Using Big Data Smarter Decision Making

Using Big Data for
Smarter Decision Making
Colin White, BI Research
July 2011
Sponsored by IBM

USING BIG DATA FOR SMARTER DECISION MAKING
“To increase competitiveness, 83% of CIOs have visionary plans that include
business intelligence and analytics.”1

“Global digital content created will increase some 30 times over the next ten
years – to 35 zettabytes).”2

Becoming an analytics-driven organization helps companies reduce costs, increase
revenues and improve competitiveness, and this is why business intelligence and
analytics continue to be a top priority for CIOs. Many business decisions, however,
are still not based on analytics, and CIOs are looking for ways to reduce time to value
for deploying business intelligence solutions so that they can expand the use of
analytics to a larger audience of users.

Big data sources are Companies are also interested in leveraging the value of information in so-called big
largely untapped by data systems that handle data ranging from high-volume event data to social media
business intelligence textual data. This information is largely untapped by existing business intelligence
applications systems, but organizations are beginning to recognize the value of extending the
business intelligence and data warehousing environment to integrate, manage, govern
and analyze this information.

This paper looks at new developments in business analytics and discusses the
benefits analyzing big data bring to the business. It also examines different types of
big data workloads and offers suggestions on how to optimize systems to handle
different workloads and integrate them into a single infrastructure for making smarter
and faster business decisions.

WHAT IS BIG DATA?
Big data is a popular buzzword, but it is important to realize that this data comes in
many shapes and sizes. It also has many different uses – real-time fraud detection,
web display advertising and competitive analysis, call center optimization, social
media and sentiment analysis, intelligent traffic management and smart power grids,
to name just a few. All of these analytical solutions involve significant (and growing)
volumes of both multi-structured3 and structured data.

Most of these analytical solutions were not possible previously because they were too
costly to implement, or because analytical processing technologies were not capable
of handling the large volumes of data involved in a timely manner. In some cases, the
required data simply did not exist in an electronic form.

1
“The Essential CIO.” IBM CIO study involving 3,018 CIOs spanning 71 countries
and 18 industries, 2011.
2
“Digital Universe Study.” IDC study sponsored by EMC, May 2010.
3
This type of data has unknown, ill-formed, or overlapping schemas. The term
unstructured is also used to describe this data, but this is misleading because most of
the data in this category does have some structure.

Copyright  2011 BI Research, All Rights Reserved. 1

Using Big Data for Smarter Decision Making

New technologies New and evolving analytical processing technologies now make possible what was
enable the analysis not possible before. Examples include:
of big data
• New systems that handle a wide variety of data from sensor data to web and
social media data.
• Improved analytical capabilities (sometimes called advanced analytics)
including event, predictive and text analytics.
• Operational business intelligence that improves business agility by enabling
automated real-time actions and intraday decision making.
• Faster hardware ranging from faster multi-core processors and large memory
spaces, to solid-state drives and virtual data storage for handling hot and cold
data.
• Cloud computing including on-demand software-as-a-service (SaaS) analytical
solutions in public clouds, and data platforms and virtualization in private clouds.

Supporting big data involves combining these technologies to enable new solutions
that can bring significant benefits to the business.

Analyzing big data Big data involves more than simply the ability to handle large volumes of data.
involves extreme Instead, it represents a wide range of new analytical technologies and business
workloads possibilities. The challenge is how to deploy these technologies and manage the
many extreme analytical processing workloads involved, while at the same time
providing faster time to value.

THE CHALLENGES OF BIG DATA AND EXTREME WORKLOADS
An enterprise data warehouse can be used to handle big data and extreme workloads,
but often it is more efficient to preprocess the data before loading it into the
warehouse. Event data from hardware sensors, for example, has more business value
if it is filtered and aggregated before using it for analysis. Usually the raw event data
is not required for historical purposes, and storage and processing costs are reduced if
it is not kept in the warehouse.

Big data and extreme The best way of handling big data and supporting the extreme processing involved is
workloads require to deploy optimized hardware and software solutions for processing different types of
optimized hardware big data workloads, and then combine these solutions with the existing enterprise
and software data warehouse to create an integrated information supply chain (see Figure 1). The
objectives of an information supply chain are to consume and integrate the many
varieties of raw source data that exist in organizations, analyze that data, and then
deliver the analytical results to business users. This information supply chain enables
the data component of what IBM calls Smarter Computing.

