Received July 30, 2016, accepted September 12, 2016, date of publication September 22, 2016, date of current

version October 15, 2016.

Digital Object Identifier 10.1109/ACCESS.2016.2611619

Queueing Characteristics of the Best Effort

Network Coding Strategy
XIANXU LI, QING CHANG, AND YONG XU
School of Electronic and Information Engineering, Beihang University, Beijing 100191, China
Corresponding author: Y. Xu. ([email protected])
This work was supported by the National Natural Science Foundation of China under Grant 61471021.

ABSTRACT Asynchronous network coding has the potential to improve wireless network performance
compared with simple routing. However, to achieve the maximum network coding gain, the encoding node
consumes a few computing and storage resources that may be unaffordable for wireless sensor networks such
as CubeSats. An analogous threshold strategy, called best effort network coding (BENC), which requires
only minimal storage resources and no computing resources, is investigated in this paper as an efficient and
convenient method of network coding. In this strategy, a new packet arrival evicts the head packet when the
queue is full to avoid excessively long waits. Moreover, in contrast to other methods that require a queue for
each flow, the BENC uses only one queue for the two coded flows. In addition, the problem of time interval
distribution for the output flow, which combines two independent flows, is investigated, and the network
coding gain is then analyzed. While the maximum coding gain requires infinite buffer capacity under two
independent Poisson arrivals with the same transmission rates, the calculation results show that the BENC
needs only 4 buffers to achieve 90% of the maximum coding gain and can reach 99% of the maximum coding
gain using 50 buffers. These results are verified by numerical simulations.

INDEX TERMS Network coding, best effort, queueing analysis, queue capacity, wireless sensor networks.

I. INTRODUCTION turn the ‘‘interference’’ into good use [6]. MAC schedul-
Network coding is a new forwarding method that breaks ing schemes, which decide whether and how intermediate
the established pattern (which regards information flow as a nodes encode packets, were considered in [7]–[9] to optimize
commodity) by mixing several packets into one coded packet the throughput performance. In the network layer, method-
using algebraic operations. These coded packets can then ologies to optimize the system’s performance under certain
deliver the contents of more than one packet in a single constraints were established taking both synchronous and
transmission and, thus, reduce transmission times through a asynchronous flows and infinite or finite buffers into account.
router (or access point), increasing throughput and decreas- One important domain of application is in two-way relay
ing power consumption [1]. Based on the maximal flow- networks with stochastic arrivals, in which a situation could
minimal cut theorem in graph theory [2], a sender and receiver occur where no packets are available for the other connection.
cannot communicate at a rate greater than the maximal The primary problem here is to design a mechanism to termi-
flow (or the minimal cut). In a butterfly network, traditional nate excessively long waits, otherwise the system becomes
‘store–and–forward’ routing cannot achieve maximal flow. unstable and the mean wait time grows unboundedly [10].
However, in their seminal work, Ahlswede et al. [3] showed Several relevant aspects have already been explored in the
that network coding can theoretically achieve the maximal literature.
flow rate. Subsequently, network coding attracted a great One methodology involves terminating waits using a prob-
deal of attention and studies proliferated rapidly. Research ability. Chieochan et al. [11] proposed a queueing model
into network coding applications can be divided into intra- with a finite buffer at the relay that could be shared by two
session [4] and inter-session [5] categories based on the independent injected source packets. Whether packets were
source nodes from which the uncoded packets were gener- transmitted or received depended on a probability pertaining
ated. To increase the coding opportunities, these two schemes to each link in every time slot. Abdelrahman et al. [12]
have been extended into different protocol layers using proposed a queuing model with infinite capacity for each
customized strategies. Physical network coding attempted to flow. The packet waits for an exponential time-out period

2169-3536 2016 IEEE. Translations and content mining are permitted for academic research only.
5990 Personal use is also permitted, but republication/redistribution requires IEEE permission. VOLUME 4, 2016
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
X. Li et al.: Queueing Characteristics of the BENC Strategy

