University of Southern California Dissertations and Theses

Understanding the characteristics of Internet traffic dynamics in wired and wireless networks

(USC Thesis Other)

Understanding the characteristics of Internet traffic dynamics in wired and wireless networks

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content UNDERSTANDING THE CHARACTERISTICS OF INTERNET TRAFFIC
DYNAMICS IN WIRED AND WIRELESS NETWORKS
by
Yongmin Choi
A Dissertation Presented to the
FACULTY OF THE GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(ELECTRICAL ENGINEERING)
December 2015
Copyright 2015 Yongmin Choi
to my mother, late father (1936 - 2011), and family
ii
Acknowledgments
First of all I would like to express my special appreciation toward my advisor and men-
tor, Professor John Andrew Silvester: I thank him for his great understanding, patience,
help and support he constantly provided to me. He also gave me chances to complete
the degree program and get a doctorate. Also I would like to thank Prof. Bhaskar Krish-
namachari, Prof. Leana Golubchik, Prof. Cauligi S. Raghavendra, and Prof. Behrokh
Khoshnevis for gladly accepting to serve on my committee.
I am indebted to many people during the long journey. While staying in the United
States between 1998 and 2003, I felt comfort with the colleagues from my homeland; to
name a few, Jun, Dongjun, and Pansop. In addition, I enjoyed my stay in Los Angeles
recently with Jongwook Woo and Seon Ho Kim. I much appreciate help from many
colleagues in KT Corporation and subcontracting firms.
Above all nothing could even come close to the support of my family. My parents
and sisters have always been there for me throughout every stage of my life. They
gave enormous encouragement and support during the hardships I had to endure. I can
never repay what they have done for me but I hope that my humble accomplishment can
compensate for some small part of their sacrifices.
Finally I am deeply indebted to my beloved wife, Young-Ju. As wife and mother of
three children, she has had to face all the tribulations during my study. Without her love
iii
and emotional support, my journey to the doctoral degree could have ended in a wreck
somewhere along the way. She is the one who takes the honor of my small achievement.
iv
Contents
Dedication ii
Acknowledgments iii
List of Figures vii
List of Tables x
Preface xi
1 Introduction 1
1.1 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.1 Multimedia Workload Characterization . . . . . . . . . . . . . 3
1.1.2 Mobile Data Traffic Surge . . . . . . . . . . . . . . . . . . . . 3
1.1.3 Application Signaling Traffic . . . . . . . . . . . . . . . . . . . 4
1.1.4 Analysis and Modeling of LTE Network Traffic . . . . . . . . . 4
1.2 Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2 Related Work 6
2.1 ISP Workload Characteristics . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Mobile Traffic Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3 Multimedia Workload Characteristics of an IP Network 10
3.1 Trends of the Broadband Internet Service Market . . . . . . . . . . . . 11
3.2 Analyzing and Modeling Workload Characteristics . . . . . . . . . . . 13
3.2.1 Content Length Distribution . . . . . . . . . . . . . . . . . . . 16
3.2.2 Session Arrival . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.2.3 Session Duration . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4 Widespread Use of Mobile Internet and Its Effect 25
4.1 Deployment of 3W Network . . . . . . . . . . . . . . . . . . . . . . . 27
4.2 Traffic Offloading in 3W Networks . . . . . . . . . . . . . . . . . . . . 30
v
4.3 Status of Mobile Data Traffic . . . . . . . . . . . . . . . . . . . . . . . 33
4.3.1 Smartphone Proliferation and Data Traffic Surge . . . . . . . . 33
4.3.2 Traffic Composition in the 3W Networks . . . . . . . . . . . . 35
4.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
5 Effect of Application Signaling Traffic 40
5.1 Traffic Composition in a WCDMA Network . . . . . . . . . . . . . . . 42
5.2 Impact of Application Signaling Traffic . . . . . . . . . . . . . . . . . 43
5.3 Efforts on Controlling the Application Signaling Traffic . . . . . . . . . 49
5.3.1 Application-Layer Solutions . . . . . . . . . . . . . . . . . . . 50
5.3.2 Network-Layer Solutions . . . . . . . . . . . . . . . . . . . . . 51
5.3.3 Future Directions of Push Notification Service . . . . . . . . . 54
5.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
6 Analysis and Modeling of LTE Network Traffic 57
6.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
6.1.1 Overview of LTE Network Architecture . . . . . . . . . . . . . 58
6.1.2 Data Set Description . . . . . . . . . . . . . . . . . . . . . . . 62
6.2 Characterizing Aggregate Traffic Dynamics . . . . . . . . . . . . . . . 65
6.2.1 Temporal Variation . . . . . . . . . . . . . . . . . . . . . . . . 66
6.2.2 Traffic V olume Distribution . . . . . . . . . . . . . . . . . . . 67
6.3 Modeling V oice Calls and Data Flows . . . . . . . . . . . . . . . . . . 76
6.3.1 V oice Call . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
6.3.2 Data Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
6.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
7 Conclusion 84
Reference List 87
vi
List of Figures
3.1 Trend in number of broadband internet subscribers and network capacity
during 1999 - 2008, KT Corporation. . . . . . . . . . . . . . . . . . . . 12
3.2 Growth in the number of subscribers for IPTV and V oIP services. (Top)
IPTV (Bottom) V oIP . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.3 Complementary CDF of content length . . . . . . . . . . . . . . . . . . 17
3.4 Quantile plot of content length vs. exponential distribution. X quantiles
correspond to the VOD contents length and y-axis denotes the inverse
of the cumulative distribution function for exponential distribution. . . . 18
3.5 The Poissonness plot for VOD service request arrivals . . . . . . . . . . 19
3.6 The Poissonness plot for V oIP service request arrivals . . . . . . . . . . 20
3.7 Complementary CDF of VOD service time . . . . . . . . . . . . . . . 21
3.8 Quantile plot of VOD service time vs. exponential distribution. X quan-
tiles correspond to the VOD service time andy-axis denotes the inverse
of the cumulative distribution function for exponential distribution. . . . 22
3.9 Complementary CDF of V oIP service time . . . . . . . . . . . . . . . . 23
3.10 Quantile plot of V oIP service time vs. lognormal distribution. X quan-
tiles correspond to the V oIP service time andy-axis denotes the inverse
of the cumulative distribution function for lognormal distribution. . . . 24
4.1 Overview of 3W network architecture . . . . . . . . . . . . . . . . . . 32
4.2 Trend of the number of WCDMA subscribers and the percentage of
smartphone users . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.3 Monthly mobile internet traffic volume of the 3W networks . . . . . . . 36
4.4 Number of subscribers accessing mobile internet through the 3W networks 37
vii
4.5 Monthly traffic volume per user for the 3W networks . . . . . . . . . . 38
5.1 Traffic composition based on the type of application: (Top) Traffic vol-
ume (Bottom) HTTP requests . . . . . . . . . . . . . . . . . . . . . . . 44
5.2 RRC state diagram in UE or network . . . . . . . . . . . . . . . . . . . 46
5.3 Relationship between the number of smartphones/feature phones and
the number of packet service RRC attempts . . . . . . . . . . . . . . . 47
5.4 Snapshot image of the traffic monitoring system showing PS RRC attempt
surge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.5 Comparison of signaling traffic by mobile applications and push notifi-
cation service: (Top) Push method by application itself (Bottom) Intro-
duction of push server . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
6.1 Overview of LTE network architecture . . . . . . . . . . . . . . . . . . 59
6.2 Traffic collection area. Traffic load is represented by colors. . . . . . . . 62
6.3 Daily variation of voice calls . . . . . . . . . . . . . . . . . . . . . . . 67
6.4 Daily variation of data traffic volume . . . . . . . . . . . . . . . . . . . 68
6.5 Complementary CDF plot for V oLTE calls per device . . . . . . . . . . 69
6.6 Complementary CDF plot for 3G voice calls per device . . . . . . . . . 70
6.7 Complementary CDF plot for LTE data flows . . . . . . . . . . . . . . 71
6.8 Complementary CDF plot for LTE data traffic: (Top) Packet (Bottom)
Byte . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
6.9 Histogram for V oLTE calls per RU . . . . . . . . . . . . . . . . . . . . 73
6.10 Probability plot for V oLTE calls per RU . . . . . . . . . . . . . . . . . 74
6.11 Histogram for 3G voice calls per RU . . . . . . . . . . . . . . . . . . . 75
6.12 Probability plot for 3G voice calls per RU . . . . . . . . . . . . . . . . 76
6.13 Traffic composition with regard to application type for the whole week . 77
6.14 Variation of traffic composition with regard to application type for the
whole week . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
6.15 Poissonness plot for voice call arrival: (Top) 3G CSFB (Bottom) V oLTE 79
viii
6.16 Probability plot for voice call duration: (Top) 3G CSFB (Bottom) V oLTE 80
6.17 Probability plot of data flow arrival process . . . . . . . . . . . . . . . 81
6.18 Complementary CDF plot for data flow duration . . . . . . . . . . . . . 82
ix
List of Tables
4.1 Coverage and offered service type of 3W networks . . . . . . . . . . . 30
x
Preface
The objective of this dissertation is to characterize and model traffic dynamics of wired
and wireless internet for supporting effective deployment and operations of commercial
communications networks. As the internet has become the prevalent network infrastruc-
ture for integrating broadcasting, telecommunications, and data communications, mul-
timedia services driven by internet service providers (ISPs) have proliferated. Mobile
internet services have become popular as mobile devices such as smartphones or tablets
are pervasively used. As a result, we have seen persistent growth of internet traffic in
wired and wireless networks.
To cope with the growth of the data traffic volume generated by these mobile devices
network operators need to increase network capacity, and design and operate their net-
works effectively. To increase the network capacity according to the forecasted traffic
demand is not always cost-effective as the traffic usage patterns can change in unantic-
ipated ways with the development of new applications. Furthermore it is not easy for
mobile network operators to increase their network capacity according to the forecasted
demand due to scarce radio frequency spectrum. The network operators are required
to build their networks efficiently and optimize them according to the characteristics of
usage and traffic behavior. To this end, the first step is to understand the spatial and
temporal patterns of the internet traffic carried by their networks.
xi
In this dissertation, we perform extensive traffic collection, analysis and modeling
of wired and wireless internet traffic to enhance network deployment and operational
efficiency. We collected traffic datasets from a commercial network operator covering
diverse services, times and locations. We characterize the variability and dynamics of
internet traffic to identify inherent characteristics and combine statistical and empiri-
cal methodologies to model the identified traffic behaviors using appropriate parametric
models. We also verify the effectiveness of suggested solutions to address several prob-
lems recently encountered in wired and wireless networks.
Our results are helpful to understand traffic behavior in wired and wireless networks
and to demystify traffic characteristics in commercial networks. The statistical models
we found can also be used to find generative models for synthetic traffic generators in
simulation or analytical studies. We believe that the solutions verified in this research
can be applied to today’s networks including long term evolution (LTE) networks.
Y . Choi
Los Angeles, California
December 2015.
xii
Chapter 1
Introduction
The internet is an essential part of our daily life and continues to expand and connect
to physical objects, or ”things.” As a result, internet traffic is rapidly increasing in both
wired and wireless networks. Multimedia applications are one of the main causes of
traffic increase. The proliferation of mobile devices such as smartphones and tablets
combined with ubiquitous access to cellular networks has changed the usage pattern of
data communications. According to forecasts by Cisco [Cis15], global IP traffic will
grow at a rate of 23 percent from 2014 to 2019 and consumer internet traffic will be the
major contributor to this traffic growth. Among all IP traffic, data traffic from mobile
users will increase the fastest at a compound average growth rate of 61 percent between
2014 and 2019.
• Global IP traffic has increased more than fivefold in the past 5 years, and will
increase nearly threefold between 2014 and 2019,
• Consumer internet video traffic will be 80 percent of all consumer internet traffic
by 2019,
• Globally mobile data traffic will increase 11-fold between 2014 and 2019,
• Global mobile data traffic will grow three times faster than fixed IP traffic from
2013 to 2018.
In addition, global mobile data traffic, 4 percent of total IP traffic in 2014, is expected
to be 14 percent of total IP traffic by 2019.
1
In addition to this data traffic surge, mobile networks have inherent characteristics
that are different from fixed broadband networks. The widespread use of mobile devices
using third-generation (3G) and Long-Term Evolution (LTE) networks has led to the
development of various applications that take advantage of the always-on internet con-
nectivity provided by these networks. These emerging mobile applications supporting
real-time communications are characterized by frequent network connections to receive
status updates or messages. However, mobile networks use a scarce resource, frequency
bandwidth, for their physical channels. They are characterized by limited bandwidth,
higher latency, and non-permanent communication channels. The frequency bandwidth
of mobile networks is strictly limited and shared among users. The number of mobile
devices or applications using always-on connectivity is increasing, and the network load
from these applications can be large enough to disturb normal operations of mobile net-
works.
The objective of this research is to characterize and model internet traffic dynamics
in wired and wireless networks. To cope with the explosive growth of data traffic and
support efficient network operations, network operators need to design and manage their
network architectures accordingly. The first step to achieve this is to understand the
spatial and temporal patterns of internet traffic carried by their networks. However,
traffic characteristics in commercial operating networks are not well known because
relevant datasets are rarely publicly available. Based on extensive traffic collection from
a large commercial operating network covering diverse services, times and locations,
we perform statistical analysis and modeling of wired and wireless internet traffic. We
characterize the variability and dynamics of internet traffic in various dimensions and
explore the inherent characteristics. We combine statistical and empirical methodologies
to model the identified traffic behavior using appropriate parametric models.
2
1.1 Contributions
The studies described in this dissertation include the following specific research prob-
lems.
1.1.1 Multimedia Workload Characterization
As the internet is used for integrating broadcasting, telecommunications, and data com-
munications, multimedia services have been introduced by internet service providers
(ISPs). As a result multimedia traffic comprises a substantial part of internet traffic.
While there are extensive studies about Web and on-demand or live-streaming work-
loads on the internet, multimedia traffic in commercial IP networks has not been studied
as well. To understand the multimedia workload characteristics provided by ISPs, we
collect service log traffic datasets for Internet Protocol Television (IPTV) and V oice over
IP (V oIP) services based on data from a commercial operator in Korea. By carrying out
statistical analysis and modeling of these datasets, we suggest session-level traffic mod-
els for IPTV and V oIP services [CScK11].
1.1.2 Mobile Data Traffic Surge
The development of mobile devices such as smartphones and tablets combined with
ubiquitous mobile networks has led to the widespread use of mobile internet services.
Mobile data traffic represents an increasingly large part of the internet traffic as the
number of smartphone subscribers increases. Globally there has been explosive increase
of mobile data traffic in recent years. Since mobile networks make use of a scarce
resource, i.e. radio frequency spectrum, it is difficult to provision required network
capacity according to the forecasted demand. Thus traffic offloading from mobile data
networks has been proposed to alleviate the impact of traffic surge. We carried out
3
empirical studies on a strategic solution for traffic offloading. We explored traffic data
from a mobile network operator (MNO) in Korea and the effect of multiple network
provision upon the network usage pattern of its subscribers. Specifically we observed
the effect of traffic offloading in a WCDMA network by provisioning WiFi hotspots and
mobile WiFi routers using a WiMAX network [CJyP
+
11].
1.1.3 Application Signaling Traffic
The widespread use of mobile devices using 3G and LTE networks has led to the devel-
opment of various applications taking advantage of the always-on internet connectivity.
These applications make use of signaling messages such as keep-alive or ping requests to
maintain the always-on connectivity. Although the traffic volume of signaling messages
is not large, sending frequent messages can incur a large amount of related signaling
