Content uploaded by K. Robert Lai
Author content
All content in this area was uploaded by K. Robert Lai on Nov 09, 2015
Content may be subject to copyright.
Photon Netw Commun (2013) 25:156–165
DOI 10.1007/s11107-013-0399-x
Genetic expression programming: a new approach for QoS traffic
prediction in EPONs
Jhong-Yue Lee ·I-Shyan Hwang ·Andrew Tanny Liem ·
K. Robert Lai ·AliAkbar Nikoukar
Received: 16 April 2012 / Accepted: 5 April 2013 / Published online: 24 April 2013
© Springer Science+Business Media New York 2013
Abstract The Ethernet passive optical network is being
regarded as the most promising for next-generation optical
access solutions in the access networks. In time division mul-
tiplexing, passive optical network technology (TDM-PON)
and the dynamic bandwidth allocation (DBA) play a crucial
key role to achieve efficient bandwidth allocation and fairness
among subscribers. Therefore, the traffic prediction in DBA
during the waiting time must be put into the account. In this
paper, we propose a new prediction approach with an evo-
lutionary algorithm genetic expression programming (GEP)
prediction incorporated with Limited IPACT referred as GLI-
DBA to tackle the queue variation during waiting times as
well as to reduce the high-priority packet delay. Simulation
results show that the GEP prediction in DBA can reduce
the expedited forwarding (EF) packet delay, shorten the EF
queue length, enhance the quality of services and maintain
the fairness among the optical network units (ONUs). We
conducted and evaluated the detail simulation in two dif-
ferent scenarios with distinctive traffic proportion. Simula-
tion results show that the GLI-DBA has EF packet delay
J-Y. Lee (B
)·I-S. Hwang ·A.T. Liem ·K. R. Lai ·A. Nikoukar
Department of Computer Science and Engineering,
Yuan-Ze University, Chung-Li32003, Taiwan
e-mail: jylee@saturn.yzu.edu.tw
A.T. Liem
e-mail: s999112@mail.yzu.edu.tw
K. R. Lai
e-mail: krlai@saturn.yzu.edu.tw
A. Nikoukar
e-mail: s1009117@mail.yzu.edu.tw
I-S. Hwang
Department of Information Communication, Yuan-Ze University,
Chung-Li32003, Taiwan
e-mail: ishwang@saturn.yzu.edu.tw
improvement up to 30% over dynamic bandwidth allocation
for multiple of services (DBAM). It also shows that our pro-
posed prediction scheme performs better than the DBAM
when the number of ONUs increases.
Keywords Genetic expression programming ·Ethernet
passive optical networks ·Dynamic bandwidth allocation ·
Quality of services ·Traffic prediction
1 Introduction
With the rapid development of various network services,
users’ requirements for access network bandwidth and qual-
ity of service (QoS) are increasing. To this end, the required
technologies should be inexpensive, simple, scalable and
capable of delivering Triple-Play-Services to the end users
over a single network that are widely deployed for the
access networks. Optical fibers have indisputably become
the communication media that satisfies the ever-increasing
bandwidth hungry applications such as high-quality audio-
visuals, data transfer and services that are offered and QoS
required via the Internet. Among many types of broad-
band access networks, the Ethernet passive optical networks
(EPONs) appear to be an attractive solution for the next-
generation broadband access networks to support full service
to the end users [1,2]. Moreover, the EPON has become an
evolving access network technology that delivers a low-cost
method of deploying optical access lines between a central
office (CO) and the end users [3].
An EPON consists of an optical line terminal (OLT) which
is located at the central office and multiple optical network
units (ONUs) near the end user’s sites. The OLT connects
a group of associated ONUs over the point-to-multipoint
topologies to deliver broadband packets and reduces costs
123
Photon Netw Commun (2013) 25:156–165 157
relative to the maintenance power. In addition, the OLT has
the entire channel bandwidth to broadcast the control mes-
sages and data packets to each ONU because of the direc-
tional properties of the splitter or coupler. However, in the
upstream direction, the entire ONUs must share the com-
mon transmission channel toward the OLT, and only a single
ONU may upload upstream data in its transmission timeslot
to avoid data collisions. Therefore, the IEEE 802.3ah stan-
dard has developed a multipoint control protocol (MPCP).
The MPCP relies on two Ethernet control messages,
GATE and REPORT, to allocate bandwidth to each ONU.
The GATE message is used by the OLT to allocate the
upstream transmission window to each ONU. The REPORT
message is used by ONUs to report its local queue length
to the OLT. After receiving all the report messages from
entire ONUs, the OLT executes a dynamic bandwidth allo-
cation (DBA) to calculate and allocate the time slots for
each ONU. The MPCP does not specify any specific DBA
algorithm, but simply provides a framework for the vari-
ous DBA algorithms to be implemented. Moreover, several
approaches were studied, such as wavelength division mul-
tiplexing (WDM) and time division multiplexing (TDM).
Between these two approaches, the TDM is currently con-
sidered as more cost-effective than the WDM [4].
In TDM, the DBA plays a key role to provide more effi-
cient bandwidth allocation for each ONU to share network
resources and offers the better quality of services for the end
users. The DBA assigns the bandwidth dynamically based on
queue state information received from ONUs. In the tradi-
tional DBA scheme, the OLT will execute bandwidth alloca-
tion after receiving all REPORT messages from entire ONUs,
which leads to the waiting time problem. The waiting time
is the time interval between sending the REPORT message,
and the last bursts out of the buffered frames of the times-
lots that potentially increases the packet delay because more
frames from end users are queued at ONUs. Thus, frames
that arrived during the waiting times have to be delayed to
the next transmission cycle, although the upstream channel
still has free timeslots. Consequently, reducing the waiting
time in DBA schemes becomes one of the important issues
to address in order to improve bandwidth utilization.
Previously, DBA without prediction mechanism, such as
LBA was studied in [5]. The LBA can prevent the upstream
channel monopolization by limiting the maximum timeslots,
Bmax, to each ONU, which can be specified by the service
level agreement (SLA). When the reported queue size is less
than Bmax, the OLT grants the requested bandwidth; other-
wise, Bmax is granted. The LBA has been poor utilization and
restricts aggressive competition for the upstream bandwidth,
especially under non-uniform traffic [6]. Moreover, in [7], the
authors classify the requests of different ONUs into lightly
loaded and heavily loaded according to the amount of traf-
fic requests. In each transmission cycle, several ONUs may
have fewer traffic requests to transmit, which is less than the
minimum guaranteed bandwidth (referred as lightly loaded
ONUs), while other ONUs may have a larger traffic request
to transmit, thus need more bandwidth (referred as heavily
loaded ONUs). As a result, one feasible method involving
prediction is to grant more bandwidth than the requested by
lightly loaded ONUs to improve overall fairness.
Currently, various DBA algorithms with the predictive
scheme have been proposed [5–16] to tackle the waiting
time problem, such as credit-based [5,8,9], linear-based [6,9,
10], proportion-based [6,11], waited-based [12], QoS-based
[8,11,13–15]and nonlinear-based [16]. The prediction-based
schemes are studied to decrease the packet delay and allocate
more time slots efficiently. In these schemes, the traffic pre-
dicted index is employed to update the allocated bandwidth to
meet the QoS requirements. Hence, while the OLT allocates
the request bandwidth to ONUs, a credit will be added into
the time slots of each ONU, as a result the frame that arrives
during the waiting time is expected to be transmitted within
the current time slots. An accurate traffic predictor index is
required to avoid over or under estimation, which will result
in longer packet delay and reduce the network performance.
The credit-based bandwidth allocation (CBA) takes some
precedent transmitted frames into consideration [5,8,9], and
it adds a credit into the requirement of each ONU when
the OLT allocates the upstream bandwidth. The bandwidth
granted to each ONU can be calculated as Bgrant =Bqueue +
C, where Bgrant is the bandwidth granted to an ONU, Bqueue
denotes the frames queue in the buffer and Cis the credit
index that could be a constant or linear credit. The CBA grants
the requested window plus a credit that is proportional to the
requested window. Therefore, some packets can be expected
to transmit with the current grant time slots, hence the aver-
age packet delay can be reduced.
Furthermore, the DBA with multiple services (DBAM)
[12] is a waited-based prediction LBA that executes predic-
tion according to the linear estimation credit. The linear esti-
mation credit of each ONUiis obtained according to the ratio
of the ONUiwaiting time during the interval nover the time
length of the current interval n. In other words, the frames
arriving in the waiting time of interval (n+1) are estimated
from the information of interval n. The OLT allocates the
time slots for multiple services among ONUs according to
each required bandwidth and the SLA limits. As a result, the
DBAM can improve the packet delay in uniform traffic flows.
However, in non-uniform traffic flows, the DBAM prediction
model suffers serious inaccuracy because of high variation’s
traffic in some ONUs. These aforementioned schemes are
unable to improve the differentiated services (DiffServ) and
also unable to tackle the queue size inconsistency problem.
Finally, a DBA with neural network prediction approach
was proposed in [16]. A pipeline recurrent neural network
(PRNN)/extended recursive least square (ERLS) update pre-
123
158 Photon Netw Commun (2013) 25:156–165
dictor is adopted accurately to predict the arrival frames at
ONU during the next prediction interval in each cycle; thus,
the bandwidth allocation is expected to be up-to-date and
accurate [16].
In this paper, a new approach to tackle the waiting time
issue with the genetic expression programming (GEP) for the
network traffic prediction is proposed. The GEP is a novel
adaptive evolutionary algorithm that offers a technique to
mine the relationship between the future and the historical
data directly, which is proposed by Candida Ferreira [17].
