wireless sensor network

5IEEE Network • July/August 2014 0890-8044/14/$25.00 © 2014 IEEE
ur world is generating data at a speed faster
than ever before. In 2010, 5 exabytes (1018
bytes, or 1 billion gigabytes) of data were cre-
ated every two days, exceeding the total
amount of information that was created by human beings
from the dawn of civilization to 2003.1 Till 2020, over 40
Zettabytes (1021 bytes) of data would be created, replicated,
and consumed.2 With the overwhelming amount of data pour-
ing into our lives, from anywhere, anytime, and any device, we
are undoubtedly entering the era of big data.
Big data brings big value. With advanced big data analyzing
technologies, insights can be acquired to enable better deci-
sion making for critical development areas such as health
care, economic productivity, energy, and natural disaster pre-
diction, to name but a few. For example, by collecting and
analyzing flu related keyword searches, Google has developed
the Flu Trends service to detect regional flu outbreaks in near
real time. Specifically, Google Flu Trends collected historical
search frequency data of 50 million common keywords each
week from 2003 to 2008. Then a linear model was used to
compute the correlation coefficient between each keyword
search history data and the actual influenza-like illness history
data obtained from the Centers for Disease Control and Pre-
vention (CDC) in the United States. After that, the keywords
with the highest correlation coefficients were picked out, and
their instant search frequencies were aggregated to predict
future flu outbreaks in the United States. With big data in
keyword searches, Google Flu Trends is able to detect flu out-
breaks over a week earlier than CDC, which can significantly
reduce the loss caused by the flu and even save lives. Another
example comes from the United Parcel Service (UPS), who
equips its vehicles with sensors to track their speed and loca-
tion. With the sensed data, UPS has optimized its delivery
routes and cut its fuel consumption by 8.4 million gallons in
2011.3 It has been reported that big data analytics is among
the top five catalysts that help increase U.S. productivity and
raise the gross domestic product (GDP) in the coming years.4
In general, these big data are stored in data warehouses
and processed in powerful data centers with massive intercon-
nected server nodes. There have been significant studies on
knowledge discovery and mining over the data and high-per-
formance parallel computing tools for big data (e.g., MapRe-
duce). But big data are not just born and reside there — they
are transmitted from a variety of sources and are to be uti-
lized at a variety of destinations for broad purposes. The life
cycle of big data consists of multiple stages ranging from gen-
O
Abstract
Big data, with their promise to discover valuable insights for better decision mak-
ing, have recently attracted significant interest from both academia and industry.
Voluminous data are generated from a variety of users and devices, and are to be
stored and processed in powerful data centers. As such, there is a strong demand
for building an unimpeded network infrastructure to gather geologically distributed
and rapidly generated data, and move them to data centers for effective knowl-
edge discovery. The express network should also be seamlessly extended to inter-
connect multiple data centers as well as interconnect the server nodes within a
data center. In this article, we take a close look at the unique challenges in build-
ing such a network infrastructure for big data. Our study covers each and every
segment in this network highway: the access networks that connect data sources,
the Internet backbone that bridges them to remote data centers, as well as the ded-
icated network among data centers and within a data center. We also present two
case studies of real-world big data applications that are empowered by network-
ing, highlighting interesting and promising future research directions.
Building a Network Highway for
Big Data: Architecture and Challenges
Xiaomeng Yi, Fangming Liu, Jiangchuan Liu, and Hai Jin
O
Xiaomeng Yi, Fangming Liu, and Hai Jin are with the Services Computing
Technology and System Lab, Cluster and Grid Computing Lab in the
School of Computer Science and Technology, Huazhong University of Sci-
ence and Technology.
Jiangchuan Liu is with Simon Fraser University.
1 “Gold Rush: The Scramble to Claim and Protect Value in the Digital
World,” Deloitte Review, 2011
2 “The Digital Universe in 2020: Big Data, Bigger Digital Shadows, and
Biggest Growth in the Far East,” IDC: Analyze the Future, 2012.
3 “Big Data in Big Companies,” SAS Research Report, 2013.
4 “Game Changers: Five Opportunities for US Growth and Renewal,”
McKinsey Global Institute Report, 2013.
LIU2_LAYOUT.qxp_Layout 1 7/17/14 1:49 PM Page 5