traffic in mobile networks. We perform empirical studies on the impact of signaling
traffic in mobile networks. We also suggest the introduction of MNO based push notifi-
cation servers to reduce the signaling traffic load [ChYK
+
14].
1.1.4 Analysis and Modeling of LTE Network Traffic
The persistent increase of mobile data traffic and need for faster speed has led to the
rapid deployment of LTE networks worldwide. The fast dissemination of LTE devices
contributes to the continuous increase of mobile data traffic. Since LTE networks have
only recently been deployed, the behaviors and characteristics of LTE traffic are not
yet well understood. We investigate device-level traffic behavior and characterize the
aggregate traffic behavior of LTE devices in a crowded area of an MNO in Seoul, Korea.
We suggest session-level traffic models for voice calls and data flows of LTE devices.
4
1.2 Publications
The chapters of this dissertation are based in part on the following publications.
• Yongmin Choi, John A. Silvester, Hyun-chul Kim. Analyzing and Modeling of
Workload Characteristics in a Multi-service IP Network. IEEE Internet Comput-
ing, 2011.
• Yongmin Choi, Hyun Wook Ji, Jae-yoon Park, Hyun-chul Kim, John A. Silvester.
A 3W Network Strategy for Mobile Data Traffic Offloading. IEEE Communica-
tions Magazine, 2011.
• Yongmin Choi, Cha-hyun Yoon, Young-sik Kim, Seo Weon Heo, John A. Sil-
vester. The Impact of Application Signaling Traffic on Public Land Mobile Net-
works. IEEE Communications Magazine, 2014.
• Yongmin Choi, John A. Silvester. Analyzing and modeling traffic characteristics
and user behavior in a commercial LTE network. In preparation.
5
Chapter 2
Related Work
In this chapter we survey previous work in traffic characterization in wired and wire-
less networks. We first review workload characteristics for the fixed internet which is
related with our work in multimedia workload characterization. Then we survey work
on mobile data traffic characteristics relevant to our analysis and modeling of LTE net-
work traffic.
2.1 ISP Workload Characteristics
For the last decade, video on demand (V oD) and voice over IP (V oIP) have been cited
as examples of new Internet applications by networking researchers. However, because
of the lack of a large-scale real-world dataset, researchers are typically inclined to rely
on simulated models to drive their design and developmental efforts, and only a few
in-depth studies exist on characterizing the workload of those services.
Recently, several studies have characterized peer-to-peer television (P2P-TV) as
P2PV oD traffic on the Internet to understand the benefits of peer assistance in alleviating
the bandwidth requirements on servers [JS07], [VGK
+
07], [HLL
+
07]. In [CKR
+
09] the
authors explored how users access videos on YouTube. The authors in [HLR07] used the
V oD server logs of Microsoft Network (MSN) Video and characterized the V oD traffic
to explore the potential for peer-assisted V oD service.
However, because those P2P-TV , P2P-V oD, YouTube, and MSN Video systems are
accessed through the general Internet which is often heavily loaded, system response to
6
user interaction is often delayed, probably causing users to demand less interaction than
they might desire. To exactly understand how much interaction users want, we need
a dataset collected in a more responsive, well-provisioned, and large-scale operational
environment.
In other work, the authors in [CRC
+
08], [QGL
+
09b], [QGL
+
09a] measured and
modeled user activities and channel popularity or switching dynamics in a large-scale
Internet Protocol television (IPTV) system. However, what they focused on were live
(multicast) TV sessions, not V oD ones, where user access patterns have a direct impact
on V oD servers.
Of all the work that exists in this area, the work of the authors in [YZZZ06] is the
closest work to our own. They conducted an extensive empirical analysis of access
patterns and user behavior in a large centralized V oD system at China Telecom. They
collected more than 21 million V oD requests from 1.5 million broadband Internet users
(with 512 Kbps to home users), covering a total of over 6,700 unique video files for
219 days, from May to December 2004. Interestingly, a key finding of their study was
that the arrival distribution of user requests increasingly resembles a Poisson distribution
(whereλ = 15) as the arrival rates gets higher (when the number of requests exceeds 5
per second). We found similar results in our studies, for example, see Figure 3.5, where
λ = 14.09.
While they performed their measurement study using a V oD system that was one or
two magnitudes smaller than ours in terms of the number of requests (70 to 120 kilobytes
versus 6.7 megabytes in requests per day), users access link bandwidth (512 kilobytes
per second versus 5 to 100 megabytes per second), video content provided (6,700 versus
124,800 unique video files), and quality of video content (encoded and played back at
200 or 300 Kbps versus 2 to 10 Mbps), we found very similar results confirming their
results.
7
We also found that the V oIP call arrival process, even in a full-IP network, can still
be modeled as a Poisson process and that call durations have a long tail. This conforms
to the recent results of measurement work performed at a commercial Internet service
provider (ISP) backbone link in Italy [BMPR10].
2.2 Mobile Traffic Analysis
There have been several measurement studies focusing on the performance of data traffic
from the view point of individual subscriber devices. These studies lack the global
view of a network and a broader analysis of subscriber behaviors. Some representative
works in this field are measurement studies on 2G and 3G networks [PPJP05], [RU05],
[LSM
+
08], [JGG
+
04], [YKH08], [TLL08].
There are previous works based on large scale measurement and analysis carried out
in 3G networks. The authors in [PSBD11] conducted measurement analysis of network
resource usage and subscriber behavior using a dataset collected from a nationwide 3G
cellular network. The measurement scale of this paper is similar to our measurement in
that the dataset tracks a million subscribers over thousands of base stations. However,
the subscriber behaviors and network usages are measured in a 3G network while our
measurement is carried in 3G and LTE networks together. The authors in [KNZG10]
carried out profiling of mobile users’ web browsing behavior using co-clustering algo-
rithm. While the web traffic occupies a large part of mobile traffic and web browsing
behavior is important, a more comprehensive understanding of usage patterns is required
from a network point of view.
In [SJLW11] the authors analyzed internet traffic dynamics and proposed traffic
models in a large cellular network. They used a week-long aggregated flow level mobile
device traffic data from a mobile network operator’s core network. They characterized
8
mobile data traffic patterns based on device type and applications and proposed traffic
models for the volume distribution of application traffic and the volume dynamics of
aggregate internet traffic. Moreover, the authors in [SJL
+
13] characterized cellular net-
work performance during crowded events such as sports games or disastrous situations.
In this paper, they illustrated the relationship between changes in crowded events and
voice and data performance degradation. In [HQG
+
13], the authors studied the interac-
tions among applications, transport protocol, and the radio layers in LTE network. They
showed that various inefficiencies in TCP over LTE occurred such as undesired slow
start because 4G LTE networks have lower latency and higher bandwidth than predeces-
sor networks.
An interesting measurement-driven analysis of mobile data traffic is to consider sub-
scriber mobility. In addition to temporal variations, spatial variations in different parts
of the networks can be informative in network planning. In [PSBD11] the authors ana-
lyzed subscriber mobility based on the number of base stations visited and the radius of
gyration. They observed the periodic nature of human mobility and tendency of return-
ing periodically to the same location. Furthermore, the majority of subscribers have a
low level of mobility and the probability distribution function for the radius of gyra-
tion of subscribers is well approximated with a truncated power-law. In [CGW
+
08],
[HW05], [PSJ03], the authors studied human mobility from cellular network data. The
authors in [CGW
+
08] used the voice call records over a six-month period of 100,000
anonymized mobile phone users, and the authors in [HW05] analyzed data traffic traces
from a major regional CDMA2000 cellular network. The conclusions from these works
are that the overall human mobility is limited and human trajectories show a high degree
of temporal and spatial regularity.
9
Chapter 3
Multimedia Workload Characteristics
of an IP Network
The internet has become the prevalent network infrastructure for integrating broadcast-
ing, telecommunications, and data communications. Internet service providers (ISPs)
are striving to expand their market beyond broadband internet services. Consequently,
multimedia services such as Internet Protocol television (IPTV) and voice over IP (V oIP)
are widely used in many countries, and the number of subscribers is increasing quickly.
While the number of broadband internet subscribers in Korea is 19.1 million, the number
of IPTV subscribers is approximately 10.8 million and V oIP subscribers is 12.4 million
as of December 2014. At KT Corporation, the number of broadband internet, IPTV , and
V oIP subscribers is 8.1 million, 5.8 million, and 3.4 million respectively.
As multimedia services have proliferated on the internet, multimedia traffic com-
prises a substantial part of internet traffic. Recent studies show a notable shift in internet
traffic from Web to multimedia content [ZDJC09]. There are extensive studies about
Web [GDS
+
03] and on-demand or live-streaming workloads on the internet [V AJ
+
02].
For IPTV multicast traffic, the authors in [IMM07] present analysis results on packet-
level traffic in a commercial IP network in Italy. For video-on-demand (V oD) systems,
the authors in [YZZZ06] report user behavior and content access patterns for user arrival
rate, session duration, and content popularity based on service log data.
In this chapter we deal with multimedia workload characteristics in an integrated
services internet that provides commercial IPTV and V oIP services. We first survey
10
the evolution of broadband internet services and trends in disseminating multimedia
applications at KT Corporation. To verify the workload characteristics for multimedia
applications, we collect service log data and find traffic models for request arrivals,
session durations, and content access patterns. In the case of IPTV , V oD service requests
directly occupy network bandwidth, whereas watching a multicast live TV channel does
not if a multicast tree already exists (which is the case in our network). Thus, we only
focus on V oD traffic in this chapter.
We provide analysis and modeling results in this area primarily because workload
characterization is one of the most fundamental steps in IP network engineering for
understanding and creating an efficient network. Traffic characteristics in commercial
multiservice IP networks, however, are not well known, and the necessary datasets from
an operational network are not publicly available [HP09]. Our findings on workload
characteristics for IP multimedia applications will be helpful for dimensioning proper
capacity in network planning and other network engineering practices.
3.1 Trends of the Broadband Internet Service Market
In looking at trends, we first survey the evolution of broadband internet service at KT
Corporation. Figure 3.1 shows trends in the number of broadband internet subscribers
and change of network capacity during the past decade. (The network capacity is given
by the sum of bandwidths of links that connect subscribers and the internet backbone.)
As we can see, the number of subscribers grew rapidly in the early 2000s. While the
growth rate in the number of subscribers slowed as the internet service market became
saturated, network capacity continued to increase exponentially from 2005 to 2010.
With a flat-rate charging system, revenue for an ISP is proportional to the number of
subscribers. If the number of subscribers only grows a little, then the revenue increase is
11
Figure 3.1: Trend in number of broadband internet subscribers and network capacity
during 1999 - 2008, KT Corporation.
restricted, especially in a highly competitive and saturated market. Meanwhile, the rapid
increase in network capacity requires a large amount of investment. Thus, we ascertain
that a discrepancy exists between the growth rate of network capacity and the number of
subscribers. This means that ISPs have to invest more in the network compared to what
they expect to earn in revenue from additional subscribers.
The increase of internet traffic is typically attributed to the proliferation of file-
sharing applications and multimedia services. File-sharing applications such as Bit-
Torrent and eDonkey are popular worldwide and contributed to the drastic growth of
internet traffic between 2000 and 2007. In addition to peer-to-peer (P2P) file sharing
applications, client-server applications known as Web disks (for example, Microsoft’s
OneDrive or Google Drive) are widely used in Korea. Recently, multimedia services
such as IPTV have been commercially deployed in many countries. ISPs introduce
12
these multimedia services to compete with cable or media companies and to expand
their market beyond broadband internet service. KT Corporation introduced IPTV ser-
vice in 2007, and a massive transition in telephony service from plain old telephone
service (POTS) to V oIP began in 2009.
IPTV provides live TV channels, on-demand content, and value-added services such
as short message service or Web/email access. For the purpose of workload analysis, we
consider only V oD service that occupies distinct bandwidth per session. V oIP offers the
same functionalities as the conventional POTS. However, it can provide various func-
tions such as video telephony. As POTS is rapidly being replaced by internet telephony,
the number of V oIP subscribers is increasing quickly. Figure 3.2 shows the increase of
IPTV and V oIP subscribers at KT.
3.2 Analyzing and Modeling Workload Characteristics
Now that we have reviewed some of the trends, we consider some statistical analysis
results on workload characteristics for IPTV/V oD and V oIP services. We can character-
ize this workload at different levels such as user, session, or protocol. We are interested
in modeling the workload characteristics at the session level. A session corresponds
to a V oD request for IPTV or a voice/video call for V oIP. Modeling the session-level
behavior requires specifying session arrivals and duration.
We use service log information from one day in April 2009 to analyze multimedia
workload characteristics. These service logs were generated for every V oD request or
V oIP call. These logs were originally recorded for billing purposes; thus, they are fairly
accurate. The number of logged requests during the time interval studied is more than
a million, so the datasets are sufficiently large for statistical analysis. The logs for V oD
13
Figure 3.2: Growth in the number of subscribers for IPTV and V oIP services. (Top)
IPTV (Bottom) V oIP
14
service contain the following information per request (customer identities, IP addresses,
and content names were anonymized for privacy considerations):
• type of service (media transmission),
• request start time and end time,
• subscriber ID,
• media server and set-top box IP addresses,
• content name, and
• encoding rate of content (megabits per second, or Mbps).
The logs for V oIP service contain the following information per each request:
• type of service (voice or video),
• request start time and end time,
• terminal equipment ID,
• phone number of calling/called terminal equipment,
• IP address of calling terminal equipment, and
• data rate of the call (kilobits per second, or Kbps).
From these service logs, we make a session level description of V oD and V oIP ser-
vices containing the start time, duration, required bandwidth, and size of transferred
data. The number of subscribers for IPTV and V oIP services were 709,778 and 504,649,
respectively. The number of logged requests for IPTV and V oIP were 6,743,035 and
3,909,345, respectively. The number of subscribers for the internet service was 6.8 mil-
lion in April 2009, and the subscription ratios of IPTV and V oIP services were 10.5 and
7.5 percent of the total number of subscribers for internet service, respectively.
15
3.2.1 Content Length Distribution
Because the holding time of V oD service is closely related to its content length, we
examined the distribution of content length. The total number of V oD items was 124,796
in April 2009. V oD content falls into one of two categories based on the encoding rate:
standard definition (SD) and high definition (HD). The data rate of SD content is less
than 3 Mbps, while HD content is about 10 Mbps. Most of the content offered was
SD (102,145), with less than 20 percent of it was HD (22,651). The average length of
SD content was 27.84 minutes and that of HD content was 36.96 minutes. The average
length of overall V oD content is 29.49 minutes. The maximum content length is 260.48
minutes.
The length of V oD content ranges from a few seconds to 4 hours and 20 minutes.
The short content items mostly consist of commercials and preview programs. Figure
3.3 shows the complementary cumulative distribution function (CDF) of content length
distribution. We also draw a quantile (or probability) plot in which the theoretical quan-
tiles are plotted against the order statistics for the given data (see Figure 3.4). Thus, we
plot thex
(i)
on one axis and
F
−1

