-
have a detailed knowledge of the network topology which could be tive reducers that perform analytical operation to reduce the amount of
modeled as G = (N, L), N is the set of programmable sensor nodes laid data. The purpose of this section is to select the most suitable node as
on the monitoring region and L is the set of all directed links. reducer dynamically at run-time, which performs reducing operation,
Sensor nodes are mounted with different types of sensors such as such as max, sum, top-k, and so on, for a particular sensor type in the
temperature, pressure, humidity as shown in Fig. 2. Let 𝑠 ∈ [1, 𝑆] be given time-slot [26]. Thereby, the overall network performance could
identifier of the type of sensor. Given a link 𝑙𝑖 ∈ 𝐿, nodes 𝑛𝑖 and be leveraged.
3
S.S.G. Shiny et al. Computer Networks 193 (2021) 108094
4.1. Formulation of game in control plane The objective of the nodes (𝑉𝑖 ) is to find an optimal action from the
action profile in order to maximize the utility function.
The SDN controller has the complete knowledge about the capa-
bilities of sensor nodes in the entire network. Thanks to the SDN 𝑉𝑖 (𝑎𝑖 (𝑡), 𝑎−𝑖 (𝑡)) = 𝑀𝑎𝑥{𝑢𝑖 (𝑎𝑖 (𝑡), 𝑎−𝑖 (𝑡))} (9)
controller which runs the game model and updates the flow tables of If a node decides not to be a reducer, when it actually can, and
the sensor nodes in the data plane, enabling them to forward their data no other nodes also has chosen to be a reducer, then the payoff for
to respective reducers. that node would be zero. The solution for a repeated game model
The repeated game, represented as 𝐺(𝑇 , 𝛽), comprises of game 𝐺 that involves 𝑚 number of sensors is acquired when each sensor node
repeatedly played for 𝑡 + 1 periods commencing from period 𝑡 = 0 to responses best against the other node’s decision. This is called Nash
period 𝑡 = 𝑇 , where 𝑇 denotes the sum of periods in the game and 𝛽
equilibrium.
= [0,1) is the discount factor. Let 𝐺 = (𝑁, (𝐴)𝑖∈𝑁 , (𝑢𝑖 )𝑖∈𝑁 ) is a strategic
game, where 𝑁 = {𝑛1 , 𝑛2 , … , 𝑛𝑚 } be the finite set of sensor nodes, 𝐴
Theorem 1. The strategy profile 𝜁 = (𝛿1 , 𝛿2 , … , 𝛿𝑁 ) is a Nash equilibrium
is the action set for any player 𝑛𝑖 ∈ 𝑁, and 𝑢𝑖 is the payoff of the
for the game 𝐺 = (𝑁, (𝐴)𝑖∈𝑁 , (𝑢𝑖 )𝑖∈𝑁 ) with finite number of nodes and the
player 𝑛𝑖 ∈ 𝑁. To formulate the reducer selection problem, it is must
action profile 𝑎∗ = (𝑎1 , 𝑎2 , … , 𝑎𝑁 ) ∈ 𝐴.
to formulate an utility function appropriate for selecting suitable nodes
as reducers for each sensor type, 𝑠𝑖 . Each component in the game is
Proof. Given the strategy profile 𝜁 = (𝛿1 , 𝛿2 , … , 𝛿𝑁 ) for all 𝑛𝑖 ∈ 𝑁, and
defined below.
𝑎𝑖 ∈ 𝐴, we define,
1. Feasible strategies: Strategies are information available to the
node on the basis of which the desired action of the node is chosen. 𝛤𝑖 , 𝑎𝑖 (𝛿) = 𝑚𝑎𝑥{0, 𝑢𝑖 (𝑎𝑖 , 𝛿−𝑖 ) − 𝑢𝑖 𝛿}. (10)
The strategies considered in this game are residual energy, degree
of the node, and distance to the sink. The strategy of a node 𝑛𝑖 is The proof proceeds by constructing a continuous function 𝑓 ∶ 𝑆 → 𝑆
called feasible (𝜉𝑖 ), only when the energy constraint is faced before a such that each fixed point of 𝑓 is a Nash equilibrium.