Supporting extreme workloads is not a new challenge for the computing industry.
Business transaction processing systems have supported extreme transaction
workloads ever since the advent of data processing. As new business needs arose,
organizations employed custom and optimized transaction processing systems to
handle application workloads that pushed the boundaries beyond what could be
handled by more generalized technologies. Airline reservation systems, bank ATM
systems, retail point-of-sale terminals, financial trading systems and mobile phone



systems, are all examples of these types of applications. More recently, applications
for handling sensor networks that track items with RFID tags can be added to the list.

Figure 1.
The information
supply chain

In recent years there has been a similar trend toward analytical processing reaching
the performance limitations of more generalized systems. Today, multi-terabyte data
warehouses are no longer the exception and analytical processing workloads are
becoming increasingly more complex. The result is that once again organizations
require optimized systems to meet the challenges of extreme workloads. The industry
response has been to offer packaged hardware and software solutions that are
optimized for analytical processing.

Satisfying business agility requirements is another important factor in supporting
analytical processing. In today’s fast paced business environment, organizations need
to make faster decisions, and this agility is important to business success. In the case
of fraud detection, for example, real time action is required. Not all business
decisions have to be made in real time, but for many organizations the ability to act in
a few seconds or minutes, rather than hours or days, can be a significant financial and
competitive advantage.

MANAGING AND ANALYZING BIG DATA
Data volume and The main challenges of big data and extreme workloads are data variety and volume,
variety, and workload and analytical workload complexity and agility. Each of these can be viewed from
complexity and the perspective of the information supply chain shown in Figure 1 above.
agility are the
challenges of big Input to the information supply chain consists of the raw source data required for
data analysis. For the past two decades most business analytics have been created using
structured data extracted from operational systems and consolidated into a data
warehouse. Big data dramatically increases both the number of data sources and the
variety and volume of data that is useful for analysis. A high percentage of this data
is often described as multi-structured to distinguish it from the structured operational
data used to populate a data warehouse. In most organizations, multi-structured data
is growing at a considerably faster rate than structured data.

There are two main techniques for analyzing big data – the store and analyze
approach, and the analyze and store approach.



Store and Analyze Approach
Traditional data The store and analyze approach integrates source data into a consolidated data store
warehousing is used before it is analyzed. This approach is used by a traditional data warehousing system
to produce data to create data analytics. In a data warehousing system, the consolidated data store is
analytics usually an enterprise data warehouse or data mart managed by a relational or
multidimensional DBMS. The advantages of this approach are improved data
integration and data quality management, plus the ability to maintain historical
information. The disadvantages are additional data storage requirements and the
latency introduced by the data integration task.

Two important big data trends for supporting the store and analyze approach are
relational DBMS products optimized for analytical workloads (often called analytic
RDBMSs, or ADBMSs) and non-relational systems (sometimes called NoSQL
systems) for processing multi-structured data. A non-relational system can be used to
produce analytics from big data, or to preprocess big data before it is consolidated
into a data warehouse. Certain vendors in the search and content management
marketplaces also use the store and analyze approach to create analytics from index
and content data stores.

Analytic RDBMSs (ADBMSs)
Analytic RDBMSs An analytic RDBMS is an integrated solution for managing data and generating
improve price analytics that offers improved price/performance, simpler management and
performance for administration, and time to value superior to more generalized RDBMS offerings.
creating data Performance improvements are achieved through the use of massively parallel
analytics processing, enhanced data structures, data compression, and the ability to push
analytical processing into the DBMS.

ADBMSs can be categorized into three broad groups:4 packaged hardware and
software appliances, software-only platforms, and cloud-based solutions.

Packaged hardware and software appliances fall into two sub-groups: purpose-
built appliances and optimized hardware/software platforms. The objective in both
cases is to provide an integrated package that can be installed and maintained as a
single system. Depending on the vendor, the dividing line between the two sub-
groups is not always clear, and this is why in this article they are both categorized as
appliances.

A purpose-built appliance is an integrated system built from the ground up to provide
good price/performance for analytical workloads. This type of appliance enables the
complete configuration, from the application workload to the storage system used to
manage the data, to be optimized for analytical processing. It also allows the solution
provider to deliver customized tools for installing, managing and administering the
integrated hardware and software system.