i−0.5
n

,
on the other axis, where F
−1
(·) denotes the inverse of the CDF for the hypothesized
distribution. Figure 3.4 shows the probability plot of V oD content length versus an
exponential distribution. The distribution of V oD content length shows some deviation
from an exponential distribution when the length is larger than 3,000 seconds.
3.2.2 Session Arrival
Service log information records arrivals of V oD requests or V oIP calls on a scale of
seconds. Thus, there are several concurrent calls or service requests during busy hours,
16
10
1
10
2
10
3
10
4
10
5
10
−15
10
−10
10
−5
10
0
Service time (sec)
Pr[X>x]
Complementary CDF of VOD contents length
Figure 3.3: Complementary CDF of content length
and the accurate computation of interarrival times is difficult. We use a Poissonness
plot that can verify how closely discrete data follows a Poisson distribution [MM08].
It also serves to highlight which points might be incompatible with the model. The
Poissonness plot is drawn with discrete variables obtained by counting the number of
times something occurs. In our case, the number of occurrences corresponds to the
number of request arrivals in a given time interval. The counts are denoted as k, with
k = 0,1,...,L, whereL is the maximum observed value for the discrete variable in the
dataset. Then the total number of observations in the sample is
N =
L
X
k=0
n
k
,
17
0 2000 4000 6000 8000 10000 12000 14000 16000
0.25
0.5
0.75
0.9
0.95
0.99
0.995
0.999
0.9995
0.9999
Data
Probability
Probability plot for Exponential distribution
Figure 3.4: Quantile plot of content length vs. exponential distribution. X quantiles
correspond to the VOD contents length andy-axis denotes the inverse of the cumulative
distribution function for exponential distribution.
wheren
k
represents the number of observations that are equal to the countk.
A basic Poissonness plot is drawn by plotting the count values k on the horizontal
axis and
ϕ(n
k
) = ln(k!n
k
/N)
on the vertical axis. These are plotted as symbols, similar to the quantile plot. If the
given data follow a Poisson distribution, then the Poissonness plot follows a straight
line. Systematic curvature in the plot would indicate that these data are inconsistent
with a Poisson distribution. The values forϕ(n
k
) tend to have more variability whenn
k
18
0 10 20 30 40 50
0
20
40
60
80
100
120
140
1
1
1
Number of Occurrences − k
φ (n
k
)
Figure 3.5: The Poissonness plot for VOD service request arrivals
is small. Figure 3.5 shows the Poissonness plot for the V oD request arrivals. We observe
that the given data follow a Poisson distribution quite well. The average arrival rate is
14.09 requests per second.
Figure 3.6 shows the Poissonness plot for V oIP connection arrivals. We observe that
the given data follows a Poisson distribution in this figure. Most of the V oIP service con-
nections are voice calls; video calls comprise only 0.06 percent of all the connections.
Thus, we do not discriminate among the V oIP service connections and analyze them as
a whole. In Figure 3.6, we observe that V oIP connection requests can be modeled by a
Poisson distribution. The average arrival rate is 26.27 connections per second.
19
0 20 40 60 80 100 120 140 160 180
0
100
200
300
400
500
600
1
1
Number of Occurrences − k
φ (n
k
)
Figure 3.6: The Poissonness plot for V oIP service request arrivals
3.2.3 Session Duration
We next look into the holding time distribution of V oD service. The total number of
V oD requests for this particular day was 6,743,035, which is the largest number of such
requests collected in a day appearing in the scientific literature, to the best of our knowl-
edge. We sub-sampled down to a million logs in our analysis for easier and faster data
processing. We believe that 1 million logs are sufficiently large to get statistically mean-
ingful results. Among these service logs, we eliminate logs for abnormal termination
(when service duration is recorded as 0 seconds). The number of abnormal termination
20
10
1
10
2
10
3
10
4
10
5
10
−15
10
−10
10
−5
10
0
Service time (sec)
Pr[X>x]
Complementary CDF of IPTV VOD service duration
Figure 3.7: Complementary CDF of VOD service time
logs is 24,532 (2.34 percent of the total logs). Additionally, there are also some excep-
tionally long holding times, with some taking up to 24 hours (46 logs). Though this
might happen in reality, it is rare and does not contribute significantly to the determina-
tion of distribution, so we restrict the V oD service time to the maximum of the contents
length. The mean holding time is 22.50 minutes.
Figure 3.7 shows the complementary CDF of holding time. As we can see, the
holding time distribution resembles the content length distribution. Additionally, we
show the quantile plot of V oD service time versus exponential distribution in Figure 3.8.
21
0 2000 4000 6000 8000 10000 12000 14000 16000
0.25
0.5
0.75
0.9
0.95
0.99
0.995
0.999
0.9995
0.9999
Data
Probability
Probability plot for Exponential distribution
Figure 3.8: Quantile plot of VOD service time vs. exponential distribution. X quantiles
correspond to the VOD service time and y-axis denotes the inverse of the cumulative
distribution function for exponential distribution.
The total number of V oIP service connections was 3,909,345, but as before, we used
only 1 million logs in the analysis. The mean holding time is 1.78 minutes and the
maximum is 256.18 minutes. We show the complementary CDF of holding time for
V oIP service in Figure 3.9. In Figure 3.10, we also show the quantile plot versus log-
normal distribution. In these Figures, we find that the holding time for V oD service
can be approximated by an exponential distribution of up to 4,000 seconds. Contrary
to the traditional model in teletraffic theory, the holding time for V oIP calls can be
approximated by a log-normal distribution rather than an exponential distribution.
22
10
1
10
2
10
3
10
4
10
5
10
−7
10
−6
10
−5
10
−4
10
−3
10
−2
10
−1
10
0
Service time (sec)
Pr[X>x]
Complementary CDF of IPTV VOD service duration
Figure 3.9: Complementary CDF of V oIP service time
3.3 Summary
From the statistical analysis results, we found that the service request process follows a
Poisson distribution for V oD and V oIP services. The session duration process for V oD
can be approximated by an exponential distribution for the average length of the contents
range (around 30 minutes). We also discovered that the session duration process for
V oIP calls has a long tail. We believe that our findings on the workload characteristics
for IP multimedia applications can be used to dimension capacity in network planning
and other network engineering practices.
23
10
0
10
1
10
2
10
3
10
4
10
5
0.0001
0.05
0.1
0.25
0.5
0.75
0.9
0.95
0.99
0.999
0.9999
Data
Probability
Probability plot for Lognormal distribution
Figure 3.10: Quantile plot of V oIP service time vs. lognormal distribution. X quantiles
correspond to the V oIP service time and y-axis denotes the inverse of the cumulative
distribution function for lognormal distribution.
24
Chapter 4
Widespread Use of Mobile Internet and
Its Effect
In recent years, there has been a big change in people’s way of using data communica-
tions services due to the availability of broadband wireless networks and the develop-
ment of mobile devices such as smartphones and tablets. In this and the following chap-
ter, we explore the impact of widespread use of mobile devices. Compared with wired
networks, mobile networks have very different constraints and features such as limited
frequency bandwidth and mobility management. We focus on two issues in mobile
networks that have gathered much attention from mobile network operators (MNOs)
recently: the rapid growth in data traffic and the effect of signaling traffic.
Mobile data traffic represents an increasingly large share of internet traffic. AT&T
acknowledged a 5000 percent growth in wireless data traffic in three years (2006 - 2009)
since the iPhone launch [Meg09]. The growth of cellular data traffic was forecasted to
outstrip fixed-line internet traffic by a factor of 10 in 2009 - 2014 [Woo09]. Most of this
traffic is generated by smartphones, with easy and quasi-ubiquitous access to streaming
media for video and audio, sales of which surpassed desktop PC sales including tablets
in 2011 [Can12].
Unfortunately, this explosive surge in mobile data traffic has caused unprecedented
pressure on the limited spectrum of the current third-generation (3G) cellular networks,
pushing them to capacity limits in many geographical areas. Moreover, signaling traffic
causes a significant burden on the 3G core networks as the multitudinous smartphone
25
applications proliferate. Many smartphone applications such as mobile messengers and
social networking applications frequently communicate with application servers or peers
in order to keep their connections alive. In conjunction with these always-on character-
istics, smartphones cause a lot of signaling traffic in the 3G networks.
The cost of supporting the exponentially increasing mobile data traffic is becoming
unsustainable: the rising cost of service delivery is outrunning the increased revenues.
Subscribers, particularly in metropolitan areas, are experiencing deteriorating 3G qual-
ity during peak times because the network is congested with the high demand [Wor09].
In response to the increase of data traffic, both the industry and research communities
have been moving quickly to address the rapid traffic growth with a low-cost viable
solution, 3G data traffic offloading [Han09], [LRL
+
10], [BMV10], [Mee09]; reducing
the pressure on 3G spectrum by using complementary network technologies to deliver
data originally targeted for cellular networks.
In this chapter, we briefly review various mobile data offloading solutions and then
present a snapshot of the real-world case of 3G data offloading supported by KT Corpo-
ration, whose wireless market share is 31.6 percent with 16 million subscribers in Korea
[Kor]. KT has recently implemented and deployed the 3W - wideband code-division
multiple access (WCDMA), WiMAX (called wireless broadband [WIBRO] in Korea),
and WiFi network strategy for efficient handling and offloading of mobile data traffic,
which is discussed in detail later.
We highlight the main contributions from this chapter:
• We briefly review various mobile data offloading solutions, including direct tun-
neling, gateway offloading, traffic policing/shaping based offloading, and WiFi or
femtocell offloading.
• We present a high-level summary of the usage statistics, traffic composition, and
growth trends of KTs 3W wireless networks, WCDMA, WiFi, and WIBRO, in
26
2010, in terms of the number of subscribers and the amount of total traffic. Due
to the proliferation of smartphones (in particular, the Apple iPhone) and intro-
duction of flat-rate pricing, we have observed a tenfold data traffic surge in our
WCDMA network in 12 months, from around 0.9 to 9.1 Gb/s on average. The
total amount of mobile traffic served by all three networks increased from around
3 Pbytes/month to 5.2 Pbytes/month in the same period.
• We also present how much traffic an average user has generated in each of the
three different mobile networks. In 2010, the average monthly traffic per sub-
scriber in our WCDMA, WiFi, and WIBRO increased from 100 to 450 Mbytes,
from 80 to 160 Mbytes, and from 7 to 11 Gbytes, respectively. The results clearly
indicate that strategically deploying multiple different types of wireless networks
addressing different needs/scenarios of mobile users is an effective way to handle
the huge amount of 3G mobile data traffic.
The rest of this chapter is organized as follows. We describe the deployment of
3W networks in KT Corporation. We briefly review existing mobile data offloading
solutions and discuss how mobile data traffic offloading is achieved in 3W networks.
We show how the 3W networks of KT were used in 2010, in terms of the number of
subscribers and the amount of total traffic, as well as per-user generated traffic. We then
conclude the chapter.
4.1 Deployment of 3W Network
This section describes the evolution and deployment of 3W wireless networks in KT
Corporation. KT Corporation launched its 3G cellular service in 2007. Although it
operated a nationwide 2G CDMA network, a new service based on 3G WCDMA was
launched first in Korea to provide high-speed data and multimedia services and global
27
roaming. Since the 2G cellular mobile telecommunications service started in 1996, the
life of facilites was close to its end. Moreover, the frequency band of 2G service was
1.8 GHz, which was used only in Korea. Hence it was preferred to deploy a new mobile
communications network rather than to upgrade its 2G cellular network to Evolution-
Data Only (EV-DO).
The 3G WCDMA service started in conformance with Third Generation Partnership
Program (3GPP) Release 4, and was upgraded based on Releases 5 and 6 to furnish
high-speed downlink/ uplink packet access (HSDPA/HSUPA) functionalities. The num-
ber of subscribers for the 2G and 3G cellular services together exceeded 16 million in
December 2010, and the mobile telecommunications market share was 31.6 percent in
Korea [Kor].
WIBRO service provides portable mobile broadband connectivity across cities and
countries through a variety of devices. WIBRO is the Korean name for the IEEE 802.16e
(mobile WiMAX) international standard. The first commercial WIBRO service started
in the Seoul metropolitan area in June 2006. Service coverage was expanded to 19
neighboring cities near Seoul and five metropolitan areas in 2008 and 2010, respectively.
The coverage was expanded to 59 local cities in 2011; thus, the service is now provided
in 84 cities in Korea. The pricing policy of WIBRO is usage-based, and subscribers can
use up to 50 Gbytes a month for $25.
Most subscribers use the WIBRO service with laptop or netbook computers that are
equipped with internal modems or external USB dongles. Other types of terminals are
also available such as smartphones, portable media players, or navigation devices. While
the number of WIBRO subscribers has not changed a lot, the introduction of a wireless
LAN access point (AP) using the WIBRO as a backhaul network contributed to wide
service dissemination and increase of traffic. WIBRO routers, called egg terminals, were
28
installed in public transportation such as subway or railway trains and buses. Personal
WIBRO routers providing WiFi connectivity while in motion are also available.
WiFi service was first introduced in the early 2000s with the growth of broadband
internet service. KT provided public hotspot zones and home APs to its broadband
subscribers for an additional charge ($10 for unlimited data). The number of WiFi
subscribers did not achieve economy of scale because the service coverage was limited.
The total number of hotspot zones was 13,000 by 2010.
However, WiFi hotspot zones turn out to be an important network infrastructure as
smart devices equipped with WiFi connectivity are widespread. Although WiFi provides
small coverage areas and has limited mobility, it can offload data traffic from 3G cellular
networks, as exemplified in the case of AT&T [Mee09]. KT started to build WiFi hotspot
or street zones in 2010, and the number of hotspot zones exceeds 40,000. It planned to
expand the number of public hotspot zones to 100,000 in 2011. The WiFi hotspot zones
are available for free to smartphone subscribers of KT.
We summarize the strategic positioning of 3W networks based on coverage and
available terminal equipment. The 3G cellular network offers nationwide coverage for
the most diverse devices such as smartphones, feature phones, tablets, and even note-
book computers through tethering. The WIBRO network covers only urban areas (84
major cities) and provides highspeed data service up to 30 Mb/s. Users can access the
WIBRO network with internal or external modems and WiFi-enabled devices such as
egg terminals. WiFi covers indoor hotspots and street zones, and provides the highest
bandwidth of the three networks.
29
Table 4.1: Coverage and offered service type of 3W networks
Type Coverage Offered service
WCDMA Nationwide Voice and low volume data
WIBRO 84 cities High-speed data
WiFi Indoor hotspot High-speed data
4.2 Traffic Offloading in 3W Networks
To handle the increasing mobile data traffic more efficiently, current 3G networks con-
tinue to be enhanced and optimized. Before we discuss the traffic offloading in 3W
networks, we briefly review offloading solutions. Several solutions are available in the
market: serving general packet radio service (GPRS) support node (SGSN) direct tun-
neling, offloading gateway, traffic policing/shaping-based, and WiFi/femtocell offload-
ing.
Direct tunneling enables the SGSN to establish a direct user plane tunnel between a
radio network controller (RNC) and a gateway GPRS support node (GGSN). Through
the tunnel, user data traffic can bypass the SGSN to reduce the loads. However, signaling
traffic loads in the control plane are still handled with the SGSN. A more recent advance-
ment over direct tunneling is SGSN and GGSN gateway offloading, where an offloading
gateway is inserted between the RNC and SGSN. The gateway monitors radio access
network application part (RANAP) control traffic and detects if a session is requested
for an internet bound session or an operator provided service (i.e., walled garden). If the
session is for a walled garden service, the gateway does nothing. Otherwise, if the ses-
sion is bound for the internet and the service provider sees no incremental opportunity
to add value, the gateway lets traffic bypass the SGSN and GGSN, relieving not only
the SGSN but also the GGSN from traffic loads, and forwards the traffic to the nearest
internet peering point.
30
Traffic policing/shaping-based offloading solutions typically exploit deep packet
inspection (DPI) technology, with which operators can inspect packet payloads (i.e.,
application-level user data) to identify certain classes of traffic and then perform various
preplanned operations on the traffic (e.g., compress and/or cache the contents).
WiFi or femtocell offloading reduces the pressure on 3G spectrum by using alterna-
tive networks like WiFi or femtocells (small micro sites of cellular coverage) when pos-
sible for transferring data [Han09], [LRL
+
10], [BMV10]. WiFi has naturally emerged
as a low-cost viable solution for operators to handle the data traffic surge in 3G networks,
due to the built-in WiFi capabilities of smartphones on the market. Most smartphones
with WiFi capability are currently configured by default to give higher priority to WiFi
over the cellular interface for data transmissions [LRL
+
10]. In addition to reducing
pressure on 3G spectrum, it has also been reported that WiFi offloading can lower the
cost of data transfers by 70 percent [Mee09].
Next, we discuss how traffic offloading is attained among 3W networks. An
overview of 3W network configuration is shown in Figure 4.1. Smartphone subscribers
of KT can use WCDMA packet service (PS) or WiFi networks to access mobile data
sevice. As explained in the previous section, users access the internet with WiFi through
two backhaul networks: the conventional wired network and WIBRO.
Mobile data traffic offloading depends on users discretion, turning on WiFi capa-
bility in their mobile devices. There are applications that compel users to make use of
WiFi. Some applications restrict downloading data larger than 20 30 Mbytes or watch-
ing streaming videos to WiFi only, and thereby traffic offloading is encouraged. While
WiFi offers more bandwidth than WCDMA, it is not always preferred to use WiFi for
reasons such as the inconvenience of turning WiFi on/off, limited coverage and mobility,
and, more important, flat-rate pricing for unlimited WCDMA data usage. However, the
WiFi traffic generated from smartphones is increasing as more WiFi hotspot or street
31
Internet
UE
RNC GGSN SGSN Node B
RAS
WIBRO
Router
Router
Switch
RNC: Radio Network Controller
SGSN: Serving GPRS Support Node
GGSN: Gateway GPRS Support Node
RAS: Radio Access Station
ACR: Access Control Router
AP: Access Point
UE: User Equipment
AP
Concentration
Node
ACR
Figure 4.1: Overview of 3W network architecture
zones are built. As we see in the next section, the total amount of WIBRO traffic is also
increasing, while the number of subscribers has not changed a lot. We believe it is an
indication of widespread use of the WIBRO network including egg terminals. To fully
utilize the traffic offloading with WiFi, KT introduced an access network discovery and
selection function (ANDSF) in order to inform users of WiFi availability in 2012.
Although WiFi networks are available free to smartphone subscribers of KT Cor-
poration, authentication and authorization is required to access WiFi. The authentica-
tion method varies depending on the types of devices. Smartphones are verified with
medium access control (MAC) or universal subscriber identity module (USIM) authen-
tication, while laptops and netbook computers use ID/password or MAC authentication.
Using the authentication information, we discriminate traffic used in this chapter based
on terminal types.
32
4.3 Status of Mobile Data Traffic
In this section, we investigate the traffic variations in 3W networks and the offloading
effect of WiFi. First, we discuss the trend of smartphone proliferation and how it con-
tributes to the surge of data trafic in the WCDMA network. We next observe the mobile
traffic composition among 3W networks and show evidence to substantiate the traffic
offloading effect of WiFi.
4.3.1 Smartphone Proliferation and Data Traffic Surge
Figure 4.2 shows the trend of smartphone proliferation in KT Corporation. Before the
iPhone release in November 2009, only a small number of subscribers had been using
smartphones such as Windows Mobile or Blackberry phones. The introduction of Apple
iPhone changed the mobile communications market and set off the massive transition
from feature phones to smartphones. By December 2010, the number of smartphone
users exceeded 6 million, which was approximately 12 percent of mobile communica-
tions service subscribers in Korea [Kor]. By December 2010 the number of smartphone
users subscribed to KT was 2.7 million, accounting for approximately 19 percent of total
subscribers, which was much higher than the average (12 percent) in Korea [KT ].
Smartphone users generate more data traffic than feature phone users. While fea-
ture phone users make use of data services provided by mobile network operators such
as short, long, or multimedia message service and walled garden portal service, smart-
phone users utilize a broad range of data services such as email, web browsing, social
networking, and video streaming. Consequently, smartphones contribute to the data
traffic surge in the WCDMA network.
In the case of our WCDMA network, the data traffic increase was tenfold in 2010
when the traffic is measured at an interface connecting GGSNs to the internet backbone.
33
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
11.0
11.5
12.0
12.5
13.0
13.5
14.0
14.5
15.0
Million
Number of 3G subscribers Percentage of smartphone users
Figure 4.2: Trend of the number of WCDMA subscribers and the percentage of smart-
phone users
The monthly average traffic at peak times increased from 0.9 Gb/s in December 2009
to 9.1 Gb/s in December 2010. It is clear that the increase of mobile internet traffic
is proportional to the number of smartphone users. However, the dramatic increase in
mobile data traffic is heavily influenced by the pricing policy. The increasing rate of
data traffic became steep when flat-rate pricing for unlimited data usage was introduced
in September 2010. Thus, the arrival of smartphones combined with flat-rate pricing led
to the tenfold data traffic surge in a year.
34
4.3.2 Traffic Composition in the 3W Networks
Mobile traffic composition among the 3W networks is shown in Figure 4.3. In case of
WIBRO and WiFi, the data is given by the total traffic generated from other devices
besides smartphones. In addition, the WiFi traffic is collected according to the APs sup-
plied by KT Corporation (hotspot zones and home APs). Thus the WIBRO and WiFi
data mostly consist of traffic occurred from traditional mobile devices such as note-
books. In this figure, the total mobile traffic increases during 2010 as diverse mobile
devices become widespread. The net increase in mobile data traffic is attributed to
WCDMA and WIBRO traffic, while WiFi traffic does not vary that much. The propor-