particular decision about the action is taken. So, (𝜉𝑖 ) is a set containing Let 𝛿𝑖 (𝑎𝑖 ) represents that the node 𝑛𝑖 will take an action 𝑎𝑖 under
feasible strategies for the node 𝑛𝑖 . Out of all the available feasible strategy 𝛿𝑖 . We define a function 𝑓 ∶ 𝑆 → 𝑆 by 𝑓 (𝑠) = 𝛿 ′ , where:
strategies, the node 𝑛𝑖 chooses a strategy 𝛿𝑖 ∈ 𝜉𝑖 to play the game. A 𝛿𝑖 (𝑎𝑖 ) + 𝛤𝑖 , 𝑎𝑖 (𝛿)
strategy profile 𝜁 = (𝛿1 , 𝛿2 , … , 𝛿𝑁 ) is a set of strategies chosen by each 𝛿𝑖′ (𝑎𝑖 ) = ∑ . (11)
𝑏𝑖 ∈𝐴 𝛿𝑖 (𝑏𝑖 ) + 𝛤𝑖 , 𝑏𝑖 (𝛿)
node in 𝑁. 𝜁 includes only one strategy for every node 𝑛𝑖 .
At any period 𝑡, the history of game ℎ(𝑡) is defined as the set of Now, (11) can be rewritten as:
past actions at all periods before 𝑡. For instance, (2) gives the history 𝛿𝑖 (𝑎𝑖 ) + 𝛤𝑖 , 𝑎𝑖 (𝛿)
of games at time, 𝑡 = 0 which is null and (3) shows the history of actions 𝛿𝑖′ (𝑎𝑖 ) = ∑ . (12)
1 + 𝑏𝑖 ∈𝐴 𝛤𝑖 , 𝑏𝑖 (𝛿)
taken by a node at time period greater than or equal to 1.
Let 𝑓 ∶ 𝑆 → 𝑆 be a continuous function. Then according to
𝑓 𝑜𝑟 𝑡 = 0, ℎ(0) = ∅ (2) Brouwer’s fixed point theorem [27], 𝑓 has a fixed point. If 𝜁 is
assumed as a Nash equilibrium, then all 𝛤 ’s will be 0. This makes 𝜁 a
𝑓 𝑜𝑟 𝑡 ≥ 1, ℎ(𝑡) = {𝑎(0), … , 𝑎(𝑡 − 1)} (3) fixed point of the function 𝑓 . Since 𝛤 ’s are 0, the denominator of (12)
For each node 𝑛𝑖 , the strategy profile is defined as: becomes 1 and it results as:
The strategy profile of all other nodes except node 𝑛𝑖 is: By linearity of expectation, there exist at least one action support
(𝑎′𝑖 ) for 𝛿𝑖 , for which 𝑢𝑖 , 𝑎′𝑖 (𝛿) ≤ 𝑢𝑖 (𝛿). Therefore, the expression given in
𝛿−𝑖 = [(𝛿𝑗 )𝑗∈𝑁 ∀𝑗 ≠ 𝑖]. (5) (13) becomes:
In the given strategy profile 𝜁, if a node 𝑛𝑖 desires to be a reducer, it 𝛿𝑖′ (𝑎𝑖 ) = 𝛿𝑖 (𝑎′𝑖 ). (14)
incurs a cost (𝐶𝑖 ). It is calculated using transport energy cost, caching
cost, and processing cost. Since 𝛿 is a fixed point of 𝑓 , 𝛿𝑖′ (𝑎′𝑖 ) = 𝛿𝑖 (𝑎′𝑖 ). Substituting the same in
(14), we get,
𝐶𝑖 = 𝐸 𝑡𝑥 + 𝐸 𝑟𝑥 + 𝐸 𝑝𝑟 + 𝐸 𝑐𝑎 (6)
𝛿𝑖′ (𝑎𝑖 ) = 𝛿𝑖′ (𝑎′𝑖 ). (15)
2. Action profile: In each iteration of the game, the sensor nodes
must select an action from the action set 𝐴 based on the history of From the definition of 𝛤 , this could only happen when no node in
actions taken by the node at time 𝑡′ and the strategy 𝛿𝑖 . Action profile the game can maximize its utility by moving its strategy. Therefore, we
𝑎∗ = (𝑎1 , 𝑎2 , … , 𝑎𝑁 ) ∈ 𝐴 is the set of actions chosen by nodes in 𝑁, conclude that the strategy profile, 𝜁 is a Nash equilibrium. □
where 𝑎𝑖 is the action proposed by node 𝑛𝑖 . The action chosen by node
𝑛𝑖 at time 𝑡 is given as, 𝑎𝑖 (𝑡) = 𝛿𝑖 (ℎ(𝑡)). 5. Routing via respective reducer protocol
3. Utility function: Given the action profile, 𝑎∗ such that 𝑎∗ =