Many of these products were developed initially by small vendors and targeted at
specific high-volume business area projects that are independent of the enterprise
data warehouse. As these appliances have matured and added workload management

4
Some vendors simply call all of these appliances, but this is oversimplified and
confusing. Although the three groups overlap, each of them has its own unique
architecture, strengths and weaknesses, and analytical use cases.



capabilities, their use has expanded to handle mixed workloads and in some cases
support smaller enterprise data warehouses. Large and well-established vendors have
acquired several of these solutions. An example is the IBM Netezza TwinFin
appliance.5

The success of these purpose-built appliances led to more traditional RDBMS
vendors building packaged offerings by combining existing products. This involved
improving the analytical processing capabilities of the software and then building
integrated and optimized hardware and software solutions. These solutions consist of
optimized hardware/software platforms designed for specific analytical workloads.
The level of integration and optimization achieved varies by vendor. In some cases,
the vendor may offer a choice of hardware platform. An example of this type of
approach is the IBM Smart Analytics System.

IBM Smart Analytics The IBM Smart Analytics System is intended for extending the performance and
System and IBM scalability of an enterprise data warehouse, whereas the IBM Netezza TwinFin
Netezza are examples appliance is used to maintain a separate data store in situations where it is
of optimized unnecessary, impractical, or simply not cost effective to integrate the data into an
systems enterprise data warehouse for processing. The IBM Netezza TwinFin is also used to
analyze data extracted from an enterprise data warehouse, either to offload an
extreme processing workload, or to create a sandbox for experimental use and/or
complex ad hoc analysis.

A software-only platform is a set of integrated software components for handling
analytical workloads. These platforms often make use of underlying open source
software products and are designed for deployment on low-cost commodity
hardware. The tradeoff for hardware portability is the inability of the product to
exploit the performance and management capabilities of a specific hardware
platform. Some software platforms are available as virtual images, which are useful
for evaluation and development purposes, and also for use in cloud-based
environments.

Cloud-based solutions offer a set of services for supporting data warehousing and
analytical application processing. Some of these services are offered on public
clouds, while others can be used in-house in private cloud environments. The
underlying software and hardware environment for these cloud-based services may
be custom built, employ a packaged hardware and software appliance, or use the
capabilities of a software-only platform. The role of cloud computing for business
intelligence and data warehousing is discussed in more detail later in this paper.

Non-Relational Systems
Non-relational A single database model or technology cannot satisfy the needs of every organization
systems are used to or workload. Despite its success and universal adoption, this is also true for RDBMS
process multi- technology. This is especially true when processing large amounts of multi-structured
structured data data and this is why several organizations with big data problems have developed
their own non-relational systems to deal with extreme data volumes. Web-focused
companies such as Google and Yahoo that have significant volumes of web
information to index and analyze are examples of organizations that have built their

5
IBM acquired Netezza Corporation in September, 2010.



own optimized solutions. Several of these companies have placed these systems into
the public domain so that they can be made available as open source software.

Non-relational systems are useful for processing big data where most of the data is
multi-structured. They are particularly popular with developers who prefer to use a
procedural programming language, rather than a declarative language such as SQL,
to process data.6 These systems support several different types of data structures
including document data, graphical information, and key-value pairs.

Hadoop with One leading non-relational system is the Hadoop distributed processing system from
MapReduce is one of the open source Apache Software Foundation. Apache defines Hadoop as “a
the leading non- framework for running applications on a large hardware cluster built of commodity
relational systems hardware.” This framework includes a distributed file system (HDFS) that can
distribute and manage huge volumes of data across the nodes of a hardware cluster to
provide high data throughput. Hadoop uses the MapReduce programming model to
divide application processing into small fragments of work that can be executed on
multiple nodes of the cluster to provide massively parallel processing. Hadoop also
includes the Pig and Hive languages for developing and generating MapReduce
programs. Hive includes HiveQL, which provides a subset of SQL.

Hadoop MapReduce is intended for the batch processing of large volumes of multi-
structured data. It is not suitable for low-latency data processing, many small files, or
the random updating of data. These latter capabilities are provided by database
products such as HBase and Cassandra that run on top of Hadoop.

Several companies offer commercialized open source or open core versions of
Hadoop for handling big data projects. IBM’s InfoSphere BigInsights product fits
into this category.