before forwarding a packet without coding. Ding et al. [13] is investigated using simulated independent exponential input
considered a two-way relay network with Bernoulli asymmet- processes with the same parameters in Section IV. Finally,
ric arrival processes with a finite capacity. The relay node conclusions are presented in Section V.
forwards one uncoded packet with a probability relevant to
the buffer length.
Some strategies have been designed to deal with the long
wait issue. The COSE strategy was proposed in [14], in which
Poisson arrival rates were assumed. A fixed time interval
was configured to maintain a coding opportunity. Packets left FIGURE 1. Queue model of an encoding node.
uncoded would be released from the waiting queue and the
process restarted. Paramanathan et al. [15] tried to guarantee II. BEST EFFORT NETWORK CODING MODEL
each flow 50% of its buffer by dropping the oldest packets to Fig. 1 shows the queuing model of an encoding node, which
manifest a high coding gain under a short-term asymmetric consists of two independent data flows named flow 1 and
rate. They implemented this mechanism on their platform. flow 2, respectively, one queue model and one output flow.
Several publications have pursued optimization prob- The data packet from each flow arrives after the time inter-
lems that combine performance and cost objectives. val T whose distribution is exponential with λ1 and λ2 ,
Yuan et al. [16] provided an optimal trade-off between cod- respectively. The packets arrive at the encoding node and
ing opportunities and packet-delay with n input flows, but are served to generate the output flow. The queue model in
this method required complex computing to calculate the the encoding node is divided into two parts. The first part is
probability matrix. The authors in [17] and [18] devised the waiting queue, which has a finite buffer length, N , and the
a mechanism that forwarded uncoded packets based on a second is a service queue that stores and forwards packets.
cost threshold such as the energy consumption required for In the second part of this queuing model, the classical queue-
transmission by the relay. Several relevant policies such ing theory of GI/GI/1/N is able to analyze the performance
as cost threshold were considered, and the results of opti- theoretically. The network coding occurs in the first part, for
mizing objectives relevant to delay or cost were presented. which there is not yet sufficient theoretical analysis.
Ciftcioglu et al. [19] and Davri et al. [20] moved the buffers Therefore, this study focuses on the first part in Fig. 1.
from the relay node to the source nodes to enhance the Packets from flow 1 and flow 2 will be processed in the
throughput of the MAC channel. Zohdy et al. [21] focused waiting queue using the BENC strategy to produce flow 3.
on efficiency gains and took delay-related performance into Then, flow 3 becomes the input flow to the second part, where
account as an optimization constraint rather than as an it is transmitted with the general service distribution to pro-
objective. duce the output flow leaving the encoding node. Therefore,
Few works have concentrated on the statistical character- the prime task here is to obtain the probability distribution
istics of the output yielded by network coding. The output of flow 3.
distribution of the time intervals for synchronous network
coding scenarios was presented in [22]. The theoretical A. THE BENC STRATEGY
results showed that the distribution was an asymptotical expo- In the BENC strategy, if the buffer is full or a matched packet
nent when the rates of the two independent exponential input arrives, then a packet should leave the waiting queue. The
flows were the same. Alsebae et al. [23] investigated the specific data flow diagram for BENC is shown in Fig. 2. The
output distribution of an M/D/1 queueing model under a relay maintains two virtual, parallel, finite buffers; each vir-
synchronous network coding scenario. In their study, only tual buffer accommodates packets received from each source.
encoded packets were included in the output flow. By ‘‘virtual’’ we mean that one real buffer is sufficient in
The contributions of the present paper are twofold. First, to a real system. When a new packet arrives, BENC checks
achieve an efficient asynchronous network coding method, whether the virtual buffer of the other flow is empty. If the
an analogous threshold queueing strategy with finite queu- other buffer is not empty, the newly arrived packet is encoded
ing capacity, called Best Effort Network Coding (BENC), is with a packet in the other flow and departs the waiting queue
investigated. The BENC needs few storage resources and no immediately. However, if the other buffer is empty when the
computing resources to achieve network coding gains close new packet arrives, it will enter its own virtual buffer to wait
to the theoretical maximum, and it is easily deployed on for a future encoding opportunity. In addition, if its own buffer
encoding nodes. Second, network coding output statistical is full, the head packet leaves the waiting queue, opening one
characteristics, such as time interval distribution, mean wait- slot for the new arrival. Conversely, when a flow’s virtual
ing queue length and delay, are also investigated. buffer is not full, packets already in the queue remain in the
This study is organized as follows: the queuing model and queue. Based on the BENC strategy, packets will not be lost
the BENC strategy are presented in Section II. Queueing from the waiting queue when the queue is full.
characteristics such as the time interval distribution of output When neither of virtual queues are empty, the network
for BENC and its expectation, as well as the mean waiting coded packet will be generated and depart the queue imme-
queue length are studied in Section III. Then, the performance diately, therefore, there is practically only one queue which

VOLUME 4, 2016 5991

 X. Li et al.: Queueing Characteristics of the BENC Strategy

and flow 2, respectively. Based on the strategy described in

Section II-A, the transition diagram is illustrated in Fig. 4. For
the BENC strategy, we have k ∗ l = 0, which means at least
one buffer is empty. To save storage resources, a real system
requires only one buffer to accommodate packets from two
input flows.