(𝑎1 , 𝑎2 , … , 𝑎𝑁 ) and the discount factor (𝛽), the utility of the node 𝑛𝑖 at In this section, a solution to optimally route packets to sink node
period t is computed as: so as to minimize the overall communication cost in the network is
∑ discussed. In traditional WSN, the data is aggregated in the sink node
𝑢𝑖 (𝑎𝑖 (𝑡), 𝑎−𝑖 (𝑡)) = 𝛽 𝑡 𝜌𝑖𝑗 𝑟𝑗𝑠 𝜔𝑗 − 𝐶𝑖 , (7)
or the nearest aggregator node or the most convenient aggregator node.
where 𝜔𝑗 is the total expected load on the node 𝑛𝑗 , 𝜌𝑖𝑗 is a binary But in routing via respective reducer (RVRR) protocol, data must be
variable that takes a value of 1 when node 𝑛𝑖 generates a value for node forwarded to the appropriate reducer responsible for the corresponding
𝑛𝑗 . 𝛽 is the discount factor that denotes the patience of the nodes. In sensor type, even though reducers of other sensor types exist near them.
other words, how much a future payoff is valued at the current period.
It is represented as 𝛽(0 < 𝛽 ≤ 1). Higher the discount factor, the greater 5.1. Routing path for RVRR protocol
the patience of the nodes and the higher probability of survival in the
next phase. Note that, if 𝛽 is 0, then it is actually not a repeated game. For the optimal in-network processing of RVRR protocol, it is nec-
{
1, if node 𝑛𝑖 generates value for node 𝑛𝑗 , essary for the nodes to forward their data packets towards their des-
𝜌𝑖𝑗 = (8) ignated reducer as well as the control packets towards the sink node
0, otherwise.
4
S.S.G. Shiny et al. Computer Networks 193 (2021) 108094
5
S.S.G. Shiny et al. Computer Networks 193 (2021) 108094
reduced packets from reducer nodes to sink node is given in (18) as:
Let 𝑝𝑖𝑠 be the probability that node 𝑛𝑖 is generating value for sensor
∑
𝑟 ∑
𝑘
type 𝑠𝑖 . 𝑟𝑗𝑠 is the binary variable that represents whether node 𝑛𝑗 is the 𝐶𝑜𝑠𝑡(𝑗↔𝑠𝑛) = 𝜙𝑗 𝑟𝑗𝑠 𝐷𝑗𝑠𝑛 . (18)
corresponding reducer for task 𝑠𝑖 . 𝐷𝑖𝑗 is the distance in terms of hops 𝑗=1 𝑠=1
between node 𝑛𝑖 and 𝑛𝑗 . Furthermore, the value of 𝐷𝑖𝑗 is calculated
(𝑠 ) The overall communication cost in the network is the sum of
at the controller which is aware of the network topology. While 𝑋̃ 𝑛𝑖 𝑖
communication cost for transporting data packets from the sensor
is the data traffic load from sensor nodes 𝑛𝑖 , 𝜃 is its own data. The
nodes to the reducer and the communication cost of transporting
communication cost of transporting data packets from sensor nodes in
resultant data from reducer to sink node. Therefore, the overall cost
the network towards their designated reducers is given in (17) as:
(𝐶𝑜𝑠𝑡𝑜𝑣𝑒𝑟𝑎𝑙𝑙 ) of transporting the data packets of all sensor types from
∑
𝑚 ∑
𝑟 ∑
𝑘
(𝑠 ) sensor nodes in the entire network is given in (19) as:
𝐶𝑜𝑠𝑡(𝑖↔𝑗) = (𝑋̃ 𝑛𝑖 𝑖 𝑝𝑖𝑠 𝑟𝑗𝑠 𝐷𝑖𝑗 + 𝜃). (17)
𝑖=1 𝑗=1 𝑠=1 ∑
𝑟 ∑
𝑘 ∑
𝑚
(𝑠 )
𝐶𝑜𝑠𝑡(𝑜𝑣𝑒𝑟𝑎𝑙𝑙) = 𝑟𝑗𝑠 [𝜙𝑗 𝐷𝑗𝑠𝑛 + (𝑋̃ 𝑛𝑖 𝑖 𝑝𝑖𝑠 𝐷𝑖𝑗 + 𝜃)]. (19)
