A-Tree: Distributed Indexing of Multidimensional Data For Cloud Computing Environments

2011 THIRD IEEE INTERNATIONAL CONFERENCE ON COUD COMPUTING TECHNOLOGY AND SCIENCE
A-Tree: Distributed Indexing of Multidimensional

Data for Cloud Computing Environments
Andreas Papadopoulos Dimitrios Katsaros
Department of Computer Communications Engineering
University of Thessaly, Greece
[email protected], [email protected]
AbstractAbstractEfficient querying of huge volumes of multidimensional data stored in cloud computing systems has become
a necessity, due to the widespread of cloud storage facilities. With
clouds getting larger and available data growing larger and larger
it is mandatory to develop fast, scalable and efficient indexing
schemes. In this paper, we propose the A-tree, a distributed
indexing scheme for multidimensional data capable of handling
both point and range queries, appropriate for cloud computing
environments. A performance evaluation of the A-tree against
the state-of-the-art competitor attests its superiority, achieving
significantly lower latencies.
Distributed index, multidimensional data, point query, range

query, query processing, cloud computing
I. I NTRODUCTION
Cloud computing and data centers are facing unprecedented
challenges due to huge amount of data and number of users
that must be handled. Several thousands of computers, terabytes of data and several millions of users comprise a typical
cloud computing system, offered as SaaS, PaaS, or IaaS [1].
Every user allocates resources for his needs on demand from
the infinite cloud, and pays only for what it was really used.
Users and companies must as well consider the privacy of their
data and cost for the services that will be used and select the
most appropriate solution for their needs [2]. The amount of
stored data and the rate of querying them, calls for new data
structures which can satisfy the needs of a cloud system. The
majority of current cloud storage systems e.g., Googles GFS
and BigTable [3], Hadoops HDFS, and Amazons DYNAMO
are based on key-value pairs, and therefore they can only
support point queries. Though this type of query is not rich
enough to fulfill the needs of cloud users; more complex
queries such as range queries are needed. Answering this
type of queries becomes more complicated since the queried
data are multidimensional in nature and spread among several
cloud nodes. Even though the past literature on databases
and distributed systems is full of data structures capable of
dealing with point and range queries for multidimensional
data, the cloud environments poses new challenges that make
these solutions inappropriate. First of all, cloud systems are
distributed over wide areas even across different countries
with a (usual) two-level hierarchy consisting of master and
slave nodes and therefore the centralized database solutions
are not an option.
978-0-7695-4622-3/11
26.002011IEEEDOI10.1109/CloudCom.2011.61
Secondly, proposals that are based on peer-to-peer overlay

structures, such as CAN [4], P-Grid [5], BATON [6], are
similarly not very efficient since they posses one or more of the
following drawbacks: they do not support multidimensional
data, or they require time-consuming, communication-hungry
and careful balancing operations, or they do not differentiate among nodes. Finally, the recently proposed cloud-aware
distributed structures, such as the EEMINC [14], incur high
latencies. In order to design a high performance distributed
index for cloud environments, we must use a cost and space
efficient indexing scheme capable of answering queries with
low latency. Specifically, this article makes the following
contributions: A new distributed indexing structure for cloud
computing environments, the A-tree1, is described, which is
capable of answering both point and range queries. It is based
on the combination of R-tree [16] and Bloom filters [17].
We describe algorithms that distributed the index nodes to
the cloud nodes, as well as well as the relevant insertion and
deletion algorithms. A performance evaluation of the proposed structure against the competing state-of-the-art structure
is conducted, which attest the superiority of the new structure.
The rest of the article is organized as follows: Section II
describes the relevant work; in Section III, we introduce
a request framework to describe the basic concepts of this
work. Section IV and V provide the details of the local and
global data structure that comprise the A-tree. Specifically, we
describe the update strategy, how R-tree nodes are selected
for being indexed in the global index, how updates are built
and sent, and we also describe the process of update handling
and the construction of global index. Section VI provides
the experimental evaluation of the A-tree and its comparison
against the state-of-the-art EEMINC. Our experimentations
have shown that A-tree is a scalable, distributed, fast and
space efficient index for multidimensional data supporting
querying, insertion and deletion of records. Finally, Section
VII concludes the present article.
II. R ELATED W ORK
Firstly, we need to describe the A-trees constituent structures, namely Bloom filter and R-tree, and then the relevant
cloud indexing structures. Bloom filter is a bit array representing a set of items. For each item in a data set the hash
values of independent hash functions are calculated and for
every value the corresponding bit is set to 1. Using Bloom
2011 THIRD IEEE INTERNATIONAL CONFERENCE ON COUD COMPUTING TECHNOLOGY AND SCIENCE
fig1.png
Fig. 1. Example of a Bloom filter.
filter a point query can be answered in O(1) with a very small

probability of false positives. Figure 1 demonstrates a simple
Bloom filter with size ten bits and two hash functions are
used: modulo three and modulo ten. Initially all values are
zero and the diagram shows which bits turn on due to the
insertion of elements 57, 83 and 94. Querying whether element
38 belongs to the set, we calculate both hash functions and
we notice that the eighth bit is not turned on, which means
that 38 does not belong to the set. R-tree is a widely used data
structure for multidimensional data. R-tree is the extension of
B-tree to multidimensional data. Each node covers an area
in the multidimensional space, usually represented as a hyper
bounding box. For non-leaf nodes, this bounding box covers
all of its childrens bounding boxes. Leaf nodes contain pointers
to the data. A lot of varieties have been studied including R+tree and R*-tree with most of them trying to minimize overlap
between nodes coverage. Figure 2 demonstrates a simple example of an R-tree in two dimensional space using rectangles
to handle space partitions. R-tree is extensively being used
on many indexing systems including geographical information
systems to provide spatial index operations in the cloud [7]. A
solution to provide query services over peer-to-peer networks
was the Distributed Hash Table (DHT) and some variances [8].
Another relevant data structure is a scalable distributed B-tree
[9]. With these data structures though, we can only answer
point queries and not range queries over multidimensional
data. As the Map-Reduce framework is widely used on cloud
systems, a decent index structure was proposed [10]. Although
this solution is claimed to be suitable for large datasets,
the Map-Reduce framework itself adds a significant overhead
during processing, which makes it inappropriate for extremely
huge datasets. In addition efforts to support standard SQL
statements with high performance global index have been
made [11]. Other relevant indices include the RT-CAN [12]
and the BR-tree [13]. Although these structures can answer
both point and range queries, their drawback is that

A-Tree: Distributed Indexing of Multidimensional Data For Cloud Computing Environments

Uploaded by

Document Informationclick to expand document information

Copyright:

Available Formats

A-Tree: Distributed Indexing of Multidimensional Data For Cloud Computing Environments

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A-Tree: Distributed Indexing of Multidimensional Data For Cloud Computing Environments

Uploaded by

Copyright:

Available Formats

2011 THIRD IEEE INTERNATIONAL CONFERENCE ON COUD COMPUTING TECHNOLOGY AND SCIENCE

A-Tree: Distributed Indexing of Multidimensional

Distributed index, multidimensional data, point query, range

Secondly, proposals that are based on peer-to-peer overlay

filter a point query can be answered in O(1) with a very small

You might also like