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/ACCESS.2017.2706328, IEEE Access

Towards Green IoT Networking: Performance

Optimization of Network Coding Based
Communication and Reliable Storage
Jian Li, Yun Liu, Zhenjiang Zhang, Jian Ren, and Nan Zhao

Abstract—Internet of things (IoT) is expanding its outreach to for processing the data. Upon these components are various
almost every aspect of our daily life. By utilizing network coding applications such as e-transportation, e-heath, smart home
in Internet of things, the IoT energy consumption can be reduced. and so on. Communication networks such as 4G and 5G
Thus it is worthwhile studying and improving the applications
in Internet of things where network coding is incorporated. In networks [9] interconnect these major components.
this paper, we optimize the performance of network coding based
communication and reliable storage in two important components
of Internet of things, including the IoT core network, where data
Applications: e‐transportation, e‐health, smart home...
is sensed and transmitted, and the distributed cloud storage,
where the data generated by the IoT core network is stored.
First, we propose an adaptive network coding (ANC) scheme Cloud  Distributed 
IoT core network
in the IoT core network to improve the transmission efficiency. computing cloud storage
We demonstrate the efficacy of the scheme and the performance
advantage over existing schemes through simulations. Next we
Communication networks
introduce the optimal storage allocation problem in the network
coding based distributed cloud storage, which aims at searching
for the most reliable allocation that distributes the n data
components into N data centers, given the failure probability p Fig. 1. High level view of Internet of things
of each data center. Then we propose a polynomial-time optimal
storage allocation (OSA) scheme to solve the problem. Both the The IoT core network is responsible for generating the
theoretical analysis and the simulation results show that the data for Internet of things. Smart devices sense various data
storage reliability could be greatly improved by the OSA scheme.
and send out the data through the networks constituted by
these devices [10], [11]. Because the smart devices are mostly
Index Terms—Internet of things, wireless sensor networks, battery-driven, plenty of researches aim at devising energy-
distributed cloud storage, green networking.
efficient schemes to prolong the operation time of the network,
such as in [12], [13], [14]. Moreover, since the communication
I. I NTRODUCTION of the IoT core network is largely through wireless, the packet
loss may be high due to fadings and interferences, which
I NTERNET of things (IoT) [1], [2] is an integral part in
today’s development of smart city. People could remotely
access and interact with a wide range of devices integrated
would bring unnecessary energy consumptions. Thanks to the
emerging of software defined wireless networking [15], we can
with sensors, from home appliances, wearable electronics to apply sophisticated algorithms to improve the communication
environmental monitors. With such enormous coverage poten- quality of the IoT core network and conserve energy.
tial in our daily life, IoT with reduced energy consumption During the operation of Internet of things, data collected
(the ’green’ attribute) has attracted more and more attention. from a vast number of sensors in the IoT core network could
In recent years, energy-efficient networking and computing [3] explode. The distributed cloud storage is the best candidate
have been extensively studied from many perspectives, such to safely and reliably store these data. The distributed data
as the framework design [4], the algorithm design [5] and the storage architecture model distributes the database to multiple
resource reusing [6]. servers in many locations across the participating network in
The high level view of Internet of things is shown in the storage cloud. Each location is directly and independently
Fig. 1, which includes IoT core network for data sensing plugged into the Internet. If something unexpected happens
and transmission, distributed cloud storage [7] for storing to the data in one location, generally only a small amount of
the data generated by the core network, cloud computing [8] backed up data is impacted. Thus the data could be recovered
with much less energy consumption, using the data stored in
Jian Li, Yun Liu, Zhenjiang Zhang are with the School of Electronic and the rest of the locations. Besides distributing the data, many
Information Engineering, Beijing Jiaotong University, Beijing 100044, China.
Email: {lijian,liuyun,zhjzhang1}@bjtu.edu.cn researches also study proactive approaches to ensure the data
Jian Ren is with the Department of Electrical & Computer Engineering, availability such as in [16].
Michigan State University, East Lansing, MI 48824, United States. Email: Network coding provides a trade-off between communica-
[email protected]
Nan Zhao is with the School of Telecommunications Engineering, Xidian tion capacity and computational complexity in network envi-
University, Xi’an, Shanxi 710071, China. Email: [email protected] ronment by enabling the intermediate relay nodes to encode

2169-3536 (c) 2016 IEEE. Translations and content mining are permitted for academic research only. Personal use is also 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/ACCESS.2017.2706328, IEEE Access