Once the reducers get data from all the sensor nodes that generated
𝑗=1 𝑠=1 𝑖=1
value for task 𝑠𝑖 to which it is designated as reducer, it executes its
aggregation function, which is an average function in this work. Let 𝜙𝑗 The overall communication cost in the SDWSN network while using
be the number of packets sent from reducer nodes after executing the RVRR protocol is calculated as in (18). The cost of transporting
reducer function towards the sink node 𝑠𝑛. 𝐷𝑗𝑠𝑛 be the distance between packets from sink node to the controller is negligible. Hence it is not
reducer node and the sink node. The cost incurred for transporting considered for calculation.
6
S.S.G. Shiny et al. Computer Networks 193 (2021) 108094
Table 2 that lies between the transmitter and the receiver. If the distance (d)
Simulation parameters. between the transmitter of sending node and receiver of receiving node
Parameters Values is lesser than the predefined threshold (𝑑𝑡ℎ ), then the free space (𝑑 2
Area 1000 m × 1000 m power loss) model is used; or the multipath fading model (𝑑 4 power
Number of nodes 250 loss) is used. There are 𝑚 nodes distributed on the given network area.
Transmission range 100 m
If there are 𝑘 tasks, then there are on average 𝑚∕𝑘 nodes executing the
MAC protocol IEEE 802.15.4
Data packet length 32 Bytes task 𝑠𝑖 . Each reducer dissipates energy for transmitting and receiving
Initial energy 5 J signals and aggregating the data. Therefore, the energy required for
𝐸𝑑𝑖𝑠 50 nJ/bit the reducers to transmit 𝑏 bits of data is given in (20) as:
𝐸𝑚𝑝 0.0013 pJ/bit/m4
𝐸𝑓 𝑠 10 pJ/bit/m2 𝐸𝑗 = 𝑏𝐸𝑑𝑖𝑠 (𝑚∕𝑘 − 1) + 𝑏𝐸𝑎𝑔𝑔 (𝑚∕𝑘) + 𝑏𝐸𝑑𝑖𝑠 + 𝑏𝐸𝑚𝑝 𝑑 4 , (20)
Data rate 250 kbps
Buffer size of sensor nodes 120 data packets where 𝐸𝑡𝑥 is the energy required for transmission of packets in the form
Simulation time 300 s of bits, 𝐸𝑑𝑖𝑠 is the energy spent when the transceiver circuit operates
and is set to 50 nJ/bit. 𝐸𝑓 𝑠 and 𝐸𝑚𝑝 are the amplifier energy of 𝑑 2
power loss and 𝑑 4 power loss models respectively and depends on an
acceptable bit rate error. The value of 𝐸𝑓 𝑠 is 10 pJ/bit/m2 and the value
6. Performance evaluation of 𝐸𝑚𝑝 is 0.0013 pJ/bit/m4 . 𝐸𝑎𝑔𝑔 is the energy used for aggregation and
is set as 5 nJ/bit/signal. The energy utilized in common nodes is given
In this section, the performance of RVRR protocol is evaluated with in (21) as:
regard to communication cost, network lifetime, energy consumption,
and average end-to-end delay using the network simulator (ns-3). RVRR 𝐸𝑖 = 𝑏𝐸𝑑𝑖𝑠 + 𝑏𝐸𝑓 𝑠 𝑑 2 . (21)
is compared to reducer selection using integer linear programming Since each common node must transmit the data only to its respective
(ILP) method [10], a SDN based framework that leverages SDN-WISE reducer, the energy dissipation follows a free space power loss model.
protocol to employ appropriate nodes as reducers using ILP, SDN- The energy utilized by common nodes are lower than the reducer
WISE [28], a SDN based routing protocol that carries chain based nodes. Note that, nearly every node in the network would have a
in-network processing and R-LEACH [30], a traditional routing protocol chance to become the reducer node at one point in time during the
for WSN which elects cluster head based on residual energy to carry out simulation. This allows the energy to be distributed uniformly through-
in-network processing. out the network. Some of the simulation parameters used for this
experiment are given in Table 2.
6.1. Experimental setup
6.2. Performance metrics
We considered a SDWSN environment, where programmable sensor
Various performance metrics such as communication cost, network
nodes [31] are deployed randomly in 1000 m x 1000 m network area.
lifetime, energy consumption, and end-to-end delay are considered
These sensor nodes adopt the architecture of SDN-WISE protocol and
for the realization of proposed system. A brief explanation about the
its packet types. For it is the sink node that has the interface to be
performance metrics is discussed as follows:
connected to the controller, all sensor nodes must communicate with
Communication cost: The evaluation measures for cost are based on
the controller only through the sink node. The controller is placed
the analysis given in Section 5. The overall communication cost for
outside the WSN and does not suffer from power shortages. When the
transporting data packets to the sink node after carrying out in-network
experiment begins, all sensor nodes send their neighbor table infor- processing is given in (19).
mation to the controller using Report packet, which makes it possible Network lifetime: It is the time period from the start of nodes de-
for the controller to keep account of the locations and the capability ployment to the moment the entire network is deemed non-functional.
of different sensor nodes. The controller uses this information to play For the monitoring applications of SDWSN that requires periodic data
the game in the control layer and to choose the reducers (one for each collection, the network lifetime is defined as, the duration between the
sensor type) as explained in Section 4. Once the reducers are chosen, start of network’s operation and the time the very first node dies.
the necessary rules are sent as a 𝑝𝑎𝑐𝑘𝑒𝑡_𝐼𝑛 message to get installed in Energy consumption in the network: It is the total amount of en-
the flow tables of all the sensor nodes in the network. ergy consumed by the nodes throughout 1000 m x 1000 m region of
A basic energy consumption radio model for sensor nodes and sink space for packets transmission, reception, computation and idle condi-
nodes with respect to path losses [32] is considered for the system tion. The energy spent by reducers and common nodes are calculated
setup. The channel models are decided according to the spacial distance according to (20) and (21) respectively.
7
S.S.G. Shiny et al. Computer Networks 193 (2021) 108094
8
S.S.G. Shiny et al. Computer Networks 193 (2021) 108094
Fig. 11. Comparison of network lifetime. Fig. 12. Impact of payload on end-to-end delay.
when the network scale is small and progressively decreases with because RVRR uses two paths, namely the in-network processing path
larger networks. Because the energy dissipation is evenly distributed for data packets, and the common path for control packets. Thus, the
on the network space in RVRR, the nodes will remain functional for packets need not wait in the input queue for a longer time. Moreover,
a prolonged period of time. Moreover, some of the energy consuming the use of different paths for control and data packets helps to ensure
functionalities such as game model execution and appropriate reducer proper channel utilization. This serves as an added reason for data
selection, are performed in the SDN controller. This greatly conserves packets to reach the destination faster. In case of reducer selection using
the energy of the sensor nodes. It is to be noted that, although the ILP method, non-negligible time is needed for the view update as well
cluster head rotation is enabled in R-LEACH, there are many control as to re-evaluate the suitable reducers. Precisely, in dynamic changing
messages (Advertisement, Join-Request to name a few) to be trans- environment, the complexity is affordable and delay is significant.
ferred within the network, which makes the nodes to drain their energy Thus, RVRR decreases the end-to-end delay by 11.70% over reducer
sooner than in RVRR. On average, RVRR protocol prolongs the lifespan selection using ILP method, 18.53% over SDN-WISE and 23.96% over
of network by 19.45% compared to reducer selection using ILP method, R-LEACH.
39.16% compared to SDN-WISE and 54.60% compared to R-LEACH.
Fig. 12 shows the average delay suffered by the nodes for varying 7. Conclusion
number of payloads in the network. In critical applications of WSN,
delay is a major concern. Minimal delay is achieved only when the In this paper, an energy efficient routing protocol, routing via
load is well balanced across the entire network. The result shows that respective reducers (RVRR) has been introduced for software defined
the packets of RVRR suffers less delay than the other protocols. This is wireless sensor networks (SDWSN) that does in-network processing to
9
S.S.G. Shiny et al. Computer Networks 193 (2021) 108094
significantly reduce the overall communication cost of the network. To [15] T. Behera, S.K. Mohapatra, U.C. Samal, M.S. Khan, M. Daneshmand, A. Gandomi,
this context, repeated game model is proposed to dynamically select Residual energy based cluster-head selection in WSNs for IoT application, IEEE
Internet Things J. 6 (3) (2019) 5132–5139, http://dx.doi.org/10.1109/JIOT.
the appropriate node as reducer responsible for executing the reducer
2019.2897119.
function and send the resultant data to the sink node. Furthermore, [16] Y. Liao, H. Qi, W. Li, Load-balanced clustering algorithm with distributed self-
latency is significantly decreased in the network by directing data organization for wireless sensor networks, IEEE Sens. J. 13 (2013) 1498–1506,
packets, and control packets in different paths. The experimental results http://dx.doi.org/10.1109/JSEN.2012.2227704.
revealed that RVRR performs well in SDWSN environment in terms [17] A. Attiah, M.F. Amjad, M. Chatterjee, C. Zou, An evolutionary routing game for
energy balance in wireless sensor networks, Comput. Netw. 138 (2018) 32–43,
of communication cost, energy distribution, and delays. Precisely, the
http://dx.doi.org/10.1016/j.comnet.2018.03.032.
quality of sensing is guaranteed by the transfer of sufficient data to [18] P. Chithaluru, R. Tiwari, K. Kumar, AREOR: Adaptive ranking based energy
the sink, thereby benefiting the application users. Since, the count of efficient opportunistic routing scheme in wireless sensor network, Comput. Netw.
reducers is limited to one per sensing task, RVRR protocol shows better 162 (2019) 106863, http://dx.doi.org/10.1016/j.comnet.2019.106863.
results for small network size. As part of future work, RVRR protocol [19] W. Zhang, L. Li, G. Han, L. Zhang, E2hrc: An energy-efficient heterogeneous ring
clustering routing protocol for wireless sensor networks, IEEE Access 5 (2017)
is ought to be leveraged for extended network size by selecting more
1702–1713, http://dx.doi.org/10.1109/ACCESS.2017.2666818.
than one reducer per sensing task. [20] R. Huang, X. Chu, J. Zhang, Y.H. Hu, Energy-efficient monitoring in software
defined wireless sensor networks using reinforcement learning: A prototype,
CRediT authorship contribution statement Int. J. Distrib. Sens. Netw. 11 (2015) 360428, http://dx.doi.org/10.1155/2015/
360428.
[21] R. Wang, Z. Zhang, Z. Zhang, Z. Jia, ETMRM: An energy-efficient trust manage-
S. Suja Golden Shiny: Methodology, Software, Validation, Investi-
ment and routing mechanism for SDWSNs, Comput. Netw. 139 (2018) 119–135,
gation, Data curation, Writing - original draft. S. Sathya Priya: Formal http://dx.doi.org/10.1016/j.jnca.2015.05.002.
analysis, Writing - review & editing. K. Murugan: Formal analysis, [22] W. Xiang, N. Wang, Y. Zhou, An energy-efficient routing algorithm for software-
Resources, Conceptualization, Supervision. defined wireless sensor networks, IEEE Sens. J. 16 (2016) 7393–7400, http:
//dx.doi.org/10.1109/JSEN.2016.2585019.
[23] N. Kumar, D.P. Vidyarthi, A green routing algorithm for IoT-enabled software
Declaration of competing interest
defined wireless sensor network, IEEE Sens. J. 18 (2018) 9449–9460, http:
//dx.doi.org/10.1109/JSEN.2018.2869629.
The authors declare that they have no known competing finan- [24] D. Zeng, P. Li, S. Guo, T. Miyazaki, J. Hu, Y. Xiang, Energy minimization in
cial interests or personal relationships that could have appeared to multi-task software-defined sensor networks, IEEE Trans. Comput. 64 (2015)
influence the work reported in this paper. 3128–3139, http://dx.doi.org/10.1109/TC.2015.2389802.
[25] D. Zeng, P. Li, S. Guo, T. Miyazaki, J. Hu, Y. Xiang, A task scheduling algorithm
considering game theory designed for energy management in cloud computing,
References Future Gener. Comput. Syst. 105 (2020) 985–992, http://dx.doi.org/10.1016/j.
future.2017.03.024.
[1] H. Tahaei, F. Afifi, A. Asemi, F. Zaki, N.B. Anuar, The rise of traffic classification [26] E. Fasolo, M. Rossi, J. Widmer, M. Zorzi, In-network aggregation techniques
in IoT networks: A survey, J. Netw. Appl. 154 (2020) 102538, http://dx.doi.org/ for wireless sensor networks: a survey, IEEE Wirel. Commun. 14 (2007) 70–87,
10.1016/j.jnca.2020.102538. http://dx.doi.org/10.1109/MWC.2007.358967.
[2] E. Dave, How the Next Evolution of the Internet Is Changing Everything, White [27] P.V. Subrahmanyam, Brouwer’s fixed-point theorem, in: Elementary Fixed Point
Paper, Cisco, 2011. Theorems, 2018, pp. 219–243.
[3] A. Noronha, R. Moriarty, K. O’Connell, N. Villa, Attaining IoT Value: How To [28] L. Galluccio, S. Milardo, G. Morabito, S. Palazzo, SDN-WISE:. Design, SDN-WISE:
Move from Connecting Things To Capturing Insights, White Paper 154, Cisco, Design prototyping and experimentation of a stateful SDN solution for wireless
2014. sensor networks, in: IEEE Conference on Computer Communications (INFOCOM),
[4] P. Guo, J. Cao, X. Liu, Lossless in-network processing in WSNs for domain- 2015, pp. 513–521, http://dx.doi.org/10.1109/INFOCOM.2015.7218418.
specific monitoring applications, IEEE Trans. Ind. Inf. 13 (2017) 2130–2139, [29] N. Dinh, T. Gu, Y. Kim, Rendezvous cost-aware opportunistic routing in
http://dx.doi.org/10.1109/TII.2017.2691586. heterogeneous duty-cycled wireless sensor networks, IEEE Access 7 (2019)
[5] P. Guo, X. Liu, J. Cao, S. Tang, Lossless in-network processing and its routing 121825–121840, http://dx.doi.org/10.1109/ACCESS.2019.2937252.
design in wireless sensor networks, IEEE Trans. Wirel. Commun. 16 (10) (2017) [30] T.M. Behera, S.K. Mohapatra, U.C. Samal, M.S. Khan, M. Daneshmand, A.H. Gan-
6528–6542, http://dx.doi.org/10.1109/TWC.2017.2724516. domi, Residual energy-based cluster-head selection in WSNs for IoT application,
[6] A. Lara, A. Kolasani, B. Ramamurthy, Network innovation using openflow: A IEEE Internet Things J. 6 (2019) 5132–5139, http://dx.doi.org/10.1109/JIOT.
survey, IEEE Commun. Surv. Tutor. 16 (2014) 493–512, http://dx.doi.org/10. 2019.2897119.
1109/SURV.2013.081313.00105. [31] C. Ding, L. Shen, Design and implementation of programmable nodes in software
[7] H.I. Kobo, A.M. Abu-Mahfouz, G.P. Hancke, A survey on software-defined defined sensor networks, in: IEEE 85th Vehicular Technology Conference (VTC
wireless sensor networks: Challenges and design requirements, IEEE Access 5 Spring), 2017, pp. 1–5, http://dx.doi.org/10.1109/VTCSpring.2017.8108545.
(2017) 1872–1899, http://dx.doi.org/10.1109/ACCESS.2017.2666200. [32] B. Zhu, E. Bedeer, H.H. Nguyen, R. Barton, J. Henry, Improved soft-k-means clus-
[8] H.I. Kobo, A.M. Abu-Mahfouz, G.P. Hancke, SD-WISE: A software-defined wire- tering algorithm for balancing energy consumption in wireless sensor networks,
less sensor network, Comput. Netw. 159 (2) (2019) 84–95, http://dx.doi.org/10. IEEE Internet Things J. 7 (2020) http://dx.doi.org/10.1109/JIOT.2020.3031272.
1016/j.comnet.2019.04.029.
[9] T. Luo, H. Tan, T.Q.S. Quek, Sensor openflow: Enabling software-defined wireless
sensor networks, IEEE Commun. Letters 16 (2012) 1896–1899, http://dx.doi.org/
10.1109/LCOMM.2012.092812.121712.
[10] A.G. Anadiotis, G. Morabito, S. Palazzo, An SDN-assisted framework for optimal S. Suja Golden Shiny received her B.E. degree from Anna
deployment of mapreduce functions in WSNs, IEEE Trans. Mob. Comput. 15 University, Chennai, India, in 2011, and the M.E. degree
(2016) 2165–2178, http://dx.doi.org/10.1109/TMC.2015.2496582. from St. Peter’s University, Chennai, India, in 2014. She is
[11] P. Maheshwari, A.K. Sharma, K. Verma, Energy efficient cluster based routing currently pursuing the Ph.D. degree with the Department of
protocol for WSN using butterfly optimization algorithm and ant colony opti- Computer Science and Engineering, Anna University, Chen-
mization, Ad Hoc Netw. 110 (2021) 102317, http://dx.doi.org/10.1016/j.adhoc. nai, India. Her research interests include wireless sensor
networks, software defined networks and Internet of Things.
2020.102317.
[12] W.B. Heinzelman, A.P. Chandrakasan, H. Balakrishnan, Improved W-LEACH
decentralized protocol with game theory, Procedia Comput. Sci. 175 (2020)
548–553, http://dx.doi.org/10.1016/j.procs.2020.07.078.
[13] D. Lin, Q. Wang, An energy-efficient clustering algorithm combined game
theory and dual-cluster-head mechanism for WSNs, IEEE Access 110 (2019)
49894–49905, http://dx.doi.org/10.1109/ACCESS.2019.2911190.
[14] J. García, M.V. Veelen, In and out of equilibrium i: Evolution of strategies
in repeated games with discounting, J. Econom. Theory 161 (2016) 161–189,
http://dx.doi.org/10.1016/j.jet.2015.11.007.
10
S.S.G. Shiny et al. Computer Networks 193 (2021) 108094
S. Sathya Priya is an Associate Professor in School of K. Murugan completed his Masters in Computer Science,
Computing Sciences, Hindustan Institute of Technology and at National Institute of Technology, Tamil Nadu, India. He
Science, Chennai, India. She received her Ph.D in Informa- received his Ph.D degree from Anna University, Chennai,
tion and Communication Engineering from Anna University, India. He is currently working as Professor at Ramanu-
Chennai, India, M.E. in Computer Science and Engineering jan Computing Centre, Anna University, Chennai, India.
from Annamalai University,Chidambaram, Tamil Nadu, In- He has published and presented papers in highly reputed
dia. She has a professional experience of around 13 years journals and conferences. He is a life member of IETE,
and her area of expertise include IoT, Wireless Networks, ISTE, and CSI. His area of interest includes Wireless Net-
Machine Learning and Blockchain. She has seven Interna- works, MANET Routing, Internet of Things (IoT), SDN and
tional Journal Publications and more than 10 International Cognitive networks.
Conference Publications to her credit.
11