FIGURE 4. State transition diagram for the Markov chain.

FIGURE 2. The best effort network coding (BENC) strategy. III. PROBLEM FORMULATION
The probability distribution for the inter-arrival time of input
flows is exponential with λ1 and λ2 , respectively. The random
variable T̂ represents the time interval of flow 3.

A. STEADY STATE PROBABILITY

The column for the steady state probability, W, is defined as:
h iT
W = P ωN1 , . . . , P ω11 , P ω0 , P ω11 , . . . , P ωN2
Although the capacity of the relay buffer is N , the state
of this system is 2N + 1. Let ωi1 , ωi2 and ω0 denote the
state (i, 0), (0, i), and (0, 0), respectively, where
i = 1, 2, . . . , N .
1 and 2 represent sets of ωi1 and ωi2 . Hence,
 = 1 , ω0 , 2 is the state set of this model. The steady
state probability, W, can be calculated as follows:
FIGURE 3. Illustration of the BENC strategy. WT 3 = WT , (1)
where 3 is the one-step transition probability matrix as shown
stores data of flow 1 or flow 2 at any given time. To clarify at the top of the next page.
the scheme, Fig. 3 shows an example. The first part of the The result of the steady state probability is expressed as:
figure shows that a new arrival will enter the queue when the n o λjN +i λN −i
j 3−j
buffer is available. The second part shows that a new arrival P ωi = , j = 1, 2 (2)
2N
will evict the head of the queue when the buffer is full. The P
λ1 λ2
2N −k k
third part shows that a new arrival of a different type will free k=0
one slot in the queue through network coding. Subsequently, λN
1 λ2
N
the two encoded packets depart immediately. P {ω0 } = , (3)
2N
λ2N λ2
−k k
P
1
B. THE BUFFER CAPACITY OF BENC k=0

This study does not consider the time required to gen- where i ranges from 1 to N .
erate an encoding packet (which will depart the queue
immediately) in the following analysis. Furthermore, only B. THE EXPECTATION OF TIME INTERVAL
two independent input flows are considered. The capacity The distribution of the time interval for flow 3 can be
of both virtual buffers is N . The elements (k, l) represent written as:
n o X n o
the system state, which consists of the lengths of the virtual P T̂ > t = P T̂ > t ω P {ω} . (4)
buffers. Here, k and l stand for the buffer lengths of flow 1 ω∈

5992 VOLUME 4, 2016

X. Li et al.: Queueing Characteristics of the BENC Strategy

λ1
 λ2 
 λ1 + λ2 λ1 + λ2
 λ1 λ2


 0 0 
 λ1 + λ2 λ1 + λ2
 
λ1 λ2

 
 0 
3=
 λ1 + λ2 λ1 + λ2 
.. .. ..

. . .
 
 
λ1 λ2 
 

0 0
λ1 + λ2 λ1 + λ2 
 

 λ1 λ 2

λ1 + λ2 λ1 + λ2

The conditional probability where the time interval T̂ is C. THE EXPECTATION OF NETWORK
greater than t can be calculated as follows: CODING TIME INTERVAL
N −i
k The output of flow 3 consists of both network coded packets
n
j
o
−(λ1 +λ2 )t
X λj t
P T̂ > t ωi = e , j = 1, 2 (5) and uncoded packets. The time interval between network
k!
k=0 coded packets is defined as TNC . The cardinality of the state
N
set  is 2N + 1 when the buffer length is N for each flow.
!
n o
−(λ1 +λ2 )t
X (λ1 t)k + (λ2 t)k
P T̂ > t ω0 = e −1 , The buffer length state changes when a new arrival packet is
k!
k=0 processed. It is only when the state is ωN1 or ωN2 that departing
(6)
packets could be uncoded. Otherwise, departing packets will
where i ranges from 1 to N . always be encoded. The conditional probability that TNC is
The conditional probability can be expressed in vector greater than t is calculated as follows:
form as follows: n
j
o
P TNC > t| ωi = e−λ3−j t , j = 1, 2 (11)
8
 n o n o n o T P {TNC > t|ω0 } = e−λ1 t + e−λ2 t − e−λ1 t e−λ2 t , (12)
P T̂ > t ωN1 , . . . , P T̂ > t ω11 , P T̂ > t ω0 ,
= n o n o  . where i ranges from 1 to N − 1.
P T̂ > t ω1 , . . . , P T̂ > t ωN
2 2
The probability that TNC is greater than t is the sum of
(7) the conditional probability multiplied by the steady state
Then, Eqn. (4) can be simplified to probability, as expressed in Eqn. (13).
NC > t|ωi PNC ωi
1
n o X n o N −1 P T
  1 !
P T̂ > t = P T̂ > t ω P {ω} = WT 8.
X
(8) P {TNC > t} =
+ P TNC > t|ωi2 PNC ωi2
 