the incoming packets before forwarding them. Since network Thanks to the advantages of SDWN, it is much easier
coding can improve the throughput and robustness of the to implement algorithms which can improve network perfor-
network, the unnecessary energy consumption due to high mance into IoT core network. In [18] the authors propose
loss rate communication or high failure rate storage can be to combine network coding and software defined networking,
saved. Application of network coding in Internet of things can where the code rate of the network coding is fixed. Although
contribute to the ‘green IoT networking’. With such desirable this approach can improve the communication throughput, the
merit, in this paper we optimize the performance of network strategy is not flexible to cope with the changing channel
coding based communication and reliable storage in Internet qualities in wireless environments. In this paper, we will show
of things. The main contributions of this paper are: that the transmission efficiency of the IoT core network can
• We propose an adaptive network coding scheme (ANC be greatly improved through our adaptive network coding
scheme) for the IoT core network and demonstrate that scheme, where the code rate of the network coding can be
the scheme can improve the transmission efficiency and dynamically adjusted in a centralized manner with the global
the performance is better than existing schemes. view of the whole network.
• For the distributed cloud storage utilizing network coding 2) Distributed Cloud Storage: The volume of all global
that stores the data generated by the IoT core network, data will be boosted dramatically with the developing of
we introduce the optimal storage allocation problem and Internet of things, where there will be hundreds of thousands
propose an optimal storage allocation (OSA) scheme. of sensors deployed to create more and more data. To the year
Simulation results show that the storage reliability can 2020, the amount of data would grow to 40 zettabytes. How
be greatly improved. to properly store the data has become a major challenge in
Internet of things.
The paper is organized as follows: in Section II we briefly
Traditional centralized data center is not suitable for the
review the IoT core network and the distributed cloud storage.
context of Internet of things. If something unexpected happens
The concept of network coding and its advantages in commu-
such as power outage or military actions, the precious data
nication and storage are also introduced in Section II. Next
stored in the data center could be lost and unrecoverable. To
we propose and analyze our adaptive network coding scheme
ensure a high reliability of the data storage, a typical solution is
for the IoT core network in Section III. After that we study
to store the data across multiple servers in the distributed cloud
the optimal storage allocation problem in the distributed cloud
storage. The main idea is that instead of storing the entire data
storage utilizing network coding in Section IV. At last is the
in one server, we can split the data into n data components
Conclusion.
and store the components separately. The original data can
be recovered only when the required (threshold) number of
II. P RELIMINARIES AND R ELATED W ORK components, say k, are collected. The storage efficiency is
much higher than simply replicating the data over multiple
A. Internet of Things
servers. The distributed cloud storage can also increase data
The objective of Internet of things is to equip everything availability while reducing network congestion, thus leading
related to human beings with smart chips integrating sensors, to increased resiliency. A popular approach is to employ an
actuators and transceivers. Smart devices equipped with smart (n, k) maximum distance separable (MDS) code, such as the
chips within a certain range can communicate with each other Reed-Solomon (RS) code in the Total Recall system [19]. In
and form networks. These networks can be further connected later sections, we will show that by applying network coding
to the Internet through proper interconnecting. In this paper, in the distributed cloud storage we can further improve the
we will mainly focus on the IoT core network and the performance of data storage.
distributed cloud storage which stores the data generated by
the IoT core network as shown in Fig. 1.
1) IoT Core Network: IoT core network consists of the B. Network Coding
smart devices mentioned above and the networks among these In this section, we will briefly introduce the concept of
devices. Several challenges of the IoT core network exists. network coding and its advantages in improving the communi-
First, it lacks of a unified infrastructure and protocol stack. cation throughput and the distributed cloud storage reliability,
Second, the monitor and control of the network lacks of which could eventually contribute to the ‘green IoT network-
flexibility. Third, the functionality of the network cannot be ing’. Network coding was first introduced in the seminal paper
changed without reprogramming the smart devices when the by Ahlswede et al. in [20]. By allowing the intermediate relay
application environment changes. nodes to encode the incoming packets, the network could
To overcome these shortcomings, software defined wireless achieve the maximum multicast capacity.
networking (SDWN) [15] was proposed based on the paradigm A network is equivalent to a directed graph G = (V, E),
of software defined networking (SDN) [17]. In the context where V represents the set of vertices corresponding to the
of SDWN, the network elements in the data plane are smart network nodes and E represents all the directed edges between
devices which act as both end users and switches. The data vertices corresponding to the communication link. The start
flow is separated from the control flow. We can easily change vertex v of an edge e is called the tail of e and written as
the network behaviors through exchange of the control flow v = tail (e), while the end vertex u of an edge e is called the
among smart devices. head of of e and written as u = head(e). For a source node

2169-3536 (c) 2016 IEEE. Translations and content mining are permitted for academic research only. Personal use is also 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/ACCESS.2017.2706328, IEEE Access

Data Center 1 Data Center 2
x1 Sink
node Node1 Node2 Node3 Node4
R1 u1+3u2 2u1+u2 5u1+2u2 u1+4u2
x1 x1 x1+x2
Source u3+3u4 2u3+u4 5u3+2u4 u3+4u4
node
R3 Data Center 1 Data Center 2
x2 x2 x1+x2
Sink
x2
node
R2 u1+3u2+u3+3u4 New Node3 u1+4u2+u3+4u4
3u1+4u2+3u3+4u4
Fig. 2. A simple example of network coding 2u1+u2+2u3+u4
2u1+7u2+2u3+7u4

u, there is a set of symbols X(u) = (x 1, . . . , x k ) to be sent. Fig. 3. network coding based distributed cloud storage
Each of the symbol is from the finite field GF (2m ), where m
is a positive integer. For a link e between intermediate nodes
r 1 and r 2 , written as e = (r 1, r 2 ), the symbol ye transmitted file using the same (n, k) code to recover the encoded part
on it is the function of all the ye0 such that head(e 0 ) = r 1 . of the file stored in the failed node. This approach is a waste
And ye can be written as: of bandwidth because the whole file has to be downloaded to
X recover a fraction of it.
ye = βe0,e · ye0, (1)
To overcome this drawback, Dimakis et al. [25] introduced
e0 :head(e0 )=r1
the conception of (n, k, d, α, β, B) regenerating code based on
in which the encoding coefficients βe0,e ∈ GF (2m ). For a sink the network coding. In the context of regenerating code, the
node v, there is a set of incoming symbols ye0 (e 0 : tail (e 0 ) = contents stored in a failed node can be regenerated by the
v) to be decoded. replacement node through downloading β help symbols from
1) Network Coding in Communication: The main idea of each of d helper nodes. This regeneration is identical to the
network coding can be illustrated through Fig. 2. Assume the encoding process of the intermediate nodes in network coding.
capacity of all the edges is C, the capacity of this network The bandwidth consumption for the failed node regeneration
is 2C according to the max-flow min-cut theorem. Only by could be far less than the whole file. Thus the energy con-
encoding the incoming packet symbols x 1, x 2 at node R3, this sumption in data regeneration could be greatly reduced.
network can achieve the maximum capacity. Since node R3 In [25], a tradeoff between the regeneration bandwidth
only need to send out one symbol x 1 + x 2 instead of two γ = d β and the storage requirement α was derived based on
symbols, the energy consumption of R3 in data transmission network coding theory and two extreme points were found:
could be reduced by 50%. minimum storage regeneration (MSR) point in which the
In [21], [22] the authors have shown that linear codes storage parameter α is minimized:
with random selected coefficients are sufficient to achieve the !
B Bd
multicast capacity by coding on a large enough field. Sink (α M SR, γ M SR ) = , , (2)
nodes that have received more linear independent encoded k k (d − k + 1)
symbols than the original symbol generated by the source and minimum bandwidth regeneration (MBR) point in which
nodes can easily decode the original symbols by solving a set the bandwidth γ is minimized:
of linear equations. Moreover, it has been demonstrated that !
2Bd 2Bd
network coding can improve the communication throughput. (α M BR, γ M BR ) = , . (3)
As an example, the authors in [23] have applied the principles 2kd − k 2 + k 2kd − k 2 + k
of random network coding to the context of peer-to-peer (P2P) Fig. 3 is an illustrative example of regenerating code with
content distribution, and have shown that file downloading parameters n = 4, k = 2, d = 3, α = 2, β = 1, B = 4. 4
times can be reduced. Thus in this paper we propose to apply symbols u1, u2, u3, u4 are stored in 4 storage nodes, and can
adaptive random linear network coding in the IoT core network be retrieved from any 2 of the storage nodes. A failed node
to improve the network transmission efficiency. In [24], the can be regenerated by downloading 1 symbol each from the
authors develop an OpenCoding protocol to improve the 3 remaining nodes. Here we suppose node 3 fails. For the
network throughput through intra-flow network coding for storage systems simply employing RS code, 4 symbols have
wireless mesh networks. to be downloaded first to decode the original symbols. Then
2) Distributed Cloud Storage Utilizing Network Coding: we have to encode the 4 decoded symbols again to regenerate
When a storage node in the distributed cloud storage network the symbols in the failed node 3. So the bandwidth needed
that employing (n, k) RS code (such as Total Recall [19]) fails, for repairing the failed node 3 is 4. For the regenerating code
the replacement node connects to k nodes and downloads the solution in Fig. 3, by linearly combing the 3 downloaded
data of the same amount as the whole file first to decode the symbols u1 + 3u2 + u3 + 3u4 , 2u1 + u2 + 2u3 + u4 and
original file. Then the replacement node encodes the original u1 + 4u2 + u3 + 4u4 into 2 symbols 3u1 + 4u2 + 3u3 + 4u4 and