Which DBMS To Use When?
Organizations will It is important to realize that generalized relational DBMSs, analytic RDBMSs, and
use both analytic non-relational data systems are not mutually exclusive. Each approach has its
RDBMSs and non- benefits, and it is likely that most organizations will employ some combination of all
relational systems three of them.

When choosing a system for handling big data and extreme processing, several
factors have to be considered: the volume of data, the variety of data, and the
complexity and agility requirements of the analytical workloads. Figure 2 positions
the different approaches with respect to these factors. To provide an integrated
analytical infrastructure these approaches must coexist and interoperate with each
other. This is why vendors such as IBM are delivering connectors that allow data to
flow between the different systems.

The database technologies shown in Figure 2 are capable of supporting big data from
the perspective of data volume, data variety and analytical workload complexity. To
provide business agility, data must move along the information supply chain at a pace
that matches the action time requirements of the business.

6
The term NoSQL is frequently used to describe these systems, but this misleading
because some of them support a subset of SQL. A non-relational system is a better
term to use.



Figure 2.
The three dimensions
Analytical workload
of big data complexity and agility

most of the data
is structured

Data volume

ADBMS

co-existence

most of the data
Universal is multi-structured
RDBMS Non-
Relational
System

Data variety

Action time is best explained in terms of the information supply chain (see Figure 1)
where the input is raw source data and the output is business analytics. When
processing data for business decision making the data has to be collected for analysis,
analyzed, and the results delivered to the business user. The user then retrieves the
results and decides if any action is required. There is always a time delay, or latency,
between a business event occurring and the time the business user acts to resolve an
issue or satisfy a requirement. This action time varies from application to application
based on business needs. For a fraudulent credit card transaction it is the time taken
from the credit card being used to it being rejected for fraud.

Reducing action time helps make organizations more agile. The business benefits
obtained, however, must be balanced against the IT costs of supporting that level of
service. In general, the closer the action time gets to real-time the higher the costs of
supporting that action time.

Agility can be To reduce action time in the store and analyze approach, data has to be gathered and
improved by analyzed faster, and business users need to make faster decisions and take faster
analyzing data as it actions. In some cases this may not be possible for technology or cost reasons.
flows through Another way of reducing action time is to analyze the data as it flows through
systems operational systems, rather than integrating it into a data store before analyzing it.
This not only reduces action time, but also reduces storage, administration and
security resources requirements. This can be thought of as an analyze and store
approach. This approach analyzes data in motion, whereas the store and analyze
approach analyzes data at rest.



Analyze and Store Approach
The analyze and store approach analyzes data as it flows through business
processes, across networks, and between systems. The analytical results can then be
published to interactive dashboards and/or published into a data store (such as a data
warehouse) for user access, historical reporting and additional analysis. This
approach can also be used to filter and aggregate big data before it is brought into a
data warehouse.

There are two main ways of implementing the analyze and store approach:

Process analytics • Embedding the analytical processing in business processes. This technique
can be created by works well when implementing business process management and service-
embedding analytical oriented technologies because the analytical processing can be called as a service
processing in from the process workflow. IBM supports this style of processing in its
business processes WebSphere product set. This technique is particularly useful for monitoring and
analyzing business processes and activities in close to real-time – action times of
a few seconds or minutes are possible here. The process analytics created can
also be published to an operational dashboard or stored in a data warehouse for
subsequent use.
Stream analytics can • Analyzing streaming data as it flows across networks and between systems.
be created by This technique is used to analyze data from a variety of different (possibly
analyzing data as it unrelated) data sources where the volumes are too high for the store and analyze
flows through approach, sub-second action times are required, and/or where there is a need to
networks and across analyze the data streams for patterns and relationships. To date, many vendors
systems have focused on analyzing event streams (from trading systems, for example)
using the services of a complex event processing (CEP) engine, but this style of
processing is evolving to support a wider variety of streaming technologies and
data. IBM’s InfoSphere Streams product, for example, supports a stream-
processing engine that creates stream analytics from many types of streaming
data such as event, video and GPS data.

The benefits of the analyze and store approach are fast action times and lower data
storage overheads because the raw data does not have to be gathered and
consolidated before it can be analyzed.

Figure 3 shows how systems for analyzing in-motion data can be combined with the
existing business intelligence and data warehousing environment and data coming
from non-relational systems such as Hadoop.

THE ROLE OF CLOUD COMPUTING
Cloud computing, like several of the other technologies discussed in this paper,
covers a wide range of capabilities. One important and growing use of cloud
computing is to deploy an integrated on-premises computing environment, or private
cloud. This allows the many different workloads involved in both operational and
analytical processing to co-exist and be consolidated into a single system. Some of
these workloads may run as virtualized machines, whereas others may run as native
systems.



A private cloud can The private cloud approach provides a consolidated, but flexible system that can
be used to scale to support multiple workloads. This solution, however, is not suited to all types
consolidate multiple of analytical processing. Not only does it take longer to implement a private cloud
workloads than a standalone appliance, but also it is often difficult to tune and optimize a cloud
environment consisting of operational workloads as well as complex analytical
workloads. These latter workloads are often best suited to an appliance approach,
rather than a cloud-computing environment. It is therefore important to match
business requirements and performance needs to the right type of technology solution
and budget.

Figure 3.
Integrating stream
and process stream Stream processing stream
data engine analytics
analytics into the
traditional data filtered data
warehousing
environment
business Business processes with process
transactions embedded analytical services analytics

data changes analytics
operational
data data
Data integration &
management warehouse
other
data
analytics & filtered data analytics

Non-relational Business intelligence data
applications applications analytics

BIG DATA STORAGE CONSIDERATIONS
Several new Many organizations are struggling to deal with increasing data volumes, and big data
technologies help simply makes the problem worse. To solve this problem, organizations need to
companies maximize reduce the amount of data being stored and exploit new storage technologies that
storage utilization improve performance and storage utilization. From a big data perspective there are
three important directions here:
• Reducing data storage requirements using data compression and new physical
storage structures such as columnar storage. These technologies may be
implemented in hardware and/or software. IBM, for example, has real-time
hardware data compression on several of its disk systems that compress data
before storage and decompress it on retrieval. This approach offloads the
overheads of data compression from application processing systems. Note also
that several non-relational products support data compression and columnar
storage to reduce the overheads of storing big data.
• Improving input/output (I/O) performance using solid-state drives (SSDs).
SSDs are especially useful for mixed analytical and enterprise data warehouse



workloads that involve large amounts of random data access. Sequential
processing workloads gain less performance benefit from the use of SSDs.
• Increasing storage utilization by using tiered storage to store data on different
types of devices based on usage. Frequently used hot data can be managed on
fast devices such as SSDs, and less frequently accessed cold data can be
maintained on slower and larger capacity hard disk drives (HDDs). This
approach enables the storage utilization of the HDDs to be increased since they
don’t have to be underutilized to gain performance. System software moves the
data between different storage types based on usage. The location of the data is
transparent to applications. The use of SSDs and tiered storage can significantly
improve both performance and throughput.
All of these technologies can be used to improve the price/performance of data
storage in big data environments, and as organizations plan for big data they need to
reevaluate their storage strategies to take advantage of these developments.

CONCLUSIONS
Big data adds several new high-volume data sources to the information supply chain.
Several new and enhanced data management and data analysis approaches help the
management of big data and the creation of analytics from that data. The actual
approach used will depend on the volume of data, the variety of data, the complexity
of the analytical processing workloads involved, and the responsiveness required by
the business. It will also depend on the capabilities provided by vendors for
managing, administering, and governing the enhanced environment. These
capabilities are important selection criteria for product evaluation.

Big data involves An enhanced information supply chain, however, involves more than simply
more than just implementing new technologies. It requires senior management to understand the
technology – senior benefits of smarter and timelier decision making, and what IBM calls Smarter
management needs Computing. It also requires the business to make pragmatic decisions about the
to understand the agility requirements for analyzing data and producing analytics given tight IT
benefits of smarter budgets. The good news is that many of the technologies outlined in this paper not
decision making only support smarter decision making, but also provide faster time to value.

Note that brand and product names mentioned in this paper may be the trademarks
or registered trademarks of their respective owners.

About BI Research
BI Research is a research and consulting company whose goal is to help companies understand and exploit new
developments in business intelligence, data management, and collaborative computing.


Using Big Data Smarter Decision Making

More Related Content

What's hot (16)

Viewers also liked (6)

Similar to Using Big Data Smarter Decision Making (20)

More from IBM India Smarter Computing (20)

Recently uploaded (20)

Using Big Data Smarter Decision Making