tion of WiFi traffic is reduced to approximately 50 percent of the total mobile network
traffic in December 2010, while it was more than 70 percent in December 2009. On
the other hand, the proportion of WCDMA traffic has increased to 30 percent of total
mobile network traffic, while it was less than 5 percent in December 2009. This figure
shows that the WCDMA network plays an increasingly important role in 3W networks
when smart devices are pervasive and flat-rate pricing is available.
Figure 4.4 shows the monthly average number of subscribers accessing each of the
3W networks. The number of WCDMA and WiFi subscribers corresponds to the smart-
phone users accessing WCDMA and WiFi networks, respectively. The smartphones
users accessing WiFi is computed with the AAA (Authentication, Authorization and
Accounting) servers which have the information on types of devices. However, the num-
ber of WIBRO subscribers is not related with the number of smartphone users accessing
WIBRO. In this figure we observe that users accessing the mobile internet make use of
WCDMA mostly. The widespread use of WCDMA in accessing mobile internet can
be inferred from coverage of the WCDMA network and diverse devices such as smart-
phones, feature phones, tablets, and even notebook computers with tethering. The num-
ber of subscribers accessing WiFi increased since August 2010 as WiFi hotspots were
35
0
1000
2000
3000
4000
5000
6000
TB/Month
Total WCDMA WIBRO WiFi
Figure 4.3: Monthly mobile internet traffic volume of the 3W networks
massively deployed. While the number of WIBRO users stays smaller than the other
two networks, it has increased by 40 percent in 2010, which also indicates widespread
usage of mobile internet in Korea.
Next, we examine monthly traffic volume per subscriber of the 3W networks (Figure
4.5). The WiFi traffic volume occurred from smartphone users is available since June
2010. In case of WCDMA, the monthly traffic volume per user in January 2010 is not
conformance with the trend of increase. It can be explained with the characteristics of
users in January 2010 since early adopters might purchase smartphones and use them
heavily. While most mobile internet users access the WCDMA network, the monthly
traffic volume per user is small compared to that of WIBRO. However, monthly data
traffic volume per WCDMA user rapidly increased after the introduction of flat-rate
pricing in September 2010. Monthly traffic volume per WiFi user is small compared to
36
0.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
4.0
Million
WCDMA WIBRO WiFi
Figure 4.4: Number of subscribers accessing mobile internet through the 3W networks
that of WCDMA but also increased in late 2010. Monthly traffic volume per WIBRO
user shows the characteristics of the subscribers of the service. The WIBRO subscribers
represent a heavy group of mobile internet service users, and the traffic volume per user
has recently exceeded 10 Gbytes/mo.
In summary we categorize the positioning of the 3W networks as follows:
• WCDMA offers nationwide coverage for voice and low-volume data service.
• WIFI covers indoor hotspots or street zones for high-speed data service.
• WIBRO covers 84 urban cities for highspeed data service of heavy users. offload-
ing.
37
0
2
4
6
8
10
12
0
50
100
150
200
250
300
350
400
450
500
WIBRO (GB)
WCDMA, WiFi (MB)
WCDMA WiFi WIBRO
Figure 4.5: Monthly traffic volume per user for the 3W networks
4.4 Summary
This chapter presents a quantitative study of mobile data traffic surge and a strategic
solution implemented for traffic offloading by KT Corporation. As smartphones includ-
ing the Apple iPhone have become pervasive in Korea, we observed a tenfold mobile
data traffic surge in our WCDMA network in 2010. Traffic offloading of the WCDMA
network is necessary; thus, several solutions have already been proposed, such as SGSN
direct tunneling, gateway offloading, and traffic policing/shaping based on deep packet
inspection. In this article we survey how deploying multiple mobile networks can be
effective in sharing the traffic load of the WCDMA network. We discuss the usage trend
and growth of 3W (WCDMA, WIBRO, and WiFi) network traffic in KT Corporation,
and how users access each network. We also investigate the composition of mobile data
38
traffic generated by smartphones. Traffic offloading is expected to be increasingly indis-
pensable as the available radio spectrum is diminished for mobile communications, and
carrier aggregation in heterogeneous networks will be used in the near future.
39
Chapter 5
Effect of Application Signaling Traffic
The widespread use of mobile devices using third-generation (3G) and Long-Term Evo-
lution (LTE) networks has led to the development of various applications that take
advantage of the always-on internet connectivity provided by these networks. Instant
messenger (IM) or social network service (SNS) like Facebook and Twitter are some
examples of this class of new mobile applications. Traditional internet applications,
such as web surfing and file transfer, are characterized by usage patterns that have dis-
tinct active and inactive phases. An active phase is a period in which several bursts of
packets are transmitted, while an inactive phase is characterized by no data transmission
during a sustained time period. The traffic pattern of recent and emerging applications
that rely on always-on connectivity is quite different. Since the emerging mobile appli-
cations support real-time communications services, they are often constantly running in
background mode to receive status updates or messages from other parties. Thus, the
applications continuously generate short signaling messages such as keep-alive and ping
requests to maintain the always-on connectivity.
Although the traffic volume of keep-alive messages is not large, frequent short mes-
sages can incur a large amount of related signaling traffic in the mobile network. In 3G
or LTE networks, the user equipment (UE) and radio access networks keep the radio
resource control (RRC) states. The UE stays in RRC Connected mode when it trans-
mits or receives data during active periods and stays in RRC Idle mode during inactive
periods. To send even a small data packet, the UE changes the RRC state to the RRC
Connected mode prior to transmission. This RRC radio state change generates a lot of
40
signaling messages, resulting in a rapid increase in traffic loading. The amount of sig-
naling traffic leads to two major problems: rapid drainage of the mobile devices battery
and a signaling traffic surge in the mobile network. In [GJKV13], the authors focused
on the issues of the energy impact on the mobile device. In this chapter, we focus on the
signaling impact of these applications on public land mobile networks (PLMNs).
The signaling traffic surge, or so-called signaling storm, due to the rapid growth in
use of these applications is having a serious impact on mobile network performance.
The frequent RRC state change leads to increased signaling overhead over the air inter-
face and through the core elements of a mobile network. The effect of signaling traffic
loading gets more severe for the core network as the number of UE devices connected
to the core network elements increases. Several mobile network operators (MNOs) have
experienced severe service outage or degraded network performance due to the increase
of application signaling traffic [Gab12], [Reu12], [Sah12]. Furthermore, the stability
of the network can also be impacted by signaling traffic when there is an application
server failure or outage. If an application server is unexpectedly out of service, all the
clients of the server lose their connections and try to restore them. When the server is
reinstated, simultaneous reconnection tries from all the clients occur. This phenomenon
is similar, in many ways, to a distributed denial of service (DDoS) attack, resulting in a
traffic overload in the core network.
This chapter explores these signaling traffic issues in mobile networks. We first ana-
lyze mobile data traffic characteristics based on the type of application and show how the
traffic characteristics of emerging mobile applications are responsible for an increase of
RRC state change attempts. We then discuss how the frequent RRC attempts impact the
network traffic load due to increased signaling messages. We present solutions and stan-
dardization efforts to reduce application signaling. Finally, we make some concluding
remarks.
41
5.1 Traffic Composition in a WCDMA Network
In this section, we analyze the traffic composition in the commercial wideband code-
division multiple access (WCDMA) network of KT Corporation, one of the major
MNOs in Korea. We demonstrate that the emerging mobile applications generate a large
proportion of HTTP requests compared to their proportion of traffic volume. We also
discuss why frequent application signaling is used in the emerging mobile applications.
The traffic data of KTs WCDMA network was collected at the interface connecting
the mobile data backbone network and the internet (G
i
interface in Third Generation
Partnership Project [3GPP] specification) in August 2011. We used the proprietary net-
work monitoring system, WCDMA Network Traffic Analysis System (WNTAS), for
collecting and analyzing mobile data traffic. It collects traffic by mirroring all the links
at the G
i
interface and then collects traffic log data based on the source/destination IP
addresses, port numbers, and transport layer protocols (TCP and UDP). It further ana-
lyzes the TCP sessions having port numbers 80 and 8080 as HTTP traffic and records
HTTP transactions for each TCP session based on HTTP header information.
At the time of the study, the number of subscribers for WCDMA service was 16
million. The number of smartphone users was 6 million, corresponding to 38 percent
of the total subscribers. The monthly traffic volume was 5175 Tbytes. The proportion
of HTTP traffic was approximately 80 percent of the total data traffic. Since many
smartphone applications are web-based, HTTP traffic constitutes most of the total data
traffic. The HTTP traffic is classified according to the type of application identified
with the URL field in the HTTP header. To accurately classify the traffic with URL,
we keep a database of website information, which is frequently updated and verified.
In this chapter, we classify mobile applications in five categories: media, application
download, web surfing, IM, and other.
42
Figure 5.1 shows the composition of HTTP traffic based on the application type.
Media traffic is generated by music and video download/streaming services, which
account for a large fraction of the explosive increase in mobile network traffic volume.
The application download category refers to the traffic used to download applications
from Apples Appstore, Google Play Store (formerly Android Market), and similar appli-
cation download sites. This category accounts for the second largest volume of traffic.
The web surfing category accounts for the traffic used by browsers such as Apples Safari,
Googles Android browser, and other web browsers. In the IM category, only the traffic
related to the most widely used mobile messenger in Korea (KakaoTalk) is included.
The remaining traffic is classified as other.
We observe that the large majority of the traffic in mobile networks is related to
media streaming and application download services. While the proportions of traffic
volume for web surfing and IM are small, the proportions of associated HTTP requests
for them are much larger. This phenomenon is consistent with the usage pattern of
smartphones: users frequently send short messages, and messenger clients constantly
send keep-alive messages to maintain their connections to the application server. While
the share of IM traffic volume is only 2 percent, the share in the associated number of
HTTP requests is more than 24 percent. Although the large increase in mobile data
traffic volume has been widely recognized, the traffic load from short (but frequent)
signaling messages has lately become a major concern for MNOs, as explained in the
following section.
5.2 Impact of Application Signaling Traffic
In this section, we briefly explain the basic keep-alive mechanisms and review the oper-
ation of RRC, which is responsible for assigning radio resources between the UE and
43
50.52%
18.29%
18.49%
2.14%
10.56%
Media
App Download
Web
Messenger
Others
4.23%
2.72%
46.34%
24.60%
22.10%
Media
App Download
Web Browsing
Messenger
Others
Figure 5.1: Traffic composition based on the type of application: (Top) Traffic volume
(Bottom) HTTP requests
the network. We then investigate the message sequences exchanged between the UE and
the network to send a short keep-alive message (when in the idle state), and the resulting
signaling traffic load in WCDMA or LTE networks.
In many networks, including internet access networks, Network Address Translators
(NATs) or other forms of middle-boxes are deployed. In a system using a NAT, a local
network uses IP addresses within a private IP address range, and the local network is
connected to the outside internet with a publicly routable IP address. As the UE sends
44
data traffic from the local network, the NAT translates the local IP address to the pub-
licly routable address by modifying the IP address and TCP/UDP port numbers. The
NAT needs to keep the state of the active connections so that inbound packets can be
routed to the correct local addresses and applications. The NAT uses expiration timers
to remove unneeded connection state table entries, which are recreated whenever there
is traffic. If a chatty application needs to be accessed from outside the local network, the
expiration of connection state during an idle period could be a problem. In order to reset
the expiration timers in the NAT, many protocols use a keep-alive mechanism, such as
having the UE send a dummy packet on a regular basis when there is no traffic to send.
The RRC states of the UE for packet data transmission are illustrated in Figure
5.2. When the UE powers on, it indicates its existence to the network by performing
attachment or registration procedure. Roughly speaking, UE can be in either of two
different operating modes: RRC connected and RRC idle mode. In connected mode,
the radio link is kept active, allowing the UE to transmit or receive traffic. In idle mode,
the radio link is released to reduce power consumption. Thus, prior to data transmission
the UE must first change state to the connected mode.
To be more specific, the UE can be in one of four RRC states: Idle, PCH (Cell PCH,
URA PCH), Cell FACH, and Cell DCH [3GP08]. In the Idle state the UE does not
maintain the RRC connection with the network, so connection must be re-established
before any data transmission. In PCH state the UE is RRC connected with the network.
No user data can be sent, and the UE only checks the paging information. In Cell FACH
state the UE is in RRC connected mode and can communicate with the network using
the common or shared radio channels. This state is ideal when the UE transmits or
receives short data packets. In Cell DCH state the UE is connected to the network with
a dedicated channel or a high-speed downlink shared channel (HS-DSCH). This state
is ideal for the exchange of large data packets; however, a lot of signaling traffic is
45
Figure 5.2: RRC state diagram in UE or network
necessary to change RRC state from Idle to Cell DCH as a dedicated channel must be
established prior to data transmission.
Application signaling traffic, such as keep-alive messages, causes frequent switching
between Idle/PCH state and Cell FACH/Cell DCH state. The result is that a significant
amount of radio resource is consumed, and there is an increased chance of packet colli-
sion on the common channels if several UE devices try to access the server simultane-
ously. To make matters worse, this frequent switching of RRC states generates increased
control signal processing loads in network entities such as the radio network controller
(RNC) and serving gateway support node (SGSN).
As discussed in the previous section, the traffic composition in a WCDMA network
resulting from signaling traffic from real-time communication services generates a sub-
stantial portion of the HTTP request traffic. The increase in signaling traffic is incurred,
46
-
5
10
15
20
25
30
35
40
45
-
2
4
6
8
10
12
14
16
Billion
Million
Smartphone Feature Phone No. of PS RRC attempts
Figure 5.3: Relationship between the number of smartphones/feature phones and the
number of packet service RRC attempts
in part, by the application signaling messages such as keep-alive or ping requests. Based
on our measurement with commercial devices, the most popular mobile messenger in
Korea (KakaoTalk) sent keep-alive messages every 10 minute in 2011. There is another
mobile application (Stock Radar) that sends more than 120 keep-alive messages per
hour. The average number of RRC attempts per smartphone user was 237 times a day
in July 2011, while that of a feature phone user was only two times a day.
In summary, pervasive dissemination of smartphones and widespread use of real-
time communication applications contribute to the rapid increase of RRC attempts for
data calls, as shown in Figure 5.3. As opposed to iOS, in which the Apple Push Notifica-
tion service (APNs) is mandatory for all applications, this phenomenon is more promi-
nent in Google Android smartphones in which each application may maintain its con-
nection individually; so each application installed in UE may send separate keep-alive
messages. The total volume of keep-alive messages grows rapidly in proportion as more
applications are installed in the UE.
47
Figure 5.4: Snapshot image of the traffic monitoring system showing PS RRC attempt
surge
In addition to the amount of signaling traffic, the traffic pattern is also an important
factor for the network load. Figure 5.4 shows the signaling traffic caused by a server
failure. In this figure, each curve corresponds to the number of RRC attempts in a SGSN.
The sum of all the curves is the number of RRC attempts in KTs WCDMA data network
(i.e., packet service [PS] RRC attempts). If an application server is unexpectedly out of
service or malfunctions, all the clients of the server lose their connections and try to
restore them. When the server is reinstated, simultaneous reconnection tries from all the
clients happen. This phenomenon is similar to a DDoS attack, and overload in the core
network happens. (We see an example of such a surge of PS RRC attempts between 5
a.m. and 6 a.m. in April 2010 in Figure 5.4.) Several MNOs have reported network
failures or degraded performance arising from application server problems [Gab12],
[Reu12], [Sah12].
48
5.3 Efforts on Controlling the Application Signaling
Traffic
Compared to fixed broadband wireline networks, wireless networks are characterized
by limited bandwidth, higher latency, and non-permanent communication channels. As
we have seen, it makes sense to limit or control application signaling traffic to alle-
viate the radio resource and processing loads in wireless networks. The mobile envi-
ronment also puts constraints on the resources available to mobile device applications.
Ideally, smartphone application developers would consider the characteristics and con-
straints of mobile environments, but developers are not always familiar with the details
of the mobile environment and wireless network issues. This has led to the concept of a
network-friendly application architecture.
The signaling load in mobile networks is associated with several operational char-
acteristics of the emerging mobile applications [3GP11]. The client of emerging appli-
cations receives data or updates from the application server and other parties by either
client-driven polling or connection-oriented push methods (e.g., persistent connections
with enabling server-initiated data delivery.) With the polling approach, the client occa-
sionally connects with its server and downloads data. If the polling interval is too short,
unnecessary traffic is generated, whereas too long of a polling interval causes delayed
response time.
In the connection-oriented push method, the server sends a push notification to
inform the client of receiving data or status updates. The client establishes a persis-
tent connection (e.g., via HTTP, a TCP-based protocol such as XMPP, or other propri-
etary TCP- based protocols), and either the client or the server regularly sends signal-
ing messages (e.g., keep-alive or ping requests.) Sending signaling messages regularly
maintains always-on connectivity and prevents expiration of the IP session allocated to
49
the mobile device. Maintaining always-on connectivity facilitates fast receipt of data or
updates from the server since the call setup delay to (re-)establish the PS session is not
required. However, frequent transmission of signaling messages not only occupies the
radio resources but increases the processing load in the WCDMA network.
This section describes related activities to reduce the application signaling traf-
fic. A Korean standards organization and the Global System for Mobile Communica-
tions Association (GSMA) have developed some guidelines for application development
[Tel11, GSM12]. In this section, we briefly review the recommendations in these guide-
lines, discuss the use of a push notification server (or push proxy gateway) to manage
connections between UE and application server, and present the effect of an MNOs push
notification server in Korea.
5.3.1 Application-Layer Solutions
We first discuss push and polling methods recommended by the Telecommunications
Technology Association (TTA), a national standardization body in Korea. MNOs in
Korea participated in TTA to establish guidelines for application signaling procedures
in 2011 [Tel11].
In the push method, the guideline recommends adjusting the intervals of sending
signaling messages appropriately. To keep always-on connectivity to the application
server, the client of an application on UE should send keep-alive messages before the
IP session of the UE expires. The holding time of an IP session of UE is determined
by MNOs. The typical value of session holding time is on the order of an hour. Thus,
application signaling traffic can be reduced by adjusting the time interval of sending
keep-alive messages to the session holding time provisioned by the mobile network.
However, duplicated signaling by multiple applications in a mobile device should also
be controlled appropriately.
50
To avoid network load from simultaneous reconnect attempts from UE devices, the
guidelines suggest regulation of the connection restoration method. When the connec-
tions between application clients and servers are lost, the clients should try to reconnect
to the servers at random times. If users make reconnection attempts manually, the guide-
line suggests not making attempts too frequently.
In the polling method, an application client occasionally connects with its server and
downloads data. Since too short polling intervals result in signaling traffic overload, the
guideline recommends the following polling methods. First, the default time interval
for polling should be at least one hour. Second, to avoid simultaneous polling of many
clients, the polling time of each client should be different. Since the email service
is considered to require more real-time communication in the guideline, it can use a
polling time of less than one hour.
GSMA also provides guidelines for smartphone application developers [GSM12].
It suggests key principles to bear in mind when developing applications for mobile
devices. In addition to the application signaling issues we have discussed, it provides
detailed tips for developing mobile applications to make them smarter. The guideline
describes seven requirements of an ideal mobile application and provides detailed spe-
cific tips for Android, iOS, and Windows mobile phones.
5.3.2 Network-Layer Solutions
Since the push method makes a connection only when there are data or status updates, it
is more efficient than polling in which periodic connection attempts are made regardless
of data or updates from servers. However, to keep always-on connectivity between
application clients and servers, it is necessary to send signaling messages on a regular
basis. If each application maintains its connection independent of other applications, the
signaling messages increase in proportion to the number of applications that are in use in
51