ω∈ i=1
The probability distribution of the random variable T̂ is + P {TNC > t|ω0 } PNC {ω0 } . (13)
denoted by FT̂ (t), which is defined as follows:
n o However, the steady state probability (PNC ) is no
FT̂ (t) = P T̂ < t = 1 − WT 8. (9) longer the system steady state probability presented
in Eqns. (2 and 3). PNC can be easily calculated by steady
Then the expectation of the random variable T̂ can be state probability using the following equations:
obtained as shown below: n o
j
h i n o P ωi
j
E T̂ PNC ωi = , j = 1, 2
−1
NP
P ωk + P ωk + P {ω0 }
Z ∞  1  2

d h i
= t· 1 − WT 8 dt k=1
t=0 Z dt
∞   (14)
d
= WT · t · − 8 dt P {ω0 }
t=0 dt PNC {ω0 } = , (15)
−1
NP
λ1 − λ2 P ωk1 + P ωk + P {ω0 }
  2

= 2N +1
λ − λ2N +1 k=1
 1 2
where i ranges from 1 to N − 1.

 λ1 λ2N
N −1
2 N +1 λN1
−1
−λN 2
−1 
+λ2 The conditional probability and the steady state probability

 

(λ ) N λ λ

+λ −
 

1 2 1 2
can be expressed in vector form as follows:
 
λ λ N +2 N +2
 2 2

N

+λ +λ (λ +λ ) −λ −λ

1 2 1 2
· +λN λ2
−1 N −1 1 2 1 2 . T
PNC ωN1 −1 , · · · , PNC ω1 , PNC {ω0 } ,
   1
1 N +1

 (λ 1 +λ 2 ) 
 WNC = ,
+ λ2 λ1 +λN +1 λ1 −λ2
N −1 2N N −1 N −1
PNC ω2 , · · · , PNC ω2

 
 
 −λ t 1 −λ t −λ t N −1

 

 T
N 1 λ1 −λ2 e 2 , · · · , e 2 , e 1 +e−λ2 t −e−λ1 t e−λ2 t ,
  
(λ1 +λ2 )
GNC = −λ1 t .
(10) e ,··· ,e 1 −λ t

VOLUME 4, 2016 5993

 X. Li et al.: Queueing Characteristics of the BENC Strategy

Then, Eqn. (13) can be simplified to E. MEAN PACKET WAIT TIME

Little’s Law holds if a queueing process is regenerative
P {TNC > t} = WTNC GNC , (16) [24]. The BENC process is regenerative, thus, the mean wait
j j
time, WQ , for each flow can be calculated by LQ = λj WQ ,
and the expectation of the random variable TNC can be
where j is 1, 2.
obtained as follows:
−1
NP k
λN −1−k
λk−i
j λ3−j
i
P
E[TNC ] j
j k=0 i=0
WQ = λN ,
Z ∞
d h i
j j = 1, 2. (19)
= t· 1 − WTNC GNC dt 2N
λ2N λ2
−k k
P
t=0 dt 1
Z ∞   k=0
T d
= WNC · t · − GNC dt
t=0 dt IV. PERFORMANCE ANALYSIS AND SIMULATION
−λ1 2N +1
−λ1 λ2 +λN
2N
1
+1 N
2 −λ1 λ2
λ2 +λ1 λ2N N N +1 +λ2N +1
2 A. THE NETWORK CODING GAIN
=
 h i
(λ1 +λ2 ) −λ2N λ λ2N The expectation rate, λ̂, of flow 3 is λ̂ = 1/ E T̂ . The


