Performance Analysis on Gain Prediction of Dual-Radio Aggregation System in Heterogeneous Network

Abstract

Nowadays, most of the smartphones or mobile devices are already equipped with multiple interfaces. In general, the utility of multiple interfaces concurrently will escalate the device’s performance of speed communication. In spite of that, an issue called performance anomaly when a new station (STA) with multiple interfaces aggregation and low transmission rate desires to connect with an access point (AP) exists. We can assume that a new STA connects to an AP which already has several associated STAs. Before turning an additional interface on for that new STA, we are obliged to consider the effect of this activity to the whole network performance. To optimize the utilization of aggregated multiple interfaces and diminish the performance anomaly (i.e throughput degradation for all STAs) all at once, a novel mechanism to predict gain (or throughput, equally) of each STA in the network based on its transmission rate is proposed. The current number of STAs, predicted transmission rate of dual interfaces from a new STA and each STA’s time duration to
send the packets are examined to determine the expected gain. After the investigation based on MATLAB simulation, we can optimize the utilization of a new STA’s additional interface to achieve better average throughput in the network up to 10% and additional interface activation up to 92.2% of the simulation
time while avoiding the gain anomaly of the whole network concurrently.

Full text

Performance Analysis on Gain Prediction of
Dual-Radio Aggregation System in Heterogeneous
Network
Yuris Mulya Saputra Rifqy Hakimi
Dept. of Electrical Engineering and Informatics School of Electrical Engineering and Informatics
Vocational College Institut Teknologi Bandung
Universitas Gadjah Mada Bandung, Indonesia
Yogyakarta, Indonesia rifqyhakimi@stei.itb.ac.id
ym.saputra@ugm.ac.id
Abstract—Nowadays, most of the smartphones or mobile
devices are already equipped with multiple interfaces. In general,
the utility of multiple interfaces concurrently will escalate the
device’s performance of speed communication. In spite of that,
an issue called performance anomaly when a new station (STA)
with multiple interfaces aggregation and low transmission rate
desires to connect with an access point (AP) exists. We can assume
that a new STA connects to an AP which already has several
associated STAs. Before turning an additional interface on for
that new STA, we are obliged to consider the effect of this activity
to the whole network performance. To optimize the utilization
of aggregated multiple interfaces and diminish the performance
anomaly (i.e throughput degradation for all STAs) all at once, a
novel mechanism to predict gain (or throughput, equally) of each
STA in the network based on its transmission rate is proposed.
The current number of STAs, predicted transmission rate of dual
interfaces from a new STA and each STA’s time duration to
send the packets are examined to determine the expected gain.
After the investigation based on MATLAB simulation, we can
optimize the utilization of a new STA’s additional interface to
achieve better average throughput in the network up to 10%
and additional interface activation up to 92.2% of the simulation
time while avoiding the gain anomaly of the whole network
concurrently.
Index Terms—Aggregation, wireless network, multi-radio com-
munication, performance anomaly, throughput
I. INTRODUCTION
High-speed data services in wireless networks have been
a major demand especially for today’s user satisfaction. Ex-
panding the transmission capacity of wireless communication
links by using additional frequency spectra (i.e concurrent
use of multiple frequency channels or aggregation) may fulfill
these requirements. This solution has already been operated by
advanced wireless systems (e.g., carrier aggregation in LTE-
Advanced, channel bonding in IEEE802.11n) and it outper-
forms a single aggregation-capable interface by improving re-
silience against a failure of individual interfaces [1]. Moreover,
the market opportunity of dual-band (2.4 and 5 GHz) WiFi in
multiple interfaces using 802.11n and 802.11ac continues to
deploy. We may know that most smartphones available today
in the market are equipped with multiple wireless networking
interfaces such as cellular radio and WiFi. We can let a laptop
has multiple wireless interfaces as well by purchasing and
installing additional interfaces like USB wireless adaptors to
run Multipath Transmission Control Protocol (MPTCP) [2].
Not only user devices, but also some of access network
components are already equipped with multiple interfaces;
e.g., dualband WiFi access point (AP) and femtocell base
station (BS) having a WiFi AP within the same box.