2169-3536 (c) 2016 IEEE. Translations and content mining are permitted for academic research only. Personal use is also 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/ACCESS.2017.2706328, IEEE Access

2u1 + 7u2 + 2u3 + 7u4 , we can regenerate a new node 3 that

has the same function as the failed node 3. In the repairing
process, only 3 symbols need to be downloaded. Thus the End
repair bandwidth is saved by 25%. In the later section, we Internet Monitor/Control users

will introduce a storage allocation problem for regenerating

code in the distributed cloud storage and propose an optimal
storage allocation scheme that can achieve the highest possible data plane
reliability.
Smart 
devices

III. A DAPTIVE N ETWORK C ODING IN THE I OT CORE

NETWORK

In this section, we will show our adaptive network coding

(ANC) scheme in the IoT core network. The size of the data to Report packets  Adjust  network 
be transmitted in the IoT core network may be larger than the receiving  coding 
size limit of a single packet, such as new firmwares to update status strategies
the smart devices on-air. So the data needs to be divided into
data fragments first then transmitted in multiple packets with SWDN
Data path
Controller
one data fragment per packet. A node has to correctly receive Control path
enough linearly independent packets to reassemble the original control plane
data.
software defined IoT core network

A. Limitations of existing works

Fig. 4. ANC Scheme for the IoT core network
Since the communication between smart devices are through
wireless channel and there may be various fadings and inter-
ferences in the channel, some of the nodes may experience B. ANC Scheme for the IoT core network
packets loss in the communication. When network coding is To overcome the limitations mentioned above, we propose
not utilized, retransmission is a common method to mitigate an adaptive network coding (ANC) scheme to further improve
the packets loss. In some cases, certain packets may get lost the transmission efficiency of the IoT core network with
most of the time so these packets have to be retransmitted SDWN, as illustrated in Fig. 4. In this example, end users
many times until they are correctly received. Thus the overall could communicate with the IoT core network with SDWN
transmission efficiency will be low. Here the transmission through Internet for monitoring/control purposes. The data
efficiency is defined as the ratio between the minimum number transmission inside the IoT core network could benefit from
of the packets needed to reassemble the original data and the our ANC scheme. In the figure, the IoT core network formed
number of total packets transmitted from the source node and by the smart devices could be the smart appliance network
the intermediate nodes. When network coding is utilized, a at home, the surveillance camera network on streets or the
node can retrieve the original data as long as the node can emission detecting network in factories, etc. Here we only
correctly receive enough number of packets. The entire trans- include the data plane and control plane of the SWDN to
mission of the data will not be affected by lacking of certain show the main idea of the ANC scheme.
particular packets. So the overall transmission efficiency will For the data plane, the source node will send out linear
be higher. combinations of the original packets. Each intermediate node
However, there are still limitations for simply applying the will perform random linear network coding. The incoming
network coding in the IoT core network, where the number of packets will be linearly combined using random coefficients
encoded packets to be generated and sent in the intermediate then sent out to succeeding nodes. The code rate r of the
nodes is predetermined [18]. If too few packets are generated, network coding is defined as the ratio of the number of
the sink node may not even be able to collect enough packets encoded packets to the number of incoming packets. And the
to decode the original data. If too many packets are generated, code rates of the network coding will be automatically adjusted
the transmission efficiency will be low. Moreover, the fact that by the SWDN controller mentioned below. The sink nodes will
the quality of wireless channel is changing over time makes the decode the original packets after receiving enough number of
situation even worse. As an example, when the channel quality linearly independent packets.
becomes better and the packet loss rate goes lower, some Meanwhile, for the control plane, the smart devices will
of the encoded packets will be useless and the transmission report packets receiving statistics to the SDWN controller
efficiency could be higher. The encoding strategy should be periodically through the control path. Based on the information
able to dynamically adjusted according to the transmission reported, the SDWN controller will dynamically adjust the
conditions. network coding strategies to eliminate unnecessary transmis-

2169-3536 (c) 2016 IEEE. Translations and content mining are permitted for academic research only. Personal use is also 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/ACCESS.2017.2706328, IEEE Access

2) Intermediate Node Algorithm of the ANC scheme:

δi,1 δi,2 …... δi,n data symbols For each coding group, the intermediate node will open a
encoding vector Δi packet data hi/g i receiving buffer to store the incoming fresh packets from the
nodes designated by the SDWN controller for encoding. The
intermediate node will also record all the encoding vectors
Fig. 5. ANC packet format
received in the incoming packets. A packet is called a fresh
packet if its encoding vector is linearly independent from all
sions. If the packet loss becomes higher around some node, of the previously received packets’. In order to get a trade-
more encoded packets will be generated in the corresponding off between the packet diversity and communication delay,
intermediate nodes. If the packet loss becomes lower, the the intermediate node will encode the incoming fresh packets
number of encoded packets will be decreased. Since the received during a preset interval τ which is measured by a
SDWN controller has the global information of the network, timer then clear the receiving buffer and wait for the next
this centralized control will be more effective. incoming fresh packet to restart the timer and the buffering. At
1) Source Node Algorithm of the ANC scheme: In the the end of each time interval, the encoding of the fresh packets
source node, the data to be transmitted will be fragmented in the receiving buffer is performed. For better illustration,
into data packets with equal length. Every n data packets we can split each of the nτ fresh packets in the receiving
will form a coding group, in which random linear network buffer into the encoding vector ∆i and data gi (1 ≤ i ≤ nτ ).
coding will be performed. For the purpose of clarity, in the nτ is the number of fresh packets in the receiving buffer.
paper we assume that there is only one coding group. For each The intermediate node will send out drnτ e encoded packets
packet hi in the coding group, there will be an encoding vector using Algorithm 2, where r is the code rate determined by the
∆i = [δi,1, δi,2, . . . , δi,n ] (δi, j ∈ GF (2m ), 1 ≤ i, j ≤ n) attached SWDN controller. At the same time, the intermediate node
in front to indicate which packets participate in the encoding will report the receiving and the sending of the packets to the
of hi . GF (2m ) denotes the finite field with 2m elements where SWDN controller.
m ∈ {8, 16, 32, 64, . . . } is determined by the symbol size. For
an uncoded packet hi , the elements in the encoding vector will Algorithm 2 ANC scheme - intermediate node
be all-zero except δi,i = 1. The packet format is illustrated in . the network coding code rate r is determined/updated by
Figure 5. The source node will perform Algorithm 1 to send the SWDN controller
out the encoded packets gi (1 ≤ i ≤ drne) where r is the for i = 1 → drnτ e do
code rate determined by the SWDN controller and drne is the if i ≤ nτ then
ceiling operation to get the smallest integer that is larger than repeat generate a random vector Ri = [r i,1, r i,2,
or equal to rn. . . . , r i,nτ ]
until Ri is linearly independent from all the R j ,
Algorithm 1 ANC scheme - source node 1 ≤ j < i (except for i = 1)
. the network coding code rate r is determined/updated by else
the SWDN controller generate a random vector Ri = [r i,1, r i,2, . . . , r i,nτ ]
for i = 1 → drne do end if
if i ≤ n then . multiply each symbol of ∆ j by r i, j
repeat generate a random encoding vector ∆i = ∆i0 ⇐ nj=1
P τ
r i, j ∆ j
[δi,1, δi,2, . . . , δi,n ]
. multiply each symbol of g j by r i, j
until ∆i is linearly independent from all the ∆ j ,
gi0 ⇐ nj=1
P τ
r i, j g j
1 ≤ j < i (except for i = 1)
f g
else send out ∆i0 ||gi0 , where“||” is the concatenation oper-
generate a random encoding vector ∆i = [δi,1, δi,2, ation
. . . , δi,n ] end for
end if report the number of received packets from each of the
. multiply each symbol of packet data h j by δi, j other nodes/the number of sent out packets to the SWDN
gi ⇐ nj=1 δi, j h j
P
  controller
send out ∆i ||gi , where“||” is the concatenation oper-
ation
end for In algorithm 2, the intermediate node generates drnτ e nτ -
report the number of sent out packets to the SWDN con- dimensional vectors Ri (first nτ vectors are linearly inde-
troller pendent) and uses the vector elements r i,1, r i,2, . . . , r i,nτ as
coefficients to generate and send out recoded packets from
In algorithm 1, the source node generates drne n- the received packets g1, . . . , gnτ . The corresponding encoding
dimensional encoding vectors ∆i (first n encoding vec- vectors are processed the same way.
tors are linearly independent) and uses the vector elements 3) Sink Node Algorithm of the ANC scheme: Once the sink
δi,1, δi,2, . . . , δi,n as coefficients to generate and send out node receives n linearly independent packets, it can solve
encoded packets from the uncoded packets h1, . . . , hn . the following equation to decode the original packets data

2169-3536 (c) 2016 IEEE. Translations and content mining are permitted for academic research only. Personal use is also 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/ACCESS.2017.2706328, IEEE Access

simulation, the leftmost node in Fig. 4 tries to send data to the

1.8 code rate = 1.1 rightmost two nodes. The qualities of the channels between
code rate = 1.15 the intermediate nodes are chosen randomly. We calculate
the case without network coding
transmission efficiency ratio to

code rate = 1.2

code rate = 1.25 the transmission efficiencies under different numbers of total
1.6 ANC scheme
without network coding
original data packets (can be viewed as one coding group
for network coding). For performance comparison, we also
1.4 simulate the cases for no network coding (pure retransmis-
sion) and network coding with predetermined code rates as
1.2 in [18]. To make the comparison more clear, we normalize
the transmission efficiencies for the cases with fixed network
coding and the ANC scheme by the transmission efficiency for
1
the case without network coding. From the simulation results,
we can see that the transmission efficiency for the case with
0.8 network coding becomes higher than the case without network
5 10 15 20 25
number of original data packets coding with the increasing of the number of original data
packets. And the ANC performs best among all the cases.
Fig. 6. Performance of the ANC scheme It can also be seen that the performance gain of the ANC
scheme will increase when the number of total original data
packets becomes larger.
h1, h2, . . . , hn :
 h1   g1   ∆1 
IV. O PTIMAL S TORAGE A LLOCATION IN THE
 h2  =  g2  .  ∆2 
     
 ...   ...   ...  (4) D ISTRIBUTED C LOUD S TORAGE U TILIZING N ETWORK
      C ODING
hn  gn  ∆ n  In this section we first introduce a storage allocation
Then h1, h2, . . . , hn can be concatenated to restore the original problem for the distributed cloud storage utilizing network
data. The sink node will also periodically report the packets coding. Then we propose the optimal storage allocation (OSA)
receiving status to the SWDN controller. scheme. We also show the performance of the optimal storage
4) SDWN Controller Algorithm of the ANC scheme: Since allocation scheme.
the SDWN controller receives the packets sending/receiving
status from each of the nodes in the IoT core network A. Storage Allocation Problem
periodically , it can adjust the code rate of the network coding
In the storage allocation problem S, the data is encoded
for each of the nodes accordingly. Suppose η i is the number
with an (n, k, d, α, β, B) regenerating code, so there will be n
of packets sent by node i, Ni is the set of succeeding nodes
encoded parts. There are N (N < n) data centers in total to
receiving the packets from node i, and η (i) j is the number of store these parts, each with a failure probability of p. If a data
packets received by node j ∈ Ni . The code rate r i of node i
center fails, all the data stored in the data center will be lost. If
can be determined by
the total number of encoded parts in the remaining data centers
ηi
ri = , (5) is less than k, the original data cannot be recovered any more.
max j ∈Ni (η (i) j ) Since there are more encoded parts than the data centers, there
where max() is the operation to select the maximum element. will be different allocation strategies of the encoded parts with
Besides the code rate, since the SDWN controller has all of different storage reliabilities. For the problem S, we try to find
the topology information, for each node i, it can specify the out the allocation strategy with the lowest failure probability
succeeding relay nodes to receive the packets sent from node among all the possible allocation strategies.
i. Definition 1. A set S with N elements n1, n2, . . . , n N (ni >
ni = n.
PN
0, 1 ≤ i ≤ N) is a valid allocation if i=1
Algorithm 3 ANC scheme - SDWN controller
for each of the source node or intermediate node i do Definition 2. For an allocation strategy S, the failure prob-
calculate r i according to the packet sending/receiving ability P is defined as the probability that the original data
status of node i and Ni cannot be recovered given the failure probabilities of individ-
send r i to update the code rate of node i through the ual data centers.
control path The problem S can be formulated as:
end for
find the allocation S among all the valid allocations,
X X
C. Performance Evaluation of the ANC scheme such that P *. ni ≥ n − k +/ is minimal.
In Fig. 6 are the simulation results of the ANC scheme. ∀S j ⊆S ,ni ∈S j -
The simulation is carried out in the NS-2 platform [26]. In the (6)

2169-3536 (c) 2016 IEEE. Translations and content mining are permitted for academic research only. Personal use is also 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/ACCESS.2017.2706328, IEEE Access

As an example, for the regenerating code in Fig. 3, n = 4 Algorithm 4 OSA scheme - stage I
encoded parts are stored in N = 2 data centers. Suppose Input: the number of encoded parts n and the number of
the failure probability of each data center is p = 0.01. Two storage centers N
storage allocation strategies are shown in the figure. For the Output: all the valid allocations S(n, N, l), (1 ≤ l ≤ P(n, N ))
first allocation strategy S = {3, 1} (blue data centers with dash 1: function F INDA LL A LLOCATIONS (n, N)
lines), 3 encoded parts are stored in data center 1 and 1 encode 2: for i = 1 → n do
part is stored in data center 2. It is easy to calculate the failure 3: P(i, 1) ⇐ 1
probability of this allocation strategy is 0.01. For the second 4: S(i, 1, 1) ⇐ i
allocation strategy S = {2, 2} (orange data centers with solid 5: for j = 2 → N do
lines), 2 encoded parts are stored in each of the two data 6: if i ≥ j then
centers. The failure probability of this allocation strategy is 7: if i − j < j then
0.0001, which is much lower than that of the first strategy. 8: P(i, j) ⇐ P(i − 1, j − 1)
9: S(i, j, l) ⇐ S(i − 1, j − 1, l) ∪ {1}, for
1 ≤ l ≤ P(i, j)
B. Optimal Storage Allocation Scheme
10: else
In this section, we will show our optimal storage allocation 11: P(i, j) ⇐ P(i − 1, j − 1) + P(i − j, j)
(OSA) scheme to solve the storage allocation problem. The 12: S(i, j, l) ⇐ S(i − 1, j − 1, l) ∪ {1}, for
OSA scheme includes two stages: the first stage is to find all 1 ≤ l ≤ P(i − 1, j − 1)
out all the possible valid allocations S and the second stage 13: S(i, j, P(i − 1, j − 1) + l) ⇐ S(i − j, j, l)+
is to calculate the failure probability P for each S. Then we 1, for all 1 ≤ l ≤ P(i − j, j)
can output the allocation with the lowest failure probability 14: end if
through comparison. 15: end if
1) Stage I: Find out All the Possible Valid allocations S: 16: end for
The naive approach to find out all the possible valid S is to 17: end for
search all the possible combinations of n1, n2, . . . , n N such that 18: end function
i=1 ni = n. However, this approach will take exponential time
PN
thus is not practical. In our OSA scheme, we first change this
problem into an integer partition problem [27]: to allocate the l th valid allocation out of the P(n, N ) allocations for n
n encoded parts into N storage centers is the same as to encoded parts and N storage centers.
partition an integer n into N parts. Take n = 7, N = 3 as
an example, there are 4 ways to partition 7 into 3 parts: Proof. Algorithm 4 calculates S(i, j, l) (1 ≤ l ≤ P(i, j)) for
{1, 1, 5}, {1, 2, 4}, {1, 3, 3} and {2, 2, 3}, which also consist all 1 ≤ j ≤ N from i = 1 to i = n through a bottom-up manner
the possible valid allocations. Then we can solve the integer and we can get S(n, N, l) (1 ≤ l ≤ P(n, N )) for i = n, j = N.
partition problem using dynamic programming based on the For each i, line 3 to line 4 first calculate P(i, 1) = 1 and
following recurrence equation: S(i, 1, 1) = i, corresponding to the case of allocating i encoded
data parts into one data center. Then for each j = 2, . . . , N,
P(n, N ) = P(n − 1, N − 1) + P(n − N, N ), (7) there will be two cases:
• Line 8 to line 9 correspond to the case with i − j < j,
where P(i, j) is the total number of ways of partitioning integer where at least one storage node will be allocated only
i into j parts. The first part of equation (7) is the subproblem 1 encoded data part. The second part of equation (7)
where at least one 1 exists in the partition and the second part does not exist. So the number of ways of allocating i
of the equation is the subproblem where no 1 exists in the encoded data parts into j storage nodes will be equal
partition. Thus the solution to the original problem perfectly to that of allocating i − 1 encoded data parts into j − 1
incorporates these two subproblems, which make it feasible to storage nodes: P(i, j) = P(i − 1, j − 1). And each of the
solve using dynamic programming. We propose Algorithm 4 valid allocations S(i, j, l) will be the union of each already
to find out all the possible valid allocations S. In the algorithm, calculated allocations S(i − 1, j − 1, l) with the set {1}.
we use S(i, j, k) to represent the k th valid allocation out of the • Line 11 to line 13 correspond to the case with i − j ≥ j,
P(i, j) allocations for i encoded parts and j storage centers. ∪ where P(i, j) is the summation of two previously calcu-
is the union operation between two sets. The addition between lated parts as shown in equation (7). The computation of
a set S and a number x is defined as the additions between the first part and the corresponding valid allocations is
every element of the set and the number: the same as in line 8 to line 9. The second part is the
S + x := {ni + x|ni ∈ S for 1 ≤ i ≤ N }. (8) number of ways of allocating i − j encoded data parts
into j storage nodes P(i − j, j), where each of the storage
After the execution of the algorithm, we can get all the possible node will be allocated at least 2 encoded data parts. Thus
valid allocations S(n, N, k) (1 ≤ k ≤ P(n, N )). It is easy to each of the valid allocations S(i, j, P(i − 1, j − 1) + l) will
see that the algorithm runs in polynomial time. be each of the already calculated allocations S(i − j, j, l)
Theorem 1. Algorithm 4 can output all the valid allocations plus 1 as defined in equation (8).
S(n, N, l) for 1 ≤ l ≤ P(n, N ), where S(n, N, l) represents 

2169-3536 (c) 2016 IEEE. Translations and content mining are permitted for academic research only. Personal use is also 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/ACCESS.2017.2706328, IEEE Access

value summation results of the subsets S j(i) and count the total
(7,1) (7,2) (7,3)
number of S j(i) which have that summation value. Then for
(6,1) (6,2) (6,3) the subsets that have summation results larger than n − k, we
(5,1) (5,2) (5,3) can calculate the corresponding failure probability according
to equation (10). In the algorithm, T, L, R represent three
(4,1) (4,2) (4,3) auxiliary lists for subset summation. For a auxiliary list X,
(3,1) (3,2) (3,3) we use X .length to denote the number of elements of the
(2,1) (2,2) list, X .index to denote the current index number of the list,
VX ( j) to denote the value of j th element in X, and CX (i, j) to
(1,1) denote the total number of subsets that have the same element
number i and the same summation value VX ( j). Although the
Fig. 7. The calculation of Algorithm 4 for n = 7, N = 3 total number of subsets is 2 N , Algorithm 5 is a polynomial
time algorithm:
Fig. 7 illustrates the algorithm for n = 7 encoded data Theorem 2. The complexity of Algorithm 5 is O(nN ).
parts and N = 3 data centers. Each (i, j) pair represent the
calculation of P(i, j) and S(i, j, l). The pairs without shades Proof. Since the summation of a valid allocation S itself is the
are calculated using line 8 to line 9 (the first case) while the largest in all the summations of the subsets of S, the element
pairs in shades are calculated using line 11 to line 13 (the number T .length in T cannot exceed n. Through the merge
second case). The solid lines correspond to the first part of of subsets with the same summation values, each of the N
equation (7) and the dashed lines correspond to the second for-loops has the complexity O(n). So the total complexity is
part. From the figure we can clearly see that (7, 3) can be O(nN ). 
efficiently calculated using the results of (6, 2) and (4, 3),
which have already been calculated the same way as illustrated Theorem 3. Algorithm 5 can output the failure probability PS
in Fig. 7. for the input allocation.
2) Stage II: Calculate the Failure Probability P for Each
Valid Allocation S: After we get all the possible valid alloca- Proof. In line 3 we initialize the auxiliary list L with an
tions S, we can calculate the failure probability PS for each of empty element 000, representing the summation result of 0
them. The goal function of equation (6) can be further written element of the input allocation S. Line 4 to line 31 calculate
as: the summations of every subset of the input allocation. At the
X X beginning of each round i of the for loop i = 1, . . . , n, the
PS = P *. ni > n − k +/ auxiliary list L is the list containing the summation results
∀S j ⊆S ,ni ∈S j of every subset of the first l (0 ≤ l < i) elements of the
- (9)
X input allocation S. Line 7 to line 8 calculate the auxiliary list
= p S j | (1 − p) N − | S j | ,
|
R by adding the new element ni to L: R = L + ni . Since
ni >n−k
P
∀S j ⊆S, s.t. n i ∈S j
the first element in L is the empty 000, CR (1, 1) will be 1,
where p is the failure probability of each storage center, indicating that the total number of subsets that have 1 element
 S  is the number of elements in subset S . If we try to and summation value ni is 1. Then the rest value of CR (l, j)
 j j
directly calculate PS for every subset S j ∈ S, the order of will be CL (l − 1, j) for 2 ≤ j ≤ L.length because of the
the number of !subsets to be calculated ! will be approximate addition of ni to L. The elements of allocation S are sorted
PN N N in non-descending order, thus the elements in both L and R
to ≈ 2 N , where   denotes the number of
| S j | =1  S j   S j  are also in non-descending order. From line 10 to line 29,
 
 S -combinations of the set S, thus making it infeasible to we merge the elements of the auxiliary lists L and R into
 j a temporary auxiliary list T one by one, following the rules
calculate in practice.
In the second stage of the OSA scheme (Algorithm 5), we below:
propose to change the exhaust search problem into a number • If the value of the current element VL (L.index) in L is
counting problem. More specifically, for each i (1 ≤ i ≤ N), equal to the current element VR (R.index) in R, add the
we count the total number of subsets S j(i) such that S j(i) denotes value into T. The corresponding counter CT (l, T .length)
the subsets with exactly i elements and the summation of every is equal to the sum of the two counters: CT (l, T .length) =
element in S j(i) is larger than n − k: CL (l, L.index) + CR (l, R.index) for 1 ≤ l ≤ i.
• If the value of the current element VL (L.index) in L is

N    smaller than the current element in R, add the element
  VL (L.index) into T. Set the counter CT (l, T .length) to
X  X 

(i)
PS = ni > n − k   pi (1 − p) N −i .
  
 Sj |  (10)
CL (l, L.index) for 1 ≤ l ≤ i.
 
 
i=1   ni ∈S (i)
 
j
  • If the value of the current element VR (R.index) in R is
In Algorithm 5, we first calculate the summations of every smaller than the current element in L, add the element
subset, which can be viewed as a variant of the subset-sum VR (R.index) into T. Set the counter CT (l, T .length) to
problem [28]. For each i (1 ≤ i ≤ N), we merge the same- CR (l, R.index) for 1 ≤ l ≤ i.

2169-3536 (c) 2016 IEEE. Translations and content mining are permitted for academic research only. Personal use is also 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/ACCESS.2017.2706328, IEEE Access

