Big Data with Cloud Computing: an insight on the

Big Data with Cloud Computing: an insight on the
Overview
Big Data with Cloud Computing:
an insight on the computing
environment, MapReduce,
and programming frameworks
Alberto Fernández,1∗ Sara del Río,2 Victoria López,2
Abdullah Bawakid,3 María J. del Jesus,1 José M. Benítez2 and
Francisco Herrera2,3
The term ‘Big Data’ has spread rapidly in the framework of Data Mining and
Business Intelligence. This new scenario can be defined by means of those
problems that cannot be effectively or efficiently addressed using the standard
computing resources that we currently have. We must emphasize that Big Data
does not just imply large volumes of data but also the necessity for scalability,
i.e., to ensure a response in an acceptable elapsed time. When the scalability
term is considered, usually traditional parallel-type solutions are contemplated,
such as the Message Passing Interface or high performance and distributed
Database Management Systems. Nowadays there is a new paradigm that has
gained popularity over the latter due to the number of benefits it offers. This
model is Cloud Computing, and among its main features we has to stress its
elasticity in the use of computing resources and space, less management effort,
and flexible costs. In this article, we provide an overview on the topic of Big
Data, and how the current problem can be addressed from the perspective of
Cloud Computing and its programming frameworks. In particular, we focus
on those systems for large-scale analytics based on the MapReduce scheme
and Hadoop, its open-source implementation. We identify several libraries and
software projects that have been developed for aiding practitioners to address this
new programming model. We also analyze the advantages and disadvantages of
MapReduce, in contrast to the classical solutions in this field. Finally, we present
a number of programming frameworks that have been proposed as an alternative
to MapReduce, developed under the premise of solving the shortcomings of this
model in certain scenarios and platforms.© 2014 John Wiley & Sons, Ltd.
How to cite this article:
WIREs Data Mining Knowl Discov 2014, 4:380–409. doi: 10.1002/widm.1134
INTRODUCTION
W
e are immersed in the Information Age where
vast amounts of data are available. Petabytes
∗ Correspondence
1 Department
2 Department
to: [email protected]
of Computer Science, University of Jaen, Jaen, Spain
of Computer Science and Artificial Intelligence, University of Granada, Granada, Spain
3 Faculty of Computing and Information Technology—North Jeddah, King Abdulaziz University, Jeddah, Saudi Arabia
Conflict of interest: The authors have declared no conflicts of
interest for this article.
380
of data are recorded everyday resulting in a large
volume of information; this incoming information
arrives at a high rate and its processing involves
real-time requirements implying a high velocity; we
may find a wide variety of structured, semi-structured,
and unstructured data; and data have to be cleaned
before the integration into the system in order to
maintain veracity.1 This 4V property is one of the most
widespread definitions of what is known as the Big
Data problem,2,3 which has become a hot topic of
interest within academia and corporations.
© 2014 John Wiley & Sons, Ltd.
Volume 4, September/October 2014
WIREs Data Mining and Knowledge Discovery
Big Data with Cloud Computing
The current explosion of data that is being generated is due to three main reasons4 : (1) hundreds of
applications such as mobile sensors, social media services, and other related devices are collecting information continuously; (2) storage capacity has improved
so much that collecting data is cheaper than ever, making preferable to buy more storage space rather than
deciding what to delete; (3) Machine Learning and
information retrieval approaches have reached a significant improvement in the last years, thus enabling
the acquisition of a higher degree of knowledge from
data.5,6
Corporations are aware of these developments.
Gaining critical business insights by querying and
analyzing such massive amounts of data is becoming a
necessity. This issue is known as Business Intelligence
(BI),7,8 which refers to decision support systems that
combine data gathering, data storage, and knowledge
management with analysis to provide input to the
decision process.9 Regarding the former issues, a new
concept appears as a more general field, integrating
data warehousing, Data Mining (DM), and data
visualization for Business Analytics. This topic is
known as Data Science.10,11
The data management and analytics carried out
in conventional database systems (and other related
solutions) cannot address the Big Data challenges:
data size is too large, values are modified rapidly,
and/or they do no longer satisfy the constraints of
Database Management Systems (DBMS). According
to this fact, new systems have emerged to solve the
previous issues: (1) ‘Not Only SQL’ (NoSQL) systems
that modify the storage and retrieval of key/value
pairs for interactive data serving environments12 and
(2) systems for large-scale analytics based on the
MapReduce parallel programming model,13 Hadoop
being the most relevant implementation.14
These two approaches are under the umbrella
of Cloud Computing.15–17 Cloud Computing has
been designed to reduce computational costs and
increase the elasticity and reliability of the systems.18
It is also intended to allow the user to obtain various services without taking into consideration the
underlying architecture, hence offering a transparent scalability. The basis of Cloud Computing is the
Service-Oriented Architecture,19 which is designed to
allow developers to overcome many distributed organization computing challenges including application
integration, transaction management, and security
policies.
The advantages of this new computational
paradigm with respect to alternative technologies are
clear, especially regarding BI.20 First, cloud application providers strive to give the same or better
Volume 4, September/October 2014
service and performance as if the software programs
were locally installed on end-user computers, so the
users do not need to spend money buying complete
hardware equipment for the software to be used.
Second, this type of environment for the data storage
and the computing schemes allows companies to get
their applications up and running faster. They have
a lower need of maintenance from the Information
Technology department as Cloud Computing automatically manages the business demand by dynamically assigning resources (servers, storage, and/or
networking) depending on the computational load
in real time.21
Being a hot topic, there are currently a high
number of works that cover the Big Data problem. But
there is also a lack of an unifying view in order to fully
understand the nature of this issue, and the ways to
manage it. New approaches emerge every day and it is
hard to follow this trending topic. In particular, there
is not a clear ‘guide’ for a new user whose aim is to get
introduced to the solutions to this problem. According
with the above, the main contributions of this article
are summarized as follows:
(1) An introduction to the ‘Big Data’ is given. The
significance of addressing properly Big Data in
DM and BI applications is stressed.
(2) We will show that Cloud Computing platforms
(e.g., Amazon’s EC2, Microsoft Azure, and so
on) enable researchers to perform Big Data
analysis in a very flexible way and without
much fixed costs. But we also point out that
Big Data technologies can also be deployed in
noncloud clusters of computers.
(3) New approaches that have been developed for
achieving scalability in Big Data are described
in detail. We focus on the MapReduce programming model, and the NoSQL approach for
handling data and queries.
(4) A critical evaluation regarding advantages and
drawbacks of these new approaches, with
respect to classical solutions, is given. Additionally, a compilation of the milestones on the
topic are summarized and analyzed. Furthermore, new tools and platforms that have been
designed as alternative to MapReduce will be
enumerated.
(5) Finally, the discussion developed in this review
is aimed for helping researchers to better understand the nature of Big Data. The recommendations given may allow the use of available
resources, or the development of new ones for
addressing this type of problems.
© 2014 John Wiley & Sons, Ltd.
381
wires.wiley.com/widm
Overview
In order to reach the previous objectives, this
article is structured as follows. First we provide a
definition of Big Data, linking these concepts to
DM and BI, and showing how this problem can
be addressed. Next section introduces the main concepts of Cloud Computing, presenting an architecture approach to develop Big Data solutions on such
platforms. Afterwards, we focus on the MapReduce
programming framework as the most prominent solution for Big Data, describing its features and comparing it with some classical parallel approaches, as
well as enumerating several limitations of this model.
We then enumerate several alternatives to MapReduce
that have been designed in order to overcome its performance under certain work scenarios. The lessons
learned from this review are given in the next section.
Finally, the main concluding remarks are presented.
ON THE SIGNIFICANCE OF BIG DATA
IN BUSINESS INTELLIGENCE
In this section, we will first introduce what it is
understood as Big Data. Then, we will establish the
relationship between DM and BI for the sake of better
understanding the significance of both facets with
respect to scalability. Finally, we will present several
guidelines that are necessary to address, in a proper
way, the Big Data problem.
What is Big Data?
Recently, the term of Big Data has been coined
referring to those challenges and advantages derived
from collecting and processing vast amounts of data.22
This topic has appeared as organizations must deal
with petabyte-scale collections of data. In fact, in
the last 2 years we have produced 90% of the total
data generated in history.23 The sources of such huge
quantity of information are those applications that
gather data from click streams, transaction histories,
sensors, and elsewhere. However, the first problem for
the correct definition of ‘Big Data’ is the name itself,4
as we might think that it is just related to the data
Volume.
The heterogeneous structure, diverse dimensionality, and Variety of the data representation, also have
significance in this issue. Just think about the former
applications that carry out the data recording: different software implementations will lead to different
schemes and protocols.24
Of course it also depends on the computational
time, i.e., the efficiency and Velocity in both receiving
and processing the data. Current users demand a
‘tolerable elapsed time’ for receiving an answer. We
382
must put this term in relationship with the available
computational resources, as we cannot compare the
power of a personal computer with respect to a
computational server of a big corporation.3
Finally, one main concern with applications that
deals with this kind of data is to maintain the Veracity
of the information. This basically refers to the data
integrity, i.e., avoiding noise and abnormalities within
the data, and also the trust on the information used to
make decisions.25,26
All these facts are known as the 4V’s of Big
Data,1 which lead to the definition given by Steve Todd
at Berkeley Universitya :
Big data is when the normal application of current
technology does not enable users to obtain timely,
cost-effective, and quality answers to data-driven
questions.
We must point out that additional definitions
including up to 9V’s can be also found, adding terms
such as Value, Viability, and Visualization, among
others.27
The main challenge when addressing Big Data is
associated with two main features28 :
• The storage and management of large volumes
of information. This issue is related to DBMS,
and the traditional entity-relation model. Commercial systems report to scale well, being able
to handle multi-petabyte databases, but in addition to their ‘cost’ in terms of price and hardware
resources, they have the constraint of importing data into a native representation. On the
other hand, widely adopted open-source systems,
such as MySQL, are much more limited in terms
of scalability than their commercial analytics
counterparts.
• The process for carrying out the exploration
of these large volumes of data, which intends
to discover useful information and knowledge
for future actions.23 The standard analytical
processing is guided by an entity-relation scheme,
from which queries were formulated using the
SQL language. The first hitch of these type
of systems is the necessity of preloading the
data, as stated previously. Additionally, there
is not much support for in-database statistics
and modeling, and many DM programmers may
not be comfortable with the SQL declarative
style. Even in the case that engines provide these
functionalities, as iterative algorithms are not
easily expressible as parallel operations in SQL,
they do not work well for massive quantities of
data.
© 2014 John Wiley & Sons, Ltd.
Volume 4, September/October 2014
WIREs Data Mining and Knowledge Discovery
Big Data with Cloud Computing
In summary, there are several conditions that
must be taken into account in order to consider a
problem within the Big Data framework. First of all,
and referring to the 4Vs’ properties, a threshold for
the quantity of information that is being processed,
and the time constraints for giving an answer, must
be established. These two concepts are also closely
related. For example, if we address the fingerprint
recognition application,29,30 there is a limit for the
number of fingerprints we can manage in the database
for providing an accurate answer within a short period
of time, i.e., tenths of a second or few seconds.
But, how do we set this limit? The answer is
unclear as what was ‘big’ years ago, can now be considered as ‘small’. Therefore, for a clear definition of
Big Data we must also include which technology is
necessary to solve the problem. Suppose a major sales
enterprise, which aims to adjust the unit pricing for
a collection of items based on demand and inventory.
Clearly, this firm will need a computational technology beyond a standard cluster of machines with a
relational database and a common business analytics
product. Now, if we consider a project of similar ambitions within the domain of a retailer company, the
application could easily be completed using existing
databases and ETL tools. The latter cannot be categorized as Big Data project, according to our definition.
Finally, Big Data is about the insight that we
want to extract from information. There are many
well-known applications that are based on Cloud
Computing such as email servers (Gmail), social media
(Twitter), or storage sharing and backup (Dropbox).
All this software manage high volumes of data, where
fast responses are essential, and with information
coming at a high rate in a semi-structured or unstructured way. They must also face the veracity in the
information; however, they are not intrinsically considered Big Data.
The key here is the analysis that is made for
knowledge and business purposes, what is known
as Data Science.10,11 This speciality include several
fields such as statistics, Machine Learning, DM, artificial intelligence, and visualization, among others.
Hence, Big Data and Data Science are two terms with
a high synergy between them.31 Some well-known
examples include e-Sciences32 and other related scientific disciplines (particle physics, bioinformatics, and
medicine or genomics) Social Computing33 (social network analysis, online communities, or recommender
systems), and large-scale e-commerce,34,35 all of which
are particularly data-intensive.
Regarding the former, many Big Data challenges
proliferate nowadays for encouraging researchers to
put their efforts in solving these kind of tasks.
Volume 4, September/October 2014
As examples, we may refer to the ‘Data Mining
Competition 2014’,b which belongs to the Evolutionary Computation for Big Data and Big Learning Workshop (a part of the well-known GECCO
conference), and the three ‘DREAM9’c challenges
opened by Sage Bionetworks and DREAM, linked
to the International Biomedical Commons Congress
and the RECOMB/ISCB Systems and Regulatory
Genomics/DREAM Conference.
Data Science: Data Mining as a Support
for Business Intelligence Applications to Big
Data
With the establishment of the Internet, business in
the 21st century is carried out within a digital environment, which is known as e-economy or digital
economy. Information is a most valuable asset of corporations, but we must take into account that there is
a handicap in their operations: each functional area of
a company manages their own data. Because of this,
it is necessary to integrate all information systems in
the corporation, not only for processing this information in an efficient manner, but also to create some
‘business intelligence’ that can be used for all activities.
The BI concept is not a technology by itself, but
rather a collection of information systems that works
in a coordinate way. It includes several blocks such as
data warehouse systems, Data Mining systems, Online
Analytical Processing systems, knowledge-based systems, query and report tools, and dashboards, as
depicted in Figure 1.
Organizations are looking for new and effective
ways of making better decisions that allow them to
gain competitive advantage. Therefore, they are in a
need of some fundamental principles that guide the
extraction of information. We must point out that
decisions made by corporations using BI rely on the
current activities from which they must capture the
information. Among others, we must stress social
media, user behavior logs on websites, transactional
data, and so on. This implies huge volumes of data to
be exploited and analyzed, and the trend is that this
vast amount of information will continue growing.
Here is where the concept of Data Science comes
up, which encompasses the application of DM techniques in Big Data problems,36,37 as explained above.
DM38,39,6 consists of identifying valid and novelty patterns, which are potentially useful. It is built up by
means of a set of interactive and iterative steps, where
we can include preprocessing of data, the search of
patterns of interest with a particular representation,
and the interpretation of these patterns.
Principles and techniques of Data Science and
DM are broadly applied across functional areas in
© 2014 John Wiley & Sons, Ltd.
383
wires.wiley.com/widm
Overview
Reports &
queries
Dashboards
Analysis of the
information details
ETL processes
External data
Data
mining
Data warehouse
(data mart)
Statistics
Trends and behavioral analysis
Projections
Integrated business model
Information repository
Metadata
Multidimensional analysis
Analysis of FCE
Analysis of summarized data
FIGURE 1 | Business Intelligence structure.
business. Many firms have differentiated themselves
strategically thanks to the former, sometimes to the
point of evolving into DM companies. We present
several examples on how companies would benefit
from this:
• The operational risk for setting prices for a
retailer grows according to the time that is spent
to recalculate markdown prices and tailor prices
to individual stores and shopper profiles.
• For large investment banking firms, the speed
and agility in the operations can give significant
competitive advantages. As they need to recalculate entire risk portfolios, a good efficiency
in the data treatment can enable the obtaining
of fine-tune responses to changing interest rates,
exchange rates, and counterpart risks.
• Insuring companies, and specifically actuaries,
rely heavily on using historical data to predict
future behavior or creating premium rates to
price products. The growing volumes of available
data limit the company at using only a subset of
this information to generate pricing models.
• Finally, we must mention the case of unplanned
growth in the corporate activities. The investment in infrastructure when the company
expands its activities can be severe. The contrary
case is also possible. In both scenarios, we must
stress the benefits that come from a flexible and
elastic solution.
The knowledge extraction process from Big
Data has become a very difficult task for most of
384
the classical and advanced existing techniques.40 The
main challenges are to deal with the increasing amount
of data considering the number of instances and/or
features, and the complexity of the problem.41,42 From
the previous examples, we can observe that several
key concepts such as scalability, elasticity, and flexible
pricing need to be considered in this scenario. Thus,
it is straightforward to realize about the necessity of
constructing DM solutions within a framework of the
former characteristics in order to integrate them into
BI applications.
How Can Big Data Problems be Addressed?
From the first part of this section we must recall that
there are two main design principles for addressing the
scalability in Big Data problems. First, we may refer to
very large databases for the storage and management
of the data. Furthermore, the processing and analysis
must be carried out by parallel programming models.
Regarding the first issue, standard solutions in
this field included distributed databases for intensive updating workloads,43 and parallel database
systems44 for analytical workloads. While the former
has never been very successful, the latter approach
has been extensively studied for decades. Therefore,
we may find several commercial implementations
featuring well-defined schemas, declarative query
languages, and a runtime environment supporting
efficient execution strategies. When large volumes
of data are involved, the hitch with these systems is
twofold: (1) as they must run under high hardware
requirements, it becomes prohibitively expensive
© 2014 John Wiley & Sons, Ltd.
Volume 4, September/October 2014
WIREs Data Mining and Knowledge Discovery
Big Data with Cloud Computing
when we scale to a cluster of computing elements; (2)
the underlying assumptions of relational databases,
i.e., fault tolerance, cannot be longer satisfied. Finally,
we are aware that current applications are now
managing semi-structured or even unstructured
data, which imposes another challenge for database
solutions.
According to these facts, an alternative to
relational databases has arisen. This new data management technology is known as ‘NoSQL’,12,45
which basically consists of storing the information as
‘Key-Value’ pairs, achieving a distributed horizontal
scaling. An important difference between traditional
databases and NoSQL databases is that they do not
support updates and deletes. The major reason for
the popularity of the NoSQL scheme is their flexible
data model, and the support of various types of data
models, most of these not being strict. With these
arguments, there is a clear trend in migrating to this
recent technology in Big Data applications.
Focusing on the programming models for data
analysis, common solutions are those based on parallel computing,46 such as the Message Passing Interface (MPI) model.47 Challenges at this point rely on
the data access and the ease of developing software
with the requirements and constraints of the available programming schemes. For example, typical DM
algorithms require all data to be loaded into the main
memory. This imposes a technical barrier in Big Data
problems as data are often stored in different locations
and this supposes an intensive network communication and other input/output costs. Even in the case we
could afford this, there is still the need for an extremely
large main memory to hold all the preloaded data for
the computation. Finally, there is a clear need in a
robust fault-tolerant mechanism, as it is crucial for
time-consuming jobs.
To deal with the previous issues, a new generation of systems has been established, where
MapReduce13 and its open-source implementation
Hadoop14,48 are the most representative ones in both
industry and academia. This new paradigm removes
the previous constraints for preloading data, fixed
storage schemes, or even the use of SQL. Instead,
developers write their programs under this new model
that allows the system to automatically parallelize the
execution. This can be achieved by the simple definition of two functions, named as Map and Reduce.
In short, ‘Map’ is used for per-record computation,
whereas ‘Reduce’ aggregates the output from the Map
functions and applies a given function for obtaining
the final results.
The success of these type of systems is also
related to additional constraints that have been
Volume 4, September/October 2014
considered recently. Among them, we must stress
low cost, security (considering technical problems),
simplicity (in programming, installing, maintaining),
and so on. According to the previous features, a new
computational paradigm has imposed as the answer
to all these issues. This system is Cloud Computing,
and it has been settled as the baseline environment for
the development of the aforementioned solutions for
Big Data.
The concept of Cloud Computing allows several
advantages to the perspective of deploying a huge
cluster of machines configured such that the load can
be distributed among them. The most relevant one
is to rent the computational resources when they are
strictly necessary. Hence, the cost of processing the
data will be only spent when the data are ready to be
processed, i.e., paying according to service quantity,
type, and duration of the application service. We must
be aware that, although the global amount of data
in these cases actually exceeds the limits of current
physical computers, the frequency of both the data
acquisition and the data processing can be variable,
thus stressing the goodness of Cloud Computing.
Additionally, the inner structure of this computational
paradigm in terms of data storage, data access, and
data processing make it the most reasonable solution
for addressing this problem.
CLOUD COMPUTING
ENVIRONMENTS FOR BIG DATA
Cloud Computing is an environment based on using
and providing services.49 There are different categories
in which the service-oriented systems can be clustered.
One of the most used criteria to group these systems
is the abstraction level that is offered to the system
user. In this way, three different levels are often
distinguished50 : Infrastructure as a Service (IaaS),
Platform as a Service (PaaS), and Software as a Service
(SaaS) as we can observe in Figure 2.
Cloud Computing offers scalability with respect
to the use of resources, low administration effort,
flexibility in the pricing model and mobility for the
software user. Under these assumptions, it is obvious
that the Cloud Computing paradigm benefits large
projects, such as the ones related with Big Data and
BI.51
In particular, a common Big Data analytics
framework52 is depicted in Figure 3. Focusing on
the structure of the data management sector we may
define, as the most suitable management organization
architecture, one based on a four-layer architecture,
which includes the following components:
© 2014 John Wiley & Sons, Ltd.
385
wires.wiley.com/widm
Overview
d
En
Throughout the following subsections, we will
address the aforementioned levels in detail.
s
er
om
st
cu
File System and Storage
v
De
el
er
s
m
Ad
s
Sy
s
in
IAAS
FIGURE 2 | Illustration of the layers for the Service-Oriented
Architecture
• A file system for the storage of Big Data, i.e.,
a wide amount of archives of large size. This
layer is implemented within the IaaS level as it
defines the basic architecture organization for the
remaining tiers.
• A DBMS for organizing the data and access
them in an efficient way. It can be viewed in
between the IaaS and PaaS as it shares common
characteristics from both schemes. Developers
used it to access the data, but its implementation
lies on a hardware level. Indeed, a PaaS acts
as an interface where, at the upper side offers
its functionality, and at the bottom side, it has
the implementation for a particular IaaS. This
feature allows applications to be deployed on
different IaaS without rewriting them.
• An execution tool to distribute the computational load among the computers of the cloud.
This layer is clearly related with PaaS, as it is kind
of a ‘software API’ for the codification of the Big
Data and BI applications.
Data bases
Sensors
Mobiles
Web...
FIGURE 3 | Big Data framework.
386
Data sources
Extract (semi-structure/unstructure)
• A query system for the knowledge and information extraction required by the system’s users,
which is in between the PaaS and SaaS layers.
Distributed file
system (HDFS/GFS)
Data mining
Parallelize computing
(MapReduce...)
Machine learning
Data storage
(NoSQL, ...)
Statistics
Data cleaning
Network analysis
Data security
Time series
analysis
....
....
Data management
© 2014 John Wiley & Sons, Ltd.
Data analytics
Access (web services, mobiles...)
The first level is the basis of the architecture of the
Cloud system. It includes the network architecture
with many loosely coupled computer nodes for providing a good scalability and a fault-tolerance scheme.
As suggested, the system must consider a dynamic/
elastic scheme where the performance, cost, and
energy consumption of the node machines is managed
in runtime.
Following this scheme, we need to define a
whole system to manage parallel data storage of the
large files from which the software will operate on.
This is not trivial, and new file system approaches
must be considered. Among all possible solutions,
the one that has achieved a higher popularity is the
Hadoop-distributed file system (HDFS).53
HDFS is the open-source project of the Apache
Foundation that implements the Google file system, the initial solution conceived to deal with this
problem.54 An HDFS installation is composed of multiples nodes, which are divided into two classes: a master node (namenode) and a large number of fragments
storages or datanodes. Files are divided into fixed size
chunks of 64 megabytes, in a similar way than the clusters or sectors of traditional hard disk drives. Datanodes store these fragments, which are assigned with a
unique ID label of 64 bits in the namenode when it
is generated. When an application aims to read a file,
it contacts with the namenode in order to determine
where the actual data is stored. Then, the namenode
returns the corresponding block ID and the datanode, such that the client can directly contact the data
node for retrieving the data. An important feature of
the design is that data are never moved through the
op
PAAS
Visualization
SAAS
Access / application
Volume 4, September/October 2014
WIREs Data Mining and Knowledge Discovery
Big Data with Cloud Computing
Application
(block id,
byte range)
HDFS client
(file name,
block id)
(block id,
block location)
Block data
HDFS datanode
Datanode state
HDFS namenode
File namespace
Linux file system
/foo/bar
...
Block 4A3f
HDFS datanode
Instructions
to datanode
Linux file system
...
FIGURE 4 | The architecture of Hadoop-distributed file system (HDFS). The namenode (master) is responsible for maintaining the file namespace
and directing clients to datanodes (slaves) that actually hold data blocks containing user data.
namenode. Instead, all data transfer occurs directly
between clients and datanodes; communications with
the namenode only involves transfer of metadata. This
architecture is depicted in Figure 4.
In order to ensure reliability, availability, and
performance according to an expected high demand
of service, redundancy is maintained by default replicating each chunk in at least three servers of the Cloud,
this being a parameter of the system. Usually two of
them are set in datanodes that are located at the same
rack and another on a different one. HDFS is resilient
toward two common failure scenarios: individual
datanode crashes and failures in networking equipment that bring an entire rack offline. The namenode
communicates with datanodes to re-equilibrate data,
move copies, and preserve the high replication of the
data: if there are not enough replicas (e.g., due to disk
or machine failures or to connectivity losses due to
networking equipment failures), the namenode directs
the creation of additional copies; if there are too many
replicas (e.g., a repaired node rejoins the cluster), extra
copies are discarded.
In summary, the HDFS namenode has the following responsibilities14 :
• Namespace management: For a quick access,
the namenode holds in memory all information
Volume 4, September/October 2014
regarding the directory structure, the location of
the blocks and so on.
• Coordinating file operations: As pointed out
previously, communication is made directly from
clients to datanodes by the coordination of the
namenode. Files are deleted by a lazy garbage
collector.
• Maintaining overall health of the file system: The
integrity of the system is kept by the creation
of new replicas of the data blocks. When some
datanodes have more blocks than others, a rebalancing process is applied.
The main difference with respect to other file
systems is that HDFS is not implemented in the kernel
space of the operating system, but it rather works
as a process in the user space. As blocks are stored
on standard file systems, HDFS lies on top of the
operating system stack. Additionally, it is not frequent
that the stored data are overwritten or removed; in
general the files are read only or just new information
is added to them.
Finally, there are other platforms that follows
different implementations such as Amazon Simple
Storage Service (S3),55 Cosmos,56 and Sector.57,58
They aim at managing the information in a local way
in order to avoid transactions through the net that
© 2014 John Wiley & Sons, Ltd.
387
wires.wiley.com/widm
Overview
can detriment the performance when dealing with
executions on Big Data.
Database Management: Not Only SQL
The second layer is devoted to the DBMS. Relational
databases may have difficulties when processing Big
Data along a big number of servers, and keeping the
efficiency required by BI applications with traditional
database systems is complicated. Specifically, the storage format and the access to the data must be completely modified, implying the necessity of using different DBMS.
In this framework, NoSQL databases12 emerge
as a new kind of system. The main difference between
this new system and the relational model is that they
allow a horizontal scalability of the data, by means of
variations on the storage and a retrieval of key/value
pairs, as opposed to the relational model based on
primary-key/foreign-key relationships.
Below we enumerate the basic features of this
model:
• As its name indicates, they use a query language
similar to SQL, but it is not fully conformant.
As it is designed over a partitioned file system,
JOIN-style operations cannot be used.
• The ‘A.C.I.D.’ guarantees (atomicity, consistency,
isolation, and durability)59 cannot be satisfied
anymore. Usually, only an eventual consistency
is given, or the transactions are limited to unique
data items. This means that given a sufficiently
long period of time in which no changes are
submitted, all the updates will be propagated
over the system.
• It has a distributed and fault-tolerant architecture. Data reside in several redundant servers,
in a way that the system can be easily scalable
adding more servers, and the failure of one single server can be managed without difficulties.
This is the issue we stated at the beginning of this
section regarding the horizontal scalability, the
performance and the real time nature of the system being more important than the consistency.
In this way, NoSQL database systems are usually
highly optimized for the retrieval and appending operations, and they offer little functionality beyond the
register storing (key-value type), implying the necessity of ad hoc efficient join algorithms.60,61 However,
the reduced flexibility compared with relational systems is compensated by means of a significant gain in
scalability and performance for certain data models.
388
In brief, NoSQL DBMS are useful when they
work with a large quantity of data, and the nature
of these data do not require from a relational model
for their structure. No scheme is needed, the data
can be inserted in the database without defining at
first a rigid format of the ‘tables’. Furthermore, the
data format may change at any time, without stopping
the application, providing a great flexibility. Finally,
to reduce the latency and substantially enhance the
data throughput, a transparent integrated caching is
implemented in the memory system.
One of the first conceptual models of this type
of DBMS is possibly BigTable.45 This database engine
was created by Google in 2004 with the aim of
being distributed, high efficiency, and proprietary.
It is built over the Google file system and works
over ‘commodity hardware’. In order to manage the
information, tables are organized by different groups
of columns with variable dimensions, one of them
always being the timestamp to maintain the control
version and the ‘garbage collector’. They are stored
as ‘subtables’, i.e., fragments of a unique table, from
100 to 200 megabytes each, that can also be stored
in a compressed format. This disposition enables the
use of a load balancing system, i.e., if a subtable is
receiving too many petitions, the machine where it is
stored can get rid of the other subtables by migrating
them to another machine.
The Dynamo system62 is also a milestone for
NoSQL DBMSs. It was originally built to support
internal Amazon’s applications, and to manage the
state of services that have very high reliability requirements. Specifically, a tight control over the tradeoffs
between availability, consistency, cost-effectiveness,
and performance is needed. The state is stored as
binary objects (blobs) identified by unique keys, and
no operations span multiple data items. Dynamo’s
partitioning scheme relies on a variation of a consistent hashing mechanism63 to distribute the load across
multiple storage hosts. Finally, in the Dynamo system,
each data item is replicated at N hosts where N is a
parameter configured ‘per-instance’.
From these systems, several open-source implementations have been developed make these models
accessible to worldwide Big Data users. In the following, we enumerate a nonexhaustive list of examples:
• HBase64 is built over HDFS following the scheme
of the BigTable approach, thus featuring compression, in-memory operations, and Bloom filters that can be applied to columns. Tables in
HBase can serve as input and output for MapReduce jobs run in Hadoop and may be accessed
through the Java API.
© 2014 John Wiley & Sons, Ltd.
Volume 4, September/October 2014
WIREs Data Mining and Knowledge Discovery
Big Data with Cloud Computing
TABLE 1 Design Decisions of NoSQL DBMS. CP stands for Consistency and Partition tolerance, and AP stands for Availability and Partition
tolerance, regarding the CAP theorem.
System
Data Model
Consistency
CAP Options
License
BigTable
Column families
Eventually
CP
Google
Consistent
Dynamo
Key-value storage
Proprietary Lic.
Eventually
AP
Consistent
HBase
Column families
Eventually
CP
Consistent
Cassandra
Column families
MongoDB
CouchDB
Eventually
AP
Multidimensional
Eventually
Consistent
Document-oriented
Optimistically
Storage
Consistent
AP
Open source
—GNU
AP
Open source
—GNU
Document-oriented
Optimistically
Storage
Consistent
Volume 4, September/October 2014
Open source
—Apache
Table
• Cassandra65 brings together the BigTable features and the distributed systems technologies
from Dynamo. It has a hierarchical architecture
in which the database is based on columns (name,
value, and timestamp). Columns are grouped in
rows as ‘Column families’ (a similar relationship
to the one between rows and columns in relational databases) with keys that map the columns
in each row. A ‘keyspace’ is the first dimension of
the Cassandra hash and is the container for column families. Keyspaces are of roughly the same
granularity as a schema or database (i.e., a logical
collection of tables) in a relational DBMS. They
can be seen as a namespace for ColumnFamilies
and are typically allocated as one per application.
SuperColumns represent columns that have subcolumns themselves (e.g., Maps).
• In HyperTable,66 data is represented in the system as a multidimensional table of information.
HyperTable systems provide a low-level API and
the HyperTable Query Language, which allows
the user to create, modify, and query the underlying tables. The data in a table can be transformed
and organized at high speed, as computations are
performed in parallel and distributing them to
where the data is physically stored.
• MongoDB67 is another NoSQL-type scalable
DBMS. It is written in C++ with MapReduce
support for a flexible data aggregation and
processing. It also includes a document-oriented
storage, which offers simplicity and potential for JavaScript Object Notation (JSON)68
Open source
—Apache
Consistent
Hypertable
Amazon
Proprietary Lic.
AP
Open source
—Apache
(an alternative to XML) type documents with
dynamic schemes.
• CouchDB69 is a document-oriented DBMS that
enables the use of queries and indexes using
JavaScript following a MapReduce style, i.e.,
using JSON to store the data. A CouchDB document is an object that consists of named fields,
such as strings, numbers, dates, or even ordered
lists and associative maps. Hence, a CouchDB
database is a flat collection of documents where
each document is identified by a unique ID. Additionally, it offers incremental replicas with detection and resolution of bidirectional conflicts. The
CouchDB document update model is lockless and
optimistic. It is written in the Erlang language,
thus providing a robust functional programming
base for concurrent and distributed systems, with
a flexible design for its scalability and ease to
extend.
Finally, for the sake of comparison we show
in Table 1 a summary of the main properties of the
former systems.
Execution Environment
The third level is the execution environment. Owing
to the large number of nodes, Cloud Computing is
especially applicable for distributed computing tasks
working on elementary operations. The best known
example of Cloud Computing execution environment
is probably Google MapReduce13 (the Google’s implementation of the MapReduce programming model)
© 2014 John Wiley & Sons, Ltd.
389
wires.wiley.com/widm
Overview
Map
Shuffle
Reduce
File system
k1
k2
k3
v
v
v
k1
k2
k3
k1
k2
k3
k1
k2
k3
v
v
v
v
v
v
v
v
v
k1
k2
k3
v
v
v
k1
k2
k3
v
v
v
k1
v
v
v
v
v
v
k2
v
v
v
v
v
v
k3
v
v
v
v
v
v
Output
FIGURE 5 | MapReduce simplified flowchart.
and Hadoop, its open-source version.14 This environment aims at providing elasticity by allowing the
adjustment of resources according to the application,
handling errors transparently, and ensuring the scalability of the system.
This system has been designed under the following assumptions: first, all the nodes in the cloud
should be colocated (within one data center), or a high
bandwidth is available between the geographically distributed clusters containing the data. Secondly, individual inputs and outputs to the cloud are relatively
small, although the aggregate data managed and processed are very large.
As its name suggests, this programming model
is built upon two ‘simple’ abstract functions named
Map and Reduce, which are inherited from the classical functional programming paradigms. Users specify
the computation in terms of a map (that specify the
per-record computation) and a reduce (that specify
result aggregation) functions, which meet a few simple
requirements. For example, in order to support these,
MapReduce requires that the operations performed at
the reduce task to be both ‘associative’ and ‘commutative’. This two-stage processing structure is illustrated
in Figure 5.
We extend the description of this programming
scheme in a separate section, where we will establish
the properties of this approach for its use in Big Data
problems.
Query Systems for Analysis
The last tier is related to the query systems, being the
interface to the user and providing the transparency
to the other tiers of the architecture. In environments
where large databases exist, the necessity of carrying
390
out ‘complex’ queries with answers in short time
implies the use of a parallel model such as MapReduce,
as explained in the previous subsection.
In most cases, for obtaining the required information from the data several sophisticated operations
such as ‘joins’ or ‘data filtering’ need to be performed. Regarding the functional programming model
of MapReduce, this task could become quite difficult
and time-consuming to implement and it will require
highly skilled developers.
According to the previous facts, several
approaches have been developed to ease the user
to obtain knowledge from the data in NoSQL-type
databases. The goal of these systems is being able to
provide a trade-off between the declarative style of
SQL and the procedural style of MapReduce. This
will diminish the efforts of implementing data analysis applications on top of a distributed system, i.e.,
adapting them toward Online Analytical Processing
and Query/Reporting Data-Warehousing tools.
In the remainder of this section, we will present
some well-known systems that allow analysts with
strong SQL skills (but meager Java programming
skills) the use of this type of query languages:
• Hive is a platform developed by Facebook that
uses the HiveQL language, which is close to a
SQL-like scheme.70 It provides a subset of SQL,
with features subqueries in the From clause, various types of joins, group-bys, aggregations, ‘create table as select’, and so on. Given a query, the
compiler translates the HiveQL statement into a
directed acyclic graph of MapReduce jobs, which
are submitted to Hadoop for execution. Hive
compiler may choose to fetch the data only from
certain partitions, and hence, partitioning helps
© 2014 John Wiley & Sons, Ltd.
Volume 4, September/October 2014
WIREs Data Mining and Knowledge Discovery
Big Data with Cloud Computing
in efficiently answering a query. It supports all the
major primitive types, as well as collection types
such as map, list, and struct. Hive also includes
a system catalogue, i.e., a meta-store, which
contains schemas and statistics quite useful in
data exploration, query optimization, and query
compilation.
• Pig is a high-level scripting language developed
by Yahoo to process data on Hadoop.71 It is
aimed at combining the high-level declarative
queries (SQL) with the lower level programming
procedure (MapReduce). Pig (via a language
called Pig Latin) provides concise primitives for
expressing common operations that performs
a simple data transformation, i.e., projection,
selection, group, join, and so on.72 This conciseness comes at low cost: Pig scripts approach
the performance of programs directly written in
Hadoop Java.
Programs written in Pig only need to specify
a query execution plan or a dataflow graph. The
plan is compiled by a MapReduce compiler, which is
then optimized once more by a MapReduce optimizer
performing tasks such as early partial aggregation, and
then submitted for execution.
Pig has a flexible data model that allows complex
types such as set or map. Unlike Hive, stored schemas
are optional. Pig also has the capability of incorporating user define functions. Finally, it provides a debugging environment that can generate sample data to
help a user in locating any error made in a given script.
• Jaql is a functional data query language,73
designed by IBM and built upon the JSON data
model.68 Jaql is a general-purpose dataflow
language that manipulates semi-structured information in the form of abstract JSON values. It
provides a framework for reading and writing
data in custom formats, and provides support for
common input/output formats like CSVs. In the
same way as Pig and Hive, it provides significant
SQL operators such as filtering, transformations,
sort, group-bys, aggregation, and join.
Being constructed over JSON, Jaql is extendable
with operations written in several programming languages such as Javascript and Python. Regarding data
types, besides to the standard values it supports arrays
and records of name-value pairs. It also comes with a
rich array of built-in functions for processing unstructured or semi-structured data. Jaql also provides a user
with the capability of developing modules, a concept
similar to Java packages.
Volume 4, September/October 2014
• Dremel architecture74 works in a similar way as
distributed search engines do, i.e., the query is
managed by a serving tree, and it is rewritten at
each step. The result of the query is assembled by
aggregating the replies received from lower levels
of the tree. In contrast to Hive or Pig, Dremel
does not translate these queries into MapReduce
jobs.
The benefit of this model is the high efficiency
achieved by the way the data is stored. In particular, it follows a column-striped storage representation,
which enables it to read less data from secondary storage and reduce CPU cost due to cheaper compression.
An extension to a nested-column scheme is developed
for a faster access, similarly to JSON. There is an
open-source framework for Dremel, which is developed under the Apache Project, known as Drill.75
• Scope is a stack of protocols developed by
Microsoft in contraposition to the Apache
Hadoop project.76 This platform aims at merging distributed databases and MapReduce. It
is developed for structured and unstructured
data, with its own file management system and
execution engine. Data are stored in a relational format within a distributed data platform,
named Cosmos.56
The Scope scripting language resembles SQL
(as in the case of Hive) but with integrated C#
extensions. This fact allows users to write custom
operators to manipulate rowsets where needed. A
Scope script consists of a sequence of commands,
which are data manipulation operators that take one
or more row sets as input, perform some operation
on the data, and output a row set. Every row set has
a well-defined schema that all its rows must adhere
to. Users can name the output of a command using
an assignment, and an output can be consumed by
subsequent commands simply by referring to it by
name. Named inputs/outputs enable users to write
scripts in multiple (small) steps, a style preferred by
some programmers.
To summarize this subsection, we present in
Table 2 the main characteristics for the previously
introduced query models.
THE CENTRAL AXIS OF SCALABILITY:
THE MAPREDUCE PROGRAMMING
MODEL
In this section, our aim is to first go deeper on
the description of parallel computation tier and to
© 2014 John Wiley & Sons, Ltd.
391
wires.wiley.com/widm
Overview
TABLE 2 Summary of the Characteristics of the Query Systems for Big Data
System
Developed by
Language
Hive
Facebook
HiveQL
Declarative (SQL dialect)
Better suited for structured data
Pig
Yahoo!
Pig Latin
Data flow
Complex
Jaql
IBM
Jaql
Data flow
JSON, semi-structured
Dremel/Drill
Google/Apache
DrQL
Declarative (SQL dialect)
Structured and unstructured data
Scope
Microsoft
SQL/C#
Data flow
Structured and unstructured data
explain in detail the features of the MapReduce
programming model.77 Different implementations of
the MapReduce framework are possible depending on
the available cluster architecture.
We will focus on the Hadoop MapReduce
implementation14 for its wider usage and popularity
due to its performance, open-source nature, installation facilities and its distributed file system. This fact
is quite important to remark, as we may distinguish
between MapReduce (the theoretical framework) and
Hadoop MapReduce (the worldwide open-source
implementation).
In order to do so, we will first describe the
features and advantages of this programming model.
Then, we will contrast its functionality versus some
traditional approaches such as MPI and parallel
databases. Next, we will introduce those key areas of
DM that benefit the most from Big Data problems, and
we describe several libraries that support MapReduce
for solving these tasks. Finally, we will point out some
current drawbacks of this approach, stressing several
cases where researchers have reported that MapReduce is not the most appropriate solution.
Features and Advantages of Hadoop
MapReduce
The MapReduce framework13 was initially introduced
by Google in 2004 for writing massive scale data
applications. It was developed as a generic parallel
and distributed framework, which allows to process
massive amounts of data over a cluster of machines.
In this section, we will first introduce the features of
this model. Then, we will present the elements that
compose a cluster running such a system. Afterwards,
we will enumerate some of the goodness that have lead
to the success of this new programming model.
Introduction to MapReduce
The MapReduce framework is based on the fact that
most of the information processing tasks consider a
similar structure, i.e., the same computation is applied
over a large number of records; then, intermediate
392
Type of Language
Data Structures Supported
results are aggregated in some way. As it was previously described, the programmer must specify the
Map and Reduce functions within a job. Then, the job
usually divides the input dataset into independent subsets that are processed in parallel by the Map tasks.
MapReduce sorts the different outputs of the Map
tasks that become the inputs that will be processed by
the Reduce task. The main components of this programming model, which were previously illustrated in
Figure 5, are the following ones:
• The job input is usually stored on a distributed
file system. The master node performs a segmentation of the input dataset into independent
blocks and distributes them to the worker nodes.
Next, each worker node processes the smaller
problem, and passes the answer back to its master node. This information is given in terms of
<key,value> pairs which form the processing
primitives.
• The former input <key,value> pairs are split
and distributed among the available Map tasks.
Specifically, Map invocations are distributed
across multiple machines by automatically partitioning the input data into a set of M splits.
The input splits can be processed in parallel
by different machines. Then, Map functions
emit a set of intermediate <key,values> pairs
as output. Before the execution of a Reduce
function, MapReduce groups all intermediate
values associated with the same intermediate key
(<key,list(values)> and transforms them to speed
up the computation in the Reduce function.
• The intermediate values are supplied to the user’s
Reduce function via an iterator, which allows
to handle lists of values that are too large to
fit in memory. In particular, Reduce invocations
are distributed by partitioning the intermediate
key space into R pieces using a partitioning
function [e.g., hash(key) module R]. The number
of partitions (R) and the partitioning function
are specified by the user. Finally, the Reduce
© 2014 John Wiley & Sons, Ltd.
Volume 4, September/October 2014
WIREs Data Mining and Knowledge Discovery
Big Data with Cloud Computing
functions generate an arbitrary number of final
<key,values> pairs as output.
The whole process can be summarized as follows: the master node collects the answers to all the
subproblems, sorts the Map task outputs by their keys,
groups those that have the same key, and shuffles them
among the available Reduce tasks using a dynamic
scheduling mechanism. In this approach, the runtime
assigns Map/Reduce tasks to the available computation resources simplifying the optimal utilization of
heterogeneous computational resources while the initial assignment of Map tasks is performed based on
the data locality. This approach also provides an automatic load balancing for Map tasks with skewed data
or computational distributions.
An illustrative example about how MapReduce
works could be finding the average costs per year
from a big list of cost records. Each record may be
composed by a variety of values, but it at least includes
the year and the cost. The Map function extracts from
each record the pairs <year, cost> and transmits them
as its output. The shuffle stage groups the <year, cost>
pairs by its corresponding year, creating a list of costs
per year <year, list(cost)>. Finally, the Reduce phase
performs the average of all the costs contained in the
list of each year.
• The DataNode daemon works on the slave
machines of the cluster. When accessing an HDFS
file, the client communicates directly with the
DataNode daemons to process the local files corresponding to the blocks, which will be previously located by the NameNode. A DataNode
may communicate also with other DataNodes for
replicating its data blocks for the sake of redundancy.
• The JobTracker daemon is the linkage between
the parallel application and Hadoop. There is
just one JobTracker daemon running in the MasterNode. Its functionality consists of determining
which files to process, assigning nodes to different tasks, and monitoring all tasks they are
running.d
• In addition to the JobTracker, several TaskTrackers manage the execution of individual tasks on
each slave nodes. There is a single TaskTracker
per node, but each one of them can run many
Java virtual machines for parallel execution. If
the JobTracker does not receive any signal from
the TaskTracker, it assumes that the node has
crashed and it resubmits the corresponding task
to other node in the cluster.14
The Success of MapReduce
Hadoop MapReduce Cluster Configuration
A full configured cluster running MapReduce is
formed by a master-slave architecture, in which one
master node manages an arbitrary number of slave
nodes. The file system replicates the file data in multiple storage nodes that can concurrently access the
data. As such cluster, a certain percentage of these
slave nodes may be out of order temporarily. For this
reason, MapReduce provides a fault-tolerant mechanism, such that, when one node fails, it restarts automatically the task on another node. In accordance with
the above, a cluster running MapReduce includes several ‘daemons’ (resident programs) that will work in
this server, namely the NameNode, DataNode, JobTracker, and TaskTracker:
• A server that hosts the NameNode is devoted
to inform the client which DataNode stores the
blocks of data of any HDFS file. As it is related
to memory and I/O, this server does not hold
any data, neither performs any computations to
lower the workload on the machine. The node
that hosts the NameNode is the most important
one of the cluster, so if it fails, the complete
system will crash.
Volume 4, September/October 2014
We may observe a growing interest of corporations
and researchers in Big Data analysis.78 We have
pointed out that the main reason is related to the large
number of real applications that require scalable solutions such as sensor networks,79 intrusion detection
systems,80 or Bioinformatics.81–83
This success of the MapReduce framework in
this field is based on several issues that imply several
advantages for the programmer, mainly facilitating
parallel execution:
(1) Fault-tolerant service: In large clusters, machine
failures should be assumed as common, such
that active cluster nodes should be prepared for
rebalancing the load. This issue is extremely
significant for those DM tasks that require a
long time to execute, where it is quite costly to
restart the whole job.
For MapReduce, the master pings every Mapper
and Reducer periodically. If no response is received
for a certain time window, the machine is marked
as failed. Ongoing task(s) and any task completed
by the Mapper is reset back to the initial state
and reassigned by the master to other machines
© 2014 John Wiley & Sons, Ltd.
393
wires.wiley.com/widm
Overview
from scratch. Completed Map tasks are re-executed
on a failure because their output is stored on the
local disk(s) of the failed machine and is therefore
inaccessible. Completed Reduce tasks do not need to
be re-executed as their output is stored in the global
file system.
(2) Fair task distribution: The MapReduce system
will determine task granularity at runtime and
will distribute tasks to compute nodes as processors become available. Additionally, if some
nodes are faster than others, they will be given
more tasks.
(3) Move processing to the data: On grid
computing it was common to have both ‘processing’ and ‘storage’ nodes linked together
by a high-capacity interconnect. However,
current applications on Big Data imply
data-intensive workloads that are not very
processor-demanding. This causes a bottleneck
in the cluster network that degrades the system
productivity.
The MapReduce alternative is running the code
directly within the node where the data block is
located, considering that the data is spread across the
local disks of nodes in the cluster. The complex task of
managing storage in such a processing environment is
typically handled by a distributed file system that sits
underneath MapReduce.
(4) Trade latency for throughput: By definition,
data-intensive processing implies that the relevant datasets are too large to fit in memory and
must be held on disk. Instead, a sequential process of the data into long streaming operations
is carried out. Random accesses must be also
avoided for all computations.
(5) Hide system-level details from the application
developer: Distributed programming is quite
challenging to write and debug as the code
runs concurrently in unpredictable orders, and
access data in unpredictable patterns. The idea
behind MapReduce is to isolate the developer
from system-level details maintaining a separation between which processes are going to
be performed and how those computations are
actually carried out on a cluster.
Therefore, the main advantage of this framework is that the programmer is only responsible for
the former, i.e., they only need to focus on the Map and
Reduce function, everything else being common to all
394
programs. With this small constraint, all of the details
of communication, load balancing, resource allocation, job startup, and file distribution are completely
transparent to the user.
MapReduce Versus Traditional Parallel
Approaches
The ubiquitous distributed-memory MPI47 has been
the de facto standard for parallel programming for
decades. MPI supports a rich set of communication
and synchronization primitives for developing parallel
scientific applications. However, the challenge for MPI
in Big Data applications is related to the application
of the checkpointing fault-tolerance scheme, as it is
challenging at extreme scale due to its excessive disk
access and limited scalability.84,85
MPI and MapReduce are not so different
regarding their implementation. For example, the
intermediate data-shuffle operation in MapReduce is
conceptually identical to the familiar MPI_Alltoall
operation. But in spite of the features they have in
common, MPI lacks from the benefit of MapReduce
on less reliable commodity systems.86 Regarding this
fact, there have been several efforts for migrating
classical MPI-based software to MapReduce. One of
the pioneers works on the topic considered the implementation of MapReduce within a Python wrapper
to simplify the writing of user programs.87 Other
approaches have been developed under a C++ library
for its use in graph analytics.88
Despite of the previous fact, we must take into
account that any MapReduce application is preferred
to an MPI implementation only when it accomplishes
two simple conditions89 : (1) input data must be Big;
(2) the overhead due to a ‘bad’ codification of the Map
and Reduce functions should be minimal. This latter
issue refers to the use of those optimization mechanisms that are available in MapReduce,90 and also to
a proper implementation in which both functions balance in an appropriate way the workload.
Focusing on DBMSs, traditional relational
databases seem to be relegated to the status of legacy
technology. But, if these databases do not scale adequately, how is it possible that the biggest companies
are still making use of these systems? The answer is
that commercial systems can actually scale to stunning
amount of transactions per seconds. However, these
are expensive solutions and they need high-end servers
(even more expensive) to work properly. Hence, there
is an unfulfilled need for a distributed, open-source
database system.4
Nevertheless, regarding usability purposes,
DBMSs do also a great job of helping a user to
© 2014 John Wiley & Sons, Ltd.
Volume 4, September/October 2014
WIREs Data Mining and Knowledge Discovery
Big Data with Cloud Computing
maintain a dataset, adding new data over time,
maintaining derived views of the data, evolving its
schema, and supporting backup, high availability, and
consistency in a variety of ways.3 Even more: some
of the advantages enumerated for key-value storage
(NoSQL), such as schema later, or better availability
and faster response time through relaxed consistence
(as explained in the Database Management section),
can be also partially accomplished by traditional
DBMS with respect to binary large objects, and the
support of semi-structured data in XML format.
When performing some well-structured analytical tasks, the goodness of DBMS over MapReduce and
NoSQL approaches is related to the lack of indexes
and a declarative language of the latter.4 There are
many cases, especially for Map functions, for which
the function is too complicated to be expressed easily in a SQL query, such as fault-tolerant parallel
execution of programs written in higher-level languages (such as Pig Latin71 ) across a collection of input
data.90
However, this can be viewed as a double-edged
sword depending on the system’s user. For example,
Machine Learning scientists may find unnatural (if not
impossible!) to develop their scripts into a declarative
way. Hence, they might prefer more conventional,
imperative ways of thinking about these algorithms.
Other programmers may find that writing the SQL
code for each task is substantially easier than writing
MapReduce code.91
In spite of all the aforementioned issues, MapReduce excels in several factors.90 For example, a single
MapReduce operation easily processes and combines
data from a variety of storage systems. Tools based
on MapReduce provide a more conventional programming model, an ability to get going quickly on
analysis without a slow import phase, and a better
separation between the storage and execution engines.
On the contrary, the input for a parallel DBMS must
first be copied into the system, which may be unacceptably slow, especially if the data will be analyzed
only once or twice after being loaded. In particular,
a benchmark study using the popular Hadoop and
two parallel DBMSs,91 showed that the DBMSs were
substantially faster than the MapReduce system only
once the data is loaded.
According to these facts, MapReduce-based systems may offer a competitive edge with respect to traditional parallel solutions in terms of performance, as
they are elastically scalable and efficient. Specifically,
we have highlighted their goodness versus MPI and
DBMS, as MapReduce systems are able to provide
the functionality for both doing ‘traditional’ SQL-like
queries for analysis, e.g., using systems such as Hive
Volume 4, September/October 2014
or Pig, and also for automated data analysis such
as DM.
Data Processing Algorithms for Big Data
with MapReduce Open-Source Libraries
It is straightforward to acknowledge that, the more
data we have, the more insight we can obtain from
it. However, we have also pointed out that current
implementations of data processing algorithms suffer
from the curse of dimensionality and they can also
scale to certain problems. Standard DM algorithms
require all data to be loaded into the main memory,
becoming a technical barrier for Big Data because
moving data across different locations is expensive,
even if we do have enough main memory to hold all
data for computing.
Because of the former, many efforts have been
carried out to develop new approaches, or adapt previous ones, into the MapReduce parallel programming
model, as well as providing a Cloud Computing platform of Big Data services for the public.23 In general,
when coding an algorithm into MapReduce, the Map
function is commonly used as simply dividing the initial problem into different subsets. However, this process may create small datasets with lack of density.
This can be considered as a hitch for some problems
such as classification. On the contrary, the Reduce
function is considered as the most creative stage. This
is due to the fact that it is necessary to build a global
quality model associated to the problem that is aimed
to solve.
Regarding the structure of Big Data problems,
the list DM areas with this type of applications exist
essentially covers all major types. For the sake of
clarity, we focus on those that have attracted the
highest interest from researchers. Specifically, these
are related to classification, frequent pattern mining,
clustering, and recommender systems:
• Classification techniques decide the category a
given ‘object’ belongs to with respect to several
input attributes.92,93 It mainly works by determining whether a new input matches a previously
observed pattern or not, i.e., matching a learned
discrimination function.
In this area, several implementations have
been proposed for different classification architectures such as Support Vector Machines,94 Fuzzy
Rule-Based Classification Systems,95 Hyper Surface Classification,96 rough sets,97 and ensembles of
classifiers.98,99 Additionally, there are many applications in engineering that directly benefit from parallel
© 2014 John Wiley & Sons, Ltd.
395
wires.wiley.com/widm
Overview
classification models in MapReduce. Among them,
we may stress fault detection and diagnosis,100,101
remote sensing images,102 and Bioinformatics.103–106
• Frequent pattern mining, also known as association rule mining,107,108 discovers interesting relations among items (objects or variables) in large
databases. It basically aims at finding groups of
elements that commonly appear ‘together’ within
the database and that can be transformed into
rules of causality A ⇒ B, i.e., when element A
shows up, then it is likely to have also B.
For Market Basket Analysis, two extensions
from the well-known Apriori algorithm to the MapReduce framework have been proposed in Ref 109 We
must also stress the ‘Plute’ algorithm, developed for
mining sequential patterns.110
• Clustering techniques, as its name suggests, aim
to group a large number of patterns according
to their similarity.111,112 In this way, the clustering identifies the structure, and even hierarchy,
among a large collection of patterns.
Several clustering approaches have been shown
to be properly adapted to the MapReduce framework.
In Ref 113, Particle Swarm Optimization for defining the cluster parameters is considered. In Ref 114
DBCURE-MR, a parellelizable density-based clustering algorithm is proposed. Finally, the Cop-Kmeans
method with cannot-link constraints is presented in
Ref 115.
• Recommender systems are designed for inferring likes and preferences and identify unknown
items that are of interest.116,117 They are mostly
observed in online systems such as in e-commerce
services, social networks, and so on.
Among the examples of these kinds of systems,
we may stress TWILITE,118 which helps people find
users to follow for the Twitter application. We may
also find a TV program recommender system,119 and a
system based on collaborative users applied in movies
databases proposed in Ref 120.
There is a myriad of different implementations that may overwhelm any interested researcher
to address Big Data problems. Fortunately, several
projects have been built to support the implementation and execution of DM algorithms in a straightforward way. Among them, the most relevant one
is possibly the Mahout Machine Learning library.121
396
It contains scalable Machine Learning implementations written in Java and built mainly upon Apache’s
Hadoop-distributed computation project, which was
previously described in detail.
First of all, we must point out two significant
issues: (1) On the one hand, it is just a library. This
means that it does not provide a user interface, a
prepackaged server, or an installer. It is a simply
framework of tools intended to be used and adapted
by developers. (2) On the second hand, it is still
under development (currently is under version 0.9).
It is a quite recent approach and it is far from being
thoroughly documented. However, it includes a wide
amount of methods and it is continuously in expansion
with new DM approaches. Currently, it supports
the aforementioned main kind of learning tasks. It
addition, it includes dimension reduction techniques,
and other miscellaneous approaches, all of which are
summarized in Table 3.
In April 2014, Mahout has said goodbye to
MapReduce. The justification for this change is
twofold: on the one hand organizational issues, as it
was onerous to provide support for scalable ML; on
the other hand, technical considerations, as MapReduce is not so well suited for ML, mainly due to the
launching overhead, especially noticeable for iterative
algorithms, different quality of implementations,
and/or and the unsuitability of some methods to be
codified as Map and Reduce operations.
Hence, the future of Mahout is the usage of
modern parallel-processing systems that offer richer
programming models and more efficient executions,
while maintaining the underlying HDFS. The answer
to these constraints is Apache Spark,122 which will
be also described in detail in the following sections.
In this way, future implementations will use the
DSL linear algebraic operations following Scala &
Spark Bindings. These are basically similar to R
(Matlab)-like semantics, allowing automatic optimization and parallelization of programs.
Although Mahout has gained great popularity
among DM practitioners over Big Data problems,123
it is not the unique solution for this type of tasks.
In this review we would like to stress five additional
open-source approaches as an alternative to Mahout
for data processing algorithms:
• NIMBLE124 allows the development of parallel
DM algorithms at a higher level of abstraction,
thus facilitating the programmer using reusable
(serial and parallel) building blocks that can be
efficiently executed using MapReduce. Furthermore, NIMBLE provides built-in support to process data stored in a variety of formats; it also
© 2014 John Wiley & Sons, Ltd.
Volume 4, September/October 2014
WIREs Data Mining and Knowledge Discovery
Big Data with Cloud Computing
TABLE 3 DM Tasks and Algorithms Implemented in the Mahout
Software Tool Version 0.9
Type of Task
List of Algorithms
Classification
Naive Bayes/Complementary Naive Bayes
Multilayer perceptron
Random forest
Logistic regression
Hidden Markov models
Clustering
Canopy clustering
k-means clustering
Fuzzy k-means
Streaming k-means
Spectral clustering
Collaborative
filtering
User-based collaborative filtering
Item-based collaborative filtering
Matrix factorization with alternating least
squares
Weighted matrix factorization, SVD++, Parallel
SGD
Dimension
reduction
Singular value decomposition
Stochastic singular value decomposition
Principal components analysis
Lanczos algorithm
Topic models
Latent Dirichlet allocation
Miscellaneous
Frequent pattern mining
RowSimilarityJob—compute
pairwise—similarities between the rows of a
matrix
ConcatMatrices—combine two matrices—or
vectors into a single matrix
Collocations—find colocations of tokens in text
allows facile implementation of custom data formats.
• SystemML125 is similar to NIMBLE. It allows
to express algorithms in a higher-level language,
and then compiles and executes them into a
MapReduce environment. This higher-level language exposes several constructs including linear
algebra primitives that constitute key buildings
blocks for a broad class of supervised and unsupervised DM algorithms.
• Ricardo is a system that integrates R
(open-source statistical analysis software)
and Hadoop.126 The main advantage of this
approach is that users who are familiarized with
the R environment for statistical computing,
can still used the same functionality but with
Volume 4, September/October 2014
larger data, as it supports the data management
capabilities of Hadoop.
• Rhipe stands for R and Hadoop Integrated Programming Environment.127,128 Its main goals
are (1) not losing important information in the
data through inappropriate data reductions; (2)
allowing analysis exclusively from within R,
without going into a lower level language.
• Wegener et al.129 achieved the integration of
Weka6 (an open-source Machine Learning and
Data Mining software tool) and MapReduce.
Standard Weka tools can only run on a single machine, with a limitation of 1-GB memory. After algorithm parallelization, Weka breaks
through the limitations and improves performance by taking the advantage of parallel computing to handle more than 100-GB data on
MapReduce clusters.
To conclude this section, it is worth to point
out that we are attending a generational shift regarding ML libraries and software tools,130 which is
summarized in Figure 6. The first generation comprises the traditional tools/paradigms such as SAS,
R, or KNIME, which are typically only vertically
scalable. In this section we have focused on the
second generation approaches, which work over
Hadoop/MapReduce. However, we must be aware
of a third generation that involves those programming frameworks that go beyond MapReduce such
as Spark. The basis of these novel models will be
introduced in the next Sections.
We must also point out that there are also several
open-source and proprietary platforms that have been
used in business domains such as Pentaho Business
Analytics, Jaspersoft BI suite, and Talend Open Studio
among others,131 but they are out of the scope of
this work.
Beyond MapReduce: Global Drawbacks
of the MapReduce Approach
We have stressed the goodness of MapReduce as a
powerful solution designed to solve different kinds
of Big Data problems. However, it is not a panacea:
there are some scenarios in which this functional
programming model does not achieve the expected
performance, and therefore alternative solutions must
be chosen.
The first drawback is straightforward: it does not
provide any significant improvement in performance
when the work cannot be parallelized, according to
Amdahl’s law.132 But there are some other issues that
© 2014 John Wiley & Sons, Ltd.
397
wires.wiley.com/widm
Overview
Generation
1st Generation
2nd Generation
3rd Generation
Examples
SAS, R, Weka, SPSS,
KNIME, KEEL...
Mahout, Pentaho, Cascading
Spark, Haloop, GraphLab,
Pregel, Giraph, ML over
Storm
Scalability
Vertical
Horizontal (over Hadoop)
Horizontal (beyond Hadoop)
Algorithms
available
Huge collection of
algorithms
Small subset: sequential logistic
regression, linear SVMs,
Stochastic Gradient Descendent,
k-means clustering, Random
forest, etc.
Much wider: CGD, ALS,
collaborative filtering, kernel
SVM, matrix factorization,
Gibbs sampling, etc.
Algorithms
Not available
Practically nothing
Vast no.: Kernel SVMs,
Multivariate logistic Regression,
Conjugate Gradient Descendent,
ALS, etc.
Multivariate logistic
regression in general form,
k-mean clustering, etc. –
Work in progress to expand
the set of available
algorithms
Fault-tolerance
Single point of
failure
Most tools are FT, as they are
built on top of Hadoop
FT: Haloop, Spark
Not FT: Pregel, GraphLab,
Giraph
FIGURE 6 | Machine Learning software suites: a three-generational view.
should be taken into account prior to the migration
this type of systems, which are enumerated below:
• Not all algorithms can be efficiently formulated
in terms of Map and Reduce functions, as they
are quite restrictive.
In particular, among all possible deficiencies of
MapReduce, the greatest critic reported is the implementation of iterative jobs.133 In fact, many common
Machine Learning algorithms apply a function repeatedly to the same dataset to optimize a parameter. The
input to any Hadoop MapReduce job is stored on the
HDFS. Hence, whenever a job is executed, the input
has to be reload the data from disk every time. This is
done regardless of how much of the input has changed
from the previous iterations, incurring in a significant
performance penalty.
• MapReduce has also some problems in processing networked data, i.e., graph structures.134
The main difference in the processing of regular
data structures (tables) and relational models (graphs)
relies on different problem decompositions. Table
structures are simpler as the algorithm must only
process individual records (rows). However, for the
398
networked data, single processing of a graph vertex
usually requires access to the neighborhood of this
vertex. As, in most cases, the graph structure is static,
this represents wasted effort (sorting, network traffic,
among others), as it must be accessed at every iteration
via the distributed file system, showing the same
behavior as in the previous case.
• Processing a lot of small files, or performing
intensive calculations with small size data is
another issue.
As stated previously, for each MapReduce job
that is executed, some time is spent on background
tasks, incurring in high startup costs (in Hadoop, they
can be tens of seconds on a large cluster under load).
This places a lower bound on iteration time.135
• Regarding interactive analytics, it has not shown
a good behavior when processing transactions.
The hitch here is that this considers random
access to the data. MapReduce is often used to run
ad hoc exploratory queries on large datasets, through
SQL interfaces as explained in previous Sections.
Ideally, a user would be able to load a dataset of
interest into memory across a number of machines and
© 2014 John Wiley & Sons, Ltd.
Volume 4, September/October 2014
WIREs Data Mining and Knowledge Discovery
Big Data with Cloud Computing
query it repeatedly. However, with MapReduce each
query incurs significant latency (i.e., tens of seconds)
because it runs as a separate MapReduce job and reads
data from disk.122 Therefore, its usage is not adequate
for low latency data access. Even worse, in some cases
there are very high memory requirements, as all the
data used for local processing must fit in local memory.
• We must also point out security and privacy
issues, not because it is an inner problem of
MapReduce, but because it must be implemented
within a Cloud Computing environment.
Users are reluctant to the use of a public cloud
as they cannot be sure that it guarantees the confidentiality and integrity of their computations in a way
that is verifiable.25 In addition to security challenges,
there remains the issue of accountability with respect
to an incorrect behavior. If data leaks to a competitor,
or a computation returns incorrect results, it can be
difficult to determine whether the client or the service
provider are at fault.26
• Finally, we must stress the complexity of
installing and having ready a complete cluster with a complete stack of protocols, such as
the one described within the Cloud Computing
Section. Fortunately, this issue may be overcome
Mapreduce
model
DAG model
making use of Clouderae ,136 a system that provides an enterprise-ready, 100% open-source
distribution that includes Hadoop and related
projects.
A ‘PLAN-B’ FOR MAPREDUCE:
PROGRAMMING FRAMEWORKS
According to the issues that were raised in the previous
part of the section, several alternatives have been
proposed in the specialized literature. Depending on
the features of the problem we are dealing with, we
may choose the one that better suits our necessities.137
Considering the previous comments, in this
section we provide a representative list of programming frameworks/platforms that have been either
adapted from the standard MapReduce, or developed
as new framework, aiming to achieve a higher scalability for Big Data applications.130 A summary of these
tools is shown in Figure 7, where we divide each platform with respect to their paradigm and use.
In the following, we describe them according to
this division structure, i.e., horizontal and vertical.
Directed Acyclic Graph Model
The Directed Acyclic Graph (DAG) model defines the
dataflow of the application, where the vertices of the
Graph model
BSP/collective
model
Hadoop
MPI
For
iterations/
learning
HaLoop
Pregel
Twister
Giraph
Dryad
GraphLab
Spark
S4
For
streaming
GraphX
Storm
Spark streaming
Phoenix
MARS
GPU-based
GPMR
GREX
FIGURE 7 | Alternative frameworks for the standard MapReduce model.
Volume 4, September/October 2014
© 2014 John Wiley & Sons, Ltd.
399
wires.wiley.com/widm
Overview
graph define the operations that are to be performed
on the data. The ‘computational vertices’ are written
using sequential constructs, devoid of any concurrency
or mutual exclusion semantics.
The main exponent of this model is Dryad,138 a
Cloud Computing framework proposed by Microsoft
Research as a general-purpose execution environment for distributed, data-parallel applications. In
this framework, applications are modeled as directed
acyclic graphs, where each vertex is a program and the
edges represents the data channel.
As the core of MapReduce, it allows automatic
management of scheduling, distribution, and fault tolerance. DryadLINQ139 is the base programming API
for Dryad and hence it is more suitable for applications that process structured data. It generates Dryad
computations from the LINQ Language-Integrated
Query extensions to C#.
granularity for the Map tasks; (3) a new ‘Combine’
operation that acts as another level of reduction; and
(4) the implementation of a set of programming extensions to MapReduce, such as MapReduceBCast(Value
value), which facilitates sending a set of parameters,
a resource name, or even a block of data to all Map
tasks.
The disadvantage in this case is that, as it
requires data to fit into the collective memory of the
cluster in order to be effective, it cannot cope with
jobs that require the processing of terabytes of data.
Another disadvantage of Twister is its weak fault
tolerance compared to Hadoop.
• Spark122 and Spark2142 are developed to overcome data reuse across multiple computations.
It supports iterative applications, while retaining
the scalability and fault tolerance of MapReduce,
supporting in-memory processes.
Iterative MapReduce
We have pointed out that one of the main drawbacks
of MapReduce is that is not so well suited for iterative
algorithms, mainly due to the launching overhead that
is present even if the same task has been carried out
already. For this reason, there has been many efforts
to develop new approaches that can address this issue,
which we present below:
• Haloop140 is a new programming model and
architecture that provides support for iterative
algorithms by scheduling tasks across iterations
in a manner that exploits data locality, and by
adding various caching mechanisms. Its main
contribution is the reuse of Mappers and Reducers when they do the same job. It also implements planning loops, i.e., tasks are assigned to
the same nodes at every turn. By caching loop
invariants no resources are spent by reloading
repeated information several times. There is also
a local ‘Reducer’ to compare the loop condition
efficiently. It maintains the fault tolerance of the
standard MapReduce and a compatibility with
Hadoop.
• Twister141 is a distributed in-memory MapReduce system with runtime optimized operations
for iterative MapReduce computations, from
which intermediate data are retained in memory
if possible, thus greatly reducing iteration overhead.
Its main features are the following ones: (1) addition of a ‘configure’ phase for loading static data
in both Map and Reduce tasks; (2) using a higher
400
With this aim, the authors include a new abstraction model called Resilient Distributed Datasets
(RDDs), which are simply a distributed collection of
items. RDDs are fault-tolerant, parallel data structures
that let users explicitly persist intermediate results in
memory, control their partitioning to optimize data
placement, and manipulate them using a rich set of
operators. In a nutshell, it provides a restricted form
of shared memory based on coarse-grained transformations rather than fine-grained updates to shared
state. RDDs can either be cached in memory or materialized from permanent storage when needed (based
on lineage, which is the sequence of transformations
applied to the data). However, Spark does not support
the group reduction operation and only uses one task
to collect the results, which can seriously affect the
scalability of algorithms that would benefit from
concurrent Reduce tasks, with each task processing a
different subgroup of the data
Similar to Mahout for Hadoop,121 Spark implements a Machine Learning library known as MLlib,
included within the MLBase platform.143 MLlib currently supports common types of Machine Learning
problem settings, as well as associated tests and data
generators. It includes binary classification, regression,
clustering, and collaborative filtering, as well as an
underlying gradient descent optimization primitive.
Bulk Synchronous Parallel/Graph Model
The Bulk Synchronous Parallel (BSP) model144 is a
parallel computational model for processing iterative
graph algorithms. With this scheme, computations
are ‘vertex-centric’ and user defined methods and
© 2014 John Wiley & Sons, Ltd.
Volume 4, September/October 2014
WIREs Data Mining and Knowledge Discovery
Big Data with Cloud Computing
Processors
Local
computation
Communication
Barrier
synchronisation
FIGURE 8 | BSP model workflow.
algorithms proceed in supersteps concurrently, with
synchronization barriers between each one of them.
Each vertex has a state and is able to receive messages
sent to it from the other vertexes in the previous
superstep. A superstep consists of three ordered stages
(Figure 8):
(1) Concurrent computation: computation on
locally stored data.
(2) Communication: messages in a point-to-point
manner (collective).
(3) Barrier synchronization: wait and synchronize
all processors at end of superstep.
In this way, more complex graph models allow
a better representation of the dataflow of the application, i.e., cyclic models define that iterativity. For this
reason, they are quite popular for some Big Data problems, such as social networks. A notorious example is
given by Facebook to analyze the social graph formed
by users and their connections (distance between people). Below we enumerate some BSP implementations:
• Pregel145 was first developed by Google based on
the premise that there are many practical computing problems that concern large graphs, such
as the Web graph and various social networks.
Therefore, it is not based on the MapReduce
model, but it implements the BSP model. While
the vertex central approach is similar to the
MapReduce Map operation which is locally performed on each item, the ability to preserve the
state of each vertex between the supersteps provides the support for iterative algorithms. In this
implementation, all states including the graph
structure, are retained in memory (with periodic
checkpoints).
Volume 4, September/October 2014
• Giraph146 is an open-source graph-processing
framework that implements the BSP computation. It includes a Hadoop Map-only job,
from which the graph is loaded and partitioned
across workers, and the master then dictates
when workers should start computing consecutive supersteps. Therefore, it can be run on any
existing Hadoop infrastructure, providing the
API and middleware of Pregel, as well as adding
fault-tolerance, and in-memory processing.
• GraphX147 extends Spark with a new graph
API. Its main goal is to unify graph-parallel and
data-parallel computations in one system with a
single composable API.
• GraphLab148 is a project started at Carnegie Mellon University in 2009 designed to efficiently
run in both shared and distributed-memory systems. It is basically a graph-based, high performance, distributed computation framework in an
optimized C++ execution engine. Graphlab provides MapReduce-like operations, called Update
and Sync functions. The Update function is able
to read and modify overlapping sets of data,
whereas the Sync function can perform reductions in the background while other computation
is running. GraphLab uses scheduling primitives
to control the ordering in which update functions are executed. It also includes powerful ML
toolkits for topic modeling, graph analytics, clustering, collaborative filtering, graphical models,
and computer vision, among others.
Stream Processing
The management of data streams is quite important
as these types of applications are typically under the
umbrella of Big Data. In what follows, we enumerate
some of the most significant approaches for stream
processing:
• Storm149 in an open-source project designed for
real-time processing of streaming data, in contrast to Hadoop which is for batch processing.
To implement real-time computation on Storm,
users need to create different topologies, i.e., a
graph of computation created and submitted in
any programming language. There are two kinds
of node in topologies, namely, spouts and bolts.
A spout is one of the starting points in the graph,
which denotes source of streams. A bolt processes input streams and outputs new streams.
Each node in a topology contains processing
logic, and links between nodes indicate how data
should be processed between nodes. Therefore, a
© 2014 John Wiley & Sons, Ltd.
401
wires.wiley.com/widm
Overview
topology is a graph representing the transformations of the stream, and each node in the topology
executes in parallel.
• S4150 is a general-purpose, distributed, scalable,
fault-tolerant, pluggable open-source computing
platform for processing continuous unbounded
streams of data. S4 provides a runtime distributed platform that handles communication,
scheduling, and distribution across containers.
Applications are built as a graph of processing
elements, which communicate asynchronously
by sending events on streams that interconnect
them. Events are dispatched to nodes according
to their key. The implementation of a S4 job is
designed to be modular and pluggable for easily and dynamically processing large-scale stream
data.
• Spark Streaming151 is an extension of the core
Spark API that enables stream processing of live
data streams. It features ‘DStream’, represented
as a sequence of RDDs. As it is integrated
under the Spark framework, it can handle Spark’s
learning and graph-processing algorithms.
MapReduce on Graphics Proccesing Units
Although we have focused on the Cloud Computing environments for developing Big Data applications, it is also possible to address this task using
the capabilities of graphics processing units (GPUs).
Next, we show some proposals that get advantage
of this:
• Phoenix152 implements MapReduce for sharedmemory systems. Its goal is to support efficient
execution on multiple cores without burdening
the programmer with concurrency management.
Because it is used on shared-memory systems it is
less prone to the problems we encountered with
iterative algorithms as long as the data can fit into
memory.
• MARS153 was created as an alternative to
Phoenix for the execution of MapReduce in a
GPU or multi-core environment.154 The motivation is that GPUs are parallel processors with
higher computation power and memory bandwidth than CPUs, also improving at a higher
rate.155 Thus, it follows a similar scheme to
MapReduce but adapted for a general-purpose
GPU environment, which implies several technical challenges.
• GPMR156
is
MapReduce
library
that
leverages the power of GPU clusters for
402
large-scale computing. GPMR allows flexible mappings between threads and keys, and
customization of the MapReduce pipeline with
communication-reducing stages (both PCI-e and
network).
• GREX157 uses general-purpose GPUs for parallel
data processing by means of the MapReduce
programming framework. It supports a parallel
split method to tokenize input data of variable
sizes in parallel using GPU threads. It also evenly
distributes data to Map/Reduce tasks to avoid
data partitioning skews. Finally, it provides a
new memory management scheme to enhance
the performance by exploiting the GPU memory
hierarchy.
LESSONS LEARNED
In this article we have identified the problem that
arises in the management of a large amounts of fast
and heterogenous data in DM, which is known as
the Big Data Problem. This situation is present in
a great number of applications, especially for those
based on the use of BI. We have pointed out the
handicap to efficiently manage this quantity of information with standard technologies, both regarding
storage necessities and computational power. This
fact makes clear that a new paradigm must be provided in order to support the scalability of these systems: Cloud Computing.158 Although it is possible
to deploy Big Data technologies in a noncloud cluster of computers, we must excel the properties of
these systems for easing the development of these
tasks.
Throughout this article we have described the
structure and properties of a Cloud Computing environment and we have focused on the execution engine
for allowing the scalability in the computation of DM
algorithms, i.e., the MapReduce approach. From the
realization of this thorough study, we may emphasize five important lessons learned that may help other
researchers to understand the intrinsic features of this
framework:
(1) There is a growing demand in obtaining
‘insight’ from as much amount of information
as possible, but we must be careful about what
we are actually able to take on. In order to
process Big Data, we must take into account
two main requirements: (1) to deploy a proper
computational environment, either a private
or public cloud; and (2) technical staff for
developing the software needed for the DM
process.
© 2014 John Wiley & Sons, Ltd.
Volume 4, September/October 2014
WIREs Data Mining and Knowledge Discovery
Big Data with Cloud Computing
In the former case, although we acknowledge
that the majority of the machines can be built on commodity hardware (scaling-out instead of scaling-up),
the master node of the system should be robust enough
to support the required scalability, i.e., number of jobs;
and the costs of the Information Technology crew
for installing and maintaining such a cluster are not
trivial. To ease this task, the Cloudera project136 has
established as an enterprise-ready source distribution
that includes Hadoop and related projects. In the case
of a public cloud, users can easily deploy Hadoop on
Amazon EC2 and Microsoft Azure cloud computing
platforms, by using Amazon Elastic MapReduce159
and HDInsight.160 The hitch here is that we must face
security and accountability of the system, as the main
issues users are reluctant with.
However, maybe the most important point is
related to the people that must manage the data and
translate it into useful information. Currently, being
a Data Scientist is undoubtedly the most demanding
career regarding its ability in managing statistics and
Machine Learning algorithms. There is a necessity of
correctly choosing the model structure for a given
problem, tuning appropriately the parameters of the
systems, and also being able to browse, visualize,
and report the results for the understanding of the
algorithms’ results.
(2) In accordance with the above, we have stressed
the goodness of the MapReduce model as a new
methodology that can help current researchers
and data scientists for the development of fast
and scalable applications in DM and BI. Two
main advantages of this programming model
are: (1) the transparency for the programmer,
who just needs to focus on the development of
the Map and Reduce functions, and not in the
inner operation of the system (better separation
between the storage and execution engines),
and (2) the robust fault tolerance that allows
a higher scalability and reliability for long
running jobs. In addition to the former, we have
stressed the use of key/value pairs (NoSQL)
instead of relational tables in case we are
dealing with unstructured or semi-structured
data, such as text documents or XML files. This
new database model provides higher flexibility
for working this type of data.
(3) In order to fully benefit from the MapReduce features, a certain degree of expertise is
needed. Specifically, the programmer must carefully tune several factors such as providing a
good storage system (e.g., HDFS), exploiting
indexes, and the use of an efficient grouping
Volume 4, September/October 2014
and scheduling algorithm.161,162 This supposes
a constraint for the achievement of the highest
performance with the MapReduce approach.
(4) Relational (and parallel) DBMSs leverage its
features to provide the user with a simple environment for storing, accessing, and obtaining
information from data. However, they require
a slow ‘import phase’ prior to perform the data
exploration, and several users may not find conformable with a declarative/relational framework.
Nevertheless, we must not consider MapReduce
as an adversary to classical approaches for managing the Big Data problem, but as a complement
to them. In fact, it is possible to write almost any
parallel-processing task as either a set of database
queries or a set of MapReduce jobs. Parallel DBMSs
excel at efficient querying large data sets, whereas
MapReduce style systems excel at complex analytics
and ETL tasks. Neither is good at what the other does
well, such that both technologies are complementary
rather than rivals.162
In fact, we have stressed several query systems
that act as a compromise solution, such as Hive, Pig,
or Dremel, among others. They are built upon the
upper layer of the Cloud Computing stack, developed
trying to link the MapReduce framework to DBMS,
such that complex analytics can be carried out into
the DBMS with embedded queries.
(5) Finally, although the Hadoop stack has become
the de facto general-purpose, large-scale data
processing platform of choice, we must first
study the nature of the problem to solve. First of
all, it is suitable for data-intensive parallelizable
algorithms, but when facing iterative and/or
graph-based algorithms, the performance that
can be achieved is below its full potential.135
Additionally, it does not support online transactions, as the way it is implemented does not
support the random reading and writing of a
few records.
For this reason, several alternative approaches
have been developed for extending the MapReduce model to different work scenarios.130 Among
these new approaches, those related with the Spark
environment122,142,151 have gained the greatest popularity; as a example, we may refer to the recent
migration of Mahout from Hadoop to Spark. However, no single programming model or framework
can excel at every problem; there are always tradeoffs between simplicity, expressivity, fault tolerance,
© 2014 John Wiley & Sons, Ltd.
403
wires.wiley.com/widm
Overview
performance, etc. The former issue makes us recall
the argument by Lin133 : ‘If all you have is a hammer,
throw away everything that’s not a nail!’. Hence,
we must remark that although MapReduce has its
limitations, as it cannot be applied for every problem
we have, the useful current properties of this approach
should make us adopt this framework as much as we
can for Big Data problems.
CONCLUSIONS
At present, the Big Data problem in DM and BI has
been established as one of the hot topics in the areas
of business and research. Owing to the novelty of
this concept several questions yet exist. Specifically,
we shall refer to its exact definition, features, and
especially the most properly methodology to solve this
problem successfully.
According to the above, the aim when developing this work was to unify the vision of the latest state
of the art on the subject. In particular, emphasizing the
importance of this new field of work with respect to its
application in BI tasks, and exposing the use of Cloud
Computing as the right tool when compared to more
traditional solutions. This new paradigm of work is
based on the use of elastic computing resources. It
allows users to execute requests dynamically, such
that it can achieve a high degree of scalability even
with low-end hardware resources. We have described
the architecture of this model, stressing those current
implementations that can be found in each of the
component layers, i.e., the file system, the execution
layer, DBMSs based on NoSQL, and those systems at
the query level.
Within the Cloud Computing paradigm, the
main point of this work has been focused on the execution layer, and more specifically on the MapReduce
programming model. We have described the main features of this approach and its ‘pros and cons’ when
compared versus some reference parallel computational models. In addition, we have enumerated several alternatives to the standard MapReduce model
that have been proposed in the literature. These systems aim to extend the application of MapReduce
for environments where it shows certain deficiencies,
such as the implementation of iterative or graph-based
algorithms.
In summary, our main purpose was to provide
an opportunity for any interested reader in the field of
Big Data, to become familiar with the most relevant
information of this promising topic.
NOTES
a
http://stevetodd.typepad.com/my_weblog/big-data/
http://cruncher.ncl.ac.uk/bdcomp/index.pl
c
http://www.synapse.org/dream
d In Hadoop version 2.0, the functionality of JobTracker has been further subdivided into several
daemons. Notwithstanding, the overall working is
equivalent.
e
https://www.cloudera.com
b
ACKNOWLEDGMENTS
This article is funded by King Abdulaziz University, under grant no. 3-611-1434-HiCi. The authors, therefore,
acknowledge technical and financial support of KAU. This work was also partially supported by the Spanish
Ministry of Science and Technology under projects TIN2011-28488, TIN-2012-33856, and the Andalusian
Research Plans P12-TIC-2958, P11-TIC-7765, and P10-TIC-6858. V. López holds a FPU scholarship from
Spanish Ministry of Education.
REFERENCES
1. Gupta R, Gupta H, Mohania M. Cloud computing and
big data analytics: what is new from databases perspective? In: 1st International Conference on Big Data
Analytics (BDA), New Delhi, India, 2012, 42–61.
2. Agrawal D, Das S, Abbadi AE. Big data and cloud
computing: current state and future opportunities.
In: 14th International Conference on Extending
Database Technology (EDBT), Uppsala, Sweden,
2011, 530–533.
3. Madden S. From databases to big data. IEEE Internet
Comput 2012, 16:4–6.
404
4. Kraska T. Finding the needle in the big data systems
haystack. IEEE Internet Comput 2013, 17:84–86.
5. Alpaydin E. Introduction to Machine Learning. 2nd
ed. Cambridge, MA: MIT Press; 2010.
6. Witten IH, Frank E, Hall MA. Data Mining: Practical Machine Learning Tools and Techniques. Morgan
Kaufmann Series in Data Management Systems. Amsterdam: Morgan Kaufmann; 2011.
7. March ST, Hevner AR. Integrated decision support systems: a data warehousing perspective. Decis
Support Syst 2007, 43:1031–1043.
© 2014 John Wiley & Sons, Ltd.
Volume 4, September/October 2014
WIREs Data Mining and Knowledge Discovery
Big Data with Cloud Computing
9. Negash S, Gray P. Business intelligence. Commun
Assoc Inf Sys 2004, 13:177–195.
27. Zikopoulos PC, Eaton C, deRoos D, Deutsch T, Lapis
G. Understanding Big Data—Analytics for Enterprise
Class Hadoop and Streaming Data. 1st ed. New York
City (USA): McGraw-Hill Osborne Media; 2011.
10. O’Neil C, Schutt R. Doing Data Science. 1st ed.
Sebastopol, CA: O’Reilly; 2013.
28. Labrinidis A, Jagadish HV. Challenges and opportunities with big data. PVLDB 2012, 5:2032–2033.
11. Provost F, Fawcett T. Data Science for Business.
What You Need to Know about Data Mining and
Data-Analytic Thinking. 1st ed. Sebastopol, CA:
O’Reilly; 2013.
29. Peralta D, Triguero I, Sanchez-Reillo R, Herrera F,
Benítez JM. Fast fingerprint identification for large
databases. Pattern Recognit 2014, 47:588–602.
8. Watson HJ, Wixom BH. The current state of business
intelligence. Computer 2007, 40:96–99.
12. Han J, Haihong E, Le G, Du J. Survey on nosql
database. In: 6th International Conference on Pervasive Computing and Applications, (ICPCA), Port Elizabeth, South Africa, 2011, 363–366.
13. Dean J, Ghemawat S. MapReduce: simplified data
processing on large clusters. Commun ACM 2008,
51:107–113.
14. Lam C. Hadoop in Action. 1st ed. Greenwich, Connecticut (USA): Manning; 2011.
15. Buyya R, Broberg J, Goscinski A. Cloud Computing:
Principles and Paradigms. Chichester: John Wiley &
Sons; 2011.
16. Mell P, Grance T. The NIST definition of cloud
computing ( draft ) recommendations of the national
institute of standards and technology. NIST Spec Publ
2011, 145.
17. Plummer D, Bittman T, Austin T, Cearley D, Cloud
DS. Defining and describing an emerging phenomenon. Technical Report, Gartner, 2008.
18. Hummer W, Satzger B, Dustdar S. Elastic stream
processing in the cloud. WIREs Data Mining Knowl
Discov 2013, 3:333–345.
19. Papazoglou M, Van Den Heuvel W-J. Service oriented
architectures: approaches, technologies and research
issues. VLDB J 2007, 16:389–415.
20. Ouf S, Nasr M, Amr M, Mosaad M, Kamal K,
Mostafa F, Said R, Mohamed S. Business intelligence
software as a service (SAAS). In: 2011 IEEE 3rd
International Conference on Communication Software
and Networks (ICCSN), Xian, China, 2011, 641–649.
21. Velte AT, Velte TJ, Elsenpeter R. Cloud Computing: A
Practical Approach. New York City (USA): McGraw
Hill; 2010.
22. Marx V. The big challenges of big data. Nature 2013,
498:255–260.
23. Wu X, Zhu X, Wu G-Q, Ding W. Data mining with big
data. IEEE Trans Knowl Data Eng 2014, 26:97–107.
24. Schlieski T, Johnson BD. Entertainment in the age of
big data. Proc IEEE 2012, 100:1404–1408.
25. Peltier TR, Peltier J, Blackley JA. Information Security Fundamentals. Boca Raton, FL: Auerbach Publications; 2005.
26. Xiao Z, Xiao Y. Achieving accountable mapreduce in
cloud computing. Future Gener Comput Syst 2014,
30:1–13.
Volume 4, September/October 2014
30. Peralta D, Galar M, Triguero I, Benítez JM,
Miguel-Hurtado O, Herrera F. Minutiae filtering
to improve both efficacy and efficiency of fingerprint
matching algorithms. Eng Appl Artif Intel 2014,
32:37–53.
31. Waller MA, Fawcett SE. Data science, predictive analytics, and big data: a revolution that will transform
supply chain design and management. J Bus Logistics
2013, 34:77–84.
32. Hey T, Trefethen AE. The UK E-science core programme and the grid. Future Gener Comput Syst 2002,
18:1017–1031.
33. Wang F-Y, Carley KM, Zeng D, Mao W. Social computing: from social informatics to social intelligence.
IEEE Intell Syst 2007, 22:79–83.
34. Linden G, Smith B, York J. Amazon.com recommendations item-to-item collaborative filtering. IEEE Internet Comput 2003, 7:76–80.
35. Ngai EWT, Wat FKT. Literature review and classification of electronic commerce research. Inf Manage
2002, 39:415–429.
36. Mattmann CA. Computing: a vision for data science.
Nature 2013, 493:473–475.
37. Provost F, Fawcett T. Data science and its relationship
to big data and data-driven decision making. Big Data
2013, 1:51–59.
38. Cherkassky V, Mulier F. Learning from Data: Concepts, Theory, and Methods. Wiley Series in Adaptive
and Learning Systems for Signal Processing, Communications and Control Series. New York, NY: John
Wiley & Sons; 1998.
39. Tan P-N, Steinbach M, Kumar V. Introduction to Data
Mining. London (U.K.): Pearson; 2006.
40. Chen CP, Zhang C-Y. Data-intensive applications,
challenges, techniques and technologies: a survey on
big data. Inf Sci 2014, 275:314–347.
41. Assent I. Clustering high dimensional data. WIREs
Data Mining Knowl Discov 2012, 2:340–350.
42. Bacardit J, Llorà X. Large-scale data mining using
genetics-based machine learning. WIREs Data Mining
Knowl Discov 2013, 3:37–61.
43. Rothnie JBJ, Bernstein PA, Fox S, Goodman N, Hammer M, Landers TA, Reeve CL, Shipman DW, Wong
E. Introduction to a system for distributed databases
(sdd-1). ACM Trans Database Syst 1980, 5:1–17.
© 2014 John Wiley & Sons, Ltd.
405
wires.wiley.com/widm
Overview
44. DeWitt DJ, Ghandeharizadeh S, Schneider DA, Bricker
A, Hsiao H-I, Rasmussen R. The gamma database
machine project. IEEE Trans Knowl Data Eng 1990,
2:44–62.
60. Kim Y, Shim K. Parallel top-k similarity join algorithms using mapreduce. In: 28th International
Conference on Data Engineering (ICDE), Arlington,
VA, 2012, 510–521.
45. Chang F, Dean J, Ghemawat S, Hsieh WC, Wallach
DA, Burrows M, Chandra T, Fikes A, Gruber RE.
BigTable: a distributed storage system for structured
data. ACM Trans Comput Syst 2008, 26:1–26.
61. Okcan A, Riedewald M. Processing theta-joins using
mapreduce. In: Sellis TK, Miller RJ, Kementsietsidis A,
Velegrakis Y, eds. SIGMOD Conference. New York,
NY: ACM; 2011, 949–960.
46. Shafer J, Agrawal R, Mehta M. Sprint: a scalable parallel classifier for data mining. In: 22th International
Conference on Very Large Data Bases (VLDB ’96),
Mumbai, 1996, 544–555.
62. DeCandia G, Hastorun D, Jampani M, Kakulapati G,
Pilchin A, Sivasubramanian S, Vosshall P, Vogels W.
Dynamo: Amazon’s highly available key-value store.
SIGOPS Oper Syst Rev 2007, 41:205–220.
47. Snir M, Otto S. MPI-The Complete Reference: The
MPI Core. Cambridge, MA: MIT Press; 1998.
48. The Apache Software Foundation. Hadoop, an open
source implementing of mapreduce and GFS, 2012
49. Creeger M. Cloud computing: an overview. ACM
Queue 2009, 7:2.
50. Vaquero LM, Rodero-Merino L, Caceres J, Lindner
M. A break in the clouds: towards a cloud definition.
SIGCOMM Comput Commun Rev 2008, 39:50–55.
51. Shim K, Cha SK, Chen L, Han W-S, Srivastava
D, Tanaka K, Yu H, Zhou X. Data management
challenges and opportunities in cloud computing.
In: 17th International Conference on Database Systems for Advanced Applications (DASFAA’2012).
Berlin/Heidelberg: Springer; 2012, 323.
52. Kambatla K, Kollias G, Kumar V, Grama A.
Trends in big data analytics. J Parallel Distrib 2014,
74:2561–2573.
53. White T. Hadoop: The Definitive Guide. 1st ed.
Sebastopol, CA: O’Reilly; 2009.
54. Ghemawat S, Gobioff H, Leung S-T. The google file
system. In: 19th Symposium on Operating Systems
Principles, 2003, Bolton Landing, NY, 29–43.
55. Murty J. Programming Amazon Web Services—S3,
EC2, SQS, FPS, and SimpleDB: Outsource Your
Infrastructure. Ist ed. Sebastopol, CA: O’Reilly; 2008.
56. Chaiken R, Jenkins B, Larson P, Ramsey B, Shakib
D, Weaver S, Zhou J. SCOPE: easy and efficient
parallel processing of massive data sets. PVLDB 2008,
1:1265–1276.
57. Grossman RL, Gu Y. Data mining using high performance data clouds: experimental studies using sector and sphere. In: 14th ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining
(KDD), 2008, Las Vegas, NV, 920–927.
58. Grossman RL, Gu Y, Sabala M, Zhang W. Compute and storage clouds using wide area high performance networks. Future Gener Comput Syst 2009, 25:
179–183.
59. Härder T, Reuter A. Principles of transaction-oriented
database recovery. ACM Comput Surv 1983, 15:
287–317.
406
63. Karger D, Lehman E, Leighton T, Panigrahy R, Levine
M, Lewin D.. Consistent hashing and random trees:
distributed caching protocols for relieving hot spots on
the world wide web. In: 29th Annual ACM Symposium
on Theory of Computing (STOC), New York, NY,
1997, 654–663.
64. Dimiduk N, Khurana A. HBase in Action. 1st ed.
Greenwich, Connecticut (USA): Manning; 2012.
65. Lakshman A, Malik P. Cassandra: a decentralized
structured storage system. Oper Syst Rev 2010,
44:35–40.
66. Russell J, Cohn R. Hypertable. 1st ed. Key Biscayne,
FL, (USA): Bookvika publishing; 2012.
67. Chodorow K, Dirolf M. MongoDB: The Definitive
Guide Powerful and Scalable Data Storage. 1st ed.
Sebastopol, CA: O’Reilly; 2010.
68. Severance C. Discovering javascript object notation.
IEEE Comput 2012, 45:6–8.
69. Anderson JC, Lehnardt J, Slater N. CouchDB: The
Definitive Guide Time to Relax. 1st ed. Sebastopol,
CA: O’Reilly; 2010.
70. Thusoo A, Sarma JS, Jain N, Shao Z, Chakka P,
Anthony S, Liu H, Wyckoff P, Murthy R. Hive—a
warehousing solution over a map-reduce framework.
VLDB J 2009, 2:1626–1629.
71. Olston C, Reed B, Srivastava U, Kumar R, Tomkins A.
Pig latin: a not-so-foreign language for data processing. In: 2008 ACM SIGMOD International Conference on Management of data (SIGMOD), Vancouver,
Canada, 2008, 1099–1110.
72. Gates A, Natkovich O, Chopra S, Kamath P,
Narayanam S, Olston C, Reed B, Srinivasan S, Srivastava U. Building a highlevel dataflow system on top
of mapreduce: the pig experience. PVLDB 2009, 2:
1414–1425.
73. Beyer KS, Ercegovac V, Gemulla R, Balmin A,
Eltabakh MY, Kanne C-C, Özcan F, Shekita EJ. Jaql:
a scripting language for large scale semistructured
data analysis. VLDB J 2011, 4:1272–1283.
74. Melnik S, Gubarev A, Long J, Romer G, Shivakumar
S, Tolton M, Vassilakis T. Dremel: interactive analysis
of web-scale datasets. PVLDB 2010, 3:330–339.
75. Hausenblas M, Nadeau J. Apache drill: interactive
ad-hoc analysis at scale. Big Data 2013, 1:100–104.
© 2014 John Wiley & Sons, Ltd.
Volume 4, September/October 2014
WIREs Data Mining and Knowledge Discovery
Big Data with Cloud Computing
76. Zhou J, Bruno N, Wu M-C, Larson P-K, Chaiken R,
Shakib D. SCOPE: parallel databases meet mapreduce.
VLDB J 2012, 21:611–636.
77. Lin J, Dyer C. Data-Intensive Text Processing with
MapReduce. Synthesis Lectures on Human Language
Technologies. California (USA): Morgan and Claypool
Publishers; 2010.
78. Shim K. Mapreduce algorithms for big data analysis.
PVLDB 2012, 5:2016–2017.
79. Akyildiz I, Su W, Sankarasubramaniam Y, Cayirci E. A
survey on sensor networks. IEEE Commun Mag 2002,
40:102–105.
80. Patcha A, Park J-M. An overview of anomaly detection
techniques: existing solutions and latest technological
trends. Comput Netw 2007, 51:3448–3470.
81. Elo L, Schwikowski B. Mining proteomic data for
biomedical research. WIREs Data Mining Knowl Discov 2012, 2:1–13.
82. Glaab E, Bacardit J, Garibaldi JM, Krasnogor N. Using
rule-based machine learning for candidate disease gene
prioritization and sample classification of cancer gene
expression data. PLoS One 2012, 7:e39932.
83. Swan AL, Mobasheri A, Allaway D, Liddell S, Bacardit
J. Application of machine learning to proteomics
data: classification and biomarker identification in
postgenomics biology. Omics 2013, 17:595–610.
84. Cappello F, Geist A, Gropp B, Kalé LV, Kramer B,
Snir M. Toward exascale resilience. Int J High Perform
Comput Appl 2009, 23:374–388.
85. Jin H, Chen Y, Zhu H, Sun X-H. Optimizing hpc
fault-tolerant environment: an analytical approach. In:
39th International Conference on Parallel Processing
(ICPP), San Diego, CA, 2010, 525–534.
86. Jin H, Sun X-H. Performance comparison under failures of mpi and mapreduce: an analytical approach.
Future Gener Comput Syst 2013, 29:1808–1815.
87. Tu T, Rendleman CA, Borhani DW, Dror RO,
Gullingsrud J, Jensen M, Klepeis JL, Maragakis P,
Miller PJ, Stafford KA, Shaw DE. A scalable parallel
framework for analyzing terascale molecular dynamics simulation trajectories. In: ACM/IEEE Conference
on Supercomputing (SC), Austin, TX, 2008, 1–12.
92. Duda RO, Stork DG, Hart PE. Pattern Classification.
New York, NY; Chichester: John Wiley & Sons; 2000.
93. Hastie T, Tibshirani R, Friedman J. The Elements
of Statistical Learning: Data Mining, Inference and
Prediction. 2nd ed. New York, NY; Berlin/Heidelberg:
Springer; 2009.
94. He Q, Du C, Wang Q, Zhuang F, Shi Z. A parallel
incremental extreme SVM classifier. Neurocomputing
2011, 74:2532–2540.
95. López V, del Río S, Benítez JM, Herrera F.
Cost-sensitive linguistic fuzzy rule based classification
systems under the mapreduce framework for imbalanced big data. Fuzzy Set Syst in press. doi: 10.1016/
j.fss.2014.01.015.
96. He Q, Wang H, Zhuang F, Shang T, Shi Z. Parallel sampling from big data with uncertainty distribution. Fuzzy Set Syst in press. doi: 10.1016/
j.fss.2014.01.016:1–14.
97. Zhang J, Rui Li T, Ruan D, Gao Z, Zhao C. A parallel
method for computing rough set approximations. Inf
Sci 2012, 194:209–223.
98. Palit I, Reddy CK. Scalable and parallel boosting
with mapreduce. IEEE Trans Knowl Data Eng 2012,
24:1904–1916.
99. Río S, López V, Benítez J, Herrera F. On the use
of mapreduce for imbalanced big data using random
forest. Inf Sci in press. doi: 10.1016/j.ins.2014.03.043.
100. Bahga A, Madisetti VK. Analyzing massive machine
maintenance data in a computing cloud. IEEE Trans
Parallel Distrib Syst 2012, 23:1831–1843.
101. Magoulès F, Zhao H-X, Elizondo D. Development of
an rdp neural network for building energy consumption fault detection and diagnosis. Energy Build 2013,
62:133–138.
102. Wang P, Wang J, Chen Y, Ni G. Rapid processing
of remote sensing images based on cloud computing.
Future Gener Comput Syst 2013, 29:1963–1968.
103. Schatz MC. Cloudburst: highly sensitive read mapping with mapreduce. Bioinformatics 2009, 25:
1363–1369.
88. Plimpton SJ, Devine KD. Mapreduce in mpi for
large-scale graph algorithms. Parallel Comput 2011,
37:610–632.
104. Schönherr S, Forer L, Weißensteiner H, Kronenberg F,
Specht G, Kloss-Brandstätter A. Cloudgene: a graphical execution platform for mapreduce programs on
private and public clouds. BMC Bioinformatics 2012,
13:200.
89. Srinivasan A, Faruquie TA, Joshi S. Data and task
parallelism in ilp using mapreduce. Mach Learn 2012,
86:141–168.
105. Taylor RC. An overview of the hadoop/mapreduce/
hbase framework and its current applications in bioinformatics. BMC Bioinformatics 2010, 11:S1.
90. Dean J, Ghemawat S. MapReduce: a flexible data
processing tool. Commun ACM 2010, 53:72–77.
106. Wang Y, Goh W, Wong L, Montana G. Random
forests on hadoop for genome-wide association studies
of multivariate neuroimaging phenotypes. BMC Bioinformatics 2013, 14:S6.
91. Pavlo A, Paulson E, Rasin A, Abadi DJ, DeWitt
DJ, Madden S, Stonebraker M.. A comparison of
approaches to large-scale data analysis. In: 2009 ACM
SIGMOD International Conference on Management
of data, Providence, RI, 2009, 165–178.
Volume 4, September/October 2014
107. Han J, Cheng H, Xin D, Yan X. Frequent pattern
mining: current status and future directions. Data
Mining Knowl Discov 2007, 15:55–86.
© 2014 John Wiley & Sons, Ltd.
407
wires.wiley.com/widm
Overview
108. Zhang C, Zhang S. Association Rule Mining: Models
and Algorithms. Berlin/Heidelberg: Springer; 2002.
Conference on Knowledge Discovery and Data Mining, 2011, San Diego, CA, 334–342.
109. Woo J. Market basket analysis algorithms with
mapreduce. WIREs Data Min Knowl Discov 2013,
3:445–452.
125. Ghoting A, Krishnamurthy R, Pednault E, Reinwald B,
Sindhwani V, Tatikonda S, Tian Y, Vaithyanathan S.
Systemml: declarative machine learning on mapreduce.
In: 27th International Conference on Data Engineering (ICDE), 2011, Washington, DC, 231–242.
110. Qiao S, Li T, Peng J, Qiu J. Parallel sequential pattern
mining of massive trajectory data. Int J Comput Intell
Syst 2010, 3:343–356.
111. Jain AK. Data clustering: 50 years beyond k-means.
Pattern Recognit Lett 2010, 31:651–666.
112. Xu R, Wunsch D. Clustering. 1st ed. Hoboken, New
Jersey (USA): Wiley-IEEE Press; 2009.
113. Aljarah I, Ludwig SA. Parallel particle swarm optimization clustering algorithm based on mapreduce
methodology. In: 4th World Congress Nature and Biologically Inspired Computing (NaBIC), Mexico City,
Mexico, 2012, 104–111.
114. Kim Y, Shim K, Kim M-S, Sup Lee J. Dbcure-mr: an
efficient density-based clustering algorithm for large
data using mapreduce. Inf Syst 2014, 42:15–35.
115. Yang Y, Rutayisire T, Lin C, Li T, Teng F. An improved
cop-kmeans clustering for solving constraint violation
based on mapreduce framework. Fundam Inf 2013,
126:301–318.
116. Ricci F, Rokach L, Shapira B, Kantor PB, eds.
Recommender Systems Handbook. Berlin/Heidelberg:
Springer; 2011.
117. Schafer JB, Frankowski D, Herlocker J, Sen S.
Collaborative filtering recommender systems. In:
Brusilovsky P, Kobsa A, Nejdl W, eds. The Adaptive Web. Berlin/Heidelberg: Springer-Verlag; 2007,
291–324.
118. Kim Y, Shim K. Twilite: a recommendation system
for twitter using a probabilistic model based on latent
dirichlet allocation. Inf Syst 2013, 42:59–77.
119. Lai C-F, Chang J-H, Hu C-C, Huang Y-M, Chao H-C.
Cprs: a cloud-based program recommendation system
for digital TV platforms. Future Gener Comput Syst
2011, 27:823–835.
120. Zigkolis C, Karagiannidis S, Koumarelas I, Vakali A.
Integrating similarity and dissimilarity notions in recommenders. Expert Syst Appl 2013, 40:5132–5147.
121. Owen S, Anil R, Dunning T, Friedman E. Mahout in
Action. 1st ed. Manning; 2011.
122. Zaharia M, Chowdhury M, Franklin MJ, Shenker S,
Stoica I. Spark: cluster computing with working sets.
In: HotCloud, Boston, MA 2010, 1–7.
123. Ericson K, Pallickara S. On the performance of
high dimensional data clustering and classification algorithms. Future Gener Comput Syst 2013,
29:1024–1034.
124. Ghoting A, Kambadur
Nimble: a toolkit for
allel data mining and
on mapreduce. In 17th
408
P, Pednault EPD, Kannan R.
the implementation of parmachine learning algorithms
ACM SIGKDD International
126. Das S, Sismanis Y, Beyer KS, Gemulla R, Haas PJ,
McPherson J. Ricardo: integrating R and Hadoop.
In: 2010 International Conference on Management of
data (SIGMOD), Indianapolis, IN, 2010, 987–998.
127. Guha S, Hafen R, Rounds J, Xia J, Li J, Xi B, Cleveland
WS. Large complex data: divide and recombine (D&R)
with RHIPE. Stat 2012, 1:53–67.
128. Guha S, Hafen RP, Kidwell P, Cleveland W. Visualization databases for the analysis of large complex
datasets. J Mach Learn Res 2009, 5:193–200.
129. Wegener D, Mock M, Adranale D, Wrobel S.
Toolkit-based high-performance data mining of
large data on mapreduce clusters. In: IEEE International Conference on Data Mining Workshops,
Miami, FL, 2009, 296–301.
130. Agneeswaran V. Big Data Analytics Beyond Hadoop:
Real-Time Applications with Storm, Spark, and More
Hadoop Alternatives. 1st ed. London (U.K.): Pearson;
2014.
131. Wise L. Using Open Source Platforms for Business
Intelligence. Amsterdam: Morgan Kaufmann; 2012.
132. Amdahl GM. Validity of the single-processor approach
to achieving large scale computing capabilities. In:
Sprint Joint Computer Conference of the American Federation of Information Processing Societies,
Atlantic City, NJ, 1967, 483–485.
133. Lin J. Mapreduce is good enough? Big Data 2013,
1:BD28–BD37.
134. Kajdanowicz T, Kazienko P, Indyk W. Parallel processing of large graphs. Future Gener Comput Syst
2014, 32:324–337.
135. Srirama SN, Jakovits P, Vainikko E. Adapting scientific computing problems to clouds using mapreduce.
Future Gener Comput Syst 2012, 28:184–192.
136. Chen Y, Ferguson A, Martin B, Wang A, Wendell P.
Big data and internships at cloudera. ACM Crossroads
2012, 19:35–37.
137. Zhang J, Wong J-S, Li T, Pan Y. A comparison of
parallel large-scale knowledge acquisition using rough
set theory on different mapreduce runtime systems. Int
J Approx Reason 2014, 55:896–907.
138. Isard M, Budiu M, Yu Y, Birrell A, Fetterly D.. Dryad:
distributed data-parallel programs from sequential
building blocks. In: 2nd ACM SIGOPS/EuroSys European Conference on Computer Systems (EuroSys),
New York, NY, 2007, 59–72.
139. Yu Y, Isard M, Fetterly D, Budiu M, Erlingsson
L, Gunda PK, Currey J. Dryadlinq: a system for
© 2014 John Wiley & Sons, Ltd.
Volume 4, September/October 2014
WIREs Data Mining and Knowledge Discovery
Big Data with Cloud Computing
general-purpose distributed data-parallel computing
using a high-level language. In: Operating Systems
Design and Implementation, San Diego, CA, 2008,
1–14.
140. Bu Y, Howe B, Balazinska M, Ernst MD. The haloop
approach to large-scale iterative data analysis. PVLDB
2012, 21:169–190.
141. Ekanayake J, Li H, Zhang B, Gunarathne T, Bae
S-H, Qiu J, Fox G. Twister: a runtime for iterative
mapreduce. In: ACM International Symposium on
High Performance Distributed Computing (HPDC),
2010, New York, NY, 810–818.
142. Zaharia M, Chowdhury M, Das T, Dave A, Ma
J, McCauley M, Franklin MJ, Shenker S, Stoica
I. Resilient distributed datasets: a fault-tolerant
abstraction for in-memory cluster computing. In:
9th USENIX Conference on Networked Systems
Design and Implementation, San Jose, CA, 2012,
1–14.
143. Kraska T, Talwalkar A, Duchi J, Griffith R, Franklin
M, Jordan M. Mlbase: a distributed machine learning
system. In: Conference on Innovative Data Systems
Research, Asilomar, CA, 2013, 1–7.
144. Valiant LG. A bridging model for parallel computation. Commun ACM 1990, 33:103–111.
145. Malewicz G, Austern MH, Bik AJ, Dehnert JC, Horn
I, Leiser N, Czajkowski G.. Pregel: a system for largescale graph processing. In: 2010 ACM SIGMOD
International Conference on Management of data,
Indianapolis, IN, 2010, 135–146.
146. Martella C, Shaposhnik R, Logothetis D. Giraph in
Action. 1st ed. Manning; 2014.
147. Xin RS, Gonzalez JE, Franklin MJ, Stoica I. Graphx:
a resilient distributed graph system on spark. In:
GRADES’13, New York, NY, 2013, 1–6.
148. Low Y, Kyrola A, Bickson D, Gonzalez J, Guestrin
C, Hellerstein JM. Distributed graphlab: a framework
for machine learning in the cloud. PVLDB 2012,
5:716–727.
151. Zaharia M, Das T, Li H, Hunter T, Shenker S,
Stoica I. Discretized streams: fault-tolerant streaming
computation at scale. In: 24th ACM Symposium on
Operating Systems Principles (SOSP), Farmington, PA,
2013, 423–438.
152. Ranger C, Raghuraman R, Penmetsa A, Bradski G,
Kozyrakis C. Evaluating mapreduce for multi-core and
multiprocessor systems. In: IEEE 13th International
Symposium on High Performance Computer Architecture (HPCA), Phoenix, AZ, 2007, 1–6.
153. Fang W, He B, Luo Q, Govindaraju N. Mars: accelerating mapreduce with graphics processors. IEEE Trans
Parallel Distrib Syst 2011, 22:608–620.
154. Chu C-T, Kim SK, Lin Y-A, Yu Y, Bradski GR, Ng
AY, Olukotun K.. Map-reduce for machine learning
on multicore. In: Advances in Neural Information
Processing Systems 19, Twentieth Annual Conference
on Neural Information Processing Systems (NIPS),
Vancouver, Canada, 2006, 281–288.
155. Ailamaki A, Govindaraju NK, Harizopoulos S,
Manocha D. Query co-processing on commodity
processors. In: 32nd International Conference on
Very Large Data Bases (VLDB), Seoul, Korea, 2006,
1267–1267.
156. Stuart JA, Owens JD. Multi-gpu mapreduce on gpu
clusters. In: 25th IEEE International Parallel and Distributed Processing Symposium (IPDPS), Anchorage,
AK, 2011, 1068–1079.
157. Basaran C, Kang K-D. Grex: an efficient mapreduce
framework for graphics processing units. J Parallel
Distrib Comput 2013, 73:522–533.
158. Pan Y, Zhang J. Parallel programming on cloud computing platforms—challenges and solutions. J Convergence 2012, 3:23–28.
159. Schmidt K, Phillips C. Programming Elastic MapReduce. Ist ed. Sebastopol, CA: O’Reilly; 2013.
160. Nadipalli R. HDInsight Essentials. Ist ed. Birmingham: PACKT Publishing; 2013.
149. Anderson Q. Storm Real-Time Processing Cookbook.
1st ed. Sebastopol, CA: O’Reilly; 2013.
161. Jiang D, Ooi B, Shi L, Wu S. The performance
of mapreduce: an in-depth study. PVLDB 2010,
3:472–483.
150. Neumeyer L, Robbins B, Nair A, Kesari A. S4: distributed stream computing platform. In: 2010 IEEE
Data Mining Workshops (ICDMW), Sydney, Australia, 2010, 170–177.
162. Stonebraker M, Abadi D, DeWitt D, Madden S,
Paulson E, Pavlo A, Rasin A. Mapreduce and parallel dbmss: friends or foes? Commun ACM 2010,
53:64–71.
Volume 4, September/October 2014
© 2014 John Wiley & Sons, Ltd.
409
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement