Analytics over Large-Scale Multidimensional Data:

The Big Data Revolution!

Alfredo Cuzzocrea

Il-Yeol Song

Karen C. Davis

ICAR-CNR and University of Calabria

Rende, Cosenza, Italy

Drexel University
Philadelphia, PA, USA

University of Cincinnati
Cincinnati, OH, USA

[email protected]

[email protected]

[email protected]

proper databases on remote (data) servers. Due to the nave features

of big data, DaaS is the most appropriate computational data
framework to implement big data repositories [2]. MapReduce [8] is
a relevant realization of the DaaS initiative. IaaS is a provision
model according to which organizations outsource infrastructures
(i.e., hardware, software, network) used to support ICT operations.
The IaaS provider is responsible for housing, running and
maintaining these services, by ensuring important capabilities like
elasticity, pay-per-use, transfer of risk and low time to market. Due
to specific application requirements of applications running over big
data repositories, IaaS is the most appropriate computational service
framework to implement big data applications.

ABSTRACT
In this paper, we provide an overview of state-of-the-art research
issues and achievements in the field of analytics over big data, and
we extend the discussion to analytics over big multidimensional
data as well, by highlighting open problems and actual research
trends. Our analytical contribution is finally completed by several
novel research directions arising in this field, which plays a leading
role in next-generation Data Warehousing and OLAP research.

Categories and Subject Descriptors

H.2 [Database Management]: H.2.7 Database Administration Data
Warehouse and Repository

Among the collection of open problems and research challenges

deriving from the latest big data revolution, analytics over big data
play a relevant role in the context of Data Warehousing and OLAP
research. Let us focus on this research challenge in a greater detail.
Analytics can be intended as complex procedures running over
large-scale, enormous-in-size data repositories (like big data
repositories) whose main goal is that of extracting useful knowledge
kept in such repositories. Two main problems arise, in this respect.
The first one is represented by the issue of conveying big data stored
in heterogeneous and different-in-nature data sources (e.g., legacy
systems, Web, scientific data repositories, sensor and stream
databases, social networks) into a structured, hence wellinterpretable, format. The second one is represented by the issue of
managing, processing and transforming so-extracted structured data
repositories in order to derive Business Intelligence (BI) components
like diagrams, plots, dashboards, and so forth, for decision making
purposes. Actually, both these aspects are of emerging interest for a
wide spectrum of research communities, and more properly for the
Data Warehousing and OLAP research community. As a
consequence, this has generated a rich literature. At the industrial
research side, Hadoop [3] and Hive [13] are two fortunate
implementations of the ETL layer and the BI layer of big data
applications, respectively.

General Terms
Algorithms, Design, Management, Performance, Theory

Keywords
Analytics over Big Data, Analytics over Big Multidimensional Data,
Data Warehousing, OLAP

1. INTRODUCTION
Big Data refers to enormous amounts of unstructured data
produced by high-performance applications falling in a wide and
heterogeneous family of application scenarios: from scientific
computing applications to social networks, from e-government
applications to medical information systems, and so forth. Data
stored in the underlying layer of all these application scenarios have
some specific characteristics in common, among which we recall: (i)
large-scale data, which refers to the size and the distribution of data
repositories; (ii) scalability issues, which refers to the capabilities of
applications running on large-scale, enormous data repositories (i.e.,
big data, for short) to scale over growing-in-size inputs rapidly; (iii)
supporting advanced Extraction-Transformation-Loading (ETL)
processes from low-level, raw data to somewhat structured
information; (iv) designing and developing easy and interpretable
analytics over big data repositories in order to derive intelligence
and extract useful knowledge from them.

Although analytics over large-scale data repositories have been

deeply investigated recently, the problem of extending actual
models and algorithms proposed in this respect to the specific big
multidimensional data context plays a leading role, as
multidimensional data naturally marry with analytics [6].

Cloud computing is a successful computational paradigm for

managing and processing big data repositories, mainly because of its
innovative metaphors known under the terms Database as a
Service (DaaS) [10] and Infrastructure as a Service (IaaS). DaaS
defines a set of tools that provide final users with seamless
mechanisms for creating, storing, accessing and managing their

Inspired by this main motivation, in this paper we provide an

overview of state-of-the-art research issues and achievements in the
field of analytics over big data, and we extend the discussion to
analytics over big multidimensional data as well, by highlighting
open problems and actual research trends, and drawing novel
research directions in this field.

Copyright is held by the author/owner(s).

DOLAP11, October 28, 2011, Glasgow, Scotland, UK.
ACM 978-1-4503-0963-9/11/10.

101

(a) Data Source Heterogeneity and Incongruence. Very often,

data sources storing data of interest for the target analytics processes
(e.g., legacy systems, Web, scientific data repositories, sensor and
stream databases, social networks, and so forth) are strongly
heterogeneous and incongruent. This aspect not only conveys in
typical integration problems, mainly coming from active literature
on data and schema integration issues, but it also has deep
consequences on the kind of analytics to be designed.

2. ANALYTICS OVER BIG DATA: STATEOF-THE-ART

Analytics over big data repositories has recently received a great
deal of attention from the research communities. One of the most
significant application scenarios where big data arise is, without
doubt, scientific computing. Here, scientists and researchers produce
huge amounts of data per-day via experiments (e.g., think of
disciplines like high-energy physics, astronomy, biology, biomedicine, and so forth) but extracting useful knowledge for decision
making purposes from these massive, large-scale data repositories is
almost impossible for actual DBMS-inspired analysis tools.

(b) Filtering-Out Uncorrelated Data. Due to the enormous size of

big data repositories, dealing with large amount of data that are
uncorrelated to the kind of analytics to be designed occurs very
frequently. As a consequence, filtering-out uncorrelated data plays a
critical role in the context of analytics over big data, as this heavily
affects the quality of final analytics to be designed.

In response to this computational emergency, the Hadoop system

[3] has been proposed. Hadoop runs MapReduce [8] tasks over big
data, and also it makes available the Hadoop Distributed File
System (HDFS) [ 3] for supporting file-oriented, distributed data
management operations efficiently. It has been highlighted that
Hadoop is a kind of MAD system [7] meaning that (i) it is capable of
attracting all data sources (M standing for Magnetism), (ii) it is
capable of adapting its engines to evolutions that may occur in big
data sources (A standing for Agility), (iii) it is capable of supporting
depth analytics over big data sources much more beyond the
possibilities of traditional SQL-based analysis tools (D standing for
Depth). In a sense, Hadoop can be reasonably considered as the
evolution of next-generation Data Warehousing systems, with
particular regards to the ETL phase of such systems. MapReduce is
the core of Hadoop. MapReduce is a programming model with the
associated computational framework that is inspired to the
primitives Map and Reduce of functional languages. Basically, Map
partitions computational tasks into smaller computational tasks (this
involves in the partition of the target data domain as well) and
assigns to then appropriate Key, Value pairs. These smaller
computational tasks are executed very efficiently, even by
exploiting parallelism. The final result of the overall computational
task (i.e., the output, processed big data) is obtained via a Reduce
operation that combines all the values sharing the same Key value.

(c) Strongly Unstructured Nature of Data Sources. In order to

design meaningful analytics, it is mandatory that input big data
sources are transformed in a suitable, structured format, and finally
stored in the HDFS. This poses several issues that recall classical
ETL processes of Data Warehousing systems, but with the
additional challenges that data alimenting big data repositories are
strongly unstructured (e.g., social network data, biological
experiment result data, and so forth) in contrast with less
problematic unstructured data that are popular in traditional BI tools
(e.g., XML data, RDF data, and so forth). Again, here
transformations from unstructured to structured format should be
performed on the basis of the analytics to be designed, according to
a sort of goal-oriented methodology.
(d) High Scalability. High scalability of big data analytics is one of
the primer feature to be ensured for a MAD-inspired big data
analytics system. To this end, exploiting the cloud computing
computational framework seems to be the most promising way to
this end [2]. The usage of the IaaS-inspired cloud computing
computational framework is meant with the aim of achieving some
important characteristics of highly-scalable big data analytics
systems, among which we recall: (i) true scalability, i.e. the
effective scalability that a powerful computational infrastructure like
clouds is capable of ensuring; (ii) elasticity, i.e. the property of
rapidly adapting to massive updates and fast evolutions of big data
repositories; (iii) fault-tolerance, i.e. the capability of being robust
to faults that can affect the underlying distributed
data/computational architecture; (iv) self-manageability, i.e. the
property of automatically adapting the framework configuration
(e.g., actual load balancing) to rapid changes of the surrounding
data/computational environment; (v) execution on commodity
machines, i.e. the capability of scale-out on thousands and
thousands of commodity machines when data/computational peaks
occur.