1 2 +λ 1 2
2N −1 result is presented in Eqn. (20) when λ1 and λ2 are equal to λ
(λ1 + λ2 ) λ1 λ2 − λN
2N −1−i i
1 λ2
N
P
in Eqn. (10):
i=0
= . (17) 2N + 1
2N −2 λ̂ = λ . (20)
(λ1 + λ2 ) λ1 λ2 λ1 λ2
2N −2−i i
P 1 1
2N
+ 2N − 2
i=0

Assuming λ2 if greater than λ1 , according to Eqn. (17), 1) WHEN THE BUFFER IS ABSENT
with the increase of λ2 or buffer capacity N , the expectation The rate of flow 3 is λ̂ = 2λ when N is 0. Because λ1
of time interval of encoded packets tends toward 1/λ1 , due to equals λ2 , the rate of flow 3 is equal to λ1 + λ2 . In this
the rate of encoded fraction of total output flow is determined case, the process of flow 3 is the sum of the two independent
by the rate of slower input flow, which is λ1 . exponential distributed processes with parameters λ1 and λ2 .
The result of λ̂ can then be verified.
D. MEAN WAITING QUEUE LENGTH
2) WHEN THE CAPACITY OF THE BUFFER IS INFINITE
The mean waiting queue lengths, LQ1 and LQ2 , for
flows 1 and 2, respectively, can be obtained by using the The asynchronous network coding process of flow 3 will
system steady state probability as follows: degenerate to the synchronous network coding. The rate
of the asynchronous network coding asymptotically tends
N n o toward λ as the capacity of the buffer increases to infinity,
j j
X
LQ = 0 × P {ω0 } + i × P ωi which can be calculated as follows:
i=1 2N + 1
h i λ̂ = lim λ 1 = λ. (21)
λN
j
+1
N λj
N +1
− (N + 1) λN
j λ3−j + λ3−j
N +1
N →∞ + 2N − 21
2N
=   ,
(λ1 − λ2 ) λ2N
1
+1
− λ2N
2
+1 In this case, the encoding node reaches the maximum
coding gain, which is 2. Meanwhile, the result also implies
j = 1, 2.
that the output flow rate of a synchronous network coding is λ
−1
NP k
λN λk−i when the two independent Poisson processes have the same
j λ3−j
−1−k P i
j
+1 k=0 i=0 rate, λ. The same result is shown in [14] as well.
= λN
j (18)
2N
λ2N λ2
−k k
P
1 3) WHEN THE CAPACITY OF BUFFER IS N
k=0
From the preceding analysis, the maximum network coding
The mean waiting queue length varies with the rate of input gain is 2 when two independent processes have the same rate.
flow and the capacity of queue. Assuming the capacity of The BENC strategy will achieve the maximum gain when
queue is fixed, according to Eqn. (18), the queue length of the buffer capacity is infinite. However, in real networks,
flow j decreases with the increase of rate of flow 3 – j, when this theoretical maximum is obviously unreasonable. A better
the rate of the flow 3−j is higher than the rate of flow j, where j question is: How large a buffer is required to achieve 99% of
has a value of 1 or 2. The reason is as follows. The packets the maximum coding gain? The answer is shown below.
from flow 3 − j arrives faster than the packets from flow j, The network coding gain is the ratio of the number of
therefore, the packets from flow j stay in the waiting queue forwarding transmissions required to the number of transmis-
for a short time, and then encoded packets are generated sions required by the network coding strategy to deliver the
and depart the queue, as a result, the queue length of flow j packets with the same size. The output of flow 3 consists of
decreases. On the contrary, the queue length of flow 3 − j both encoded and uncoded packets. The output rate of flow 3
increases. is the sum of the network coding rate, λNC , and the uncoded

5994 VOLUME 4, 2016

X. Li et al.: Queueing Characteristics of the BENC Strategy

flow rate. The network coding gain can then be calculated by

(2λNC + λ̂ − λNC )/λ̂ = 1 + λNC /λ̂. The expectation rate
is defined as λNC = 1/(E[TNC ]). The result is presented
in Eqn. (22) when the parameters λ1 and λ2 are equal to λ
in Eqn. (17):
4N − 2
λNC = λ. (22)
4N − 1
The buffer size N is the smallest integer that is no less than
the result from solving Eqn. (23) to achieve x percent of the
maximum coding gain. Therefore, buffer sizes of 50, 10 and 4
can reach 99%, 95% and 90% of the maximum coding gain,
respectively.
 
(4N − 2) 21N + 2N − 21 FIGURE 6. Time interval distribution with N = 4.
1+ = 2 × x%. (23)
(4N − 1) (2N + 1)

B. SIMULATION
The theoretical results discussed in Section 3 such as the
expectation time interval of flow 3, the waiting queue length
for each flow and the network coding gain are investigated
in this section. In particular, the situation when λ1 equals λ2
is analyzed here. The distribution of these two input flows is
exponential. Without loss of generality, λ1 = λ2 = λ = 1 is
assumed, which can be easily achieved by properly choosing
a time scale. The duration of every simulation is at minimum
100,000 time units.

FIGURE 7. Time interval distribution with N = 10.

FIGURE 5. Time interval distribution with N = 1.

1) THE TIME INTERVAL DISTRIBUTION OF FLOW 3

The time interval distribution of flow 3 depends on λ1 , λ2 FIGURE 8. The output flow rate with two asymmetric flow rates.
and the waiting queue length N in Eqn. (10). In addition,
the hazard function of the time interval for flow 3 is no
longer a constant; therefore, flow 3 is no longer a Poisson exponential, as shown in Fig. 6 and Fig. 7. When the capacity
process [24]. The exponential distribution curves in Fig. 5, is greater than 4, the distribution is almost exponential.
Fig. 6, and Fig. 7 are obtained by curve fitting method based Fig. 8 shows the output flow rate with asymmetric source
on the simulation data. Fig. 5 shows that when the queue rates. The rate ratio of the two flows ranges from 1 to 10.
capacity is 1, the time interval frequency histogram does not The output rate is almost equal to the rate of the fastest flow,
match the exponential distribution. However, as the capacity which means that the packets from the slower flow are nearly
of waiting queue increases, the distribution is asymptotically entirely encoded.

VOLUME 4, 2016 5995

 X. Li et al.: Queueing Characteristics of the BENC Strategy

FIGURE 9. Mean waiting queue lengths. FIGURE 11. Mean waiting queue length with N = 10.

2) MEAN WAITING QUEUE LENGTHS

OF FLOW 1 AND FLOW 2
When λ1 = λ2 = λ, Eqn. (18), which calculates the mean
waiting queue lengths of flow 1 and flow 2, can be simplified
to LQ1 = LQ2 = N (N + 1)/2 (2N + 1). The theoretical result
shows that the mean waiting queue lengths of flow 1 and
flow 2 are equal and are related only to the queue capacity.
Fig. 9 shows the mean waiting queue length plotted as a
function of the queue capacity. The solid theoretical line lies
between the two simulation lines; however, the simulation
results are extremely close to the theoretical result.

FIGURE 12. Simulation of network coding gain.

3) NETWORK CODING GAIN

The network coding gain is illustrated in Fig. 12. The dia-
mond and circle lines respectively represent the simulated
and theoretical results of the network coding gain. The
simulation results show that the network coding gains are
1.798 and 1.908 when N equals 4 and 10, respectively. This
simulation verifies the theoretical result for network coding
gain. That is to say, if 90% of the maximum coding gain
is sufficient for a communication system, the length of the
FIGURE 10. Mean waiting queue length with N = 4.
waiting queue for each flow needs to be only 4. The output
Fig. 10 and Fig. 11 show the mean waiting queue lengths flow rate, λ̂, is 1.19λ when the network coding gain is 1.8.
plotted as functions of the simulation time. The queue capac- The mean result of the network coding gain from simu-
ities in Fig. 10 and Fig. 11 are 4 and 10, respectively. The lation is consistent with the theoretical analysis. Therefore,
solid lines present the theoretical results. The dashed lines the BENC strategy is an efficient and succinct way to obtain
show the results for queue 1, while the dotted lines show sufficient network coding gain at the encoding node.
the results for queue 2. The theoretical mean waiting queue
lengths are LQ1 = LQ2 = 4 (4 + 1)/2 (2 × 4 + 1) ≈ 1.1111 V. CONCLUSION
and LQ1 = LQ2 = 10 (10 + 1)/2 (2 × 10 + 1) ≈ 2.6190 In this study, the queueing characteristics of network coding
when the queue capacities are 4 and 10, respectively. The have been investigated. A novel strategy called Best Effort
simulation results show that the mean queue length oscillates Network Coding (BENC) was designed and formulated using
between 0 and the queue capacity at the beginning, but tends a Markov chain. The problem of the distribution of the time
to stabilize quickly. Hence, the system will be stable. interval for the output flow has been solved. In addition,