Principally, GEP is similar to genetic algorithm (GA) and
genetic programming (GP) in the basic general idea and
evolutionary. GEP is one computational model that discov-
ers knowledge from data and expressed it as a formula. It
employs linear strings as its chromosome rather than tree
forms, which are later expressed as nonlinear entities of dif-
ferent size and entities. Moreover, it uses a very smart method
to decode chromosome as a formula, scans the chromosome
string from left to right. Afterward, it parses a character and
generates an expression node to denote it and append it to
an expression tree from top to bottom, from left to right. The
schematic fundamental steps of GEP process begin with the
random generation of the chromosomes of a certain number
of individuals (the initial population). Afterward, these chro-
mosomes are expressed, and the fitness of each individual
is evaluated against a set of fitness cases (also called selec-
tion environment). The individuals are then selected accord-
ing to their fitness to reproduce with modification, leaving
progeny with new traits. These new individuals are, in their
turn, subjected to the same developmental process: expres-
sion of the genomes, confrontation with the selection envi-
ronment, selection according to fitness and reproduction with
modification. The process is repeated for a certain number of
generations or until a good solution has been found. Further-
more, according to [18], the GEP is considerably powerful
than GA and GP in the following senses: (a) it can solve many
problems that cannot be solved by GA and GP. (b) It is much
faster (100–60,000 times) than GA and GP on solving some
particular problems.
The predictors in GEP are set based on the traits of ONUs,
such as cycle time, waiting time, REPORT, GATE and the
history request of mean bandwidth in every cycle time.. We
conduct detailed simulation experiments using OPNET mod-
eler to study the performance of the proposed scheme and
validate its efficiency. Simulation results show that the GLI-
DBA outperforms the dynamic bandwidth allocation with
multiple services (DBAM) [12] in terms of the EF queue
length, EF packet delay, jitter, fairness and system through-
put. The rest of this paper is organized as follows. In Sect. 2,
we introduce the proposed GLI-DBA scheme. Our simula-
tion results and evaluation of the impact of traffic load with
two different scenarios are described in Sect. 3. We conclude
our work in the Sect. 4.
2 Proposed GLI-DBA
2.1 Genetic expression programming prediction
In this section, a new prediction scheme for dynamic band-
width allocation (DBA) using GEP referred as GLI-DBA is
presented. The GLI-DBA algorithm is proposed to resolve
the queue state inconsistent problem and enhance the QoS
for differentiated services in EPON system. We integrate the
GEP prediction scheme with the limited bandwidth alloca-
tion (LBA) service discipline. Our proposed algorithm not
only supports the QoS but also maintains the fairness among
ONUs in EPONs. Fairness is important because every ONU
should have an equal bandwidth. We used priority queues
(PQ) to deal with three different types of traffic allotment (i.e.,
voice, video and data services) in order to guarantee the high-
priority traffics. The reason is that without PQ, all traffic types
will be treated equally, which will lead to various problems,
particularly the EF traffic with delay sensitive constraint. Our
prediction scheme algorithm classifies the traffics into three
priority queues: expedited forwarding (EF)- highest-priority
traffic; assured forwarding (AF)-medium-priority traffic; and
best-effort (BE) low-priority traffic [5,19]. The voice that is
EF traffic requires packet delay constraint and jitter speci-
fications. Video is the AF traffic which not delays sensitive
but requires bandwidth guarantees. Lastly, the data are the
BE traffic that is not sensitive to delay or jitter. Therefore, to
achieve better system performance, the bandwidth should be
assigned to the ONUs according to the rate of these applica-
tions.
The flowchart of GLI-DBA mechanism is illustrated in
Fig. 1.TheBavailable denotes the initial available bandwidth
for ONUs. Bavailable =r×(TMax
cycle −N×Tg)−N×512,where
ris the transmission speed of the EPON in bit per second,
TMax
cycle is the maximum cycle time, Nis the number of ONUs,
Tgis guard time and REPORT message length is 64 Byte
of the EPON system. The Bmax is upper bounded by the
maximum time slot length of each ONU, which can be spec-
ified by service level agreement (SLA). Firstly, when receive
whole REPORT message of each ONU, the total available
bandwidth and minimum guaranteed bandwidth, Bavailable,
of each ONUs will be calculated. Secondly, the training data
are collected from the OPNET simulator; we perform IPACT
algorithms [5] and retrieve the target value and the predicted
values from the simulation process. The GLI-DBA executes
GEP prediction mechanism by using the DTREG simula-
tor. The process of reproduction, fitness determination and
GEP parameter selection is repeated from generation to gen-
eration, creating a genetically diverse evolving population.
We run V-fold cross-validation model and adjust the GEP
parameter based on the captured training data repeatedly,
until the completion of the fitness determination. Moreover,
limited bandwidth allocation compares the minimum guar-
123
Photon Netw Commun (2013) 25:156–165 159
Start GLI-DBA
Run the IPACT in OPNET simulator
Receive REPORT message
Initialize B
available
and calculate B
max
Grant P
i
P
i
<
B
max
Yes No
Finish GLI-DBA
Capture the training data from OPNET result
e.g: target value and predictor value
Import the training data into DTREG simulator
Run the GEP model based on
V-fold cross
validation model
No
Yes
Find the P
i
from prediction function
Grant B
max
Adjust GEP parameter
e.g:
Operator, Population
size, mutation rate...
Fitness
determination
Create training data GEP prediction mechanism
LBA mechanism
Fig. 1 Flowchart of GLI-DBA mechanism
anteed bandwidth with prediction index value of each ONU.
If Pi<Bmax, then grant Pi, otherwise grant Bmax to each
ONU.