Push Recipient
(Application)
User Device
(Smartphone)
Application 1
Application n
…
Application 2
…
Push Function
(Third-Party
Application Server)
PLMN

Push
Function
PLMN
User Device
(Smartphone)
Push User
Agent
Push Recipient
(Application)
Push Initiator
(Third-Party
Application Server)
(Push Server)
…
Application 1
Application n
…
Figure 5.5: Comparison of signaling traffic by mobile applications and push notification
service: (Top) Push method by application itself (Bottom) Introduction of push server
the UE devices. Application connection management can be simplified by introducing
a push notification server (or push proxy gateway).
The push notification server maintains a single connection to a UE device and noti-
fies it when there are data or updates to send. Since each client application no longer
has to maintain its own session, the signaling traffic is significantly reduced. Figure 5.5
compares these two push methods: notification by the application server itself (Figiure
5.5 (Top)) and use of a push notification server provided by the network operator (5.5
(Bottom)).
52
The architecture of push notification by an integrated push server is shown in Fig-
ure 5.5 (Bottom). A single connection is maintained between a UE device and the
push server regardless of the number of mobile applications installed. Apples Push
Notification service (APNs) or Googles Cloud Messaging (GCM, formerly Cloud to
Device Messaging, C2DM) are examples of push servers. While Apple mandates the
use of APNs for all applications registered at Appstore, Google does not enforce the use
of GCM. As might be expected, for a similar application environment, Android-based
smartphones tend to generate more signaling traffic than iPhones.
Since C2DM had limitations on the number of push messages to send for each appli-
cation and limitations on the supported Android OS versions (2.2 or later), MNOs in
Korea deployed their own push servers for Android smartphones starting in 2011. Fur-
thermore, any MNO can attempt to control application signaling by introducing their
own push server. To demonstrate the effectiveness of an MNO-based push server, we
exemplify SK Telecoms smart push system [Kim12]. SK Telecom commercialized the
first MNO-based push server in May 2011. It accommodated up to 7.5 million Android
smartphones in December 2011. The total number of Android smartphone users was
8.8 million, and the accommodation ratio was 85.2 percent.
The smart push system provided push notification service to many popular applica-
tions in Korea including four IM applications such as KakaoTalk. In order to use SK
Telecoms push notification service, the mobile applications were modified to utilize a
push user agent provided by SK Telecom. The smart push user agent is programmed to
send a keep-alive message every hour. It is reported that the smart push system reduced
75.7 percent of RRC attempts for the four messaging services [Kim12].
53
5.3.3 Future Directions of Push Notification Service
While the introduction of push servers in MNO helps to alleviate the signaling traf-
fic problem, the problem of duplicate push notification still arises with Android smart-
phones, since push notifications can be by an application server, by an MNO push server,
or by a GCM server, and it is necessary to maintain connections to several or all of them.
The signaling traffic increases as multiple push notification functions are used. More-
over, it becomes complicated to develop smartphone applications if MNOs push servers
have different application programming interfaces (APIs). Application providers there-
fore prefer to use GCM or their own servers over MNOs push servers. On the other
hand, MNOs prefer to use their own push servers since this gives them greater control
over the UE, and, in addition, they can develop value-added services to go along with
the push servers.
If MNOs mandate the use of their own push services, application providers need
to modify their applications according to the MNOs push APIs. To reduce the burden
of application developers, it is desirable to standardize MNOs (proprietary) push server
solutions, including APIs, message formats, functions, and procedures of an MNOs push
service. In 2011, KT proposed standardization of the interfaces between application
server and push server, push server and push client, and push client and application.
The recommendation for standardized interfaces was approved in 2013.
Additionally, it is preferable to establish a unified push notification service specifica-
tion independent of the mobile operating system to simplify the application development
environment. Web application-based push service is also being considered by the World
Wide Web Consortium (W3C). Since web applications run on a users browser, a push
notification service can be implemented regardless of the mobile operating system.
Techniques to reduce application signaling at the network layer are also considered
by the 3GPP. The Technical Specification Group-Service and System Aspects (TSG-SA)
54
is defining work items on small data and device triggering enhancements (SDDTE) to
be included in the Release 12 specification [3GP12]. Small data transfers like signaling
messages or frequent device triggering are to be transmitted efficiently with minimal
network impact (e.g., signaling overhead, network resources, and delay for reallocation).
The proposed solutions are based on the optimized handling of control plane or RRC
connections in order to avoid user plane call setup or alleviate non-access stratum (NAS)
signaling. These efforts can contribute to reduction of signaling procedures in the core
network and radio interface, but modifications to the existing interfaces and nodes will
be required [3GP12].
5.4 Summary
In addition to the explosive increase in data traffic, mobile network signaling traffic is
a major concern for network operators because of the network investment required for
processing the increased signaling traffic. As mobile devices such as smartphones and
tablets are used more and more, application signaling traffic will become increasingly
important in future mobile networks such as LTE. In this chapter, we have quantitatively
explored the traffic composition according to types of mobile applications and showed
that signaling traffic mostly arises from applications supporting real-time communica-
tion. One example is the increased signaling traffic and associated increased network
load to process RRC attempts.
To help resolve the signaling traffic issue, application signaling should be carefully
controlled. In this regard, TTA in Korea has developed a recommended specification
for the application signaling mechanisms. GSMA has also established guidelines for
the development of user- and network-friendly smartphone applications. The signal-
ing traffic load can be significantly remediated by introducing push servers that manage
55
connections between mobile devices and application servers. In addition to the UE oper-
ating system vendors push servers, MNOs have also deployed push servers, resulting in
multiple push functions in Android smartphones (with associated increased signaling
traffic.) The conclusion is that development of standards for push function APIs and
service specifications is desirable. An ideal mix of mobile-operating-system- indepen-
dent application server and operator based push function support is required.
56
Chapter 6
Analysis and Modeling of LTE
Network Traffic
Since 2010 4G Long Term Evolution (LTE) networks have started to be commercially
deployed to accommodate high-speed data services for mobile devices. As of December
2014, 364 operators have deployed LTE networks and it is expected that 460 operators
will be operating their LTE networks by the end of 2015 [Glo15a]. 4G LTE network
is distinguished by several features compared with its predecessor: no circuit switched
domain for voice communications, increased bandwidth, and reduced end-to-end delay
(or latency) [Cox12]. Since LTE was provisioned to resolve the congestion in 2G and
3G networks, the first requirement is to increase network capacity in developing related
specifications. Additionally LTE is designed to implement a packet-switched core net-
work only to reduce capital and operational expenditure of network operators. In this
case voice calls are to be transported by packet switched networks using voice over IP
(V oIP) or 2G/3G circuit-switched service. To transfer real-time interactive voice calls
with data packets, the end-to-end delay, or latency, of LTE network should be reduced
compared with the latency of previous networks (for example, 3G networks introduce
latencies of the order of 100 milliseconds for data traffic).
In this chapter, we investigate device-level traffic behavior of a commercial LTE net-
work. Mobile data traffic is increasing fast and occupies a large part of total IP traffic as
more and more smart devices are pervasively used in accessing the internet or watching
57
videos. To cope with the explosive growth of mobile data traffic and to efficiently man-
age the traffic, it is important to understand the distribution and temporal dynamics of
traffic. However, the LTE network technology is deployed only recently and there have
not been extensive studies on traffic measurement and analysis in commercial environ-
ments. Moreover the traffic data in commercially operating networks is not publicly
available because of user privacy concerns. Hence characterizing and modeling of the
traffic behavior of LTE devices is important especially for commercial environments.
We collect traffic data from the core networks of a mobile network operator. We then
characterize the aggregate traffic behaviors from LTE devices in most crowded area in
Korea. The term ”aggregate” means that we are interested in the workload generated
by an LTE device including voice calls and data flows together. For voice traffic, we
collect call detail records (CDRs) that are used for billing. The CDR contains session-
level behaviors such as call start/finish time, duration, and reason for call termination,
etc. For data traffic, we use flow-level traffic records generated with a proprietary traffic
monitoring system.
We first characterize network-wide traffic behavior such as temporal variation, traffic
volume distribution and composition with regard to application type. Then we analyze
and model arrival and duration processes of voice call and data flow arrival. We fur-
ther analyze device-level traffic to see how the performance of network is related with
workload from LTE devices.
6.1 Background
6.1.1 Overview of LTE Network Architecture
The mobile network collecting traffic data employs both 3G (WCDMA) and 4G (LTE)
mobile communications technologies that are part of Third Generation Partnership
58
Figure 6.1: Overview of LTE network architecture
Project (3GPP) specifications. Figure 6.1 illustrates the high-level architecture of the
LTE network. The LTE network has three main components, i.e. user equipment (UE),
the evolved UMTS terrestrial radio access network (E-UTRAN), and the evolved packet
core (EPC).
User Equipment
The UE can be divided into two components, namely the mobile termination (MT)
which handles all the communication functions, and the terminal equipment (TE) which
terminates the data streams. In most cases, the UE used in our study is just a single
device in which MT and TE are integrated. The universal integrated circuit card (UIUC)
in the UE runs an application known as the universal subscriber identity module (USIM),
which stores user-specific data such as the user’s phone number (international mobile
59
subscriber identity [IMSI]) and home network identity. In this study, we do not use
IMSI as identifying the subscriber for privacy reason. Instead we use anonymized con-
tract number that is used in the KT Corporation for administering the service contract.
Evolved UMTS Terrestrial Radio Access Network
The E-UTRAN handles the radio communications between the UE and the EPC. The
E-UTRAN consists of just one component, the evolved Node B (eNB) which is a base
station that controls the UEs in one or more cells. The eNB sends radio transmissions
to all its UEs on the downlink and receives transmission from them on the uplink. In
addition, the eNB controls the low-level operation of all its UEs combining the Node B
and the radio network controller (RNC) in 3G WCDMA network. The hardware of eNB
is separately implemented into digital unit (DU) and radio unit (RU) in LTE equipment.
Thus the coverage of RU is same as the base station in earlier networks while the DU
covers larger area consisting of several RUs. The number of RUs that one DU can handle
depends on the equipment manufacturer.
Evolved Packet Core
The EPC is the name of LTE core network and a main component of system architecture
evolution (SAE) which is the core architecture of 3GPP’s LTE standard. The subcompo-
nents of EPC are mobility management entity (MME), serving gateway (SGW), packet
data network (PDN) gateway (PGW) and other systems related with service provision-
ing. The MME controls the high-level operation of the UE, by sending it signaling
messages. The SGW routes and forwards user data packets between the base station
and the P-GW, while also acting as the mobility anchor for the user plane during inter-
eNB handovers. The PGW provides connectivity from the UE to external packet data
networks. Each UE is assigned to a default PGW when it first switches on, to give
60
always-on connectivity to a default PDN such as the internet. A UE may have simulta-
neous connectivity with more than one PGW for accessing multiple PDNs. The PGW
performs policy enforcement, packet filtering for each user, charging support, lawful
interception and packet screening. Another key role of the PGW is to act as the anchor
for mobility between 3GPP and non-3GPP technologies such as WiMAX and 3GPP2
(CDMA 1X and EvDO).
The EPC is characterized to support a packet-only core network as explained earlier.
It does not provide circuit service, which is traditionally used for phone calls and short
message service (SMS). 3GPP specified two solutions for voice calls in LTE networks:
(i) circuit-switched service fallback (CSFB), (ii) internet multimedia subsystem (IMS).
In CSFB, the UE changes its radio access technology from LTE to a 2G/3G technology
supporting circuit-switched services. The voice service based on IMS is called V oice
over LTE (V oLTE). In this approach the voice service is delivered as data flows within
the LTE data bearer. The implementation of V oLTE can offer operators cost and oper-
ational benefits by eliminating the need to keep separate networks for voice and data.
Since V oLTE utilizes IMS as the common service platform, V oLTE can be deployed in
parallel with video calls over LTE and RCS multimedia services including video share,
multimedia messaging, chat and file transfer.
As of April 2015, 90 operators in 47 countries are investing in V oLTE deployments,
studies or trials while 16 operators in 7 countries among them commercially launched
V oLTE-HD voice service [Glo15b]. The number of smartphones supporing V oLTE is
186 according to GSA and it increases as many device manufacturers support V oLTE
functionality. However the use of V oLTE is not popular yet among subscribers of KT
Corporation. The newly launched smartphones are equipped with V oLTE functionality
but it is necessary to enable the function to use V oLTE preferentially over 3G CSFB.
61
Low
Medium
High
Seocho
Songpa
Gangnam
• Area: 120.38 km
2
(19.8% of total Seoul area)
• Operating networks
ü 3G WCDMA / LTE
ü WIBRO (WiMAX)
ü WiFi
Figure 6.2: Traffic collection area. Traffic load is represented by colors.
6.1.2 Data Set Description
This study is based on flow-level mobile device traffic data collected from a cellular
operator’s core network. To characterize and model the behavior of mobile devices, we
collect voice and data traffic records together. This allows us to characterize the total
load generated from a mobile device and develop models that predict the bandwidth
demands in the operator’s core network over time. In case of voice calls, the flows gen-
erated by a device are distinguished by a distinct connection for each call. Contrarily
data flows are composed of a successive stream of packets based on the 5-tuple infor-
mation such as source/destionation IP addresses, port numbers, and protocol. Due to the
large volume of data and other limitations of our analytic platform, we focus our study
in one of the most crowded areas in Korea which consist of three southern districts of
the Seoul metropolitan area (Figure 6.2).
We study only the activities of mobile devices that are associated to base stations
in this area. The data covers activities during one whole week in 2014 (August 31st to
62
September 6th). During this period, several millions of distinct devices were observed
using voice and data communications services.
Voice Call