Regardless the concurrent use of multiple interfaces’ high
performance, it leads to the performance anomaly in the whole
network when a new STA’s low transmission rate (with dual
interfaces) is applied to associate with the AP. Consequently,
before turning an additional interface on for that new STA
(the terms to be used for the interfaces are primary interface
as the main interface to be always activated and additional
interface as the secondary interface to be activated/deactivated
for the rest of the paper), we must evaluate the effect of
this activity to the whole network performance. In WiFi, the
basic CSMA/CA channel access method assures that the long
term channel access probability is equal for all STAs [3]. It
specifies that when a STA applies low transmission rate, it
will deteriorate the performance of other STAs which employ
higher transmission rate. For example, if a new STA is far
from the AP then its interfaces are considered to use low
transmission rate, it is better to deactivate additional interface
(or turn it off) for maintaining good network performance,
otherwise it will degrade the throughput in the network. The
focus is how to overcome this drawback in the STA side
instead of AP side (since AP is always powered on) by using
a new mechanism to decrease the performance anomaly while
optimizing the throughput of each STA using aggregated dual
interfaces at the same time. As a summary, a question has to
be answered : when to turn on/off additional interface?.
To achieve the solution for this problem, a novel mechanism
to predict gain (or throughput, equally) of all STAs in the
network based on the predicted transmission rate is proposed.
This transmission rate prediction may depend on the current
recorded signal strength of another interface since its value
has high correlation with transmission rate used by additional
interface in the same location as described later. A non-
complex mathematical model and algorithm containing current
2016 8th International Conference on Information Technology and Electrical Engineering (ICITEE), Yogyakarta, Indonesia
978-1-5090-4139-8/16/$31.00 ©2016 IEEE

number of STAs, predicted transmission rate of dual interfaces
from a new STA and each STA’s time duration to send the
packets are defined to decide the expected current gain. In
addition, to realize the heterogeneous network, two interfaces
are applied using WiFi 802.11n at 2.4 GHz and 802.11a at
5 GHz frequency band both at the AP and STA side. From
the simulation result, the mechanism shows that 92.2% of
activated additional interface and better average throughput
in the network up to 10% are generated based on existing
STAs’ location and new STA’s predicted transmission rate
while avoiding the performance degradation in the network.
The rest of the paper is organized as follows. Section II
presents the related work. The transmission rate prediction is
provided in Section III. The explanation of new mechanism
using mathematical model and algorithm is reached out in
Section IV. Section V shows the simulation performance
results of the model. Finally, Section VI concludes the paper.
II. RELATED WORK
There have been several experimental studies of using
multiple interfaces aggregation and its consideration. [4] and
[5] showed a comparison of aggregation strategies including
power consumption in dual-radio communication of hetero-
geneous network. In [6] and [7], a mechanism that exploits
a large number of WiFi interfaces concurrently by solving
carrier-sensing and interferences between adjacent channels
was proposed. Link bonding using multiple WiFi interfaces
was tested and its performance gain was presented in [8].
[9] introduced scheduler which chooses the best interface
to be used from two interfaces based on measured through-
put and energy state only from previous time using offline
Markov decision process. Extensive experimental results of
using multiple 802.11n interfaces for high capacity and also
reveals that carrier-sensing between close channels is a serious
issue was provided by [10]. [11] introduced energy-aware
MPTCP on an Android platform by choosing the proper
combination of interfaces to save energy based on available
throughput of each interface. [12] introduced E$PA in which
an adaptive LTE/WiFi network was used with multi-attribute
cost function including energy (battery life), data usage ($),
and performance of transmission’s completion time.
Performance anomaly issue was also considered in some
works. In [3], performance degradation of the whole networks
due to lower bit rate mobile hosts anomaly by using 802.11b
was discussed. A model incorporating multi-rate STAs to
analyze performance anomaly in 802.11 was brought by [13].
Similar analysis yet different approach using the conditions
of frame error rate to calculate saturated throughput and
packet delay of STA was analyzed by [14]. [15] performed
the scheduler approach to avoid the performance anomaly
in WLAN by considering distributed channel-aware. Another
method to solve this issue was evaluated by [16] by trying
to aggregate packet using dynamic time interval. Finally, the
comparison of performance analysis on the issue in uplink and
downlink TCP/UDP environment by using 802.11n was done
by [17].
Based on those previous works, there is no mechanism
which exploits the consideration of using multiple interfaces
aggregation concurrently by determining gain prediction of
each STA based on current signal strength and predicted
transmission rate in the heterogeneous network. Hence, ob-
taining better throughput while avoiding performance anomaly
is discussed in this paper.