Several studies, like [9,12], have provided recommendations for

further improving the computational capabilities of Hadoop,
whereas [1] proposes HadoopDB, a novel hybrid architecture that
combines MapReduce and traditional DBMS technologies for
supporting advanced analytics over large-scale data repositories.
Furthermore, Starfish [11] is a recent self-tuning system for
supporting big data analytics that is still based on Hadoop but it
incorporates special features trying to achieve higher performance
by means of adaptive metaphors.
By looking at BI aspects of analytics over big data, the state-of-theart research result is represented by Hive [13], a BI system/tool for
querying and managing structured data built on top of the Hadoops
HDFS. Hive which allows us to obtain the final analytics
components (in the form of diagrams, plots, dashboards, and so
forth) from the big data processed, materialized and stored via
Hadoop. Also, Hive introduces a SQL-like query language, called
HiveQL [13], which runs MapReduce jobs immersed into SQL
statements.

(e) Combining the Benefits of RDBMS and NoSQL Database

Systems. One of the more relevant features to be achieved by big
data analytics systems is represented by flexibility, which refers to
the property of covering a large collection of analytics scenarios
over the same big data partition. In order to obtain this critical
feature, it is necessary to combine the benefits of traditional
RDBMS database systems and those of next-generation NoSQL
database systems, which propose representing and managing data
via horizontal data partitions by renouncing to fixed table schemas
and, consequentially, resource-expensive join operations [4].

3. OPEN PROBLEMS AND ACTUAL

RESEARCH TRENDS OF BIG DATA
ANALYTICS
There are a number of open problems and actual research trends
related to big data analytics. In the following, we provide an
overview on some of the most significant of them.

(f) Query Optimization Issues in HiveQL. Several open issues

arise with respect to query optimization aspects of HiveQL. Among

102

these, noticeable ones are the following: (i) moving towards more
expressive, complex aggregations, e.g. OLAP-like rather than SQLlike, hence enforcing the User Defined Function (UDF) and the
User Defined Aggregate Function (UDAF) [5] paradigms; (ii)
covering advanced SQL statements such as nested queries and
order-by predicates; (iii) incorporating data compression paradigms
in order to achieve higher performance; (iv) devising novel costbased optimizations, e.g. based on table or column statistics; (v)
integration with third-part BI tools.

(4) How to Design Complex Analytics over Hadoop-Integrated

Multidimensional Data? Multidimensional data provide add-on
value to big data analytics. In this respect, design complex analytics
over Hadoop-integrated multidimensional data plays a critical role.
Actual analytics, although quite well-developed, still do not go
beyond classical BI components, like diagrams, plots, dashboards,
and so forth, but complex BI processes of very large organizations
demand for more advanced BI-oriented decision support tools,
perhaps by integrating principles and results of different-in-nature
disciplines like statistics.

4. TURNING INTO ANALYTICS ON BIG

MULTIDIMENSIONAL DATA

(5) How To Deal with Visualization Issues Arising From Big

Multidimensional Data Analytics? Visualization issues represent a
leading problem in Data Warehousing and OLAP research. These
issues get worse when re-visited in the context of big
multidimensional data analytics, as here visualization must kept a
stronger decision-support value. More complex techniques, such as
multidimensional space exploration approaches, must be
investigated to this end.

It has been already studied that multidimensional data naturally

marry with analytics [6]. Indeed, analytics significantly extend
typical OLAP operators (e.g., roll-up, drill-down, and so forth)
hence it is natural to think of multidimensional data as an add-on
value for analytics models and methodologies. Multidimensional
data finally allow us to enhance the expressive power and the
capabilities of analytics, and actual research experiences in the
context of big data analytics are mature enough to launch a novel
paradigm for Data Warehousing and OLAP research: analytics over
big multidimensional data. Basically, this innovative paradigm aims
at integrating the classical, well-known benefits of multidimensional
data models (such as multidimensional abstractions, hierarchy-based
dimensional tables, multi-resolution fact tables, multi-way
aggregations, OLAP tools, and so forth) with analytics, in order to
achieve more powerful analytics capable of enhancing actual
models [13,11] by means of typical amenities deriving from such
multidimensional data models. In order to realize this sort of
second-generation big data revolution, it is necessary to face-off a
number of open research problems, some of which can be
summarized by the following questions.