Round 1
L 0 R 1 T 0 1
Algorithm 5 OSA scheme - stage II CL 0 CR 1 CT 0 1
Input: a valid allocation S(n, N, l), (1 ≤ l ≤ P(n, N ))
Round 2
Output: the failure probability PS of the allocation L 0 1 R 2 3 T 0 1 2 3
1: function C ALCULATE P ROBABILITY (S(n, N, l)) CL 0 1 CR 1 0 CT 0 1 1 0
2: {n1, n2, . . . , n N } ⇐ sort the allocation S(n, N, l) in 0 0 0 1 0 0 0 1
non-descending order
Round 3
3: L ⇐ {0}
L 0 1 2 3 R 2 3 4 5 T 0 1 2 3 4 5
4: . calculate summations of every subset CL 0 1 1 0 CR 1 0 0 0 CT 0 1 2 0 0 0
5: for i = 1 → N do 0 0 0 1 0 1 1 0 0 0 0 2 1 0
6: T⇐φ 0 0 0 0 0 0 0 1 0 0 0 0 0 1
7: R ⇐ L + ni , CR (1, 1) ⇐ 1
8: CR (l, j) ⇐ CL (l − 1, j), for all nonzero CL (l, j), Fig. 8. The calculation of Algorithm 5 for S = {1, 2, 2}
2 ≤ j ≤ L.length, 2 ≤ l ≤ i, i ≥ 2
9: L.index, R.index ⇐ 1
10: while L.index ≤ L.length do • Since the last element in L is smaller than some elements
11: if VL (L.index) == VR (R.index) then in R, after merging L into T, we can directly merge the
12: T ⇐ T ∪ VR (R.index) remaining elements of R into T through line 28 to line
13: for all 1 ≤ l ≤ i, CT (l, T .length) ⇐ 29.
CL (l, L.index) + CR (l, R.index) At the end of each for loop, the merged list T is assigned back
14: increase L.index, R.index by 1 to L for the next round of calculation. After N th round, list T
15: else has the summation results of all the subsets in S.
16: if VL (L.index) < VR (R.index) then Then the failure probability of S can be easily calculated
17: T ⇐ T ∪ VL (L.index) from line 34 to line 42 by counting the number of subsets
18: CT (l, T .length) ⇐ CL (l, L.index), for with the summation results larger than n − k. 
all 1 ≤ l ≤ i
19: increase L.index by 1 Fig. 8 illustrates the summations for all the subsets of S =
20: else {1, 2, 2}. For i = 1, L = {0}, CL (1, 1) = 0, R = {1}, CR (1, 1) =
21: T ⇐ T ∪ VR (R.index) 1. The merged list T = {0, 1}, CT = {0, 1}. For the second
22: CT (l, T .length) ⇐ CR (l, R.index), for round, L, CL are assigned the values of T, CT . According to
all 1 ≤ l ≤ i line 7 and line 8 of Algorithm 5, R = L + n2 = {2, 3} and
23: increase R.index by 1 CR (2, 2) = CL (1, 2) = 1. At the end of the third round, we
24: end if can get the summation results T = {1, 2, 3, 4, 5} and the counter
25: end if matrix CT , which correctly record the number of subsets that
26: end while have the same summation value. As an example, CT (2, 3) = 2,
27: oldLn ⇐ T .length indicating that there are two 2-element subsets ({n1 = 1, n2 =
28: T ⇐ T ∪ {VR (R.index), VR (R.index + 1), . . . , 2}, {n1 = 1, n3 = 2}) that have the same summation value
VR (R.length)} VT (3) = 3.
29: {CT (l, oldLn + 1), . . . , CT (l, T .length)} ⇐ 3) OSA Scheme: Based on the algorithms of the two
{CR (l, R.index), . . . , CR (l, R.length)}, 1 ≤ l ≤ i stages, we can achieve the optimal storage allocation through
30: L ⇐T Algorithm 6. And it is straightforward to see:
31: end for Theorem 4. The OSA scheme is a polynomial time algorithm.
32: PS ⇐ 0
33: . count the number of subsets with the summation
results larger than n − k Algorithm 6 OSA scheme
34: for i = 1 → N do Input: the number of encoded parts n and the number of
35: sum ⇐ 0 storage centers N
36: for j = 1 → T .length do Output: the allocation with the lowest failure probability
37: if VT ( j) > n − k then function OSA(n, N)
38: sum ⇐ sum + CT (i, j) S(n, N, l) ⇐ F INDA LL A LLOCATIONS(n, N) (1 ≤ l ≤
39: end if P(n, N ))
40: end for for l = 1 → P(n, N ) do
41: PS ⇐ PS + sum × pi (1 − p) N −i PS ⇐ C ALCULATE P ROBABILITY(S(n, N, l))
42: end for end for
43: end function output the allocation with the lowest PS
end function

2169-3536 (c) 2016 IEEE. Translations and content mining are permitted for academic research only. Personal use is also 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/ACCESS.2017.2706328, IEEE Access

10

-1
10 10 -2
OSA scheme
even allocation
OSA scheme
even allocation

-2
10 10 -3
failure probability

failure probability
10 -3 10 -4

-4
10 10 -5

10 -5 10 -6
16 18 20 22 24 26 28 30 32 4 6 8 10 12 14 16 18
k N

Fig. 9. Performance of the optimal storage allocation for different k Fig. 10. Performance of the optimal storage allocation for different number
of storage centers

C. Simulation Results for the OSA Scheme

In this section, we will show the performance of the the ANC scheme can achieve higher transmission efficiency
OSA scheme for the given regenerating code with parameters than existing schemes. Then we introduce the optimal storage
(n, k, d, α, β, B) and number of data centers N. allocation problem for the distributed cloud storage that uti-
In Fig. 9 are the simulation results for n = 45, k = lizes network coding, which stores the data generated by the
{16, 21, 26, 31}, N = 9 and p = 0.1. For performance compari- IoT core network. we propose an optimal storage allocation
son, we also plot the results for the even allocation as defined (OSA) scheme to solve the problem in polynomial time. We
in equation (11), where bn/Nc is the floor operation to get the also conduct simulations to show that the OSA scheme can
largest integer that is less or equal to n/N, mod(n, N ) is the greatly improve the storage reliability.
modulo operation to find the remainder of the division of n
by N.
R EFERENCES
bn/Nc + 1, 1 ≤ i ≤ mod(n, N )
(
ni = (11)
bn/Nc, mod(n, N ) < i < N [1] H. Yue, L. Guo, R. Li, H. Asaeda, and Y. Fang, “Dataclouds: Enabling
community-based data-centric services over the internet of things,” IEEE
The even allocation is a natural allocation scheme to store Internet of Things Journal, vol. 1, pp. 472–482, Oct 2014.
[2] J. Lin, W. Yu, N. Zhang, X. Yang, H. Zhang, and W. Zhao, “A survey
equal number of data blocks into each storage center. From on internet of things: Architecture, enabling technologies, security and
the figure we can see that the failure probability of the OSA privacy, and applications,” IEEE Internet of Things Journal, vol. PP,
scheme is about half order of magnitude lower than the even no. 99, pp. 1–1, 2017.
[3] J. Baliga, R. W. A. Ayre, K. Hinton, and R. S. Tucker, “Green cloud
allocation. Both of the probabilities will become higher when computing: Balancing energy in processing, storage, and transport,”
k increases since there is less redundancy in the distributed Proceedings of the IEEE, vol. 99, pp. 149–167, Jan 2011.
cloud storage to recover the failed storage centers. [4] Y. Chen, S. Zhang, S. Xu, and G. Y. Li, “Fundamental trade-offs on
green wireless networks,” IEEE Communications Magazine, vol. 49,
In Fig. 10 are the simulation results for n = 45, k = 21, pp. 30–37, June 2011.
p = 0.1. In this simulation, we change the number of storage [5] G. Jia, G. Han, J. Jiang, and L. Liu, “Dynamic adaptive replacement
centers N to study its impact to the failure probability. From policy in shared last-level cache of dram/pcm hybrid memory for big
data storage,” IEEE Transactions on Industrial Informatics, vol. PP,
the figure we can see that the failure probability will become no. 99, pp. 1–1, 2016.
lower when the number of storage centers increases. And the [6] Z. Zhou, Z. Cheng, L. J. Zhang, W. Gaaloul, and K. Ning, “Scientific
performance gap of the even allocation and the OSA scheme workflow clustering and recommendation leveraging layer hierarchical
will diminish with the increasing of the number of storage analysis,” IEEE Transactions on Services Computing, vol. PP, no. 99,
pp. 1–1, 2016.
centers. [7] J. Li, T. Li, and J. Ren, “Beyond the mds bound in distributed cloud
storage,” in INFOCOM, 2014 Proceedings IEEE, pp. 307–315, April
2014.
V. CONCLUSION
[8] C. Wang, K. Ren, and J. Wang, “Secure optimization computation
Applying the network coding in Internet of things could outsourcing in cloud computing: A case study of linear programming,”
IEEE Transactions on Computers, vol. 65, pp. 216–229, Jan 2016.
save energy and contribute to the ‘green IoT networking’. Mo-
[9] B. Bangerter, S. Talwar, R. Arefi, and K. Stewart, “Networks and devices
tivated by this advantage, in this paper we optimize network for the 5g era,” IEEE Communications Magazine, vol. 52, pp. 90–96,
coding based communication and reliable storage in Internet February 2014.
of things. We propose an adaptive network coding (ANC) [10] “the zigbee alliance.” http://www.zigbee.org, 2016.
[11] G. Mulligan, “The 6lowpan architecture,” in Proceedings of the 4th
scheme in the IoT core network with software defined wireless Workshop on Embedded Networked Sensors, EmNets ’07, pp. 78–82,
network (SDWN). Simulation results have demonstrated that ACM, 2007.

2169-3536 (c) 2016 IEEE. Translations and content mining are permitted for academic research only. Personal use is also 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/ACCESS.2017.2706328, IEEE Access

11

[12] G. Han, L. Liu, J. Jiang, L. Shu, and G. Hancke, “Analysis of energy-

efficient connected target coverage algorithms for industrial wireless
sensor networks,” IEEE Transactions on Industrial Informatics, vol. 13,
pp. 135–143, Feb 2017.
[13] G. Han, J. Jiang, M. Guizani, and J. J. P. C. Rodrigues, “Green routing
protocols for wireless multimedia sensor networks,” IEEE Wireless
Communications, vol. 23, pp. 140–146, December 2016.
[14] Z. Zhou, W. Fang, J. Niu, L. Shu, and M. Mukherjee, “Energy-
efficient event determination in underwater wsns leveraging practical
data prediction,” IEEE Transactions on Industrial Informatics, vol. PP,
no. 99, pp. 1–1, 2017.
[15] S. Costanzo, L. Galluccio, G. Morabito, and S. Palazzo, “Software
defined wireless networks: Unbridling sdns,” in Software Defined Net-
working (EWSDN), 2012 European Workshop on, pp. 1–6, Oct 2012.
[16] Z. Chen, G. Xu, V. Mahalingam, L. Ge, J. Nguyen, W. Yu, and C. Lu, “A
cloud computing based network monitoring and threat detection system
for critical infrastructures,” Big Data Res., vol. 3, pp. 10–23, Apr. 2016.
[17] B. A. A. Nunes, M. Mendonca, X. N. Nguyen, K. Obraczka, and
T. Turletti, “A survey of software-defined networking: Past, present,
and future of programmable networks,” IEEE Communications Surveys
Tutorials, vol. 16, no. 3, pp. 1617–1634, 2014.
[18] J. Hansen, D. E. Lucani, J. Krigslund, M. Medard, and F. H. P. Fitzek,
“Network coded software defined networking: enabling 5g transmis-
sion and storage networks,” IEEE Communications Magazine, vol. 53,
pp. 100–107, September 2015.
[19] R. Bhagwan, K. Tati, Y.-C. Cheng, S. Savage, and G. M. Voelker, “Total
recall: System support for automated availability management,” in roc.
Symp. Netw. Syst. Design Implementation, pp. 337–350, 2004.
[20] R. Ahlswede, N. Cai, S.-Y.Li, and R. Yeung, “Network information
flow,” IEEE Transactions on Information Theory, vol. 46, pp. 1205–
1216, July 2000.
[21] R. Koetter and M. Medard, “An algebraic approach to network coding,”
IEEE/ACM Transactions on Networking, vol. 11, no. 5, pp. 782–795,
2003.
[22] T. Ho, B. Leong, R. Koetter, M. Medard, M. Effros, and D. Karger,
“Byzantine modification detection in multicast networks using ran-
domized network coding,” in International Symposium on Information
Theory (ISIT) 2004, p. 144, July 2004.
[23] C. Gkantsidis and P. Rodriguez, “Cooperative security for network
coding file distribution,” in IEEE INFOCOM 2006, pp. 1–13, Apr. 2006.
[24] D. Zhu, X. Yang, P. Zhao, and W. Yu, “Towards effective intra-
flow network coding in software defined wireless mesh networks,” in
2015 24th International Conference on Computer Communication and
Networks (ICCCN), pp. 1–8, Aug 2015.
[25] A. Dimakis, P. Godfrey, Y. Wu, M. Wainwright, and K. Ramchandran,
“Network coding for distributed storage systems,” IEEE Transactions
on Information Theory, vol. 56, pp. 4539 – 4551, 2010.
[26] “Ns2 wiki.” http://nsnam.isi.edu/nsnam/index.php/Main Page, 2010.
[27] A. Nijenhuis, H. S. Wilf, and W. Rheinboldt, Combinatorial Algorithms
for Computers and Calculators. Academic Pr, 2rd ed., 1978.
[28] T. H. Cormen, C. E. Leiserson, R. L. Rivest, and C. Stein, Introduction
to Algorithms. The MIT Press, 3rd ed., 2009.

2169-3536 (c) 2016 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.