2.2 Problem definition
In the offline-DBA mechanism, the OLT will begin band-
width allocation after collecting all REPORT messages from
each ONU, thus the packet that arrived during the waiting
time will be transmitted in the next timeslot causing the queue
state inconsistent problem, shown in Fig. 2. Therefore, a pre-
dict credit is need to be added to the requested REPORT
scheme, and thus, the incoming packet during the waiting
time is expected to be transmitted within the current times-
lot. The idea behind the time series prediction is that to find
aformulafto predict a future value by its current and pre-
vious history (i.e., “history + current →future”). The key
point to discover formula fis to find a fitness function for
funder GEP frame with respect to the time series data. The
following is the problem specification for time series predic-
tion problems: Problem Specifications—Let x(t)be a time
ONU
i
OLT
packets arrival
REPORT:
ONU
i
GATE:B
grant
=
waiting time
Nth
cycle
(N+1)th
cycle
Queue:
t
1
t
2
'
i
R
−'
i
R
i
R
i
R
i
R
i
R
t
0
Fig. 2 Queue state between waiting time in EPON
series. At time point ti=t0t(0≤i≤n), the time series
data with interval tare denoted as
X=(x0,x1,...,xn)=(x(t0),x(t1),...x(tn)) (1)
The prediction problem for time series is to build a model
computation T, for arbitrary.m,(0≤m),predicts the object
data values ˆxm−xmat tmbased on the data set {xi|i<
m}, such that the relative allotment |ˆxm−xm|is as small as
possible. Equation (2) is the general model to build the model
computation T.
T=f(predictor),(2)
The GEP is implemented as an extension of the traffic
prediction mechanism for IPACT. The Tis specified as the
target set, which the mean queue length of each ONU based
on network traffic loads. The predictor set fis the traits of
ONUs, such as cycle time, waiting time, REPORT, GATE
and history mean bandwidth request in every cycle time.
Moreover, we used the common basic parameters to both
GEP and GP equally as those used by Koza [20]. Differ-
ent individuals in the initial randomly created population
have different chromosomes, defining different expression
trees. Each expression tree is evaluated using all the forecast–
observation pairs of data in the training set, and a “fitness”
of that expression tree is found using a measure of the fore-
cast error. Thus, a set of 2,000 random fitness cases chosen
from the target value were used; and a very simple func-
tion set, composed only of the seven arithmetic operator’s
F={+,−,∗,/,exp(x), sqrt(x), pi}was used. However,
in practice, more arithmetic operators can be added to the
function set to get wider range and higher accuracy, while
the complexity and computational cost are also increasing.
The real random constant from the interval [−10,10] with
the mutation rate of random constant is 0.01. Fitness can
be calculated using any appropriate statistical measure. In
this paper, the fitness function Eq. (3) used to evaluate the
performance of each evolved program is based on the mean
square error (MSE) Eq. (4) and explores the idea of a selec-
tion range and a precision. The selection range is used as
a limit for selection to operate, which the performance of a
program on a particular fitness case contributes nothing to
123
160 Photon Netw Commun (2013) 25:156–165
Tabl e 1 Simulation parameters
Parameters Value
Number of ONUs in the system 16,32
Upstream/downstream link capacity 1 Gbps
OLT-ONU distance (uniform) 10–20 km
Buffer size 10 MB
Maximum transmission cycle time 1 ms
Guard time 5 μs
Computation time of DBA 10 μs
Control message length 0.512 μs
its fitness. The precision is the limit for improvement, as it
allows the fine tuning of the evolved solutions as accurately
as necessary.
fitness =1
1+MSE
N
,(3)
MSE =
N
i=1
(Pi−Ti)2,(4)
where Piis the predicted value for the row iand Tiis the
actual target value and Nis the number of rows in the train-
ing data set. All fitness functions compute fitness scores that
range from 0.0 to 1.0. A fitness of 0.0 means the model unfits
or it is neither worthless nor not viable. A fitness score of 1.0
means the model fits the data perfectly. Thus, for a perfect
fit, the variance value term is zero and fitness is equal to 1.0.
Individuals that are more fit according to one measure are
sometimes less fit according to other measures. We chose
an optimum fitness measure by trial and error, if the fitness
value is greater than 0.9 can be considered as a good results
[21]. In our experiments, we use a selection range of 100 and
a precision of 0.01. The algorithms use a population size of
60, gene head length of 8, mutation rate of 0.0044, insertion
sequence (IS) and replication insertion sequence (RIS) trans-
position rates of 0.1, two-point and one-point recombination
rates of 0.3, recombination and gene transposition rates of
0.1 each and the linking function is addition.