The data for voice calls are call detail records (CDRs) gathered at the billing systems of
KT Corporation. As explained earlier, the voice call data consist of two types: 3G voice
CSFB and V oLTE calls. Regardless of call types, the records for voice calls have the
same information, i.e., timing, location, and device information as follows:
• call start date and time,
• call finish date and time,
• call duration,
• encrypted contract number for subscriber,
• cell identifier,
• call completion code,
• detailed information on call termination.
The service contract number for subscriber is used for discriminating each device. Thus
we characterize and model the subscribers’ behaviors using the LTE network. Since
we analyze only a short period of time, it is less likely for a subscriber to change one’s
device in this period. We presume our analysis results represent device level behaviors
also. Because of privacy concerns, we anonymize the contract number and we do not
have any information regarding calling or called party phone numbers. The call comple-
tion code is classified as completed or incomplete and the incomplete calls are further
distinguished as disconnected and not connected. The disconnected call is caused by
63
network-related issues such as barred call or handover fail etc while the not connected
calls are related with user behavior such as user busy, no response or call rejection etc.
Data Flow
The data traffic is collected by a proprietary traffic monitoring system, LTE Traffic Anal-
ysis System (LTAS). It collects traffic by mirroring all the links at the SGi interface con-
necting P-GW and other packet data networks (PDNs) and then composes flow-level
statistics based on the source/destination IP addresses, port numbers, and protocol. The
data traffic is then analyzed based on application layer protocol: HTTP and non-HTTP
traffic. HTTP traffic is further classified based on transport layer protocols. In this study,
we focus on HTTP traffic which comprises most of mobile data traffic.
Each record contained in the flow data set is a summary report of activity during
one particular flow by one mobile device. The records in the data set contain timing,
location, and device information as follows:
• flow start date and time,
• flow finish date and time,
• flow duration,
• encrypted contract number for the subscriber,
• cell identifier,
• encrypted source IP address and destination IP address,
• source and destination port numbers,
• total number of packets and bytes transferred,
• application identifier and category.
64
The records in the data set are indexed by timing information and an encrypted sub-
scriber’s identity the same as the case of voice calls. Each record in the data set also
contains a cell identifier, recognizing the radio unit (RU) that serves the device, an appli-
cation identifier, and data usage statistics for the flow including total number of packets
and total number of bytes during the flow. Applications are identified mainly according
to heuristic information. In case of HTTP traffic, the LTAS uses the URL field in the
HTTP header and identifies application type based on server IP addresses. To accu-
rately classify the traffic with URL, we keep a database of website information. As with
voice call data, we use the anonymized contract number for a subscriber to discriminate
devices. We also encrypted source IP addresses allocated to mobile devices for privacy
concerns.
Limitations of the dataset
The collected data set has two limitations described below. Firstly, the accuracy of
timestamp is 1 second. Arrival time and duration of calls or flows are recorded on order
of seconds. Thus we cannot analyze fine time scale behaviors of voice or data traffic
under one second. Secondly the LTAS does not record packet level behaviors such as
packet arrival time or packet size due to the huge volume of traffic. Hence we focus
on the characterization and modeling of flow-level traffic and the packet level behaviors
and influence of protocols are beyond this study.
6.2 Characterizing Aggregate Traffic Dynamics
This section describes aggregate traffic dynamics observed from the network perspec-
tive. To infer network load from mobile devices, we analyze the traffic arriving at the
base stations installed in the traffic collecting area. We observe temporal variation of
65
the aggregate traffic and then traffic volume distribution with respect to device, location,
and application. Due to the huge volume of traffic data, we restrict the data for analysis
considering the objective of analysis, data volume, and/or data processing time in case
of 3G voice calls and data flows. In those cases, we use the traffic data on September
4th in which the volume of data traffic was the highest during the whole week.
6.2.1 Temporal Variation
We first study the temporal dynamics of the logged traffic. We plot time-series of the
observed voice and data traffic volume per hour for the complete week in Figure 6.3
and Figure 6.4. The voice calls are mostly made by 3G CSFB although newly released
smartphones are equipped with V oLTE functionality. However, the default voice call
service is still 3G CSFB and enabling V oLTE function in the user device is required
to use V oLTE preferentially. Hence the number of V oLTE calls occupies only a small
portion of the voice calls. For data traffic volume, we plot the number of packets, bytes,
and flows together in Figure 6.4.
We clearly observe strong diurnal variations in the voice and data traffic volume.
This diurnal variation is closely associated with the weekly work pattern of people. For
the voice calls, we observe a peak every day in the afternoon. The number of voice calls
increases as the daytime starts and the peak is centered around 5 p.m. The number of
voice calls decreases as the night time passes.
For the data flows, we observe a peak every day in the afternoon. The peak is
centered around noon and lasts up to early evening. This indicates that people tend to use
their mobile devices around lunch time and evening time. In addition, we observe that
the daily peaks observed on the weekdays are higher than those observed on weekends.
66
Time
0 20 40 60 80 100 120 140 160 180
Normalized number of calls
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
3G
VoLTE
Figure 6.3: Daily variation of voice calls
6.2.2 Traffic Volume Distribution
Next we analyze the traffic volume distribution with respect to device, location, and
application. By verifying the traffic distribution in various dimensions, we show the
skewness of internet traffic distribution.
Device
Figure 6.5 shows the distribution of V oLTE calls with respect to device identifer. We
draw the complementary cumulative distribution function to observe tail behavior. A
large portion of users (35%) make one call during the whole week. The maximum
number of V oLTE calls per user is 2,260. The average number of V oLTE calls per user
67
0 20 40 60 80 100 120 140 160 180
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Flow
Packet
Byte
Figure 6.4: Daily variation of data traffic volume
is 14.918. The top 20% of users make 82.3% of the total V oLTE calls made in the period.
From the usage of V oLTE calls, we confirm the Pareto rule meaning that roughly 80%
of the effects come from 20% of the causes [Wik].
The distribution of 3G voice calls is shown in Figure 6.6. Compared with the V oLTE
calls, the number of users and the number of 3G voice calls are both larger more than
10 times. The proportion of users making one voice call during the whole week is
39% similar to the proportion for V oLTE. The maximum number of 3G calls per user is
2,196. The average number of 3G calls per user is 27.181. Contrary to the V oLTE calls,
the top 20% of users make only 69.7% of the total 3G voice calls made in the period.
Considering the average number of calls and the proportion that the top 20% of users
generate, 3G voice calls are more evenly distributed among users than V oLTE calls.
68
Number of VoLTE calls per device
10
0
10
1
10
2
10
3
10
4
Pr[X>x]
10
-6
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
Figure 6.5: Complementary CDF plot for V oLTE calls per device
The distribution of LTE data flows with respect to device identifier is shown in Figure
6.7. Considering the volume of data traffic, we restrict the data traffic of the busiest day
(September 4th) in the week. The proportion of users making one data flow during a day
is 60%. It means 60% of the users in a day rarely use their smartphones while staying
in the data collecting area or they don’t stay in the data collecting area for a long time.
The average number of data flows per user is 609.01 and the maximum number of data
flows per user is 313,671 for a day. The top 20% of users make 80.0% of the total data
flows made in the period. We confirm the Pareto rule from the usage of LTE data.
We also draw the distribution of data packets and bytes with respect to device identi-
fiers in Figure 6.8. The average number of packets per user is 34346 and the maximum
69
Number of 3G voice calls per device
10
0
10
1
10
2
10
3
10
4
Pr[X>x]
10
-7
10
-6
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
Figure 6.6: Complementary CDF plot for 3G voice calls per device
number of packets per user is 26.7 million for a day. The average number of bytes per
user is 30.2 MBytes and the maximum number of bytes per user is 23.0 GBytes for
a day. The top 20% of users make 87.2% of the total packets and 89.4% of the total
bytes. The LTE data traffic is highly skewed statistics of usage and the traffic volume in
the LTE network is dominated by a small fraction of users. However, the proportion of
traffic volume for the top 5% users is less than 60%, lower than that of previous result
[PSBD11], [SJLW11]. Thus we infer that data service in the LTE network is used more
prevalently than in the previous networks.
70
Number of data flows per device
10
0
10
1
10
2
10
3
10
4
10
5
10
6
Pr[X>x]
10
-14
10
-12
10
-10
10
-8
10
-6
10
-4
10
-2
10
0
Figure 6.7: Complementary CDF plot for LTE data flows
Location
In this section, we investigate the traffic patterns for V oLTE/3G voice calls and LTE
data flows. The location information for voice calls contains rather detailed base station
information. However, the location information for data flows does not record base sta-
tion (RU) level information and only contains DU level information. Thus the location
information of data flows covers more spacious areas than those of voice calls.
The histogram and the probability plot of V oLTE calls per base station are shown
in Figure 6.9 and 6.10, respectively. There are several thousand base stations in the
data collection area. The average number of V oLTE calls per base station is 566.89 for
the whole week while the maximum number per base station is 10,955 for the week.
71
Number of packets per device
10
3
10
4
10
5
10
6
10
7
10
8
Pr[X>x]
10
-14
10
-12
10
-10
10
-8
10
-6
10
-4
10
-2
10
0
Number of bytes per device
10
6
10
7
10
8
10
9
10
10
10
11
Pr[X>x]
10
-14
10
-12
10
-10
10
-8
10
-6
10
-4
10
-2
10
0
Figure 6.8: Complementary CDF plot for LTE data traffic: (Top) Packet (Bottom) Byte
72
Number of VoLTE calls per base station
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000 11000
Number of base stations
0
200
400
600
800
1000
1200
Figure 6.9: Histogram for V oLTE calls per RU
We find that the number of V oLTE calls with regard to location follows an exponential
distribution.
The histogram and the probability plot of 3G voice calls per base station are shown in
Figure 6.11 and 6.12, respectively. The average number of voice calls per base station is
18,978 for the whole week while the maximum number per base station is 96,088 for the
week. We find the voice calls with regard to location fits best to Rayleigh distribution.
Application
To explore data usage, we analyze the LTE data traffic with regard to type of applica-
tion. The traffic analysis system (LTAS) classifies data traffic according to the URL
73
Number of VoLTE calls per base station
1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
Probability
0.05
0.25
0.5
0.75
0.9
0.95
0.99
0.995
0.999
0.9995
Probability plot for Exponential distribution
Figure 6.10: Probability plot for V oLTE calls per RU
field of HTTP header. To identify the trend of subscribers’ data usage behavior, the sys-
tem classifies the traffic into 10 categories. Among these categories we classify mobile
applications in five categories: multimedia, application download, web surfing, instant
messenger (IM), and ”other.” The ”other” category includes email, games, P2P, and third
party V oIP traffic which are of interest to mobile network operators.
Figure 6.13 shows the composition of LTE traffic based on the application type for
the whole week. We find that smartphone users make use of web surfing most frequently
from the fact that the largest part of data flows consist of web surfing. However, mul-
timedia traffic contributes to more than a half of the traffic volume (i.e. the number of
packets or the number of bytes). Compared with previous analysis results, multimedia
74
Number of 3G voice calls per base station
×10
4
0 1 2 3 4 5 6 7 8 9
Number of base stations
0
20
40
60
80
100
120
140
160
180
200
Figure 6.11: Histogram for 3G voice calls per RU
and web traffic occupy more part of the total traffic volume as the mobile internet is
more widespread. The proportion of IM and social network service (SNS) is decreased
because the proportion of multimedia and web traffic increases. The composition of
traffic does not change throughout the week. Figure 6.14 shows the daily variation of
the traffic composition based on application type. We find that the proportion of traffic
with regard to the application type does not vary much.
75
Number of 3G voice calls per base station
×10
4
0 1 2 3 4 5 6 7 8 9 10
Probability
0.0001
0.01
0.05
0.1
0.25
0.5
0.75
0.9
0.95
0.99
0.995
0.999
0.9995
0.9999
Probability plot for Rayleigh distribution
Figure 6.12: Probability plot for 3G voice calls per RU
6.3 Modeling Voice Calls and Data Flows
In this section, we study statistical models for voice calls and data flows to find gener-
ative models for synthetic mobile traffic. We use flow-level traffic models describing
call/flow arrival and duration to represent workload characteristics. Since the time gran-
ularity of the traffic data is on a scale of one second, there are many concurrent calls
or data flows. Thus we cannot compute interarrival times for call/flow arrivals. Hence
we use Poissonness test or probability plot to find statistical models for call/flow arrival
and/or duration processes. In this section, we use the traffic of the busiest day (Septem-
ber 4th) in the whole week.
76
Figure 6.13: Traffic composition with regard to application type for the whole week
6.3.1 Voice Call
We first verify the call arrival processes of 3G CSFB and V oLTE calls. We show the
Poissonness plots for the call arrival processes in Figure 6.15. As we see in this figure,
the arrival processes of 3G and V oLTE voice calls follow a Poisson distribution well.
The average arrival rate of 3G voice calls is 74.74 (calls/sec) and that of V oLTE calls is
3.75 (calls/sec).
The call duration of 3G and V oLTE voice calls can be modeled with a lognormal
distribution as shown in Figure 6.16. The average duration of 3G voice call is 98.29
(sec) and that of V oLTE call is 92.45 (sec). The maximum duration of 3G voice call is
54,032 (sec) and that of V oLTE call is 27,005 (sec). These exceptionally long duration
voice calls are evidence of a long tail property.
77
Figure 6.14: Variation of traffic composition with regard to application type for the
whole week
6.3.2 Data Flow
In this subsection, we verify flow arrival and duration process of the LTE data traffic.
Considering the huge amount of data traffic volume, we use a one-hour time interval to
examine the flow-level model. In Figure 6.17, we draw a probability plot of the data
flow arrival process during the busiest hour in the whole week (6 p.m. and 7 p.m. on
September 4th) and see that it is well modeled by a normal distribution. Interestingly
we find the data flow arrival process follows a normal distribution during an hour period
when the flow arrival process is stationary. When the flow arrival rate is increasing or
decreasing during early morning or midnight, the flow arrival is non-stationary. As a
result, we find that the flow arrival process deviates from the normal distribution during
78
Number of 3G voice calls
0 20 40 60 80 100 120 140 160
φ (n
k
)
0
100
200
300
400
500
600
Number of VoLTE calls
0 5 10 15 20 25
φ (n
k
)
0
5
10
15
20
25
30
35
40
45
50
1
Figure 6.15: Poissonness plot for voice call arrival: (Top) 3G CSFB (Bottom) V oLTE
79
Figure 6.16: Probability plot for voice call duration: (Top) 3G CSFB (Bottom) V oLTE
80
Number of data flow arrivals
×10
4
1.3 1.4 1.5 1.6 1.7 1.8
Probability
0.0001
0.0005
0.001
0.005
0.01
0.05
0.1
0.25
0.5
0.75
0.9
0.95
0.99
0.995
0.999
0.9995
0.9999