5. CONCLUSIONS
Starting from state-of-the-art research issues and achievements in
analytics over big data, in this paper we have provided critical
discussion over open research issues and achievements arising in
this scientific field, and we have extended the discussion to the
emerging context of analytics over big multidimensional data. Open
problems and actual research trends have been highlighted, and
novel research directions have been proposed.

6. REFERENCES
[1]

(1) How To Build Multidimensional Structures On Top Of The

HDFS? This problem refers to the issue of building
multidimensional data structures on top of the structured HDFS
repositories of Haadop, as a first step towards directly integrating
multidimensional data models with analytics over big data. A
promising direction to this end consists in exploiting array-based inmemory representation methods.

[2]

[3]
[4]
[5]

(2) How To Directly Integrate Multidimensional Data Sources

Into The Hadoop Lifecycle? Hadoop populates the underlying
structured big data repositories from heterogeneous and different-innature data sources, such as legacy systems, Web, scientific data
sets, sensor and stream databases, social networks, and so forth.
Despite this, no research efforts have been devoted to the yetrelevant issue of directly integrating multidimensional data sources
into the Hadoop lifecycle, which is an exciting research challenge
for next-generation Data Warehousing and OLAP research.

[6]
[7]
[8]
[9]
[10]

(3) How To Model and Design Multidimensional Extensions of

HiveQL? In order to achieve an effective integration of
multidimensional data models with analytics over big data, the
query language HiveQL must be enriched with multidimensional
extensions as well. These extensions should take into consideration
language syntax aspects as well as query optimization and
evaluation aspects, perhaps by inheriting lessons learned in the
context of actual MDX-like languages for multidimensional data.

[11]
[12]
[13]

103

Abouzeid, A., Bajda-Pawlikowski, K., Abadi, D.J.. Rasin, A., and

Silberschatz, A. HadoopDB: An Architectural Hybrid of MapReduce
and DBMS Technologies for Analytical Workloads. PVLDB 2(1),
2009.
Agrawal, D., Das, D., and El Abbadi, A. Big Data and Cloud
Computing: Current State and Future Opportunities. Proc. of EDBT,
2011.
Apache Hadoop. http://wiki.apache.org/hadoop
Cattell, R. Scalable SQL and NoSQL Data Stores. SIGMOD Record
39(4), 2010.
Chen, Q., Hsu, M., and Liu, R. Extend UDF Technology for Integrated
Analytics. Proc. of DaWaK, 2009.
Chen, Z., and Ordonez, C. Efficient OLAP with UDFs. Proc. of
DOLAP, 2008.
Cohen, J., Dolan, B., Dunlap, M., Hellerstein, J.M., and Welton, C.
MAD Skills: New Analysis Practices for Big Data. PVLDB 2(2), 2009.
Dean, J., and Ghemawat, S. MapReduce: Simplified Data Processing on
Large Clusters. Communications of the ACM 51(1), 2008.
Dittrich, J., Quian-Ruiz, J.-A., Jindal, A., Kargin, Y., Setty, V., and
Schad, J. Hadoop++: Making a Yellow Elephant Run Like a Cheetah.
PVLDB 3(1), 2010.
Hacigumus, H., Iyer, B., and Mehrotra, S. Providing Database as a
Service. Proc. of ICDE, 2002.
Herodotou, H., Lim, H., Luo, G., Borisov, N., Dong, L., Cetin, F.B.,
and Babu, S. Starfish: A Self-Tuning System for Big Data Analytics.
Proc. of CIDIR, 2011.
Jiang, D., Ooi, B.C., Shi, L., and Wu, S. The Performance of
MapReduce: An In-depth Study. PVLDB 3(1), 2010.
Thusoo, A. Sarma, J.S., Jain, N., Shao, Z., Chakka, P. Zhang, N.,
Antony, S., Liu, H., and Murthy, R. Hive A Petabyte Scale Data
Warehouse Using Hadoop. Proc. of ICDE, 2010.