2.3 GEP experiment results
In order to get data for experiments, first we run the LBA-
IPACT [5] in the OPNET simulator and captured the traits of
ONUs as described in the Sect. 2.1. The parameters used in
the simulation are summarized in Table 1. Afterward, these
data are used for our training data in GEP. The parameters
used per run, and our experiment results are summarized in
Tables 2and 3, respectively. All the results give the value
of fitness with high accuracy. The correlation between actual
and predicted is above 0.9 for EF, AF and BE traffic predic-
Tabl e 2 Parameters used per run
Population size 60
Number of genes 4
Genes head length 8
Genes length 33
Maximum number of generations 2000
Generations without improvement 200
V-fold cross-validation 10
Mutation rate 0.044
Inversion rate 0.1
RIS transposition rate 0.1
IS transposition rate 0.1
Two-point recombination rate 0.3
One-point recombination rate 0.3
Random constant per gene 10
Random constants data type Real
Random constants range −10 to 10
Random constants Mutation rate 0.01
Number of fitness cases 200
Capture target interval time 2 ms
Tabl e 3 Experiment results
Class IPACT
EF Generations required 1748
AF Generations required 1997
BE Generations required 1754
EF Variance in input data 3687.1956
AF Variance in input data 4.9179e+006
BE Variance in input data 3.7491e+007
EF mean absolute percentage error 1.7598972
AF mean absolute percentage error 0.0089024
BE mean absolute percentage error 0.001453
tions, respectively. The following is the formulas mined by
GEP for EF, AF and BE traffic predictions in DBA accord-
ingly, where Piis the prediction index value of different traf-
fic types (i∈{EF, AF, BE}). The α1,α
1,β,γ
1,γ
2,γ
3are
the predictor variable which are produced by GEP prediction
mechanism. All prediction formulas are based on historical
allocation of bandwidth plus their related predictor of oper-
ator members.
PEF =EF
grant +α1−α2×EFrequest(5)
PAF =AFgrant +β+1
cycle_time ×AFrequest(6)
PBE =BEgrant ×(BEgrant +γ1)+γ2
×BErequest +(BEgrant −γ3)×BEgrant (7)
123
Photon Netw Commun (2013) 25:156–165 161
3 Simulation results
In this section, we present the results of our simulation experi-
ments conducted to evaluate the EF packet delay, throughput,
EF jitter performance and fairness with distinctive dynamic
bandwidth allocation (DBA) which are dynamic bandwidth
allocation with multiple services (DBAM) [12] and Genetic
Expression Programming Limited IPACT (GLI-DBA) on dif-
ferent scenarios. The system model is set up in the OPNET
simulator with one OLT and 16 or 32ONUs. The downstream
and upstream channels are both 1 Gb/s. The distance from an
ONU to the OLT is assumed to range from 10 to 20 km, and
each ONU has a finite buffer of 10Mbytes. For the traffic
model considered here, an extensive study shows that most
networks can be characterized by self-similarity and long-
range dependence (LRD). This model is utilized to gener-
ate highly bursty BE and AF traffic classes with the Hurst
parameter of 0.7, and packet sizes are uniformly distributed
between 64 and 1,518 bytes. On the other hand, high-priority
traffic (e.g., voice application) is modeled using a Poisson
distribution and the packet size is fixed to 70bytes [22]. In
order to evaluate the effect of high-priority traffic, the pro-
portion of traffic profile is analyzed in different scenarios
by simulating three distinctive scenarios, which are scenario
one: 16 ONUs (EF occupies 20 %, AF occupies 40%, BE
occupies 40 %) referred as “16_244”, (40, 30, 30 %) referred
as “16_433”, and (60, 20, 20 %) referred as “16_622”, and
scenario two: 32 ONUs (EF occupies 20 %, AF occupies
40 %, BE occupies 40 %) referred as “32_244”, (40, 30,
30 %) referred as “32_433”, and (60, 20, 20 %) referred as
“32_622”with different traffic proportions of EF, AF and BE,
respectively. The simulation parameters are summarized in
Table 1.
3.1 Packet delay
Figure 3compares the EF packet delay from ONUs to
OLT between the DBAM and GLI-DBA in scenario one and
scenario two, respectively. The packet delay dis equal to
d=dpoll +dgrant +dqueue, defined in Ref [5], packet delay
(d) consists of polling delay(dpoll), granting delay(dgrant ) and
queuing delay(dqueue). In Fig. 3a, we can observe that both
DBAM and GLI-DBA still have excellent performance of the
traffic load exceeding 90%. However, in Fig. 3b, the DBAM
with the waited-based prediction can reduce the overall EF
traffic delay unless the traffic load exceeds 90% in “32_244”
and “32_433”.The reason is that the waited-based prediction
is unable to predict the queue length inconsistencies with the
traffic loads exceeding 80%. Contrarily, the GEP prediction
scheme is not only able to guarantee the EF traffic delay
constraint below 1.5 ms [23] but also satisfies the EF traf-
fic in “32_244” and “32_433”, respectively. This is because
our proposed GLI_DBA has better prediction accuracy than
Fig. 3 Packet delay in scenario one and two: aEF packet delay for
16ONUs, bEF packet delay for 32ONUs, caverage packet delay for
16ONUs, daverage packet delay for 32ONUs
the DBAM, thus shorten the EF queue length in each ONU
buffer. Thereby, we can observe that the GLI-DBA improves
on the DBAM in terms of EF delay approximately 30 and
16 % in “32_244”, “32_433”, respectively. Nevertheless, in
“32_622”, all the schemes get saturated when the traffic loads
exceeding 80 %. Furthermore, Fig. 3c, d shows the average
packet delay for aforementioned schemes. In scenario one,
the GLI-DBA result is not as good as the DBAM. However,
123
162 Photon Netw Commun (2013) 25:156–165
when the number of ONUs and proportion of EF traffic are
increasing (i.e., in “32_622”), the GLI-DBA begins to out-
perform the DBAM. One possible reason is that the GLI-
DBA could still satisfy the EF request without sacrificing the
medium-/low-priority traffics.