IEEE Network • July/August 20146
eration, collection, aggregation, processing, and application
delivery. While big data aggregation and processing occur
mostly in data centers, the data are generated by and collect-
ed from geographically distributed devices, and the data
knowledge/service are then distributed to interested users; the
latter heavily depends on inter-data-center networks, access
networks, and the Internet backbone, as depicted in Fig. 1. As
such, networking plays a critical role as the digital highway
that bridges data sources, data processing platforms, and data
destinations. There is a strong demand to build an unimpeded
network infrastructure to gather geologically distributed and
rapidly generated data, and move them to data centers for
effective knowledge discovery. The express network should
also be extended to interconnect the server nodes within a
data center and interconnect multiple data centers, thus col-
lectively expanding their storage and computing capabilities.
Data explosion, which has been a continuous trend since
the 1970s, is not news. The Internet has long grown together
with the explosion, and has indeed greatly contributed to it.
The three Vs (volume, variety, and velocity) from today’s big
data, however, are unprecedented. It remains largely unknown
whether the Internet and related networks can keep up with
the rapid growth of big data.
In this article, we examine the unique challenges when big
data meet networks, and when networks meet big data. We
check the state of the art for a series of critical questions:
What do big data ask for from networks? Is today’s network
infrastructure ready to embrace the big data era? If not,
where are the bottlenecks? And how could the bottlenecks be
lifted to better serve big data applications? We take a close
look at building an express network infrastructure for big
data. Our study covers each and every segment in this net-
work highway: the access networks that connect data sources,
the Internet backbone that bridges them to remote data cen-
ters, and the dedicated network among data centers and with-
in a data center. We also present two case studies of
real-world big data applications that are empowered by net-
working, highlighting interesting and promising future
research directions.
A Network Highway for Big Data: Why and
Where?
With so much value hiding inside, big data have been regard-
ed as the digital oil, and a number of government funded pro-
jects have been launched recently to build up big data
analyzing systems, involving a number of critical areas ranging
from healthcare to climate change and homeland security. A
prominent example is the $200 million National Big Data
Research & Development Initiative of the United States,
announced by the White House in March 2012.5 This initia-
tive involves six departments and agencies in the United
States, and aims to improve the tools and techniques needed
to access, organize, and glean discoveries from big data. A list
of representative projects worldwide is presented in Table 1.
In industry, there is also strong interest in exploiting big
data to gain business profits. Advances in sensor networking,
cyber-physical systems, and Internet of things have enabled
financial service companies, retailers, and manufacturers to
collect their own big data in their business processes. On the
other hand, through such utility computing services as cloud
computing, high-performance IT infrastructures and platforms
that were previously unaffordable are now available to a
broader market of medium and even small companies. As a
result, more and more enterprises are seeking solutions to
analyze the big data they generate. Gartner’s survey on 720
companies in June 2013 shows that 64 percent of companies
are investing or planning to invest in big data technologies.6 A
number of big data analytic platforms have already emerged
in this competitive market. For example, Google launched
5 “Obama Administration Unveils ‘Big Data’ Initiative: Announces $200
Million in New R&D Investments,” Office of Science and Technology Pol-
icy, The White House, 2012.
6 “Survey Analysis: Big Data Adoption in 2013 Shows Substance Behind
the Hype,” Gartner, 2013.
Figure 1. Three-layered network architecture from the perspective of big data applications.
Cellular
antenna
Sensor
PDA
Mobile
phone
Laptop
PC
Gateway ISP 1
CDN
Access network
Internet backbone
Inter- and intra-
data center network
...
ISP 2WiFi
AP
Replication
big data
Content
big data
DC network
Analytical
big data
...
Sensor
big data
Mobile
big data

IEEE Network • July/August 2014 7
BigQuery, a big data service platform that enables customers
to analyze their data by exploiting the elastic computing
resources in Google’s cloud; SAP also released its in-memory
big data platform HANA, which is capable of processing large
volumes of structured and unstructured data in real time.
In both government-funded research projects and business-
oriented services, the life cycle of big data consists of multiple
stages, as illustrated in Fig. 1. At first, the user data are gener-
ated from a variety of devices and locations, which are collect-
ed by wired and wireless networks. The data are then
aggregated and delivered to data centers via the global Inter-
net. In data centers, the big data are processed and analyzed.
Finally, the results are delivered back to users or devices of
interest and utilized.
Obviously, networks play a critical role in bridging the dif-
ferent stages, and there is a strong demand to create a fast
and reliable interconnected network for the big data to flow
freely on this digital highway. This network highway concerns
not only just one segment of data delivery, but rather the
whole series of segments for the life cycle of big data, from
access networks to the Internet backbone, and to intra- and
inter-data-center networks. For each layer of the network, the
specific requirements big data transmission poses should be
satisfied.
Access networks, which are directly connected to end
devices such as personal computers, mobile devices, sensors,
and radio frequency identification (RFID) devices, lie in the
outer layer. On one hand, such raw data from fixed or mobile
devices are transmitted into the network system; on the other
hand, processed big data and analytics results are sent back to
devices and their users. With the rapid development of wire-
less networks, more data are now collected from mobile
devices that have limited battery capacity. As such, energy-
efficient networking is expected to make batteries in mobile
devices more durable. Wireless links also suffer from interfer-
ence, which results in unstable bandwidth provisioning. Big
data applications such as cinematic-quality video streaming
require sustained performance over a long duration to guar-
antee the quality of user experience, which has become a criti-
cal challenge for wireless networking.
The Internet backbone serves as the intermediate layer
that connects access networks and data center networks. Pop-
ular big data applications like photo sharing and video shar-
ing allow users to upload multimedia contents to data centers
and share them with their friends in real time. To enable
good user experience, the Internet backbone needs to for-
ward massive geologically distributed data to data centers
with high throughput, and deliver processed data to users
from data centers with low latency. As such, high-perfor-
mance end-to-end links are required for uploading, and effi-
cient content distribution networks (CDNs) are required for
downloading.
Within a data center, big data are processed and analyzed
with distributed computing tools such as MapReduce and
Dryad, which involve intensive data shuffle among servers. A
scalable, ultra-fast, and blocking-free network is thus needed
to interconnect the server nodes. Multiple geologically dis-
tributed data centers can be exploited for load balancing and
low-latency service provisioning, which calls for fast network-
ing for data exchange, replication, and synchronization among
the data centers. Moreover, inter-data-center links are leased
from Internet service providers (ISPs) or dedicatedly deployed
by cloud providers with nontrivial costs. Effective data trans-
mission and traffic engineering schemes (e.g., using software
Table 1. Representative government-funded big data projects.
Project
Begin
time
Department Goal
1000 Genomes
Project
1/2008
National Institutes of
Health
To produce an extensive public catalog of human genetic variation,
including SNPs and structural variants, and their haplotype contexts.
ARM Project 3/2012 Department of Energy
To collect and process climate data from all over the world to under-
stand Earth’s climate and come up with answers to climate change
issues.
XDATA 3/2012
Defense Advanced
Research Projects Agency
(DARPA)
To develop new computational techniques and software programs
that can analyze structured and unstructured big data sets faster and
more efficiently.
BioSense 2.0 3/2012
Center for Disease Control
and Prevention
To track public health problems and make data instantly accessible to
end users across government departments.
The Open Science
Grid
3/2012
National Science
Foundation (NSF) and
Department of Energy
To provide advanced fabric of services for data transfer and analysis to
scientists worldwide for collaboration in science discovery.
Big Data for Earth
System Science
3/2012 U.S. Geological Survey
To provide scientists with state-of-the-art computing capabilities and
collaborative tools to make sense of huge data sets and better under-
stand the earth.
Human Brain Project 2/2013 European Commission
To simulate the human brain and model everything that scientists
know about the human mind using a supercomputer.
Unique Identification
Authority
2/2009
The Indian Planning
Commission
To create a biometric database of fingerprints, photographs, and iris
scan images of all 1.2 billion people for efficient resident identifica-
tion in welfare service delivery.
7 “The Zettabyte Era—Trends and Analysis,” Cisco Visual Networking
Index white paper, 2013.

defined networking, SDN) that fully utilize the capacity of
such links are therefore expected. For instance, Google and
Microsoft both attempt to deploy SDN-based global data cen-
ter WANs to achieve fault tolerance, utilization improvement,
and policy control that can hardly be realized by traditional
WAN architectures.
In summary, to attain full speed for big data transmission
and processing, every segment of the network highway should
be optimized and seamlessly concatenated. We proceed to
investigate each part of the network system by identifying the
unique challenges that big data applications pose and analyz-
ing the state-of-the-art works that endeavor to build the net-
work highway for big data.
Access Network: Linking Sources
With the fast progress of digitalization and the development
of sensor networks, huge amounts of data are collected by all
kinds of end devices like PCs, smartphones, sensors, and GPS
devices. Meanwhile, applications like online social networks
and video streaming push rich content big data into user
devices. The access network plays a critical role here in gath-
ering such distributed data and forwarding them through the
Internet to data centers. While the last mile problem has been
addressed well with today’s high-speed home and office net-
work connections, the wireless connection remains a severe
bottleneck. Cisco predicts that traffic from wireless and
mobile devices will exceed traffic from wired devices by 2016,
which makes wireless network performance optimization of
paramount importance.
Wireless broadband technologies have evolved significantly
in recent years; but when facing data-intensive applications,
they are still insufficient to satisfy the bandwidth require-
ments. Yoon et al. exploit the wireless broadcast advantage
and use it to bridge the gap between wireless networking
capacity and the bandwidth demands of video applications.
MuVi [1], a multicast delivery scheme, is proposed to opti-
mize video distribution. By prioritizing video frames according
to their importance in video reconstruction and exploiting a
resource allocation mechanism that maximizes the system util-
ity, MuVi improves the overall video quality across all users in
a multicast group.
Multiple-input multiple-output orthogonal frequency-divi-
sion multiplexig (MIMO-OFDM) technologies, which can sig-
nificantly increase wireless capacity, have become the default
building blocks in the next generation of wireless networks.
Liu et al. [2] observe that in wireless video streaming applica-
tions, both the video coding scheme and the MIMO-OFDM
channel present non-uniform energy distribution among their
corresponding components. Such non-uniform energy distri-
bution in both the source and the channel can be exploited
for fine-grained unequal error protection for video delivery in
error-prone wireless networks. To this end, ParCast, an opti-
mized scheme for video delivery in MIMO-OFDM channels,
is proposed. It separates the video coding and wireless chan-
nel into independent components and allocates more impor-
tant video components with higher-gain channel components.
This leads to significantly improved quality for video over
wireless.
In addition to the aforementioned works, there are several
other promising approaches to improve the quality of wire-
less video streaming services. On the user’s side, an applica-
tion-aware MIMO video rate adaptation mechanism can be
deployed. It detects changes in a MIMO channel and adap-
tively selects an appropriate transmission profile, thereby
improving the quality of the delivered video. On the server’s
side, standard base station schedulers work on a fine-grained
per-packet basis to decrease the delivery delay of single
packets. However, it is insufficient to guarantee video watch-
ing experience at a coarse granularity, such as a fixed video
bit rate over several seconds, which would typically consist of
video content in thousands of packets. In response, a video
management system, which schedules wireless video delivery
at a granularity of seconds with knowledge of long-term
channel states, has the potential to further improve user
experience. Moreover, distortion and delay, which are two
important user experience metrics, conflict with each other
in wireless networks. The optimal trade-off between distor-
tion and delay in wireless video delivery systems largely
depends on the specific features of video flows. As such, a
policy that smartly balances distortion and delay according
to the features of video flows can also improve user experi-
ence.
Internet Backbone: From Local to Remote
Beyond the access network, the user or device generated data
will be forwarded through the Internet backbone to data cen-
ters. For example, in mobile cloud computing services [3],
where powerful cloud resources are exploited to enhance the
performance of resource-constrained mobile devices, data
from geologically distributed mobile devices are transmitted
to the cloud for processing. Given that the local data come
worldwide, the aggregated data toward a data center can be
enormous, which creates significant challenges to the Internet
backbone. Table 2 summarizes the Internet backbone solu-
tions for big data.
End-to-End Transmission
With the growing capacity of access links, network bottlenecks
are observed to be shifting from the network edges in access
networks to the core links in the Internet backbone. To
improve the throughput of end-to-end data transmission, path
diversity should be explored, which utilizes multiple paths
concurrently to avoid individual bottlenecks. A representative
is mPath [4], which uses a large set of geographically distribut-
ed proxies to construct detour paths between end hosts. An
additive increase and multiplicative decrease (AIMD) algo-
rithm similar to TCP is used to deal with congested proxy
paths to adaptively regulate the traffic over them, or even
completely avoid them.
Besides uploading, the big data, after being processed
by the data centers, also need to be downloaded by users
to appreciate the inside value. Downloading, however,
poses different demands. For applications like online
social networks, it is critical to deliver user required con-
tents with low latency while providing a consistent service
to all users. Wittie et al. [5] reverse engineered Facebook,
investigating the root causes of its poor performance when
serving users outside of the United States. They suggest
that this can be improved by exploring the locality of inter-
est, which, with proxy and caching, can dramatically reduce
the backbone traffic of such online social networks as well
as its access delay.
Content Delivery Network
For geologically distributed data consumers, CDNs can be
explored to serve them with higher throughput. High through-
put is typically achieved in two ways: optimizing path selec-
tion to avoid network bottlenecks, and increasing the number
of peering points. Yu et al. [6] introduce a simple model to
illustrate and quantify the benefit of them. Using both syn-
thetic and Internet network topologies, they show that
increasing the number of peering points improves the

throughput the most, while optimal path selection has only
limited contribution. Liu et al. [7] further find that video
delivery optimized for low latency or high average throughput
may not work well for high-quality video delivery that
requires sustained performance over a long duration. This
leads to an adaptive design with global knowledge of network
and distribution of clients.
To further reduce the operational cost of big data traffic
over CDN, Jiang et al. [8] suggest that the CDN infrastruc-
tures can be extended to the edges of networks, leveraging
such devices such as set-top boxes or broadband gateways.
Their resources can be utilized through peer-to-peer commu-
nications with smart content placement and routing to miti-
gate the cross-traffic among ISPs.
Table 2. A taxonomy of Internet backbone, intra- and inter-data-center solutions for big data.
Internet backbone
Approaches
Network
infrastructure
Big data
application
Goal Technique Evaluation method Overhead
mPath [4]
Internet
backbone
End-to-end
transmission
Avoid
bottleneck
AIMD algo-
rithm
Implementation on
PlanetLab
Proxy node deployment
Wittie et al. [5]
Internet
backbone
Social
networks
Reduce
service
latency
TCP proxy,
caching
Trace-driven
simulation
Cache and proxy
deployment
Liu et al. [7] CDN
Video
streaming
Improve
QoS
Bit rate
adaption
Trace-driven
simulation
Low scalability
Jiang et al. [8] CDN
Video
streaming
Reduce
operational
cost
CDN extension
Synthetic and
trace-driven
simulation
Tracker
deployment
Intra- and inter-data-center networks
Approaches
Network
infrastructure
Big data
application
Goal Technique Evaluation method Overhead
Hedera [9]
Data center
networks
Data
processing
Optimize
network
utilization
Flow
scheduling
Simulation,
implementation on
Portland testbed
Centralized control, low
scalability
FlowComb [10]
Data center
networks
MapReduce/
Dryad
Optimize
network
utilization
Flow
scheduling
Implementation on
Hadoop testbed
Monitor and transfer
demand information
Orchestra [11]
Data center
networks
MapReduce/
Dryad
Reduce job
duration
Transfer
scheduling
Implementation on
Amazon EC2 and
DETERlab
Modify the distributed
framework
RoPE [12]
Data center
networks
MapReduce/
Dryad
Reduce job
duration
Execution plan
optimization
Implementation on
Bing’s production
cluster
Pre-run jobs to acquire
job property
Camdoop [13]
Data center
network
topology
MapReduce/
Dryad
Decrease
network
traffic
Data
aggregation
Implementation on
CamCube
Special network
topology deployment
Mordia [14]
OCS
data center
network
Data
processing
Reduce
switching
delay
OCS, traffic
matrix
scheduling
Prototype
implementation
Hardware/topology
deployment
3D Beamform-
ing [15]
Wireless
data center
network
Data
processing
Flexible
bandwidth
provisioning
60 GHz
wireless links
Local testbed,
simulation
Physical antenna/
reflector deployment
NetStitcher [16]
Inter-data-
center links
Data backup
and
migration
Improve
network
utilization
Store-and-
forward
algorithm
Emulation, live
deployment
Periodical schedule
recomputation
Jetway [17]
Inter-data-
center links
Video
delivery
Minimize
link cost
Flow
assignment
algorithm
Simulation,
implementation on
Amazon EC2
Centralized controller,
video flow tracking

Data Center Networks: Where Big
Data Are Stored and Processed
Big data collected from end devices are stored
and processed in data centers. Big data applica-
tions such as data analysis and deep learning usu-
ally exploit distributed frameworks like
MapReduce and Dryad to achieve inflexibility
and scalability. Data processing in those distribut-
ed frameworks consists of multiple computational
stages (e.g., map and reduce in the MapReduce
framework). Between the stages, massive amounts
of data need to be shuffled and transferred among
servers. The servers usually communicate in an
all-to-all manner, which requires high bisection bandwidth in
data center networks. As such, data center networks often
became a bottleneck for those applications, with data trans-
fers accounting for more than 33 percent of the running time
in typical workloads. In Table 2, we summarize the state-of-
the-art solutions toward engineering better intra- and inter-
data-center networks.
Dynamic Flow Scheduling
As shown in Fig. 2, there are multiple equal-cost paths
between any pair of servers in a typical multi-rooted tree
topology of a data center network. To better utilize the paths,
Hedera [9] is designed to dynamically forward flows along
these paths. It collects flow information from switches, com-
putes non-conflicting paths for flows, and instructs switches to
reroute traffic accordingly. With a global view of routing and
traffic demands, Hedera is able to maximize the overall net-
work utilization with only small impact on data flows. Through
monitoring the flow information in switches, it schedules data
flows only after they cause congestion at certain location. To
actually avoid congestion, however, the data flows should be
detected even before they occur. This is addressed in Flow-
Comb [10], which predicts data flows effectively by monitoring
MapReduce applications on servers through software agents.
A centralized decision engine is designed to collect data from
the agents and record the network information.
Transfer Optimization in MapReduce-Like Frameworks
In addition to scheduling individual flows, optimizing the
entire data transfer has the potential to further reduce job
completion time. To this end, Orchestra [11] optimizes com-
mon communication patterns like shuffle and broadcast by
coordinating data flows. When there are concurrent transfers,
Orchestra enforces simple yet effective transfer scheduling
policies such as first-in first-out (FIFO) to reduce the average
transfer time. This is further enhanced in RoPE [12] by opti-
mizing job execution plans. An execution plan specifies the
execution order of operations in a job, as well as the degree of
parallelism in each operation. RoPE employs a composable
statistics collection mechanism to acquire code and data prop-
erties in a distributed system, and then automatically gener-
ates optimized executions plans for jobs, which reduces the
volume of data to be transferred.
Novel Topology and Hardware
Novel network topologies have also been proposed to improve
the network performance in big data processing. Costa et al.
[13] observe that in common workloads, data volumes are
greatly reduced after the progress of data processing. With this
insight, a data processing framework, Camdoop, is proposed. It
is built on a novel network topology where the servers are
directly connected, through which partial data can be aggregat-
ed on servers along the path to minimize data transmission.
Recently, optical circuit switching (OCS) has been suggest-
ed to accommodate the fast growing bandwidth demands in
data center networks. The long circuit reconfiguration delay
of OCSs, however, hinders their deployment in modern data
centers. To address this issue, Porter et al. [14] propose novel
traffic matrix scheduling (TMS), which leverages application
information and short-term demand estimates to compute
short-term circuit schedules, and proactively communicates
circuit assignments to communicating entities. As such, it can
support flow control in microseconds, pushing down the
reconfiguration time by two to three orders of magnitude. On
the other hand, wireless links in the 60 GHz band are attrac-
tive to relieve hotspots in oversubscribed data center net-
works. The 60 GHz wireless links require direct line of sight
between sender and receiver, which limits the effective range
of wireless links. Moreover, they can suffer from interference
when nearby wireless links are working. To deal with these,
Zhou et al. [15] propose a new wireless primitive, 3D beam-
forming, for data centers. By bouncing wireless signals off the
ceiling, 3D beamforming avoids blocking obstacles, thus
extending the range of each wireless link. Moreover, the sig-
nal interference range is significantly reduced, allowing nearby
links to work concurrently.
Inter-Data-Center Links
Big-data-based services such as social networks usually exploit
several geographically distributed data centers for replication
and low-latency service provision. Those data centers are
interconnected by high-capacity links leased from ISPs. Oper-
ations like data replication and synchronization require high
bandwidth transformation between data centers. It is thus crit-
ical to improve the utilization or reduce the cost for such
inter-data-center links.
Laoutaris et al. [16] observe a diurnal pattern of user
demand on inter-data-center bandwidth, which results in low
bandwidth utilization in off-peak hours. They propose Net-
Stitcher to utilize the leftover bandwidth in off-peak hours for
such non-real-time applications as backups and data migra-
tions. As illustrated in Fig. 3, NetStitcher exploits a store-and-
forward algorithm to transfer big data among data centers.
The data are split into pieces and transferred to their destina-
tion along multiple paths, each of which consists of a series of
intermediate data centers. A scheduling module decides when
and where the data pieces should travel according to the
available bandwidth.
For real-time big data applications like video streaming,
Feng et al. [17] propose Jetway, which, based on a widely used
percentile charging model, uses the qth largest traffic volume of
all time intervals during the charging period as the charging
volume. If the current time interval’s traffic volume exceeds
that of the qth percentile of previous time intervals, it will
incur additional cost. Otherwise, the already paid for band-
width will not be fully used. Jointly considering the link capac-
Figure 2. Data flow scheduling and data shuffling techniques in intra-data-
center networks.
Individual
data flow
Data shuffling
in MapReduce
Data flow
prediction
Multi-rooted tree topology
Equal-cost
Path
......

ity, bandwidth availability, delay-tolerant degree of flows, and
previous traffic volume of links, Jetway can minimize the cost
of inter-data-center links.
In Summary
The aforementioned research efforts propose solutions to
serve big data applications at different levels ranging from the
macroscopic multi-data-center level to the microscopic individ-
ual flows. Specifically, from the macroscopic level, inter-data-
center links enable multiple geologically distributed data
centers to provide such consistent reliable big data services as
video streaming and social networking with low latency. At
the single-data-center level, novel network topologies and
hardware devices provide cost-effective solutions toward a
high-performance network that interconnects all servers. At
the microscopic level, coordinated all-to-all communications
among a subset of servers in a data center mitigate network
congestion and reduce the completion times of data process-
ing jobs. In particular, for individual data flows between a spe-
cific pair of servers, dynamic flow scheduling techniques
choose proper paths for data flows to avoid network resource
competition. Nevertheless, there remain significant open ques-
tions to be addressed in the literature, such as how to sched-
ule data flows in order to meet their specific completion
deadlines, and how to provide guaranteed network perfor-
mance to multiple concurrent data processing jobs in a data
center. To achieve a comprehensive understanding of these
problems, Xu et al. [18] have investigated the state-of-the-art
research in the literature on providing guaranteed network
performance for tenants in Internet as a service (IaaS) clouds.
Big-Data-Based Networking Applications
We generally classify big data applications into two categories,
Internet applications and mobile wireless network applications,
with regard to the networking infrastructure on which they
work. For each category, we discuss the benefit and opportu-
nity that big data brings, by analyzing representative applica-
tions depicted in Fig. 4.
Internet Applications
One of the prominent big data applications closely related to
our daily lives is Netflix, which offers streaming video-on-
demand services and now takes up a third of U.S. download
Internet traffic during peak traffic hours. To support the com-
bination of huge traffic and unpredictable demand bursts,
Netflix has developed a global video distribution system using
Amazon’s cloud. Specifically, depending on customer demand,
Netflix’s front-end services are running on 500 to 1000 Linux-
based Tomcat JavaServer and NGINX web servers. These are
empowered by hundreds of other Amazon S3 and NoSQL
Cassandra database servers using the Memcached high-perfor-
mance distributed memory object caching system.
Netflix purchases master copies of digital films from movie
studios and, using the powerful Amazon EC2 cloud machines,
converts them to over 50 different versions with different
video resolutions and audio quality, targeting a diverse array
of client video players running on desktop computers, smart-
phones, and even DVD players or game consoles connected
to television. The master copies and the many converted
copies are stored in Amazon S3. In total, Netflix has over 1
petabyte of data stored on Amazon.
Thanks to insights from such big video data and the associ-
ated client behavior data as the programs they are watching,
their demographics and preferences, Netflix made a big deci-
sion of bidding over $100 million for two seasons of the U.S.
version of “House of Cards” in 2011, which turned out to be a
great success. The “House of Cards” series, directed by David
Fincher, starring Kevin Spacey, is based on a popular British
series. Big data in Netflix show that its British version has
been well watched, and the same subscribers who loved the
original BBC production also like movies starring Kevin
Spacey or directed by David Fincher. The analytical results of
big data indicate that “House of Cards” would bring signifi-
cant business value to Netflix. In fact, the series has brought
Netflix over two million new subscribers in the United States
and one million outside the United States. Netflix’s purchase
of “House of Cards” is largely regarded as a great success of
big data.
Mobile Wireless Network Applications
Powered by advanced technologies in mobile networking, sen-
sor networking, and the Internet of Things, mobile wireless
big data applications are emerging in our lives. As a typical
example of them, Nike+ provides improved service to Nike
users with wireless connected sensing devices and smart-
phones. In particular, Nike enhances its products, such as
shoes and wristbands, with built-in sensors that continuously
Figure 3. Emerging network architectures of different data centers, and multipath store-and-forward
data transfer among geo-distributed data centers.
Multi-rooted tree
topology
Novel topology
with servers
directly connected
Wireless
datacenter
network
Source
Store and forward
Multipath
data transfer
Percentile
charging
Video
delivery
Destination
Data
migration
Data backup
Optical circuit
switching
datacenter network

track the users’ movement during their workouts. The users
can install Nike+ apps in their smartphones, which collect
data from the sensors through wireless connections. The col-
lected data provide users with their instant exercise informa-
tion such as their pace, GPS position, distance moved, and
calories burned. Moreover, with mobile networks intercon-
necting users’ smartphones, the Nike+ apps enable social
interactions among users. They can share their progress and
exercising experience to cheer each other on or create groups
to go after a goal together. By now, Nike+ has become a big
data platform that collects, stores, and processes data generat-
ed from more than 18 million users.
A series of issues in wireless networking, however, should
be addressed when collecting and transferring user data in
Nike+. To provide users with real-time feedback, both sensed
data in wireless sensor networks and social interaction data in
mobile cellular networks should be synchronized to users’
smartphones frequently. Unfortunately, high-frequency data
transmission in wireless networks would result in high energy
consumption. Both wireless sensors and smartphones have
limited battery capacity. Transferring data at too high a fre-
quency would cause degradation in battery life. It is nontrivial
for wireless networks to provide an energy-efficient solution
for frequent data transmissions. In its latest product, Nike+
Fuelband, Nike+ adopts Bluetooth 4.0 wireless techniques for
data transmission. By exploiting the novel Bluetooth LE (for
low energy) protocol to synchronize sensed data to smart-
phones, Nike+ Fuelband has been made more durable.
Conclusion and Future Trends
So far we have reviewed the networking architecture and ser-
vices for big data applications. We identify the major chal-
lenges big data applications bring to networking systems and
discuss the state-of-the-art research efforts to meet the
demand of big data over networking. With the rapid growth of
big data applications, building the network system as a high-
way for big data transmission will continue to be a hot topic in
both academia and industry. We now identify some notable
future trends.
Standardization: Bridging the Fragmented World
The world of big data remains largely fragmented. Today’s big
data are generally stored and analyzed within a particular
business entity or organization to obtain insights and guide
decision making. There is a great demand to exchange these
big data for better insight discovery. For example, manufac-
turing factories may require feedback data from retailers to
discover user demands and help design new products. The
retailers, in turn, would need product-related data to set prop-
er prices for products and recommend them to target con-
sumers. Standards are thus necessary to bridge the fragmented
world. There has been active development on big data storage
and exchange. For example, the National Institute of Stan-
dards and Technology (NIST) set up a big data working group
on June 19, 2013, aimed at defining the requirements for
interoperability, reusability, and extendability of big data ana-
lytic techniques and infrastructures.
From the perspective of networking, the standards are
required to specify how to transfer big data between different
platforms. The transferred big data may contain semi-struc-
tured or unstructured data, such as pictures, audios, videos,
click streams, log files, and the output of sensors that measure
geographic or environmental information. The standards
should specify how these data should be encoded and trans-
ferred in network systems in order to facilitate network quali-
ty of service (QoS) management with low latency and high
fidelity. Moreover, as novel technologies such as software
defined networking (SDN) keep emerging in networking sys-
tems, corresponding standards are required to specify how
these technologies can be exploited for efficient big data
transmission. For example, the Open Network Foundation
(ONF) has been releasing and managing the OpenFlow stan-
dard since 2011, which defines the communications interface
between the control and forwarding layers in an SDN archi-
tecture.
Privacy and Security
With a variety of personal data such as buying preferences,
healthcare records, and location-based information being col-
lected by big data applications and transferred over networks,
the public’s concerns about data privacy and security naturally
arise. While there have been significant studies on protecting
data centers from being attacked, the privacy and security
loopholes when moving crowsourced data to data centers
remain to be addressed. There is an urgent demand on tech-
nologies that endeavor to enforce privacy and security in data
transmission. Given the huge data volume and number of
sources, this requires a new generation of encryption solutions
(e.g., homomorphic encryption).
On the other hand, big data techniques can also be used to
address the security challenges in networked systems. Network
attacks and intrusions usually generate data traffic of specific
patterns in networks. By analyzing the big data gathered by a
network monitoring system, those misbehaviors can be identi-
fied proactively, thus greatly reducing the potential loss.
Figure 4. Representative big data applications in the Internet and wireless networks: left: Netflix’s “House of Cards”; right: Nike+.
Video streaming
ServingBidding
User
feedbacks
User
preference
In 2006, Nike
released the
Nike+ platform
How can Nike+ use big data?Nike’s way to big data
EcosystemMaterial analysis
Precision
marketingProduct
improvement
In 2010, Nike
Digital Sport
division is
launched
In 2012, Nike+
Accelerator
program is
launched

Empowered by real-time big data analyzing technologies, we
expect that a series of advanced tools for identifying deep
security loopholes in large-scale and complex system could be
developed in the near future.
Acknowledgment
The corresponding author is Fangming Liu (fmliu@hust.edu.cn).
The research was supported by a grant from the National Basic
Research Program (973 program) under grant No.
2014CB347800.
References
[1] J. Yoon et al., “MuVi: A Multicast Video Delivery Scheme for 4G Cellular
Networks,” Proc. ACM Mobicom, 2012.
[2] X. L. Liu et al., “ParCast: Soft Video Delivery in MIMO-OFDM WLANs,”
Proc. ACM Mobicom, 2012.
[3] F. Liu et al., “Gearing Resource-Poor Mobile Devices with Powerful Clouds:
Architecture, Challenges and Applications,” IEEE Wireless Commun., Spe-
cial Issue on Mobile Cloud Computing, vol. 20, no. 3, 2013.
[4] Y. Xu et al., “mPath: High-Bandwidth Data Transfers with Massively-Multi-
path Source Routing,” IEEE Trans. Parallel and Distributed Systems, vol.
24, issue 10, 2013.
[5] M. P. Wittie et al., “Exploiting Locality of Interest in Online Social Net-
works,” Proc. ACM Eurosys, 2010.
[6] M. Yu et al., “Tradeoffs in CDN Designs for Throughput Oriented Traffic.”
Proc. ACM Conext, 2012.
[7] X. Liu et al., “A Case for a Coordinated Internet Video Control Plane,”
Proc. ACM Sigcomm, 2012.
[8] W. Jiang et al., “Orchestrating Massively Distributed CDNs,” Proc. ACM
Conext, 2012.
[9] M. Al-Fares et al., “Hedera: Dynamic Flow Scheduling for Data Center
Networks,” Proc. USENIX NSDI, 2010.
[10] A. Das et al., “Transparent and Flexible Network Management for Big
Data Processing in the Cloud,” Proc. ACM Hotcloud, 2013.
[11] M. Chowdhury et al., “Managing Data Transfers in Computer Clusters
with Orchestra,” Proc. ACM Sigcomm, 2011.
[12] S. Agarwal et al., “Re-optimizing Data-Parallel Computing,” Proc.
USENIX NSDI, 2012.
[13] P. Costa et al., “Camdoop: Exploiting In-Network Aggregation for Big
Data Applications,” Proc. USENIX NSDI, 2012.
[14] G. Porter et al., “Integrating Microsecond Circuit Switching into the Data
Center,” Proc. ACM Sigcomm, 2013.
[15] X. Zhou et al., “Mirror Mirror on the Ceiling: Flexible Wireless Links for
Data Centers,” Proc. ACM Sigcomm, 2012.
[16] N. Laoutaris et al., “Inter-Datacenter Bulk Transfers with NetStitcher,”
Proc. ACM Sigcomm, 2011.
[17] Y. Feng, B. Li, and B. Li, “Jetway: Minimizing Costs on Inter-Datacenter
Video Traffic,” Proc. ACM MM, 2012.
[18] F. Xu et al., “Managing Performance Overhead of Virtual Machines in
Cloud Computing: A Survey, State of Art and Future Directions,” Proc.
IEEE, vol. 102, no. 1, 2014.
Biographies
XIAOMENG YI (xiaomengyi@hust.edu.cn) is currently a Ph.D. student in the
School of Computer Science and Technology, Huazhong University of Science
and Technology, Wuhan, China. His current research interests focus on cloud
computing, modeling and optimization.
FANGMING LIU [M] (fmliu@hust.edu.cn) is an associate professor in the School
of Computer Science and Technology, Huazhong University of Science and
Technology, and he is named the CHUTIAN Scholar of Hubei Province,
China. He is the Youth Scientist of National 973 Basic Research Program Pro-
ject on Software-Defined Networking (SDN)-Based Cloud Data Center Net-
works, which is one of the largest SDN projects in China. Since 2012, he has
also been invited as a StarTrack Visiting Young Faculty at Microsoft Research
Asia (MSRA), Beijing. He received his B.Eng. degree in 2005 from the
Department of Computer Science and Technology, Tsinghua University, Bei-
jing, and his Ph.D. degree in computer science and engineering from the
Hong Kong University of Science and Technology in 2011. From 2009 to
2010, he was a visiting scholar at the Department of Electrical and Computer
Engineering, University of Toronto, Canada. He was the recipient of Best
Paper Awards from IEEE GLOBECOM 2011, IEEE IUCC 2012, and IEEE
CloudCom 2012, respectively. His research interests include cloud computing
and data center networking, mobile cloud, green computing and communica-
tions, software-defined networking and virtualization technology, large-scale
Internet content distribution, and video streaming systems. He is a member of
ACM, as well as a member of the China Computer Federation (CCF) Internet
Technical Committee. He has been a Guest Editor of IEEE Network and IEEE
Systems Journal, an Associate Editor of Frontiers of Computer Science, and
served on the Technical Program Committees for IEEE INFOCOM 2013–2015,
ICNP 2014, ACM Multimedia 2014, and IEEE GLOBECOM 2012–2014. He
served as the IEEE LANMAN 2014 Publicity Chair and GPC 2014 Program
Chair.
JIANGCHUAN LIU [SM] (jcliu@cs.sfu.ca) received his B.Eng. degree (cum laude)
from Tsinghua University, Beijing, China, in 1999, and his Ph.D. degree from
the Hong Kong University of Science and Technology in 2003, both in com-
puter science. He is currently an associate professor in the School of Comput-
ing Science, Simon Fraser University, British Columbia, Canada, and was an
assistant professor in the Department of Computer Science and Engineering at
the Chinese University of Hong Kong from 2003 to 2004. He is a recipient of
a Microsoft Research Fellowship (2000), Hong Kong Young Scientist Award
(2003), and Canada NSERC DAS Award (2009). He is a co-recipient of the
ACM Multimedia 2012 Best Paper Award, the IEEE GLOBECOM 2011 Best
Paper Award, the IEEE Communications Society Best Paper Award on Multi-
media Communications in 2009, as well as IEEE IWQoS ’08 and IEEE/ACM
IWQoS ’12 Best Student Paper Awards. His research interests are in network-
ing and multimedia, in particular, multimedia communications, peer-to-peer
networking, cloud computing, social networking, and wireless sensor/mesh
networking. He has served on the Editorial Boards of IEEE Transactions on
Multimedia, IEEE Communications Tutorial and Surveys, IEEE Internet of Things
Journal, Elsevier Computer Communications, and Wiley Wireless Communica-
tions and Mobile Computing. He is a TPC Co-Chair for IEEE/ACM IWQoS
’14 in Hong Kong.
HAI JIN [SM] (hjin@hust.edu.cn) is a Cheung Kung Scholars Chair Professor of
computer science at the Huazhong University of Science and Technology,
China. He is now dean of the School of Computer Science and Technology at
the university. He received his Ph.D. degree in computer engineering from the
Huazhong University of Science and Technology in 1994. In 1996, he was
awarded a German Academic Exchange Service fellowship to visit the Techni-
cal University of Chemnitz in Germany. He was awarded the Excellent Youth
Award from the National Science Foundation of China in 2001. He is the
chief scientist of ChinaGrid, the largest grid computing project in China, and
chief scientist of the National 973 Basic Research Program Project of Virtual-
ization Technology of Computing Systems. His research interests include com-
puter architecture, virtualization technology, cloud computing and grid
computing, peer-to-peer computing, network storage, and network security.

wireless sensor network

More Related Content

What's hot (20)

Similar to wireless sensor network (20)

More from parry prabhu (20)

wireless sensor network