You are on page 1of 9

This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2018.2797906, IEEE Internet of
Things Journal

Traffic Load Minimization in Software Defined


Wireless Sensor Networks
Guozhi Li, Songtao Guo, Member, IEEE, Yang Yang and Yuanyuan Yang, Fellow, IEEE

Abstract—The emerging software defined networking (SDN) packets are transmitted from the source to the sink while the
enables the separation of control plane and data plane and saves nodes could sleep at other time to conserve energy.
the resource consumption of the network. Breakthrough in this These aforementioned routing algorithms [7], [8] are based
area has opened up a new dimension to the design of software
defined method in wireless sensor networks (WSNs). However, on the traditional wireless sensor networks. In the scheduling
the limited routing strategy in software defined wireless sensor process, each sensor node needs to transfer its status informa-
networks (SDWSNs) imposes a great challenge in achieving tion to source nodes, and the source nodes calculate the routing
the minimum traffic load. In this paper, we propose a flow pathes based on the node status information. This requirements
splitting optimization (FSO) algorithm for solving the problem may lead to the excessive consumption of network resources
of traffic load minimization (TLM) in SDWSNs by considering
the selection of optimal relay sensor node and the transmission [9]. The recently emerging software defined networking (SDN)
of optimal splitting flow. To this end, we first establish the model [10], [11] mainly separates the control plane and data plane,
of different packet types and describe the TLM problem. We and centralizes the network control functions, opens the in-
then formulate the TLM problem into an optimization problem terfaces between control plane and data plane, and programs
which is constrained by the load of sensor nodes and the packet the network. Based on the programmable SDN controller, the
similarity between different sensor nodes. Afterwards, we present
a Levenberg-Marquardt algorithm for solving the optimization network operators can easily configure new network devices
problem of traffic load. We also provide the convergence analysis and quickly deploy new applications [12]. The SDN can be
of the Levenberg-Marquardt algorithm. Finally, we implement integrated into WSNs, called the software defined wireless
the FSO algorithm in the NS-2 simulator and give extensive sensor networks (SDWSNs), so as to address the traditional
simulation results to verify the efficiency of FSO algorithm in routing problem by decoupling the control plane from the data
SDWSNs.
plane.
Index Terms—Traffic load minimization, software defined Compared to WSNs, the SDWSN architecture relies on one
networking (SDN), flow splitting optimization (FSO) algorithm, or more SDN controllers to task the WSNs and gather the data.
wireless sensor networks (WSNs)
With the strategy of centralized control, the sensors can simply
become data plane elements forwarding and processing data
I. I NTRODUCTION [13]. WSNs are freed from network control tasks by using the
SDN control strategy. On the other hand, armed with a global
Wireless sensor networks (WSNs) consist of a great num- network view, the SDN controller can offer efficient resource
ber of nodes with limited computing, sensing, and wireless allocation and management through central topology control,
communication capabilities [1], [2]. These networks have been scheduling, routing, and network coverage and connectivity
used in a wide area of applications, such as health care, planning [14]–[16].
pollution monitoring, and target tracking systems [3]. The However, the ADA scheme in [7] does not consider the
design of WSNs is constricted by physical resources, hence, global node status information, which may result in the local
designing an efficient routing strategy is a key step towards optimum of routing algorithm. In addition, a lot of packets that
achieving the lowest resource consumption [4]–[6]. the adjacent sensor nodes gather may be similar in practical
There have been some studies on the routing algorithms in wireless sensor networks. The DGRAM algirthm in [8] does
WSNs. An attribute-aware data aggregation (ADA) scheme not take into account the packets similarity among the adjacent
in [7] consists of a packet-driven timing algorithm and a sensor nodes, which will cause sending plenty of redundant
special dynamic routing protocol, which can intentionally packets from source sensor node to sink node [17]. The
make the packets with the same attribute converge as much SDN control strategy has the advantage of global control
as possible in the WSNs with heterogenous sensors or various and flexible control. This motivates us to develop an efficient
applications. The key idea of delay guaranteed routing and routing and scheduling algorithm to achieve the minimum
MAC (DGRAM) algorithm in [8] for delay-sensitive wireless network load in SDWSNs. On the other hand, the existing
sensor network (WSN) applications is to guarantee that the routing algorithms [7], [8] are limited to schedule the packets
from source sensor node to sink node in SDWSNs.
G. Li, S. Guo and Y. Yang are with the Key Laboratory of Networks and In this paper, therefore, we study the problem of traffic load
Cloud Computing Security of Universities in Chongqing, College of Electronic minimization (TLM) in SDWSNs. Our goal is to select optimal
and Information Engineering, Southwest University, Chongqing, 400715, P.
R. China. relay sensor nodes and reduce the transmission of redundant
Y. Yang is with College of Electronic and Information Engineering, packets. Thus our routing and scheduling strategy can save
Southwest University, Chongqing, 400715, P. R. China and the Department network resources to the maximum extent by take advantage
of Electrical & Computer Engineering, Stony Brook University, Stony Brook,
NY 11794, USA. of SDN. For solving the TLM problem, we first formulate the
Corresponding author: Songtao Guo, e-mail:songtao guo@163.com. objective function of TLM problem and the constraints of the

2327-4662 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2018.2797906, IEEE Internet of
Things Journal

load of sensor nodes, and then we analyze the similarities of Đ &ŽƌǁĂƌĚŝŶŐĂƌĞĂ Đ

packets between source sensor node and relay sensor nodes.


Subsequently, we use the Levenberg-Marquardt algorithm to ϭ Ϯ ϯ ϰ
solve the formulated TLM problem. Furthermore, we propose
the flow splitting optimization (FSO) algorithm, which aims to
Ă ϱ ϲ ϳ ϴ ď
find an optimal routing path from source sensor node to sink
node, and ensure the traffic load of SDWSNs is minimum. Fi-
ϵ ϭϬ
nally, simulation results show that our FSO algorithm achieves ϭϭ ϭϮ

less redundant packet transmissions and higher packet delivery


ratio. Đ ^EĐŽŶƚƌŽůůĞƌ Ă ^ŽƵƌĐĞƐĞŶƐŽƌŶŽĚĞ
The contributions of the paper can be summarized as
ď ^ŝŶŬŶŽĚĞ ŝ dŚĞŝͲƚŚƌĞůĂLJƐĞŶƐŽƌŶŽĚĞ
follows.
• We present a traffic model to describe the flows with Fig. 1. Software defined wireless sensor networks (SDWSNs).
the different type of packets in software defined wireless
sensor networks.
• We formulate the optimization objective function for on average. We assume that packet type l belongs to a set of
achieving the minimum network traffic load, and prove packet types L, where L̃ = ∥L∥1 , i.e., l ∈ L. For instance, if
the solution of optimization problem can converge. The the size of packets are 3M and 5M , then the types of packets
optimal objective function is taken as the metric for with size 3M and 5M are corresponding to the types L1 and
selecting the appropriate relay sensor nodes. L2 , respectively. A flow in a sensor node is composed by the
• We propose an FSO algorithm to select the optimal packets of l types, as shown in Fig. 2. We assume that type-l
splitting flows and optimal transmission paths. Our FSO packets are generated according to Poisson process with rate
algorithm not only ensures the effective transmission of λl , and each packet stays in SDWSNs for a period of time
packets but also minimizes the consumption of network with exponential distribution with mean µ1l [19].
resources.
The rest of the paper is organized as follows: Section II /

gives the system model and formulates the objective function 6HQVRUQRGH ͘͘ ͘͘͘ ͘͘͘
of TLM. In Section III, we propose the FSO algorithm. Section
IV extensively evaluates the performance of FSO algorithm.  L O
͘͘͘

͘͘͘

͘͘͘
Finally, Section V concludes this paper.
6HQVRUQRGHQ ͘͘͘ ͘͘͘ ͘͘͘

II. S YSTEM M ODEL AND P ROBLEM F ORMULATION  L O


A. System Model
Fig. 2. The type of packets.
A typical software defined wireless sensor networks (SD-
WSNs) consists of c logically-centralized controllers and a For any packets type l ∈ L, let Rl denote the set of available
set of sensor nodes, V = {v1 , v2 , · · · , vn }, with n = |V |, sensor nodes, then each type-l packet must be processed by
as shown in Fig. 1. The controllers form the control plane at least one sensor nodes from Rl . We assume that Rl is
(l)
and the sensor nodes constitute the data plane of SDWSNs. nonempty for each packet type l ∈ L. Let Yi be the set
Therefore, the network topology from a view of the data plane of type-l packets routed through sensor node i, and Ỹi be
(l)
can be modeled by a graph G = (V, E), where E is the set (l) (l)
the load of type-l packets, such that Ỹi = ∥Yi ∥1 , i ∈ Rl .
of wireless links connecting sensor nodes. A SDN controller We define the traffic load of type-l packets as
manages all the sensors in a cluster. Network slicing can solve
(l)
the problem of co-tier interference and cross-tier interference Ỹ = (Ỹi ; i ∈ Rl , l ∈ L) (1)
[18]. Therefore, we divide the SDWSNs into some slices by (l) ∪ (l)
the sensor’s communication radius r, and the area of slicing We let Xi = Yi denote the set of all types of packets
l∈L
is the range with r × r. We call the slicing as forwarding area. (l)
at sensor node i in SDWSNs, therefore, X̃i indicates the load
(l) (l) ∑ (l)
Note that different sensor nodes may gather the same data of sensor node i, where X̃i = ∥Xi ∥1 = Ỹi . Let Z be
l∈L
in an adjacent environment, therefore, we only need to send the total traffic load of the network. Thus we can get
as less data as possible in source nodes, and other data can be ∑ ∑ (l)
supplemented from relay nodes. Based on the consideration, Z= Ỹi (2)
in the following, we will model the types of packets. i∈Rl l∈L

Each sensor node can transfer the requested packets to sink We define the similarity of packet sets J as the i-
node through relay nodes. The maximum memory space of dentity degree of packets between two sensor nodes, i.e.,
∩ (l)
sensor nodes in SDWSNs is C, and the minimum remaining (l) (l) |Xa(l) Xd | (l) ∩ (l)
J(Xa , Xd ) = (l) ∪ (l) , where |Xa Xd | indicates
memory space of sensor nodes is c′ , the number of packet |Xa Xd |
(l) ∪
types is l′ , l′ = cC′ , i.e., we divide the packet types into l′ parts
(l)
the number of identical packets and |Xa Xd | denotes the

2327-4662 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2018.2797906, IEEE Internet of
Things Journal

TABLE I
N OTATIONS TABLE Proof: The event ϵ can be expressed in terms of events
θij ’s as follows:
Notation Definition ∩ ∩ ∩ ∪ ∪
ω Traffic flow request ϵ = (θ11 θ12 · · · θ1k ) · · ·
n The number of sensor nodes ∩ ∩ ∩ (5)
λ Sending rate (θm1 θm2 · · · θmk )
1
µ
Transmission time
Xi
(l)
The packet set of sensor node i From (5) and by De Morgan’s laws, the event ϵ is given by
(l)
X̃i The load of sensor node i ∏
m ∩ ∩ ∩
r Communication radius J(ϵ) = 1 − J(θi1 θi2 ··· θik )
s The distance from source sensor node to sink node
i=1
J Similarity
Ti The set of sensor nodes in the i-th forwarding area ∏m ∩ ∩ ∩
T′ The set of appropriate relay sensor nodes =1− (1 − J(θi1 θi2 ··· θik ))
i=1 (6)
∏m
≥1− (1 − Jmin
k
)
total number of packets between two sensor nodes. It reflects i=1
the identity degree among packets stored in different sensor = 1 − (1 − Jmin
k
)m
(l) (l)
nodes. The larger the value of J(Xa , Xd ) is, the more the
similarity among packets is. where 1 ≤ m ≤ ⌊ rs ⌋, s denotes the distance from source