3.2 System throughput
Figure 4depicts the mean system throughput performance
of the DBAM and GLI-DBA for the EF, AF and BE traffics
in different scenario versus the traffic loads. After simulat-
ing both schemes, It can be observed that the DBAM and
GLI-DBA can perform efficient bandwidth assignments in
both scenarios; however, in scenario one, the DBAM has
a fraction higher than GLI-DBA as shown in Fig. 4a. The
DBAM throughputs are above 580Mbps, whereas GLI-DBA
throughputs are exceeding 578 Mbps. The GLI-DBA pre-
diction scheme is demonstrated to have better performance
than the DBAM, particularly in scenario two as illustrated
in Fig. 4b. The system throughputs of both schemes are
about the same level unless the traffic load exceeds 60%.
The throughput of GLI-DBA begins to increase to as high as
over 87% with the traffic load exceeding to more than 80%
in scenario two (i.e., 32_244, 32_433, and 32_622), while the
DBAM is maintained at below 82%. The reason is that the
GLI-DBA made finer estimation and allocated more properly
to the EF, AF and BE traffics as compared with the DBAM
when the numbers of ONUs are increasing.
The value of average improved percentage Pof system
throughput in GLI-DBA is defined as:
P=MeanBwthroughput
GLI−DBA −MeanBwthroughput
DBAM
MeanBwthroughput
GLI−DBA
,(8)
where MeanBwthroughput
xis the mean system throughput
for the xtype of prediction scheme. Figure 4cshows
the improvements of DBAM versus GLI-DBA. It can be
observed that the DBAM prediction performs better in the
light-load network conditions (i.e., scenario one). However,
when the numbers of ONUs and traffic load are increasing,
the GLI-DBA begins to outperform the DBAM in terms of
system throughputs. The GLI-DBA improvement exceeds 4,
7 and 13 % in “32_244”, “32_433”, and “32_622”, respec-
tively, when the traffic loads exceeding 80% as shown in
Fig. 4d. The reasons behind these facts are that when the
numbers of ONUs are expanding, the GLI-DBA has fewer
unused timeslots remainders compared to the DBAM, par-
ticularly when the EF traffic proportion is high.
3.3 Delay variance
The comparison of jitter performance with different scenarios
is depicted in Fig. 5. Delay variance, known as jitter, can
Fig. 4 System throughputs in scenario one and two: aThroughputs for
16ONUs, bthroughputs for 32ONUs, cthroughputs improvements for
16 ONUs DBAM versus GLI-DBA, dthroughputs improvements for
32ONUs GLI-DBA versus DBAM
be define as the packet delay variation of first departed EF
packets between two consecutive transmission windows and
maps the distribution property of the total EF delay sequence
for the ONU [24]. The delay variance σ2is calculated as
σ2=N
i=1
(dEF
i−¯
d)2
Nwhere dEF
iis the delay time of EF
packet iis the average delay time of EF traffic and Nis
the total number of received EF packets. It can be observed
123
Photon Netw Commun (2013) 25:156–165 163
Fig. 5 Delay variance in scenario one and two: aDelay variance for
16ONUs, bdelay variance for 32ONUs
the delay variance for EF traffic increases as the traffic load
increases. The GLI-DBA jitter with 16ONUs is not as good
as the DBAM because the EF fluctuation is more dispersed
compared to the DBAM. One possible reason is that the EF
prediction in our scheme cannot keep the EF packet delay to
be distributed evenly around its mean value, thus the total EF
sequence is not centralized. In scenario two, the GLI-DBA
has about the same level as the DBAM delay variance unless
the traffic load exceeds 90% in “32_244” and “32_433”.
However, in the “32_622” the DBAM has a very high delay
variance with the traffic load exceeding 90% compared to
the GLI-DBA. The reason is that with the proportion of EF
traffic and network load is elevated, all the above schemes
become saturated and unable to satisfy the EF traffic.
3.4 Fairness
The fairness index is the measure of the closeness of the
bandwidth allocation to the desired goals [25]. It calculated
the equivalence of the bandwidth allocation. Thus, when the
bandwidth allocated fairly to each ONU, the allotment is
equal. The fairness index f(0≤f≤1)has been addressed
in [25], which is defined as Eq. (9).
f=N
i=1G[i]2
NN
i=1G2
[i]
,(9)
Fig. 6 Fairness index in scenario one and two: aFairness index for
16ONUs, bfairness index for 32ONUs
where Nis the total number of ONUs and G[i]is the granted
bandwidth of ONUi. Jain’s fairness index f, ranging from
0 to 1, becomes 1 when all ONUs have the equivalent band-
width allocated by the OLT. Figure 6shows the comparison of
fairness in different scenarios between the DBAM and GLI-
DBA. We can see that the DBAM has less fairness index
in every scenario, due to the inaccurate predictions once the
traffic has a high variation. Therefore, it can be observed
that both schemes have less fairness index when the traffic
proportion of AF and BE are 40 % due to high-traffic varia-
tion. Additionally, it also shows that our proposed scheme is
not only suitable for handle varying traffic (i.e., AF and BE
traffics) but also fit for a stable traffic (i.e., EF traffic). Nev-
ertheless, we can observe that the fairness index is fluctuated
due to the presence of highly loaded that could decrease the
fairness index.
3.5 Bandwidth wasted
The wasted bandwidth is due to the prediction model suf-
fer serious inaccuracy. Figure 7shows the comparison of
bandwidth wasted against different scenarios between the
GLI-DBA and DBAM. Figure 7a shows that the DBAM
scheme has been more or less close to zero bandwidth wasted,
whereas the GLI-DBA has wasted exceeds to 200bits. The
fact is that, in scenario one (i.e., 16ONUs), the DBAM pre-
diction scheme seems to be always under estimation. Con-
123
164 Photon Netw Commun (2013) 25:156–165
Fig. 7 Bandwidth wasted in scenario one and two: aBandwidth wasted
for 16ONUs, bbandwidth wasted for 32ONUs
versely, the GLI-DBA shows wasted too many bandwidths
when the traffic loads range between 10 and 100%. Both
under or over estimation can induce longer packet delay and
reduce the system throughputs. Despite the GLI-DBA has
higher bandwidth being wasted compared to the DBAM, the
GLI-DBA still has a higher performance compared to the
DBAM in terms of packet delay and system throughputs. Yet,
as the number of ONUs and EF traffic proportion is increas-
ing, the DBAM demonstrated to have more bandwidth being
wasted, particularly in scenario two (i.e., “32_622”) as shown
in Fig. 7b.
4 Conclusion
In this paper, the vital factors that can improve the over-
all performance of EPON have been discussed and evalu-
ated. The proposed GLI-DBA incorporated with the GEP
has been shown to have higher performance than another
eminent DBA with a prediction scheme, the DBAM (i.e.,
waited-based prediction), in terms of EF packet delay, sys-
tem throughput and fairness. The proposed GLI-DBA can
reduce the EF packet delay even if the number of ONUs
and the proportion of EF traffic are increasing. Additionally,
the GLI-DBA can improve the DBAM system throughput
up to 7, 9 and 13 % in scenario two. The simulation results
have shown that the artificial neural network programming
has a better prediction against the differential traffic char-
acteristic and improved the system utilization compared to
the non-prediction or linear prediction programming. Finally,
although most of the ONUs is highly loaded and the traffic is
fluctuated, the GLI-DBA can still achieve a higher fairness
index in every scenario.
References
[1] Radzi, N.A.M., Din, N.M., Al-Mansoori, M.H., Mustafa, I.S.,
Sadon, S.K.: Intelligent dynamic bandwidth allocation algorithm
in Upstream EPONs. IEEE/OSA J. Opt. Commun. Netw. 2(3),
148–158 (2010)
[2] Choi, S., Park, J.: SLA-aware dynamic bandwidth allocation for
QoS in EPONs. IEEE/OSA J. Opt. Commun. Netw. 2(9), 773–781
(2010)
[3] Lam, C.: Passive optical networks: principles and practice. Aca-
demic, Burlington, MA (2007)
[4] Razavi, R., Guild, K.: Multi-constraints fuzzy-logic-based
scheduling algorithm for passive optical networks. J. Opt. Netw.
8(4), 346–357 (2009)
[5] Kramer, G., Mukherjee, B., Pesavento, G.: Interleaved polling
with adaptive cycle time (IPACT): a dynamic bandwidth distribu-
tion scheme in an optical access network. Photon. Netw.Commun.
4(1), 89–107 (2002)
[6] Hwang, I.S., Shyu, Z.D., Ke, L.Y., Chang, C.C.: A novel early
DBA mechanism with prediction-based fair excessive bandwidth
allocation scheme in EPON. Comput. Commun. 31(9), 1814–
1823 (2008)
[7] Zheng, J.: Efficient bandwidth allocation algorithm for Ethernet
Passive Optical Networks. IEEE Proc. Commun. 153(3), 464–468
(2006)
[8] Choi, S.Y., Lee, S., Lee, T.J., Chung, M.Y., Choo, H.: Double-
phase polling algorithm based on partitioned ONU subgroups for
high utilization in EPONs. IEEE/OSA J. Opt. Commun. Netw.
1(5), 484–497 (2009)
[9] Sue, C.C., Cheng, H.W.: A fitting report position scheme for
the gated IPACT dynamic bandwidth algorithm in EPONs.
IEEE/ACM Trans. Netw. 18(2), 624–637 (2010)
[10] McGarry, M., Maier, M., Reisslein, M.: Ethernet PONs: a survey
of dynamic bandwidth allocation (DBA) algorithms. IEEE Com-
mun. Mag. 42(8), 8–15 (2004)
[11] Kramer, Glen, Mukherjee, Biswanath, Dixit, Sudhir, Ye, Yinghua,
Hirth, Ryan: Supporting differentiated classes of service in Ether-
net Passive Optical Networks. J. Opt. Netw. 1(8), 280–298 (2002)
[12] Luo, Y., Ansari, N.: Bandwidth allocation for multiservice access
on EPON. IEEE Commun. Mag. 43(2), 16–21 (2005)
[13] Hwang, J., Yoo, M.: QoS-aware class gated DBA algorithm for the
EPON system, International conference on advanced technologies
for communications, pp 363–366 (2008)
[14] Assi, C.M., Ye, Y., Dixit, S., Ali, M.A.: Dynamic bandwidth allo-
cation for quality-of-service over Ethernet PONs. IEEE J. Sel.
Areas Commun. 21(9), 1467–1477 (2003)
[15] Chen, J., Chen, B., Wosinska, L.: Joint bandwidth scheduling to
support differentiated services and multiple service providers in
1G and 10G EPONs. IEEE/OSA J. Opt. Commun. Netw. 1(4),
343–351 (2009)
[16] Peng, J.W., Chang, C.J., Tien, P.L.:PRNN/ERLS-based predictive
QoS-promoted DBA scheme for upstream transmission in EPON.
Photon. Netw. Commun. 20(1), 17–26 (2010)
[17] Nadia, N., Ajith, A.: Genetic systems programming: theory and
experiences. Springer, ISBN-10: 3540298495 (2006)
[18] Zuo, J., Tang, C., Li, C., Yuan, C., Chen, A.: Time series pre-
diction based on gene expression programming. In: International
123
Photon Netw Commun (2013) 25:156–165 165
Conference on Advances in WebAge. Information Management
5 th, vol. 3129, 1, pp. 55–56 (2002)
[19] Blake, S., Black, D., Carlson, M., Davies, E., Wang, Z., Weiss,
W.: An architecture for differentiated services. IETF RFC 2475
(1998)
[20] Koza, J.R.: Genetic programming II: automatic discovery of
reusable programs. MIT Press, Cambridge, MA (1994)
[21] Bakhshaii, A., Stull, R.: Deterministic ensemble forecasts using
gene-expression programming. Weather Forecast. 24(5), 1431–
1451 (2009)
[22] Bai, X., Shami, A.: Modelling self-similar traffic for network sim-
ulation. Technical report, NetRep-2005-01 (2005)
[23] ITU-T Recommendation G.114: One-way transmission time, in
series G: transmission systems and media, digital systems and
networks (2000)
[24] Shami, A., Bai, X., Assi, C., Ghani, N.: Jitter performance in
Ethernet Passive Optical Network. J. Lightwave Technol. 43(4),
1745–1753 (2005)
[25] Jain, R., Durresi, A., Babic, G.: Throughput fairness index: an
explanation. ATM Forum/99-0045 (1999)
Author Biographies
Jhong-Yue Lee was born in Hualien, Taiwan,
ROC. He received the B.S. degree in Depart-
ment of Medical Informatics from the Tzu-
Chi University, Hualien, Taiwan, in 2007 and
the M.S. in Computer Science & Engineer-
ing, Yuan-Ze University, Taiwan in 2009 and
is currently pursuing the Ph.D. degree at the
Yuan-Ze University, Chung-Li, Taiwan. His
recent work focuses on the Dynamic Band-
width Allocation in PON and the Integration Assessment of PON with
Broadband WiMAX.
I-Shyan Hwang received the B.S. and M.S.
in electronic engineering from Chung-Yuan
Christian University, Chung-Li, Taiwan, in
1982 and 1984, respectively, and the M.S. and
Ph.D. in electrical and computer engineer-
ing from the State University of New York at
Buffalo, NY, in 1991 and 1994, respectively.
Since February 2007, he has been promoted
as a Full Professor in the Department of Com-
puter Science and Engineering at the Yuan-Ze University, Chung-Li,
Taiwan, and he is with the Department of Information Communica-
tion. His current research interests are high-speed fiber communication,
mobile computing and heterogeneous multimedia wireless networks,
integration assessment of PON with broadband WiMAX, algorithm
design and testing, and load balancing.
Andrew Tanny Liem received the B.S.
degree from the Department of Computer
Science, Adventist University of Indonesia,
Bandung, Indonesia, in 2003, and the M.T.
degree in computer science and engineering
from the Institute of Technology, Bandung,
in 2006. He is currently working toward the
Ph.D. degree at Yuan-Ze University, Chung-
Li, Taiwan. His recent work focuses on NGN
and P2P over EPONs.
K. Robert Lai received the B.S. degree from
National Taiwan University of Science &
Technology in 1980, the M.S. degree from
Ohio State University, Columbus, OH, in
1982, and the Ph.D. degree in computer sci-
ence from North Carolina State University,
Raleigh, NC, in 1992. From 1983 to 1989,
he was a senior engineer with GE Aerospace
Division, Maryland. In 1994, he joined the
Department of Computer Science & Engineering, Yuan-Ze University,
Taiwan, where he is now a professor. His current research interests are in
computational intelligence, agent technologies and mobile computing.
AliAkbar Nikoukar received the B.S. in
Mathematics and M.S. in Computer Science.
Currently, he is Ph.D. student in Computer
Science & Engineering, Yuan-Ze University,
Taiwan. His current research interests are
high-speed network, NGN and multimedia
service over fiber optic networks.
123