5996 VOLUME 4, 2016

X. Li et al.: Queueing Characteristics of the BENC Strategy

the expectations were presented in analytical form, including [20] E. C. Davri, K. Kontovasilis, and C. Skianis, ‘‘Optimized network coding
the time interval for the output flow, the network coding efficiency under QoS constraints in two-way relay networks with time-
outs,’’ in Proc. IEEE Int. Conf. Commun. (ICC), London, U.K., Jun. 2015,
time interval and the mean waiting queue length. Moreover, pp. 6157–6162.
the performance of this strategy in one particular case has [21] M. Zohdy, T. Elbatt, and M. Nafie, ‘‘Maximum throughput opportunistic
been analyzed and validated through simulation. As shown network coding in two-way relay networks,’’ in Proc. IEEE Int. Conf.
Commun. Workshop (ICCW), London, U.K., Jun. 2015, pp. 913–918.
by the simulation, the waiting queue system is stable using [22] Y. Ma, W. Li, P. Fan, and X. Liu, ‘‘Queuing model and delay analysis on
the BENC strategy. The distribution of the output flow is network coding,’’ in Proc. IEEE Int. Symp. Commun. Inf. Technol. (ISCIT),
asymptotically exponential as the queue capacity increases. Beijing, China, Oct. 2005, pp. 112–115.
[23] A. Alsebae, M. S. Leeson, and R. Green, ‘‘Performance of a network
Furthermore, network coding gain is achieved efficiently by coding queuing model with deterministic service,’’ in Proc. 9th Int. Symp.
the BENC strategy without requiring complex encoding cal- Commun. Syst., Netw. Digit. Signal Process. (CSNDSP), Manchester,
culations at the encoding node. Jul. 2014, pp. 988–993.
[24] S. M. Ross, ‘‘The exponential distribution and the poisson process,’’
in Introduction to Probability Models, 11th ed. San Diego CA, USA:
REFERENCES Academic, 2014.
[1] T. Ho and D. Lun, Network Coding: An Introduction. Cambridge, U.K.:
Cambridge Univ. Press, 2008, pp. 1–5.
[2] W. K. Chen, ‘‘Maximum Flow in Networks,’’ in Theory of Nets: Flows in
Networks. England, U.K.: Cambridge Univ. Press, 1990.
[3] R. Ahlswede, N. Cai, S.-Y. R. Li, and R. W. Yeung, ‘‘Network information
flow,’’ IEEE Trans. Inf. Theory, vol. 46, no. 4, pp. 1204–1216, Jul. 2000.
[4] S. Chachulski, M. Jennings, S. Katti, D. Katabi, ‘‘Trading structure for
randomness in wireless opportunistic routing,’’ ACM SIGCOMM Comput. XIANXU LI received the B.S. degree in com-
Commun. Rev., vol. 37, no. 4, pp. 169–180, Oct. 2007. munication and information system from the
[5] S. Katti, H. Rahul, W. Hu, D. Katabi, M. Médard, and J. Crowcroft, ‘‘XORs Beijing University of Posts and Telecommunica-
in the air: Practical wireless network coding,’’ ACM SIGCOMM Comput.
tions, Beijing, China, in 2010. He is currently
Commun. Rev., vol. 36, no. 4, pp. 243–254, Oct. 2006.
pursuing the Ph.D. degree in communication
[6] S. Zhang, S. C. Liew, and P. P. Lam, ‘‘Hot topic: Physical-layer network
coding,’’ in Proc. 12th Annu. Int. Conf. Mobile Comput. Netw. (MobiCom), and information system with Beihang University,
2006, pp. 358–365. Beijing. His research interest includes satellite
[7] P. Chaporkar and A. Proutiere, ‘‘Adaptive network coding and scheduling navigation, satellite networks with crosslinks, and
for maximizing throughput in wireless networks,’’ in Proc. 13th Annu. the applications of network coding.
ACM Int. Conf. Mobile Comput. Netw. (MobiCom), 2007, pp. 135–146.
[8] Y. E. Sagduyu and A. Ephremides, ‘‘On joint mac and network coding
in wireless ad hoc networks,’’ IEEE Trans. Inf. Theory, vol. 53, no. 10,
pp. 3697–3713, Oct. 2007.
[9] Y. P. Hsu, N. Abedini, N. Gautam, A. Sprintson, and S. Shakkottai,
‘‘Opportunities for network coding: To wait or not to wait,’’ in Proc. IEEE
Int. Symp. Inf. Theory (ISIT), Jul./Aug. 2011, pp. 791–795.
[10] X. He and A. Yener, ‘‘On the energy-delay trade-off of a two-way relay
network,’’ in Proc. 42nd Annu. Conf. Inf. Sci. Syst. (CISS), Princeton, NJ, QING CHANG received the B.S. degree in
USA, Mar. 2008, pp. 865–870. mathematics from the Beijing Normal University
[11] S. Chieochan, S. Hossain, T. Issariyakul, and D. Niyato, ‘‘Opportunistic in 1982, and the M.S. degree in mathematics and
network coding and dynamic buffer allocation in a wireless butterfly
the Ph.D. degree in communication and informa-
network,’’ in Proc. Global Telecommun. Conf. (GLOBECOM), Honolulu,
tion system from Beihang University, in 1991 and
HI, USA, Nov./Dec. 2009, pp. 1–6.
[12] O. H. Abdelrahman and E. Gelenbe, ‘‘Performance trade-offs in a network
1995, respectively. Since 2000, he has been a Pro-
coding router,’’ in Proc. 19th Int. Conf. Comput. Commun. Netw. (ICCCN), fessor with the School of Electronic and Informa-
Zurich, Switzerland, Aug. 2010, pp. 1–6. tion Engineering, Beihang University. He is the
[13] L. H. Ding, F. Yang, L. Qian, L. Liu, and P. Wu, ‘‘Cost-delay trade off Dean of Teaching-Research Section of Communi-
of network coding in asymmetric two-way relay networks,’’ in Proc. Int. cation and Information System in his school. He is
Symp. Netw. Coding, Beijing, China, Jul. 2011, pp. 1–4. the author of one book, over 30 articles, and over ten patents. His research
[14] Y. Ma, W. Li, P. Fan, K. B. Letaief, and X. Liu, ‘‘On the Characteristics of interests include satellite communication and navigation, and wireless
Queueing and Scheduling at Encoding Nodes for Network Coding,’’ Int. networks.
J. Commun. Syst., vol. 22, no. 6, pp. 755–772, Jun. 2009.
[15] A. Paramanathan, D. E. Lucani, and F. H. P. Fitzek, ‘‘On the need for buffer
optimization strategies for inter-session network coding,’’ in Proc. IEEE
Int. Conf. Commun. Workshops (ICC), Sydney, NSW, Australia, Jun. 2014,
pp. 343–348.
[16] Y. Yuan, K. Wu, W. Jia, and Y. Peng, ‘‘On the queueing behavior of inter-
flow asynchronous network coding,’’ Comput. Commun., vol. 35, no. 13,
pp. 1535–1548, Jul. 2012. YONG XU received the B.S. degree in com-
[17] Y. P. Hsu, N. Abedini, N. Gautam, A. Sprintson, and S. Shakkottai,
munication from Ji Lin University in 2005, and
‘‘Opportunities for network coding: To wait or not to wait,’’ in Proc.
the Ph.D. degree in electrical engineering from
IEEE Int. Symp. Inf. Theory (ISIT), St. Petersburg, Russia, Jul./Aug. 2011,
pp. 791–795. the Beihang University, China, in 2012. He has
[18] A. Mohapatra, G. Natarajan, S. Srinivas, and S. Alex, ‘‘Network cod- been a Post-Doctoral Fellow with the Depart-
ing decisions for wireless transmissions with delay consideration,’’ IEEE ment of Electronic Engineering, Beihang Uni-
Trans. Commun., vol. 62, no. 8, pp. 2965–2976, Aug. 2014. versity in 2012. His professional research fields
[19] E. N. Ciftcioglu, Y. E. Sagduyu, R. A. Berry, and A. Yener, ‘‘Cost-delay include communication and navigation theory and
tradeoffs for two-way relay networks,’’ IEEE Trans. Wireless Commun., technology.
vol. 10, no. 12, pp. 4100–4109, Dec. 2011.

VOLUME 4, 2016 5997