Probability plot for Normal distribution
Figure 6.17: Probability plot of data flow arrival process
the one-hour period. However, we find that the data flow arrival process in fine time
scales (minutes or seconds) again follows the normal distribution.
The duration of data flows shows a highly skewed property. Most of the data flows
(86.4%) have a short duration less than one second. The duration for the remaining flows
(13.6%) having a duration longer than one second shows long-tail property as shown in
Figure 6.18.
81
Data flow duration (second)
10
0
10
1
10
2
10
3
10
4
10
5
Pr[X>x]
10
-14
10
-12
10
-10
10
-8
10
-6
10
-4
10
-2
10
0
Complementary CDF of data flow distribution
Figure 6.18: Complementary CDF plot for data flow duration
6.4 Summary
In this chapter, we characterize and model the traffic behavior of LTE devices including
voice and data. We collect most of the behavior made with LTE devices in the busiest
area in Korea during a whole week. We believe the dataset we used in this study is
one of the most comprehensive reported on so far. We characterize voice and data
usage patterns with respect to time and location and analyze voice call and data flow
arrival process and service time distribution. We find voice call arrivals follow a Poisson
process as has traditionally been assumed. However the service time distribution for
voice calls is lognormal. In case of data flows, we find the flow arrival process follows
82
a normal distribution. The duration of data flows shows a very highly skewed property.
Most of the data flows show very short duration. Except the data flows having zero
second duration records, we find the data flow duration exhibits the long tail property.
83
Chapter 7
Conclusion
In this dissertation, we have characterized and modeled the internet traffic dynamics in
wired and wireless environments. As we have seen, the internet is now used pervasively
in our daily life and the internet traffic is increasing explosively in wired and mobile net-
works. To cope with the explosive growth of data traffic volume and effectively support
network operations, network service providers need to design and manage their network
architectures accordingly. For achieving efficient network deployment and operations,
the first step is to understand the spatial and temporal variations of internet traffic carried
by their networks.
For characterization and modeling of the dynamics of various internet traffic, we use
large-scale data sets collected from commercial wired and wireless networks. Since the
traffic datasets from commercial networks are rarely available publicly due to privacy
concerns, traffic characteristics in commercial network environments are not known
well. Furthermore, the huge volume of internet traffic makes it difficult to manipulate
and analyze it easily. In this regard, the datasets collected from commercial networks in
Korea including fixed broadband network and WCDMA and LTE networks are valuable
for understanding general users’ behaviors and characteristics.
We have analyzed traffic characteristics and variability and profiled workload char-
acteristics imposed on the networks. In a fixed broadband internet, we first find the
multimedia workload characteristics in IPTV and V oIP service traffic. From the session-
level traffic analysis and modeling, we found that request arrival in IPTV V oD and call
arrival in V oIP follow a Poisson process. The service time of V oD is approximated by an
84
exponential distribution for the average length of the VOD contents. The service time
of V oIP follows a lognormal distribution having a long tail.
In case of mobile data traffic, we analyze revolutionary changes of mobile internet
traffic and its impact on network. We also perform characterization and modeling of
mobile internet traffic in recently commercialized LTE networks. We review the effect of
traffic surges due to the proliferation of mobile devices such as smartphones and tablets
combined with ubiquitous access to the mobile networks. In addition, we analyze the
increase of application signaling traffic which causes serious load in mobile networks.
We have performed analysis and modeling of mobile data traffic in an LTE network.
In this research we focus on the total workload imposed on the network by LTE devices.
Thus we collect voice call records and flow-level data traffic from an LTE network in
Korea. We found that call arrival process for 3G voice or voice over LTE (V oLTE) fits
into a Poisson distribution. The service times of 3G voice and V oLTE calls follows a
lognormal distribution. In addition, we confirmed that the voice calls have a skewed sta-
tistical property in that most of voice calls are made by a small proportion of users. This
phenomenon is more prominent in case of newly introduced services such as V oLTE.
The proportion of V oLTE calls made by the top 20% of users (82.3%) is much higher
than that of 3G voice calls (69.7%).
We observe that data traffic shows highly-skewed statistics with regard to user and
application. Modeling the data flows, we found that the flow arrival process (flows/sec)
follows a normal distribution. The duration of data flow shows highly-skewed statistics
in that 86.4% of flows have a very short duration. On the contrary, a small proportion of
data flows having long durations correspond to a long tail property.
The datasets used in this study are comprehensive and provide a way to understand
the characteristics of traffic dynamics in wired and wireless networks. The statistical
models we present are helpful to find generative models for synthetic mobile traffic.
85
Combining voice and data traffic models we can profile the total workload from mobile
devices and mobile users. We believe that understanding the spatial and temporal pat-
terns of traffic can help to predict changes in network resource requirements and provi-
sion network engineering accordingly.
In this dissertation, we identify several interesting phenomena in mobile traffic vari-
abilities. The distribution of the number of voice calls with respect to base station shows
different characteristics depending on call type (3G voice or V oLTE). We observe that
the data flow duration changes during busy hours although the number of active devices
and the number of flows per device are similar. We want further explanation on the
causes of these variabilities to understand the traffic dynamics in mobile networks. It is
interesting to analyze long-term traffic trends including the effect of recently introduced
pricing plans offering unlimited data usage. Lastly we believe that comparing the char-
acteristics of LTE voice and data traffic with that of previous networks sheds light on
future network deployment and engineering.
86
Reference List
[3GP08] 3GPP TS 29.108. Application of the Radio Access Network Application
Part (RANAP) on the E-interface. http://www.3gpp.org/ftp/Specs/html-
info/29108.htm, Dec. 2008.
[3GP11] 3GPP TS 22.801. Study on non-MTC Mobile Data Applications impacts
(Release 12). http://www.3gpp.org/ftp/Specs/html-info/22801.htm, Dec.
2011.
[3GP12] 3GPP TS 23.887. Machine-Type and other Mobile Data
Applications Communications Enhancements (Release 12).
http://www.3gpp.org/ftp/Specs/html-info/23887.htm, Aug. 2012.
[BMPR10] Robert Birke, Marco Mellia, Michele Petracca, and Dario Rossi. Expe-
riences of V oIP Traffic Monitoring in a Commercial ISP. International
Journal of Network Management, 20:339–359, 2010.
[BMV10] Aruna Balasubramanian, Ratul Mahajan, and Arun Venkataramani. Aug-
menting mobile 3g using wifi. In ACM MobiSys, 2010.
[Can12] Canalys. Smart phones overtake client PCs in 2011.
http://www.canalys.com/newsroom/smart-phones-overtake-client-pcs-
2011, Feb. 2012.
[CGW
+
08] Julin Candia, Marta C Gonzlez, Pu Wang, Timothy Schoenharl, Greg
Madey, and Albert-Lszl Barabsi. Uncovering individual and collective
human dynamics from mobile phone records. Journal of Physics A: Math-
ematical and Theoretical, 41(22), 2008.
[ChYK
+
14] Yongmin Choi, Cha hyun Yoon, Young-Sik Kim, Seo Weon Heo, and
J.A. Silvester. The impact of application signaling traffic on public land
mobile networks. Communications Magazine, IEEE, 52(1):166–172, Jan-
uary 2014.
87
[Cis15] Cisco. Cisco Visual Networking Index: Forecast and Methodology, 2014-
2019. http://www.intellinet-tech.com/Media/PagePDF/Data Offload.pdf,
May 2015.
[CJyP
+
11] Yongmin Choi, Hyun Wook Ji, Jae yoon Park, Hyun chul Kim, and J.A.
Silvester. A 3w network strategy for mobile data traffic offloading. Com-
munications Magazine, IEEE, 49(10):118–123, Oct 2011.
[CKR
+
09] Meeyoung Cha, Haewoon Kwak, Pablo Rodriguez, Yong-Yeol Ahn, and
Sue Moon. Analyzing the video popularity characteristics of large-scale
user generated content systems. IEEE/ACM Trans. Networking, 17:1357–
1370, 2009.
[Cox12] C. Cox. An Introduction to LTE. Wiley, 2012.
[CRC
+
08] Meeyoung Cha, Pablo Rodriguez, Jon Crowcroft, Sue Moon, and Xavier
Amatriain. Watching Television over an IP Network. In ACM Internet
Measruement Conference, 2008.
[CScK11] Yongmin Choi, J. Silvester, and Hyun chul Kim. Analyzing and modeling
workload characteristics in a multiservice ip network. Internet Computing,
IEEE, 15(2):35–42, March 2011.
[Gab12] Caroline Gabriel. VoIP Signaling Crashed NTT DoCoMo; Asks Google to
help. Broadband Traffic Management, Jan. 2012.
[GDS
+
03] P. Krishna Gummadi, Richard J. Dunn, Stefan Saroiu, Steven D. Gribble,
Henry M. Levy, and John Zahorjan. Measurement, modeling, and analysis
of a peer-to-peer file-sharing workload. In SOSP, pages 314–329, 2003.
[GJKV13] Maruti Gupta, Satish Chandra Jha, Ali Taha Koc, and Rath Vannithamby.
Energy impact of emerging mobile internet applications on LTE networks:
issues and solutions. IEEE Communications Magazine, 51(2):90–97,
2013.
[Glo15a] Global mobile Suppliers Association. LTE Market Status.
http://http://www.gsacom.com/, Jun. 2015.
[Glo15b] Global mobile Suppliers Association. Mobile HD Voice zone.
http://www.gsacom.com/hdvoice/, Apr. 2015.
[GSM12] GSM Association Smarter Applications Project. Smarter Apps for
Smarter Phones. http://www.gsma.com/documents/Smarter-Apps-for-
Smarter-Phones!/22440, Feb. 2012.
88
[Han09] A. Handa. Mobile data offload for 3G networks: a white paper.
http://www.intellinet-tech.com/Media/PagePDF/Data Offload.pdf, Oct.
2009.
[HLL
+
07] Xiaojun Hei, Chao Liang, Jian Liang, Yong Liu, and Keith W. Ross. A
Measurement Study of a Large-Scale P2P IPTV System. IEEE Transac-
tions on Multimedia, 9:1672–1687, 2007.
[HLR07] Cheng Huang, Jin Li, and Keith W. Ross. Can Internet Video-on-Demand
be Profitable? In ACM SIGCOMM, 2007.
[HP09] John Heidemann and Christos Papadopoulos. Uses and challenges for net-
work datasets. In IEEE Cybersecurity Applications and Technologies Con-
ference for Homeland Security (CATCH), 2009.
[HQG
+
13] Junxian Huang, Feng Qian, Yihua Guo, Yuanyuan Zhou, Qiang Xu,
Z. Morley Mao, Subhabrata Sen, and Oliver Spatscheck. An in-depth
study of lte: Effect of network protocol and application behavior on per-
formance. In Proceedings of the ACM SIGCOMM 2013 Conference on
SIGCOMM, pages 363–374, 2013.
[HW05] Emir Halepovic and Carey Williamson. Characterizing and modeling user
mobility in a cellular data network. In Proceedings of the 2Nd ACM Inter-
national Workshop on Performance Evaluation of Wireless Ad Hoc, Sensor,
and Ubiquitous Networks, pages 71–78, 2005.
[IMM07] Kashif Imran, Marco Mellia, and Michela Meo. Measurements of multi-
cast television over IP. In IEEE Workshop on Local and Metropolitan Area
Networks, pages 170–175, 2007.
[JGG
+
04] R.M. Joyce, B. Graves, T. Griparis, I.J. Osborne, and T.M. Lee. Case
study: the capacity of a wcdma network - orange uk. In 3G Mobile Com-
munication Technologies, 2004., pages 475–479, 2004.
[JS07] Vaishnav Janardhan and Henning Schulzrinne. Peer assisted V oD for set-
top box based IP network. In ACM SIGCOMM Workshop on P2P-TV,
2007.
[Kim12] Hoojong Kim. Smartphone Applications and Network Efficiency. GSM
Association Executive Management Committee Meeting, Feb. 2012.
[KNZG10] Ram Keralapura, Antonio Nucci, Zhi-Li Zhang, and Lixin Gao. Profiling
users in a 3g network using hourglass co-clustering. In Proceedings of
the Sixteenth Annual International Conference on Mobile Computing and
Networking, pages 341–352, 2010.
89
[Kor] Korea Communications Commission (KCC). Policy and Information,
Resources, Statistical data. http://eng.kcc.go.kr/.
[KT ] KT Corporation. Investors Relations, Finantial Statements, Annual
Report. http://www.kt.com/eng/index.jsp.
[LRL
+
10] Kyunghan Lee, Injong Rhee, Joohyun Lee, Song Chong, and Yung Yi.
Mobile Data Offloading: How Much Can WiFi Deliver? In ACM CoNEXT,
2010.
[LSM
+
08] Xin Liu, Ashwin Sridharan, Sridhar Machiraju, Mukund Seshadri, and
Hui Zang. Experiences in a 3g network: Interplay between the wireless
channel and applications. In Proceedings of the 14th ACM International
Conference on Mobile Computing and Networking, MobiCom ’08, pages
211–222, 2008.
[Mee09] M. Meeker. Economy + internet trends. In Web 2.0 Summit, Oct. 2009.
[Meg09] M. Megna. AT&T faces 5,000 percent surge in traffic.
http://www.internetnews.com/mobility/article.php/3843001/ATT-Faces-
5000-Percent-Surge-in-Traffic.htm, Oct. 2009.
[MM08] W. L. Martinez and A. R. Martinez. Computational Statistics Handbook
with MATLAB. Chapman & Hall, 2008.
[PPJP05] K. Pentikousis, M. Palola, M. Jurvansuu, and P. Perala. Active goodput
measurements from a public 3g/umts network. Communications Letters,
IEEE, 9(9):802–804, Sep 2005.
[PSBD11] U. Paul, A.P. Subramanian, M.M. Buddhikot, and S.R. Das. Understanding
traffic dynamics in cellular data networks. In INFOCOM, 2011 Proceed-
ings IEEE, pages 882–890, April 2011.
[PSJ03] Pubudu N. Pathirana, Andrey V . Savkin, and Sanjay Jha. Mobility mod-
elling and trajectory prediction for cellular networks with mobile base sta-
tions. In Proceedings of the 4th ACM International Symposium on Mobile
Ad Hoc Networking and Computing, pages 213–221, 2003.
[QGL
+
09a] Tongqing Qiu, Zihui Ge, Seungjoon Lee, Jia Wang, Jun Jim Xu, and
Qi Zhao. Modeling User Activities in a Large IPTV System. In ACM
Internet Measurement Conference, 2009.
[QGL
+
09b] Tongqing Qiu, Zihui Ge, Seungjoon Lee, Jia Wang, Qi Zhao, and Jun Jim
Xu. Modeling Channel Popularity Dynamics in a Large IPTV System. In
ACM SIGMETRICS/PERFORMANCE, 2009.
90
[Reu12] Reuters. Network outage raises Olympic service concerns. Chicago Tri-
bune, Jul. 2012.
[RU05] P. Reichl and M. Umlauft. Project wisqy: a measurement-based end-to-
end application-level performance comparison of 2.5g and 3g networks. In
Wireless Telecommunications Symposium, 2005, pages 9–14, April 2005.
[Sah12] Dawinderpal Sahota. O2 outage causes concern ahead of Olympics. Tele-
coms.com, Jul. 2012.
[SJL
+
13] Muhammad Zubair Shafiq, Lusheng Ji, Alex X. Liu, Jeffrey Pang, Shobha
Venkataraman, and Jia Wang. A first look at cellular network perfor-
mance during crowded events. In Proceedings of the ACM SIGMET-
RICS/International Conference on Measurement and Modeling of Com-
puter Systems, pages 17–28, 2013.
[SJLW11] M. Zubair Shafiq, Lusheng Ji, Alex X. Liu, and Jia Wang. Characterizing
and modeling internet traffic dynamics of cellular devices. In Proceedings
of the ACM SIGMETRICS, pages 305–316, 2011.
[Tel11] Telecommunications Technology Association. Push Notification Service
for Public Land Mobile Networks, Dec. 2011.
[TLL08] Wee Lum Tan, Fung Lam, and Wing Cheong Lau. An empirical study on
the capacity and performance of 3g networks. Mobile Computing, IEEE
Transactions on, 7(6):737–750, June 2008.
[V AJ
+
02] Eveline Veloso, Virg´ ılio A. F. Almeida, Wagner Meira Jr., Azer Bestavros,
and Shudong Jin. A hierarchical characterization of a live streaming media
workload. In Internet Measurement Workshop, pages 117–130, 2002.
[VGK
+
07] Nevena Vratonjic, Priya Gupta, Nikola Knezevic, Dejan Kostic, and
Antony Rowstron. Enabling dvd-like features in p2p video-on-demand
systems. In ACM SIGCOMM Workshop on P2P-TV, 2007.
[Wik] Wikipedia. Pareto Principle. http://en.wikipedia.org/wiki/Pareto principle.
[Woo09] N. Wood. Mobile data traffic growth 10 times faster than fixed over next
five years. http://www.totaltele.com/view.aspx?ID=448681, Sep. 2009.
[Wor09] J. Wortham. Customers angered as iPhones overload 3G.
http://www.nytimes.com/2009/09/03/technology/companies/03att.html,
Sep. 2009.
91
[YKH08] Jun Yao, Salil S. Kanhere, and Mahbub Hassan. An empirical study of
bandwidth predictability in mobile computing. In Proceedings of the Third
ACM International Workshop on Wireless Network Testbeds, Experimental
Evaluation and Characterization, pages 11–18, 2008.
[YZZZ06] Hongliang Yu, Dongdong Zheng, Ben Y . Zhao, and Weimin Zheng.
Understanding user behavior in large-scale video-on-demand systems. In
EuroSys, pages 333–344, 2006.
[ZDJC09] Min Zhang, Maurizio Dusi, Wolfgang John, and Changjia Chen. Analy-
sis of UDP traffic usage on Internet backbone links. In The 2009 Ninth
Annual International Symposium on Applications and the Internet, pages
280–281, 2009.
92