n−2
(l) (l)
In addition, each SDN controller connecting to other con- sensor node to sink node, and J(Xa , Xi ) = J(ϵ). This
trollers can freely communicate with each other, and the i=1
completes the proof.
interaction does not affect the transmission of packet on the
Minimizing the objective function can minimize the amount
data plane. We assume that there is at least one available path
of transmitted data in SDWSNs. For an SDWSN, the following
to guarantee that the traffic flow request ωl is forwarded to
constraints describe how to select the optimal path connecting
sink node. Some notations used in this paper are summarized
source sensor node and sink node.
in Table I.
d (l) (l) (l)
X̃ (t) = λl pi (X̃(t)) − µl X̃i (t), (7)
dt i
(l)
B. Problem Formulation pi (X̃(t)) ≥ 0, (8)
∑ (l)
In this part, we will formulate the problem of traffic load pi (X̃(t)) = 1. (9)
minimization (TLM) in SDWSNs, which will be constrained i∈Rl
by both the load of sensor nodes and packets similarity. Our
The constraints (7), (8), and (9), denote the load constraint
goal is to find a routing and scheduling strategy that assigns (l)
of relay sensor nodes. More specifically, X̃i (t) indicates the
a packet to a single path in SDWSNs. The objective function
load status of the relay sensor node i at current time slot,
of the TLM can be given by
and the constraint (7) states that the number of type-l packets
(l)
∑ (l) over relay sensor node i ∈ Rj increases at rate λl pi (X̃(t)),
X̃i (l)
OPT-1: min F( ) (3) and decreases at rate µl X̃i (t). On the other hand, the term
(l)
X̃i L̃ (l)
i∈Rl pi (X̃(t)) of constraint (8) denotes the probability that type-l
packet arrivals at relay sensor node i. Constraint (9) shows that
where L̃ indicates the maximum value of the gathered data in at least one relay sensor node can forward this type-l packet
sensor nodes. We consider polynomial similarity functions as to next relay sensor node, therefore, the constraint ensures that
(l)
the traffic flow request ωl can be transferred to the sink node.
(l) (l)
F (x) = x1+J(Xa ,Xd )
, 0 < J(Xa(l) , Xd ) ≤ 1, (4) The goal of optimization problem (OPT-1) is to find the op-
timal routing path with the least transmitted data in SDWSNs.
It is not difficult to observe that the function F is increasing Theorem 2: As the network transmission time t is growing,
in x and the optimization problem (OPT-1) is strictly convex. the traffic load of type-l packets will converge to ϕl = µλll .
∑ (l)
We assume that each selected relay sensor node belongs to Proof: We set z (l) (t) = X̃i (t)), follows a deter-
a forwarding area, the number of forwarding area is m and the i∈Rl
maximum number of relay sensor nodes is k in a forwarding ministic trajectory described by the following equation,
area. Meanwhile, we define the similarity status of packets d (l)
z (t) = λl − µl z (l) (t), ∀l ∈ L (10)
between source sensor node a and relay sensor node d as the dt
event θ, and the similarity status of all packets in the network which clearly implies that
as the event ϵ. Let Jmin denote the minimum similarity in
SDWSNs. z (l) (t) = ϕl + (z (l) (0) − ϕl )e−µl t , ∀l ∈ L (11)
Theorem 1: In the SDWSNs, the sum of similarities Consequently at steady state,

n−2
(l) (l)
J(Xa , Xi ) is no less than 1 − (1 − Jmin k
)m .
i=1 z (l) (∞) = ϕl , ∀l ∈ L (12)

2327-4662 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2018.2797906, IEEE Internet of
Things Journal

which means that the traffic load of type-l packets will A. Decomposed Unlimited Point Method
converge to the fixed value ϕl . This completes the proof. According to KKT conditions, all the constraints in (19)-
From Theorem 2, we can deduce the following constraint (23) can be satisfied by the y ∗ , and it is the relative extremum
∑ (l) of F (y), if and only if there exist vectors υ ∗ , ψ ∗ and η ∗ ,
X̃i ≤ ϕl , ∀l ∈ L (13)
therefore, we derive the following KKT conditions:
i∈Rl
∇y L(y ∗ , υ ∗ , ψ ∗ , η ∗ ) = 0, (19)
If we use the conventional Lagrange-dual algorithm to solve
∗ ∗ ∗ ∗
the optimization problem OPT-1, the value of L̃ will be ∇υ L(y , υ , ψ , η ) ≤ 0 (20)
calculated in each iteration, which will cause the remarkably ∗ ∗ ∗ ∗
(l) ∇ψ L(y , υ , ψ , η ) = 0 (21)

increasing of computation time. Therefore, we set yl = L̃i
∇η L(y ∗ , υ ∗ , ψ ∗ , η ∗ ) ≤ 0, (22)
to reduce the computation time. Based on the analysis, we ∑ (l)
ψi∗ ( pi (L̃yi ) − 1) = 0, ψ ∗ ≥ 0
(l)
can reformulate optimization problem (OPT-1) into a convex (23)
optimization problem (OPT-2) with respect to y as follows: i∈Rl
∑ Complementary conditions:
OPT-2: min F (yl ) (14) ∑
yl
l∈E υl∗ (1 − J(L̃ya∗ , L̃yi∗ ) − (1 − Jmin )m ) = 0 (24)
i∈Rl
subject to ∑ ∗(l)
∑ ηl∗ ( L̃yi − ϕl ) = 0 (25)
J(L̃ya , L̃yi ) ≥ 1 − (1 − Jmin ) . m
(15) i∈Rl
i∈Rl
From the problem formulation, the interior point method
∑ (l) (IPM) can be implemented based on the primal-dual method
L̃yi ≤ ϕl , ∀l ∈ L, (16)
i∈Rl
[24]. We convert the slack variables from the inequality
constraints into equality constraints by the way of IPM.
∑ (l) (l) ∑
pi (L̃yi ) = 1, (17) 1− J(L̃yd , L̃yi ) − (1 − Jmin )m + si = 0,
i∈Rl i∈Rl (26)
The problem (OPT-2) is a general form of constrained i ∈ {1, 2, · · · , n − 2}
optimization problem, and we will adopt the decomposed un- ∑
L̃yi − ϕl + s′i = 0, i ∈ {1, 2, · · · , n − 2},
(l)
(27)
limited point method (DUPM) [20] and Levenberg-Marquardt
i∈Rl
method (LMM) [21] to determine the optimal value F (y ∗ ). In
this paper, we will take F (y ∗ ) as a metric for the selection of where si ∈ {s1 , s2 , · · · , sn−2 } ≥ 0 and s′i ∈
optimal paths. {s′1 , s′2 , · · · , s′n−2 } ≥ 0, and they are slack variables. We add
the slack variables s and s′ to the inequality constraints in
(26) and (27) by using IPM, then substituting them into (24)
III. F LOW S PLITTING O PTIMIZATION A LGORITHM and (25) yields:
In this section, we first employ the DUPM to transform −νi si = 0, νi ≥ 0, si ≥ 0 (28)
the Karush-Kuhn-Tucker (KKT) condition [22] of optimization
problem (OPT-2) into the equations without any limitation in −ηi s′i = 0, ηi ≥ 0, s′i ≥ 0, (29)
the variable space, then the equivalent equations are solved by Based on these steps, the KKT conditions of equation (18)
the LMM. In the end, we present a novel routing algorithm to can be written as
deal with the TLM problem. ∑ ∑ ∑
∇F (yl∗ ) + υl ∇(1 − J(L̃ya∗ , L̃yi∗ ) − (1 − Jmin )m )
Since the optimization problem (OPT-2) is strictly convex,
l∈E l∈L i∈Rl
in this section, we use Lagrangian dual decomposition ap- ∑ ∑ (l) ∗(l)
proach [23] to solve it. We first relax constraints (15),(16) + ψl ∇( pi (L̃yi ) − 1)
and (17) by introducing Lagrangian multipliers υ, ψ and η, l∈L i∈Rl
∑ ∑ ∗(l)
respectively. Then we can obtain the partial Lagrangian of the + ηl ∇( L̃yi − ϕl ) = 0,
optimization problem. l∈L i∈Rl
∑ ∑
L(x, η, ψ, υ) = F (yl ) 1− J(L̃ya∗ , L̃yi∗ ) − (1 − Jmin )m + s2 = 0,
i∈Rl
l∈L
∑ ∑ ∑ ∗
− ϕl + s′2 = 0,
(l)
+ υl (1 − J(L̃ya , L̃yi ) − (1 − Jmin )m ) L̃yi
i∈Rl
l∈L
∑ ∑
i∈Rl ∑ (l) ∗(l)
+ ψl (
(l) (l)
pi (L̃yi ) − 1) pi (L̃yi ) − 1 = 0,
i∈Rl
l∈L i∈R
∑ ∑l (l) diag{νi 2p }sq = 0,
+ ηl ( L̃yi − ϕl ) ′ ′
l∈L i∈Rl diag{ηi 2p }s′q = 0
(18) (30)

2327-4662 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2018.2797906, IEEE Internet of
Things Journal

It is worth noting that by all tested numerical examples in [25], In the following, we prove the convergence of Levenberg-
when p = p′ = 2 and q = q ′ = 1, the OPT-2 will converge to Marquardt Algorithm.
the optimal solution. Theorem 3: If we let {ϖk } denote the sequence gener-
ated by the Levenberg-Marquardt algorithm, then we have
{dist(ϖk , ϖ∗ )} converge to 0.
B. Levenberg-Marquardt Algorithm
Proof: The following conditions hold by the proofs in
The inequality in (30) can be transformed into the for- [27].
m of equality H(ϖ) = 0 by the way of Levenberg- ∥dk ∥ ≤ c1 dist(ϖk , ϖ∗ ) (34)
Marquardt method [21], and the optimization variable is
ϖ = [y T , υ T , ψ T , η T ]T . Clearly, the constrained optimization ∥H(ϖk + ρπk dk )∥ ≤ c2 dist(ϖk , ϖ∗ )1+δ
problem OPT-2 can be transformed as an unconstrained global
optimization problem. If we let ψ(ϖ) = 12 ∥H(ϖ)∥2 , then the + |ρπk |o(dist(ϖk , ϖ∗ )) (35)
k ∗
OPT-2 can be rewritten as = o(dist(ϖ , ϖ ))
1 where o(dist(ϖk , ϖ∗ )) indicates the infinitesimal of higher
min ψ(ϖ) = ∥H(ϖ)∥2 . (31)
2 order, c1 and c2 are constants, c1 > 0, c2 > 0, δ ≥ 1.
The equation (31) cannot be solved by the equivalent opti- Based on the Levenberg-Marquardt algorithm, we can get
mal problem, therefore, we only use ψ(ϖ) as a merit function ∥ϖk + dk − ϖ∗ ∥ ≤ ∥ϖk − ϖ∗ ∥ + ∥dk ∥
in the line search process. The k-th iteration is formulated as
search direction, thus we have ≤ ∥ϖk − ϖ∗ ∥ + c1 dist(ϖk , ϖ∗ )
(36)
−1
≤ ∥(c1 + 1)ϖk − ϖ∗ ∥
d = −((M ) M + β E)
k k T k k k T
(M ) H . k
(32)
≤ b′
where M ∈ ∂H , β = ∥H ∥ and E is unit matrix. d is
k k k k k
therefore, ϖk+1 ∈ N (ϖ∗ , b), b′ ≤ b, where N (ϖ∗ , b) is the
denoted as the descent direction of H k at y k [21]. neighbourhood of point ϖ∗ .
Based on the Armijo criterion [26], we set ρ ∈ (0, 1), σ ∈ On account of H(ϖk+1 ) = H(ϖk +dk ), there is a constant
(0, 0.5) and step size coefficient αk = ρπk , πk > 0, thus we ζ, thus we have
can get
∥H(ϖk + ρπk dk )∥ ≤ ∥H(ϖk )∥ H(ϖk+1 ) ≤ ∥H(ϖk + ρπk dk )∥ + ζ∥dk ∥2 (37)
(33)
+ σρπk (g k )T dk + σρπk (hk )T dk . Combining the equation (34) with (37), we can get
where similar to [20], g k and hk are the gradient vectors of the dist(ϖk+1 , ϖ∗ ) ≤ ϑH(ϖk+1 )
∑ (l)
merit function, g(y) = ∇( L̃yi − ϕl ) and h(y) = ∇(1 − ≤ ∥ϑH(ϖk + ρπk dk )∥ + ϑζ∥dk ∥2
∑ i∈Rl
≤ ϑo(dist(ϖk , ϖ∗ )) + ϑζc21 dist(ϖk , ϖ∗ )2
J(L̃yd , L̃yi ) − (1 − Jmin )m ). We utilize the Levenberg-
i∈Rl
= o(dist(ϖk , ϖ∗ ))
Marquardt algorithm to solve the equation (30) as given in
(38)
Algorithm 1.
where ϑ is a constant, ϑ > 0, ζ > 0.
Therefore, {dist(ϖk , ϖ∗ )} converge to 0. This completes
Algorithm 1 Levenberg-Marquardt Algorithm the proof.
1: Step 1: Initialization: give ρ ∈ (0, 1), σ ∈ (0, 0.5), initial
point ϖ0 and convergence tolerance ϵ1 > 0, ϵ2 > 0, and
C. Flow Splitting Optimization Algorithm
set k := 0, p = p′ = 2 , q = q ′ = 1.
2: Step 2: Stopping criterion: calculate the gradient vector In the subsection, we describe the flow splitting optimization
g(ϖk ) and h(ϖk ), then deduce the norms ∥g(ϖk )∥ and (FSO) algorithm in detail. The algorithm aims to find an
∥h(ϖk )∥. If criterion values ∥g(ϖk )∥/∥g(ϖ0 )∥ < ϵ1 and optimal routing path from the source sensor node to the sink
∥h(ϖk )∥/∥h(ϖ0 )∥ < ϵ2 , stop and output ϖk as the node, and ensure the traffic load of SDWSNs is minimum.
approximate value of minimum value; otherwise, go to Therefore, the algorithm can guarantee that the SDWSNs
Step 3. consumes as little physical resources as possible.
3: Step 3: Compute the search direction dk by using equation The first step of the FSO algorithm is to establish the
(32). ⌊ rs ⌋ forwarding areas. We assume that s denotes the distance
4: Step 4: Armijo criterion line search for the step size by from source sensor node to sink node, dij is the distance
applying equation (33): let πk := 0 and πk ≤ πmax , then between sensor node i and sensor node j, and r indicates the
π loops from 0 to πmax and get the minimum value of transmission radius of sensor node. Therefore, the number of
πk to compute the step size ρπk . forwarding areas is no more than ⌊ rs ⌋, and the average number
5: Step 5: Update iteration point ϖ k+1 k of relay sensor nodes in each forwarding area can be given
by k = ⌊ rn s ⌋. In this step, we first send the traffic flow request
ϖk+1 = y k + ρπk dk . ωl to the SDN controller, and then the controller collects the
network status information satisfying the request ωl . After that,
Let k := k + 1 and go to Step 2.
we deliver the sensor nodes in the selected forwarding areas

2327-4662 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2018.2797906, IEEE Internet of
Things Journal

to the set Tij , such that there is at least one relay sensor node
for satisfying the ωl in each forwarding area.
The second step is to choose the appropriate relay sensor
nodes which have minimal value of function F (y ∗ ) in the
(l) (l)
set Ti . Function F (y ∗ ) = y ∗ 1+J(Xa ,Xd ) , y ∗ ≤ 1, 0 < Algorithm 2 Flow Splitting Optimization Algorithm
J(Xa , Xd ) ≤ 1, the smaller the value of F (y ∗ ) is, the
(l) (l)
Input: Incoming traffic flow request ωl , the load information
less the network resource consumption is. y ∗ is the optimal Xl in each sensor node, transmission radius r.
node load satisfying the constraints in SDWSNs, and the Output: The paths of minimal traffic load.
function F (y ∗ ) will be reduced as the similarity J(Xa , Xd )
(l) (l)
/*Step 1: determine the forwarding areas*/
increases. 1: Send traffic flow request ωl to SDN controller;
The third step is to determine the forwarding packets at 2: for i = 1 to ⌊ rs ⌋ do
each appropriate relay sensor node. Based on the calculation 3: for j = 1 to k do
⌊ rs ⌋
(l) (l) ∪ 4: if dij ≤ 2r then
of Xi = Xi − XTi+1
′ , the traffic flow request ωl is
i=i+1 5: Select the relay sensor node nij and add it into set
splitted into up to ⌊ rs ⌋ packets. Each relay node only forwards Ti ;
(l)
the received data and the data of Xi . For example, the source 6: end if
l j = j + 1;
sensor node send the data of X1 to the relay sensor node n2 , 7:
and the relay sensor node n2 forwards the data of X1 and 8: i = i + 1;
⌊ rs ⌋
∪ l 9: end for
X2 to the relay sensor node n3 , X2l = X2l − XT ′ . In 10: end for
i+1
i=3
this step, the SDWSN status will be updated after the SDN /*Step 2: select the appropriate relay sensor nodes*/
controller sends the configuration signals to the corresponding 11: Calculate the y ∗ by Algorithm (1);
forwarding nodes. 12: for i = 1 to ⌊ rs ⌋ do
The processes of FSO algorithm can be summarized in 13: for j = 1 to k do
Algorithm 2. 14: Find a relay sensor node in Ti ;
(l) (l)
15: Calculate the Similarity J(Xa , Xd ) between
(l) (l)
source sensor node and relay node, J(Xa , Xd ) =
∩ (l)
|Xa(l) Xd |
D. Time Complexity Analysis (l) ∪ and optimization function F (y ∗ );
(l)
|Xa Xd |
In order to show the efficiency of our FSO algorithm, we 16: if there are multiple relay sensor nodes to be found
analyze its time complexity in this subsection. then
17: j = j + 1;
The first step of FSO algorithm is to determine the for-
18: Compare the F (y ∗ )s between the i-th relay sensor
warding areas, which takes O(k⌊ rs ⌋) time to complete, where
node and the i + 1-th relay sensor node;
k⌊ rs ⌋ ≤ n. In the second step, the FSO algorithm needs
19: Select the relay sensor node in Ti with the mini-
to select the appropriate relay sensor nodes with minimum
mum F (y ∗ );
optimization function F (y ∗ ). The appropriate relay sensor
20: else
nodes need to compare the function F (y ∗ ) of each relay sensor
21: Deliver the relay sensor node to set Ti′ ;
nodes in each forwarding area, and there are ⌊ rs ⌋ forwarding
22: end if
areas in this network, therefore, this step consumes O(k⌊ rs ⌋)
23: i = i + 1;
time. In the end, it takes O(⌊ rs ⌋) time to calculate the sent
24: end for
packets in every sensor node. Thus the FSO algorithm needs
25: end for
to send configure signals and data flows by using O(⌊ rs ⌋+⌊ rs ⌋)
/*Step 3: select the optimal splitting data flows */
time, and take O(⌊ rs ⌋ + ⌊ rs ⌋ + ⌊ rs ⌋) time to complete this step.
26: for i = 1 to ⌊ rs ⌋ do
The time complexity of the FSO algorithm is given by ⌊ rs ⌋
(l) (l) ∪ (l)
s s s s s 27: Xi = Xi − XT ′ ;
O(k⌊ ⌋ + k⌊ ⌋ + ⌊ ⌋ + ⌊ ⌋ + ⌊ ⌋)
i+1
i=i+1
r r r r r 28: i = i + 1;
s s end for
= O(2k⌊ ⌋ + 3⌊ ⌋) (39) 29:
(l) ∪ (l)
r r 30: Send packets Xi−1 Xi to the next relay sensor node
s ′
≤ O(2n + 3⌊ ⌋) Ti+1 ;
r
31: SDN controller delivers configure signals to the corre-
where k ≤ ⌊ s ⌋ . n sponding devices.
r
Note that the complexity of the FSO algorithm is linear
to the number of sensor node n and the communication
distance s, which means that the time complexity is very low.
Therefore, the FSO algorithm is time-efficient for solving the
TLM problem in SDWSNs.

2327-4662 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2018.2797906, IEEE Internet of
Things Journal

TABLE II
PARAMETER S ETTING 100
r=50m
Parameter Description r=100m
Data Rate 250 kbit/s r=150m
The size of physical layer header 8 bytes 80

EPP (%)
The size of MAC-layer header 7 bytes
Transmission Time 192 µs
Timeout 864 µs
Simulation Time 10s 70

IV. S IMULATION R ESULTS 60


0 100 200 300 400
In this section, we evaluate our FSO algorithm using the The number of sensor nodes
NS-2 [28]. Furthermore, we utilize the percentage of effective (a) The different communication radius r.
packets (EPP) to analyze the ratio of splitting flows, and the
packet delivery ratio to obtain the delivery delay distribution 100
Optimization strategy
of packets in SDWSNs. Simple strategy

EPP (%)
A. Simulation Settings
The simulation surveillance field adopts 600m by 600m with 75
100 to 400 sensor nodes, each SDN controller manages 100
nodes, and communication radius r varies from 50m up to
200m. The forwarding field is represented by an r × r 2D
slicing with a spacing of 10m between slicing points. For each 50
set of parameters, the simulation is run for 10 rounds, where, 0 100 200 300 400
The number of sensor nodes
in each round, sensor nodes are deployed randomly in the
sensor field represented by the 2D slicing. The related network (b) Optimization strategy and simple strategy.
parameters are set in Table II.
100
Uniform distribution
Random distribution
B. Comparison of the Percentage of Effective Packets
In this subsection, we first define the percentage of effective
EPP (%)

packets (EPP), then compare the EPP between our FSO 75


algorithm and other baseline algorithms for different number
of sensor nodes.
Definition 1 (Percentage of Effective Packets): The
percentage of effective packets EP P is defined as the
ratio between the successfully transmitted packets and the 50
0 100 200 300 400
gathered packets in SDWSNs, i.e., The number of sensor nodes
∪ ⌊∪
r⌋ ⌊r⌋
s s
(l) (l) ∪ (l)
(c) Uniform distribution and random distribution.
∥X1 ( (Xi − XT ′ ))∥1
i+1
i=1 i=i+1 Fig. 3. The percentage of effective packets (EPP) comparison.
EP P = (40)
⌊ rs ⌋
∪ (l)
∥ Xi ∥1
i=1

where ⌊ rs ⌋
is the number of forwarding areas, Xi indicate
(l) proportional. The simulation result manifests that the commu-
the gathered packets in the i-th sensor node. nication radius r affects the efficiency of the FSO algorithm
We take the EPP as the metric for the forwarding efficiency to a certain extent.
of packets in SDWSNs. The smaller the EPP is, the higher the We define the Algorithm 2 that does not adopt optimization
forwarding efficiency is, vice versa. method as Simple Strategy. Accordingly, Algorithm 2 with
Fig. 3(a) shows that the evolution of the EPPs of SDWSNs optimization method is optimal strategy. We can observe from
as the number of sensor nodes grows. When the communi- Fig. 3(b) that the EPP of optimal strategy is smaller than that of
cation radius r is take as 50m, the percentage of effective simple strategy. This is because the simple strategy does not
(l) (l)
packets (EPP) is highest. This is because the number of consider the packet similarity J(Xa , Xd ) between source
forwarding areas ⌊ rs ⌋ becomes smaller as the communica- sensor node and relay nodes, i.e., the similarity of packets
tion radius increases. Accordingly, the effective relay nodes on the selected relay sensor nodes is very low, therefore,
become relatively less and the forwarded effective packets the simple strategy cannot choose the most appropriate relay
decrease remarkably. Therefore, the percentage of effective sensor nodes. This simulation results further demonstrate the
packets (EPP) and the communication radius r are inversely effectiveness of the FSO algorithm.

2327-4662 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2018.2797906, IEEE Internet of
Things Journal

Furthermore, we consider two kinds of the deployment because the designs of ADA and DGRAM algorithms are
of sensor nodes, i.e., uniform distribution and random dis- based on the traditional wireless sensor networks, thus the
tribution. We can observe from Fig. 3(c) that the EPP by selected relay sensor nodes may be locally optimal. This may
uniform distribution achieves lower EPP value than the random cause the retransmission of packets because of the packet loss.
distribution, but the two EPPs are relatively close. This is In addition, compared to the ADA algorithm, the DGRAM
because the deployment by uniform distribution is easier to algorithm will transfer more redundant packets from source
select relay nodes with higher similarity and has less effective sensor node to relay nodes, which will increase the proba-
packets. In addition, the adjacent relay sensor nodes have bility of packets loss. Therefore, the packet delivery ratio of
very high packet similarity, which implies that the number DGRAM algorithm is lowest.
of relay sensor nodes in a forwarding area is not the main
factor for determining the efficiency of the FSO algorithm. 100

Packet delivery ratio (%)


Therefore, the deployment by random distribution reflects the
less performance of the FSO algorithm. 80

60
C. Comparison of Packet Delivery Ratio
In this subsection, we compare the packet delivery ratio for 40
the different number of sensor nodes and different strategy n=100
over different communication radius. Furthermore, we com- 20 n=200
pare the packet delivery ratio of our algorithm, ADA algorithm n=300
0
[7] and DGRAM algorithm [8]. We first give the definition of 50 100 150 200
packet delivery ratio. Communication radius (m)
Definition 2 (Packet delivery ratio): Packet delivery ratio (a) The different number of sensor nodes.
is defined as the ratio between the first received packets and the
100
sum of first successfully delivered packets and retransmitted Packet delivery ratio (%)
packets in SDWSNs. The higher the packet delivery ratio is,
the higher the algorithm efficiency is, vice versa.
Fig. 4(a) shows the comparison of packet delivery ratio
for different number of sensor nodes. We can find that the 50
packet delivery ratios have jitter as the communication radius
r increases gradually. The smaller the communication radius Random strategy
is, the higher the number of forwarding areas is, therefore, Optimization strategy
each node forwards relatively less packets as the number Simple strategy
of appropriate relay sensor nodes increases. As a result, the 0
50 100 150 200
packet delivery ratio is stable when the communication radius Communication radius (m)
r is small. In addition, as the relay sensor nodes become (b) The different strategy.
more, the packet delivery ratio tends to 1. The reason is
(l) (l)
that the packet similarity J(Xa , Xd ) among different sensor 100
Packet delivery ratio (%)

nodes becomes higher when the number of relay sensor nodes


increases. 80
Furthermore, we set the communication radius r varying
60
from 50m up to 200m. We define our algorithm that does not
use the strategy of splitting flows as Random Strategy. We can 40
observe from Fig. 4(b) that the optimal strategy has the highest
packet delivery ratio as the communication radius r increases ADA
20 DGRAM
gradually. This is because the simple strategy cannot select FSO
appropriate relay sensor nodes, and the random strategy needs 0
50 100 150 200
to forward more packets to relay sensor nodes in SDWSNs. Communication radius (m)
Therefore, the two strategies has lower performance than the
(c) The different routing algorithm.
optimal strategy. In addition, we can also find that the packet
delivery ratio of the simple strategy is higher than that of the Fig. 4. Packet delivery ratio comparison.
random strategy, which shows that the effect of the simple
strategy on the packet delivery ratio is larger than that of the
random strategy in the FSO algorithm. V. C ONCLUSIONS
Fig. 4(c) shows the comparison of packet delivery ratios The main challenge of traffic load minimization (TLM) in
among FSO, ADA and DGRAM algorithms. We can observe software defined wireless sensor networks lies in the selection
from Fig. 4(c) that our FSO algorithm has higher packet of optimal paths and the transmission of less redundant pack-
delivery ratio than other two routing algorithms. This is ets. To handle this challenge, in this paper, we first introduce

2327-4662 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2018.2797906, IEEE Internet of
Things Journal

the concept of similarity to identify various types of packets [12] Q. Zheng, K. Zheng, H. Zhang, and V. C. M. Leung, “Delay-optimal
in different sensor nodes. Afterwards, we formulate the TLM virtualized radio resource scheduling in software-defined vehicular net-
works via stochastic learning,” IEEE Transactions on Vehicular Tech-
problem as the optimization problem which is constrained by nology, vol. 65, no. 10, pp. 7857–7867, 2016.
the load of sensor nodes and the similarity of different packets. [13] X. Li, X. Ge, X. Wang, J. Cheng, and V. C. M. Leung, “Energy efficiency
Furthermore, we employ the Levenberg-Marquardt algorithm optimization: Joint antenna-subcarrier-power allocation in ofdm-dass,”
IEEE Transactions on Wireless Communications, vol. 15, no. 11, pp.
to solve the TLM problem and prove its convergence. We 7470–7483, 2016.
take the optimal objective function as the metric for selecting [14] I. T. Haque and N. Abu-Ghazaleh, “Wireless software defined net-
the appropriate relay sensor nodes. And then we propose the working: A survey and taxonomy,” IEEE Communications Surveys and
Tutorials, vol. 18, no. 4, pp. 2713–2737, 2016.
flow splitting optimization (FSO) algorithm by solving the [15] X. Wang, Z. Sheng, S. Yang, and V. C. M. Leung, “Tag-assisted social-
TLM problem in SDWSNs, which aims to find an optimal aware opportunistic device-to-device sharing for traffic offloading in
routing path from the source sensor node to the sink node, mobile social networks,” IEEE Wireless Communications, vol. 23, no. 4,
pp. 60–67, August 2016.
and ensure the traffic load of SDWSNs is minimized. In the [16] X. Wang, Y. Zhang, V. C. M. Leung, N. Guizani, and T. Jiang, “D2d
end, simulation results verify the performance of our FSO big data: Content deliveries over wireless device-to-device sharing in
algorithm in terms of the percentage of effective packets and realistic large scale mobile networks,” IEEE Wireless Communications,
vol. 25, no. 1, pp. 1–10, January 2018.
packet delivery ratio. [17] X. Wang, M. Chen, T. Taleb, and A. Ksentini, “Cache in the air:
exploiting content caching and delivery techniques for 5g systems,”
ACKNOWLEDGEMENTS IEEE Communications Magazine, vol. 52, no. 2, pp. 131–139, 2014.
[18] H. Zhang, N. Liu, X. Chu, K. Long, A. H. Aghvami, and V. C. M.
This work was supported by the National Natural Science Leung, “Network slicing based 5g and future mobile networks: Mo-
Foundation of China (No. 61772432, 61772433, 61503309), bility, resource management, and challenges,” IEEE Communications
Magazine, vol. 55, no. 8, pp. 138–145, 2017.
Fundamental Research Funds for the Central Universities [19] M. Shafiee and J. Ghaderi, “A simple congestion-aware algorithm for
(XDJK2016A011, XDJK2015C010, XDJK2015D023, XD- load balancing in datacenter networks,” in IEEE INFOCOM 2016 - the
JK2016D047, XDJK 201710635069), Natural Science Key IEEE International Conference on Computer Communications, 2016,
pp. 1–9.
Foundation of Chongqing (cstc2015jcyjBX0094), Natural Sci- [20] J. H. Zheng, T. Y. Ji, M. S. Li, and Q. H. Wu, “Constrained optimization
ence Foundation of Chongqing (CSTC2016JCYJA0449), Chi- applying decomposed unlimited point method based on kkt condition,”
na Postdoctoral Science Foundation (2016M592619) and in Computer Science and Electronic Engineering Conference, 2013, pp.
87–91.
Chongqing Postdoctoral Science Foundation (XM2016002). [21] X. Li, Y. Li, and S. Zhang, “Analysis of probabilistic optimal power
flow taking account of the variation of load power,” IEEE Transactions
R EFERENCES on Power Systems, vol. 23, no. 3, pp. 992–999, 2008.
[22] Boyd, Vandenberghe, and Faybusovich, “Convex optimization,” IEEE
[1] E. Fadel, V. C. Gungor, L. Nassef, S. Almasri, S. Almasri, S. Almasri, Transactions on Automatic Control, vol. 51, no. 11, pp. 1859–1859,
and I. F. Akyildiz, “A survey on wireless sensor networks for smart 2006.
grid,” Computer Communications, vol. 71, no. C, pp. 22–33, 2015. [23] M. Chiang, S. H. Low, A. R. Calderbank, and J. C. Doyle, “Layering
[2] X. Liu, “Atypical hierarchical routing protocols for wireless sensor as optimization decomposition: A mathematical theory of network
networks: A review,” IEEE Sensors Journal, vol. 15, no. 10, pp. 5372– architectures,” Proceedings of the IEEE, vol. 95, no. 1, pp. 255–312,
5383, 2015. 2007.
[3] P. Nematollahi, M. Naghibzadeh, S. Abrishami, and M. H. Yaghmaee, [24] Y. Liu, D. Niu, and B. Li, “Delay-optimized video traffic routing
“Distributed clustering-task scheduling for wireless sensor networks in software-defined interdatacenter networks,” IEEE Transactions on
using dynamic hyper round policy,” IEEE Transactions on Mobile Multimedia, vol. 18, no. 5, pp. 865–878, 2016.
Computing, vol. PP, no. 99, pp. 865–878, 2017. [25] G. Tognola and R. Bacher, “Unlimited point algorithm for opf problem-
[4] S. Guo, C. Wang, and Y. Yang, “Joint mobile data gathering and s,” Power Systems IEEE Transactions on, vol. 14, no. 3, pp. 1046–1054,
energy provisioning in wireless rechargeable sensor networks,” IEEE 1999.
Transactions on Mobile Computing, vol. 13, no. 12, pp. 2836–2852, [26] X. Tong and M. Lin, “Semismooth newton-type algorithms for solving
2014. optimal power flow problems,” in Transmission and Distribution Con-
[5] S. Guo, Y. Shi, Y. Yang, and B. Xiao, “Energy efficiency maximization in ference and Exhibition: Asia and Pacific, 2005 IEEE/PES, 2005, pp.
mobile wireless energy harvesting sensor networks,” IEEE Transactions 1–7.
on Mobile Computing, vol. PP, no. 99, pp. 1–1, Nov. 2017. [27] L. Qi, “Convergence analysis of some algorithms for solving nonsmooth
[6] S. Guo, Y. Yang, and C. Wang, “Dagcm: A concurrent data uploading equations,” Mathematics of Operations Research, vol. 18, no. 1, pp. 227–
framework for mobile data gathering in wireless sensor networks,” IEEE 244, 1993.
Transactions on Mobile Computing, vol. 15, no. 3, pp. 610–626, March [28] T. Issariyakul and E. Hossain, “Introduction to network simulator ns2,”
2016. Springer Berlin, vol. 3, pp. 21–40, 2012.
[7] J. Zhang, F. Ren, T. He, and C. Lin, “Attribute-aware data aggregation
using dynamic routing in wireless sensor networks.” IEEE Transactions
on Parallel and Distributed Systems, vol. 24, no. 5, pp. 881–892, 2013.
[8] A. Sahoo and S. Chilukuri, “Dgram: A delay guaranteed routing and
mac protocol for wireless sensor networks,” in International Symposium
on A World of Wireless, Mobile and Multimedia Networks, 2008, pp.
1–9.
[9] N. Correia, D. Sacramento, and G. Schtz, “Dynamic aggregation and
scheduling in coap/observe-based wireless sensor networks,” IEEE In-
ternet of Things Journal, vol. 3, no. 6, pp. 923–936, 2017.
[10] M. Casado, M. J. Freedman, J. Pettit, J. Luo, N. Mckeown, and
S. Shenker, “Ethane: taking control of the enterprise,” in Conference on
Applications, Technologies, Architectures, and Protocols for Computer
Communications, 2007, pp. 1–12.
[11] N. Mckeown, T. Anderson, H. Balakrishnan, G. Parulkar, L. Peterson,
J. Rexford, S. Shenker, and J. Turner, “Openflow:enabling innovation
in campus networks,” Acm Sigcomm Computer Communication Review,
vol. 38, no. 2, pp. 69–74, 2008.

2327-4662 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

You might also like