III. TRANSMISSION RATE PREDICTION
This prediction is the main procedure to decide whether
additional interface will be activated or not. The dual interfaces
used consist of WLAN 802.11n and 802.11a standard type.
802.11n has the transmission rate from 6.5 to 130 Mbps (using
20 MHz width channel without channel bonding and up to 2
spatial streams) [18] while 802.11a has the transmission rate
from 6 to 54 Mbps (using 20 MHz width channel and one
spatial stream only) [19]. To achieve the predicted transmission
rate of each interface, first we need to understand how to
classify the data using Support Vector Machine (SVM) [20]
and then we can determine the transmission rate prediction
using signal strength parameter of each interface.
A. Support Vector Machine and libSVM
Support Vector Machine or SVM is a learning technique for
data classification. This can be realized by separating the d-
dimensional data into two or multiple classes using an optimal
hyper-plane in the feature space. The sample point near the
optimal hyper-plane can be called as support vectors for a
basis model. There are two types of SVM : linear and non-
linear discriminant function. Linear discriminant function is
utilized when the example data can be linearly separable (i.e
unnoisy data). However, example data is often the noisy data
which is not linearly separable. To cover this problem, SVM
introduces a kernel. The general idea is mapping the original
input space to some higher-dimensional feature space where
the training set becomes separable. A classification usually
contains separated data called training and testing sets. Each
data point in the training set contains class labels and several
features. From the training sets, we can produce a model based
on the operated kernel function and predict the class labels in
the testing sets with the existence of features only.
For support vectors classification implementation in SVM,
the integrated software called libSVM [20] is employed. To
use libSVM, we need to transform the original value of
example data using a simple converter C program, thus it can
be processed under libSVM environment. This process will
categorize the class label into the format of SVM package
(binary or multi classes). The next process should we do is
linearly scaling the features into [-1;+1] range for both training
and testing sets to prevent the big numeric ranges dominating
smaller numeric ranges and simplify the calculation. After
this step, the model selection based on Gaussian/Radial Basis
Function (RBF) kernel [20] for samples mapping is selected.
There are two important parameters in RBF kernel called C
and γ.Cspecifies the misclassification avoidance or cost of
misclassification while γdefines the influence of a single
training sample. To obtain the best value of Cand γ,a
grid search and cross validation method are used thus the
unknown data (i.e testing sets) can be predicted accurately
2016 8th International Conference on Information Technology and Electrical Engineering (ICITEE), Yogyakarta, Indonesia

AP
Add IF
Pri IF
Add IF
Pri IF
STA
Close
distance
AP
Add IF
Pri IF
Add IF
Pri IF
STA
Far
distance
Fig. 1. Multi-WiFi system experiment
by the classifier. Finally, the chosen Cand γvalue with the
best cross validation rate are used to train the whole training
set and they will produce the model for predicting the data
inside testing sets.
B. Correlation of signal strength and transmission rate
To obtain the predicted transmission rate of each interface,
we can observe the relationship between signal strength and
transmission rate released from STA’s interfaces. An exper-
iment using 300 samples in different location (near to far
distance from the AP) to determine the transmission rate pre-
diction (high to low rate) of an interface (additional interface)
based on the signal strength of another interface (primary
interface) was conducted by two real wireless USB adapters
802.11a/b/g/n (Netgear WNDA3100) in STA and AP side. The
illustration of the experiment is shown in Fig. 1.
After 300 samples were collected by activating both inter-
faces simultaneously, we can define the correlation between
primary interface’s signal strength and additional interface’s
transmission rate (we can let 802.11n acts as primary interface
and 802.11a acts as additional interface and vice versa) by two
prediction methods. One is achieved by using probability dis-
tribution function (PDF) of the most transmission rate utilized
for each signal strength called basic prediction and another
one is using libSVM for SVM prediction. We can spot that
each of certain 802.11n’s signal strength group has different
correlation with each of certain 802.11a’s transmission rate
and vice versa as described in Fig. 2. As shown in the
figure, transmission rates of basic prediction are more various
than those of using SVM prediction. In addition, 802.11a’s
transmission rate could not be listed after 802.11n’s signal
strength attains more than -76 dBm since connection coverage
of 802.11a is smaller than that of 802.11n due to one spatial
stream specification.
IV. PROPOSED MECHANISM
This mechanism is realized according to the current trans-
mission rates from both primary and additional interface of
−80 −70 −60 −50 −40 −30 −20
6
12
18
24
36
48
54
11n − Signal Strength (dBm)
11a − Predicted Rate (Mbps)
basic prediction
SVM prediction
−80 −70 −60 −50 −40 −30 −20
6.5
26
39
52
65
78
104
130
11a − Signal Strength (dBm)
11n − Predicted Rate (Mbps)
basic prediction
SVM prediction
Fig. 2. Signal strength vs predicted transmission rate comparison
existing STAs and the current transmission rate from primary
interface along with predicted transmission rate from addi-
tional interface of new STA associating with an AP. We may
observe each STA’s current transmission rate from the rate
adaptation recorded by the AP while new STA’s predicted
transmission rate from the correlation with the signal strength
already explained before. Then, we can predict the average
throughput (or gain) produced by each STA connected to the
AP. These information are sent to the new STA for determining
its additional interface activation. Hence, before turning an
additional interface on of a new STA, we need to predict how
much the gain and loss in the network will be. If current gain is
too small or the previous gain was still better than the current
one, we better turn additional interface off/sleep to avoid the
network loss, otherwise we may turn it on.
The idea of this mechanism is each STA has different
channel access or data transmission duration based on its aver-
age transmission rate for both interfaces. This average comes
from the scheduling method called round-robin scheduling
(weight 1:1) in which a transmission flow can be separated
into multiple interfaces [8]. We may assume that the number
of STAs is STAtot =Nnow with each existing STA ihas its
transmission rates R1,i and R2,i coming from interfaces’ rate
adaptation. Then, we can consider time duration Tifor each
STA in the network with fixed packet size Lfor each interface
of each STA as below.
Ti=2L
1
2(R1,i +R2,i),0≤i<N
=4L
(R1,i +R2,i)
(1)
When a new STA associates to the network with primary
interface (PIF,i=N)’s transmission rate R1,i=Ncoming from
rate adaptation and additional interface (AIF,i=N)’s predicted
transmission rate Rpredict,i=Ncoming from basic or SVM
prediction (based on primary interface’s signal strength S1,i
between training and testing stage of libSVM), the number of
STAs now is STAtot =N+1 stations and time duration to
send packets for new STA is
Ti=N=4L
(R1,i=N+Rpredict,i=N)(2)
Then, the total time duration Ttot for all STAs (including new
STA) is
2016 8th International Conference on Information Technology and Electrical Engineering (ICITEE), Yogyakarta, Indonesia

Ttot =
N

i=0
Ti,0≤i<N+1 (3)
and average time duration ˆ
Tbecomes
ˆ
T=Ttot
STAtot
(4)
Finally, average throughput ˆμcount for all STAs in the network
for each time (counter) is calculated as below.
ˆμcount =2L
ˆ
T(5)
When counter value increases each time, transmission rate
for each STA will change. Consequently, the total time dura-
tion and average throughput value will vary as well. At the end,
we can compare the previous and current predicted average
throughput to decide the activation of additional interface for
new STA. This detail explanation can be transformed into the
following gain prediction algorithm.
V. P ERFORMANCE EVALUATION
Two main observations are conducted using MATLAB
simulation to get the gain prediction before and after new STA
comes to the network. No interference between this observed
network and other networks in the same location is defined
to predict the gain correctly. The decision to choose 802.11n
as a primary interface and 802.11a as an additional interface
is triggered by the coverage of 802.11n is larger than that of
802.11a thus new STA is still able to associate to the AP with
802.11n only when 802.11a is out of coverage.
First, expected average throughput (or predicted gain,
equally) in the network before and after new STA comes is
processed with several considerations. Then, the distribution
of additional interface activation is observed to know when
to use dual interfaces simultaneously to optimize the average
throughput and reduce the performance anomaly in the net-
work as analyzed below.
A. Expected average throughput
To determine this metric, existing STAs’ and new STA’s
different location based on their current signal strength are
considered. In this way, new STA will predict when to start
activating additional interface differently based on existing
STAs’ scattering in the network. Five existing STAs in the
network are used in the simulation. Then, 4 (four) different lo-
cations as stated by signal strength of existing STAs in separate
simulation are remarked. In addition, 6 (six) different locations
according to the primary interface’s signal strength (from -70
to -20 dBm with 10 dBm interval) and additional interface’s
predicted transmission rate of new STA are observed as well.
In detail, each of existing STAs will send the information about
their current data transmission duration based on their current
transmission rates to the AP with size of 1024 byte for each
packet. Then, AP will send these transmission duration lists
to new STA for predicting the average transmission duration
and expected throughput for all STAs including new STA in
Algorithm 1 Gain Prediction
1: STAtot ←N
2: count ←0
3: ˆμ0←0
4: ST ART ←0{simulation starts}
5: STOP ←τ{simulation ends}
6: for START to STOP do
7: if STAtot >N then
8: STAtot ←N+1
9: else
10: STAtot ←N
11: end if
12: Ttot ←0
13: for i=0to STAtot −1do
14: PIF,i ←1{primary interface ACTIVE}
15: if i=Nthen
16: record R1,i=N{from rate adaptation}
17: Rpredict,i=N(S1,i=N)←R2,i=N
18: Ti=N←4L/(R1,i=N+Rpredict,i=N)
19: else
20: AIF,i ←1{additional interface ACTIVE}
21: record R1,i {from rate adaptation}
22: record R2,i {from rate adaptation}
23: Ti←4L/(R1,i +R2,i)
24: end if
25: Ttot ←Ttot +Ti
26: end for
27: count ←count +1
28: ˆ
T←Ttot/S T Atot
29: ˆμcount ←2L/ ˆ
T{average throughput for all STAs}
30: if STAtot >N then
31: if ˆμcount <ˆμcount−1then
32: AIF,i=N←0{additional interface SLEEP/OFF}
33: else
34: AIF,i=N←1{additional interface ACTIVE}
35: end if
36: end if
37: end for
the system. The existing STAs were scattered to the location
which has signal strength -70 to -30 dBm (Location A), -30
to -20 dBm (Location B), -50 to -40 dBm (Location C), and
-70 to -60 dBm (Location D), respectively as illustrated in the
Fig. 3.
To predict the average throughput in each location of the
network, existing STAs’ and new STA’s interfaces are both
assumed to be active. The prediction consists of gain before
and after new STA comes. Previous average throughput is
remarked as before new STA comes to the network condition
while current predicted average throughput is defined as after
new STA comes to the network. The whole results can be seen
in Fig. 4, Fig. 5, Fig. 6, and Fig. 7. The location distance from
the AP can be referred to the signal strength and transmission
rate value. Lower signal strength relates to lower transmission
rate and indicates that the distance between STA and AP are
long while higher signal strength refers to higher transmission
rate and means that STA has short distance from the AP.
2016 8th International Conference on Information Technology and Electrical Engineering (ICITEE), Yogyakarta, Indonesia

AP
-20dBm
-30dBm
-40dBm
-50dBm
-60dBm
-70dBm
Location A
Location B
Location C
Location D
Fig. 3. Location of existing STAs for simulation
Basic SVM
20
30
40
50
60
70
80
90
Prediction
Average Throughput (Mbps)
(a) Before new STA comes
−70 −60 −50 −40 −30 −20
20
30
40
50
60
70
80
90
Signal strength of new STA (dBm)
Average Throughput (Mbps)
(b) After new STA comes
Basic prediction
SVM prediction
Fig. 4. Throughput prediction when existing STAs have signal strength -70
to -30 dBm
In Fig. 4, five STAs are spread throughout the network
coverage between -70 dBm and -30 dBm to realize the fairness
of the transmission rate scattering among the existing STAs.
When new STA is in signal strength less or equal than -
60 dBm, the predicted current average throughput is worse
than previous one. It strongly implies that new STA has
too low transmission rate thus it leads to the degradation of
whole network’s gain. On the other hand, when new STA
is closer to the AP (signal strength is more than -50 dBm),
the performance of predicted throughput is better than that of
previous one. In this case, the addition of new STA’s additional
interface will be beneficial for the whole network. This trend
is also almost similar to the performance of Fig. 6.
Compared to Fig. 5 in which all existing STAs are very close
to the AP, new STA needs to have equal or better transmission
rates than those of existing STAs’ interfaces to produce better
predicted throughput than previous one. Fig. 7 has the best
performance in new STA’s dual interfaces activation since
predicted current throughput has better performance due to
higher transmission rates used in most new STA’s locations.
However, the average throughput is very low compared to
other cases since all of the existing STAs has very low
transmission rate thus deteriorating the gain significantly. In
all graphs, we can conclude that the performance of SVM
prediction is always better than that of basic prediction due to
higher transmission rate prediction in lower signal strength.
B. Additional interface activation distribution
As a matter of fact, the final purpose of the gain prediction is
to optimize the utility of dual interfaces concurrently for higher
speed of data transmission while reducing the performance
Basic SVM
20
30
40
50
60
70
80
90
Prediction
Average Throughput (Mbps)
(a) Before new STA comes
−70 −60 −50 −40 −30 −20
20
30
40
50
60
70
80
90
Signal strength of new STA (dBm)
Average Throughput (Mbps)
(b) After new STA comes
Basic prediction
SVM prediction
Fig. 5. Throughput prediction when existing STAs have signal strength -30
to -20 dBm
Basic SVM
20
30
40
50
60
70
80
90
Prediction
Average Throughput (Mbps)
(a) Before new STA comes
−70 −60 −50 −40 −30 −20
20
30
40
50
60
70
80
90
Signal strength of new STA (dBm)
Average Throughput (Mbps)
(b) After new STA comes
Basic prediction
SVM prediction
Fig. 6. Throughput prediction when existing STAs have signal strength -50
to -40 dBm
anomaly. From the previous result, each existing STAs’ loca-
tion provided different predicted average throughput metric.
Then, we can discover the distribution of dual interfaces or
single interface utility as shown in Fig. 8 when new STA
moves from far to near distance from the AP (from signal
strength -70 up to -20 dBm with 1 dBm interval). When
existing STAs are in Location D, additional interface is on
up to 92. 2% of the simulation time while in Location B, only
19.6% of the simulation time additional interface is active.
In Location C, the performance has balance distribution of
activated/deactivated additional interface.
To decide when additional interface of new STA is acti-
vated during the simulation, we can observe Fig. 9. Different
location where the existing STAs are in, determine different
result in additional interface activation based on current signal
strength. From the graph, Location C has the most balancing
Basic SVM
20
30
40
50
60
70
80
90
Prediction
Average Throughput (Mbps)
(a) Before new STA comes
−70 −60 −50 −40 −30 −20
20
30
40
50
60
70
80
90
Signal strength of new STA (dBm)
Average Throughput (Mbps)
(b) After new STA comes
Basic prediction
SVM prediction
Fig. 7. Throughput prediction when existing STAs have signal strength -70
to -60 dBm
2016 8th International Conference on Information Technology and Electrical Engineering (ICITEE), Yogyakarta, Indonesia

A B C D
0
10
20
30
40
50
60
Location of existing STAs
Interfaces distribution (samples)
(a) Basic prediction
Dual interfaces
Single interface
A B C D
0
10
20
30
40
50
60
Location of existing STAs
Interfaces distribution (samples)
(b) SVM prediction
Dual interfaces
Single interface
Fig. 8. Interfaces utility distribution in each location
−70 −60 −50 −40 −30 −20
1
2
Signal strength of new STA (dBm)
Interface utility
(a) Basic prediction
Location A
Location B
Location C
Location D
−70 −60 −50 −40 −30 −20
1
2
Signal strength of new STA (dBm)
Interface utility
(a) SVM prediction
Location A
Location B
Location C
Location D
Fig. 9. Interfaces utility trend when new STA moves from far to near distance
from the AP
additional interface utility when the transition happens in
the middle of the simulation. Additional interface in SVM
prediction is activated earlier than in basic prediction as ex-
pected. Furthermore, we can analyze that the further distance
the existing STAs are from the AP, the sooner new moving
STA from low to high signal strength can activate additional
interface especially when it utilizes higher transmission rates
and vice versa.
VI. CONCLUSIONS AND FUTURE WORK
A new mechanism to predict gain of each STA in the
heterogeneous network based on the transmission rate and time
duration was presented. The expected average throughput be-
fore and after new STA comes in the network and distribution
of additional interface activation were analyzed by considering
different location for the existing STAs. This was obtained
by observing the current signal strength of primary interface,
predicted transmission rate of additional interface, and time
duration of each STA. Two prediction-based methods called
basic and SVM prediction were also defined to compare which
prediction has better performance. According to the simulation
results, it was shown that the proposed mechanism with SVM
prediction obtained better current average throughput in the
network up to 10% and additional interface activation up to
92.2% of the simulation time while avoiding the gain anomaly
of the whole network concurrently when new STA moved
from far to near distance from the AP. It was assumed that
the interference of other networks and other issues (i.e power
consumption and delay) were ignored to simplify the trend of
the mechanism, therefore it requires further study to realize
the implementation for future work.
REFERENCES
[1] IEEE Standard for Local and Metropolitan Area Networks - Link
Aggregation, IEEE Computer Society Std., 2008.
[2] TCP Extensions for Multipath Operation with Multiple Addresses, IETF
RFC 6824 Std., 2013.
[3] M. Heusse, F. Rousseau, G. Berger-Sabbatel, and A. Duda, “Perfor-
mance anomaly of 802.11b,” in Proc. of IEEE INFOCOM, 2003.
[4] Y. M. Saputra and J.-H. Yun, “Where to aggregate?: a comparative
study of aggregation strategies in multi-radio system,” in Proc. of IEEE
APWCS, 2013.
[5] Y. M. Saputra, H. Hwang, and J.-H. Yun, “A comparative study of
aggregation schemes for concurrent transmission over multiple WLAN
interfaces,” Journal of The Institute of Electronics and Information
Engineers, vol. 51, no. 7, pp. 1444–1451, 2014.
[6] C.-L. Tsao and R. Sivakumar, “On effectively exploiting multiple
wireless interfaces in mobile hosts,” in Proc. of CoNEXT, 2009.
[7] S. Kakumanu and R. Sivakumar, “Glia: a practical solution for effective
high datarate wifi-arrays,” in Proc. of ACM MobiCom, 2009.
[8] A. Jayasuriya, S. Aust, P. Davis, A. Yamaguchi, and S. Obana, “Aggre-
gation of Wi-Fi links: When does it work?” in Proc. of IEEE ICON,
2007.
[9] C. Pluntke, L. Eggert, and N. Kiukkonen, “Saving mobile device energy
with multipath TCP,” in Proc. of ACM MobiArch, 2011.
[10] S. Lakshmanan, J. Lee, R. Etkin, S.-J. Lee, and R. Sivakumar, “Realizing
high performance multi-radio 802.11n wireless networks,” in Proc. of
IEEE SECON, 2011.
[11] Y.-s. Lim, Y.-C. Chen, E. M. Nahum, D. Towsley, and R. J. Gibbens,
“How green is multipath TCP for mobile devices?” in Proc. of ACM
AllThingsCellular, 2014.
[12] W. Lee, J. Koo, S. Choi, and Y. Park, “E$PA: Energy, usage ($),
and performance-aware LTE-WiFi adaptive activation scheme for smart-
phones,” in Proc. of WoWMoM, 2014.
[13] D.-Y. Yang, T.-J. Lee, K. Jang, J.-B. Chang, and S. Choi, “Performance
enhancement of multirate IEEE 802.11 WLANs with geographically
scattered stations,” IEEE Transactions on Mobile Computing, vol. 5,
no. 7, pp. 906–919, 2006.
[14] K. Kang, “Performance anomaly of the IEEE 802.11 DCF in different
frame error rate conditions,” Journal of Information Process System,
vol. 8, no. 4, pp. 739–748, 2012.
[15] L. Tavanti, R. G. Garroppo, and S. Giordano, “Preventing the perfor-
mance anomaly of IEEE 802.11 networks with a distributed channel-
aware scheduler,” in Proc. of IEEE WOWMOM, 2007.
[16] T. Razafindralambo, L. I. Isabelle Guerin Lassous, and S. Fdida,
“Dynamic packet aggregation to solve performance anomaly in 802.11
wireless networks,” in Proc. of ACM MSWiM, 2006.
[17] C. W. S. O. Yoshiki Hashimoto, Masataka Nomoto and T. Kato,
“Experimental analysis on performance anomaly for download data
transfer at IEEE 802.11n wireless LAN,” in Proc. of ICN, 2016.
[18] M. S. Gast, 802.11n: A Survival Guide. O’Reilly, 2012.
[19] M. Gast, 802.11 Wireless Networks The Definitive Guide 2nd Edition.
O’Reilly, 2005.
[20] C.-C. Chang and C.-J. Lin, “Libsvm: A library for support vector
machines,” Tech. Rep., 2013.
2016 8th International Conference on Information Technology and Electrical Engineering (ICITEE), Yogyakarta, Indonesia