Abstract (if available)

Abstract The objective of this dissertation is to characterize and model traffic dynamics of wired and wireless internet for supporting effective deployment and operations of commercial communications networks. As the internet has become the prevalent network infrastructure for integrating broadcasting, telecommunications, and data communications, multimedia services driven by internet service providers (ISPs) have proliferated. Mobile internet services have become popular as mobile devices such as smartphones or tablets are pervasively used. As a result, we have seen persistent growth of internet traffic in wired and wireless networks. ❧ To cope with the growth of the data traffic volume generated by these mobile devices network operators need to increase network capacity, and design and operate their networks effectively. To increase the network capacity according to the forecasted traffic demand is not always cost-effective as the traffic usage patterns can change in unanticipated ways with the development of new applications. Furthermore it is not easy for mobile network operators to increase their network capacity according to the forecasted demand due to scarce radio frequency spectrum. The network operators are required to build their networks efficiently and optimize them according to the characteristics of usage and traffic behavior. To this end, the first step is to understand the spatial and temporal patterns of the internet traffic carried by their networks. ❧ In this dissertation, we perform extensive traffic collection, analysis and modeling of wired and wireless internet traffic to enhance network deployment and operational efficiency. We collected traffic datasets from a commercial network operator covering diverse services, times and locations. We characterize the variability and dynamics of internet traffic to identify inherent characteristics and combine statistical and empirical methodologies to model the identified traffic behaviors using appropriate parametric models. We also verify the effectiveness of suggested solutions to address several problems recently encountered in wired and wireless networks. ❧ Our results are helpful to understand traffic behavior in wired and wireless networks and to demystify traffic characteristics in commercial networks. The statistical models we found can also be used to find generative models for synthetic traffic generators in simulation or analytical studies. We believe that the solutions verified in this research can be applied to today’s networks including long term evolution (LTE) networks.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Efficient data collection in wireless sensor networks: modeling and algorithms

PDF

Cooperation in wireless networks with selfish users

PDF

Modeling intermittently connected vehicular networks

PDF

Optimal resource allocation and cross-layer control in cognitive and cooperative wireless networks

PDF

Aging analysis in large-scale wireless sensor networks

PDF

A protocol framework for attacker traceback in wireless multi-hop networks

PDF

Scheduling and resource allocation with incomplete information in wireless networks

PDF

Dynamic routing and rate control in stochastic network optimization: from theory to practice

PDF

Learning, adaptation and control to enhance wireless network performance

PDF

Relative positioning, network formation, and routing in robotic wireless networks

PDF

Gradient-based active query routing in wireless sensor networks

PDF

Enabling virtual and augmented reality over dense wireless networks

PDF

Improving user experience on today’s internet via innovation in internet routing

PDF

Detecting and characterizing network devices using signatures of traffic about end-points

PDF

Achieving efficient MU-MIMO and indoor localization via switched-beam antennas

PDF

Algorithmic aspects of energy efficient transmission in multihop cooperative wireless networks

PDF

Optimizing task assignment for collaborative computing over heterogeneous network devices

PDF

Machine learning for efficient network management

PDF

Design of cost-efficient multi-sensor collaboration in wireless sensor networks

PDF

Magnetic induction-based wireless body area network and its application toward human motion tracking

Asset Metadata

Creator Choi, Yongmin (author)

Core Title Understanding the characteristics of Internet traffic dynamics in wired and wireless networks

School Viterbi School of Engineering

Degree Doctor of Philosophy

Degree Program Electrical Engineering

Publication Date 10/14/2017

Defense Date 08/04/2015

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag fixed broadband network,long term evolution (LTE) network,OAI-PMH Harvest,traffic characteristics,user behavior,wideband code division multiple access (WCDMA) network,wireless network

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Silvester, John A. (committee chair), Golubchik, Leana (committee member), Krishnamachari, Bhaskar (committee member)

Creator Email yongminc@hotmail.com,yongminc@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c40-190107

Unique identifier UC11279444

Identifier etd-ChoiYongmi-3915.pdf (filename),usctheses-c40-190107 (legacy record id)

Legacy Identifier etd-ChoiYongmi-3915-1.pdf

Dmrecord 190107

Document Type Dissertation

Format application/pdf (imt)

Rights Choi, Yongmin

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA