A Report

On

Parallel Databases in Database Management

System

by

Chirayu Marathe 123A1062

Nancy Maruthuvar 123A1063

Swarika Maurya 123A1064

T.E. Computer Engineering

A Report submitted for Partial fulfilment of the requirement of Database

Systems & Advanced Database Management
 CERTIFICATE

This is to certify that the work on the project titled "Parallel Databases in Database
Management Systems" has been carried out by the following students, who are
bonafide students of SIES Graduate School of Technology, Mumbai, in partial
fulfilment of the syllabus requirement in the subject, "Advance Database
Management Systems" in the Academic Year 2025-26:

1. Chirayu Marathe 123A1062

2. Nancy Maruthuvar 123A1063

3. Swarika Maurya 123A1064

Project Guide: Amita Suke

 Abstract

A parallel database is a special kind of database system designed to handle very

large amounts of data and process many tasks at the same time. These systems
use several computers, processors, or disks working together, so they can take big
jobs and break them into smaller pieces that run at the same time instead of one
after another. The main techniques that make this possible are data partitioning,
which means splitting the data into parts and storing them on different machines,
and parallel query processing, which lets the system answer questions by using
all its parts at once. There are different ways to connect these computers: shared-
memory systems let all processors use the same memory, which makes sharing
information easy but gets slower as more computers are added. Shared-nothing
systems give each processor its own memory and disk, making it easy to add more
machines and keep the system fast and reliable, even if one part fails. Because of
these features, parallel databases work well for applications that have to sort
through huge amounts of information quickly, such as business analytics,
scientific research, or cloud services. This report explains how parallel databases
use these techniques and design choices to deliver fast and reliable data
management for today’s data-heavy world.

iii
CONTENTS
Section Page

Abstract iii
1. Introduction 1
2. Core Concepts of Parallel Databases 2
2.1 Key Objectives: Performance and Scalability
2.2 Execution Parallelism
3. Techniques 3
3.1 Data Partitioning 3
3.2 Parallel Query Processing 4
3.3 Parallel Database Architecture 6
4. Conclusion 8

References 9

iv
 1. Introduction

In today’s world, the amount of data generated by businesses, scientific research,

social media, and many other fields is growing very fast. Traditional database
systems, which often rely on a single computer or server to store and process data, are
becoming less effective for handling these large data sets. The reason is that these
systems have limits in how fast they can access data and how much data they can
manage before slowing down too much.

To solve these problems, parallel database systems have been developed. These
systems use multiple computers, processors, and storage devices working together as
a team to manage and process data. Instead of processing one task at a time, parallel
databases split large tasks into smaller parts and process many of these parts at the
same time on different machines. This way, they can finish complex queries and data
operations much faster than traditional systems.

Parallel databases are built using special techniques that help divide and manage data
and workloads efficiently. One important technique is data partitioning, which means
breaking large collections of data into smaller pieces that are spread across various
machines. This helps balance the work so that no single machine gets overloaded and
slows the whole system down. Another key technique is parallel query processing,
which allows different parts of a data request or query to be handled at the same time
by different processors, speeding up the overall operation.
The design of parallel databases is also based on different architectures. In a shared-
memory system, many processors share the same memory space, making it easy to
communicate but limiting how many processors can be used effectively because they
compete for the same resources. On the other hand, a shared-nothing system gives
each processor its own private memory and storage. These systems communicate over
a network and can add more machines easily when the data grows. This makes them
highly scalable and fault-tolerant because if one machine fails, others can continue
working without major problems.

Because of their ability to handle large-scale data and complex requests quickly,
parallel database systems are now widely used in many areas. These include big data
analytics, cloud computing, financial systems, and scientific research. Understanding
how parallel databases work and their key techniques is essential for developing and
managing modern data systems that need to deliver fast and reliable results in our
data-driven world.

1
 2. Core Concepts of Parallel Databases

2.1 Key Objectives: Performance and Scalability

The primary objectives driving the design of parallel databases are high
performance and scalability. Performance is measured by a drastic reduction in
query execution time, enabling complex analytical queries to return results in
seconds or minutes rather than hours. Scalability refers to the system's ability to
maintain this performance as the data volume and user load increase. This is
achieved through scale-up (adding power to existing machines) and, more
importantly, scale-out (adding more machines to the cluster), allowing the system
to grow with the workload.

2.2 Execution Parallelism

Execution parallelism is the fundamental principle that enables these objectives. It
is a divide-and-conquer strategy where a large task, such as processing a complex
SQL query, is broken down into smaller, independent sub-tasks. These sub-tasks
are then executed concurrently across the available processors in the system. This
approach ensures that the total work is completed significantly faster than if it were
processed sequentially by a single CPU, effectively turning one slow operation into
many fast ones.

2
 3. Techniques

3.1 Data Partitioning

Data partitioning is a method used in parallel databases to split a large set of data
into smaller, manageable pieces called partitions. Each partition contains a subset of
the total data and is usually stored on a separate machine or processor. This division
helps the system handle big amounts of data more efficiently by spreading the
workload across multiple servers.

Why is Data Partitioning Important?

• Improves Performance: By working with smaller pieces of data, each server
or processor can retrieve and process data faster than if it had to handle
everything alone.
• Supports Scalability: As data grows, new partitions can be added on new
machines. This allows the system to handle more data without slowing down.
• Balances Load: By dividing data, the workload is shared across many nodes,
preventing any single machine from becoming a bottleneck.
• Enhances Availability: If one machine or partition fails, others can continue
working, which improves the overall reliability of the database.
Types of Data Partitioning
1. Horizontal Partitioning (Sharding):
This divides data by rows. For example, in a table of customers, customers
from different regions might be stored in different partitions. Each partition
contains all columns but only some rows. This is useful for distributing very
large tables across many servers and for queries that look up data by region
or category.
2. Vertical Partitioning:
This splits the data by columns. For instance, a customer table might have a
partition with frequently accessed columns like customer ID and name, and
another with less commonly used columns like purchase history or customer
notes. This approach reduces the amount of data read when only certain
columns are needed.
3. Functional Partitioning:
Sometimes data is split based on function or business logic. For example, in
an online store, user login data might be kept separate from product
inventory data, allowing each part to be managed individually according to
its needs.
3
How Data Partitioning Works
Every partition needs a way to decide which data goes where. This is done using
a partition key — a specific column or set of columns used to assign data to a
partition. For example, in horizontal partitioning by region, the key could be the
customer’s region. When a query asks for data, the system looks at the key to know
which partition to search, avoiding unnecessary scanning of all partitions.
Benefits of Data Partitioning
• It allows parallel processing, where multiple partitions work at the same time,
speeding up query responses.
• It supports easier data management by enabling independent updates,
backups, and recovery for each partition.
• It optimizes storage resources by placing frequently accessed partitions on
faster storage devices while using cheaper ones for less critical data.
In summary, data partitioning is a crucial technique that allows parallel databases to
handle large data volumes effectively by breaking them into logical pieces
distributed across multiple machines, improving speed, scalability, and reliability.

3.2 Parallel Query Processing

Parallel query processing is a method used by parallel databases to make queries run
faster by splitting and handling them at the same time on multiple processors or
machines. Instead of answering a database request step by step on a single computer,
the database breaks the request into smaller jobs that can run together. This helps
finish complex queries much more quickly, especially when dealing with large
amounts of data.

How Parallel Query Processing Works

When a user sends a query to a parallel database, the system divides the query into
smaller parts. These parts might be things like reading data from tables, joining
different tables, sorting, or calculating results. Each part runs on a different processor
or machine simultaneously. After all the parts are done, the database combines the
results and sends the final answer back to the user.

Types of Parallelism in Query Processing

There are several ways parallel databases can split tasks among machines:
• Intra-query Parallelism: A single query is divided into many smaller tasks
that are executed at the same time. This technique focuses on speeding up a
single, complex query by allowing its constituent operations to be processed in
4
 parallel. For instance, a query involving a scan, a join, and an aggregation can
have these three operations executed concurrently on different processors, with
data flowing between them in a pipelined fashion. This is crucial for achieving
fast response times for complex analytical queries (OLAP)
• Inter-query Parallelism: Different queries from different users can run
simultaneously on different machines. This allows the system to handle many
requests at once, increasing overall throughput. Inter-query parallelism
involves executing multiple different, independent queries at the same time,
with each query running on its own processor. This does not make any single
query faster, but it dramatically increases the system's overall transaction
throughput—the number of queries it can complete in a given time window.
This is essential for high-concurrency environments like online transaction
processing (OLTP) systems.
• Intra-operator Parallelism: Even within a single step of a query, such as a
table scan or join operation, the work can be divided across processors to speed
up that step. It is also known as partitioned parallelism, this is a specific form
of intra-query parallelism where a single database operation, or "operator," is
parallelized. The most common example is a parallel table scan, where the data
of a single large table, which has been partitioned across multiple disks, is
scanned by multiple processes simultaneously, with each process scanning only
its local partition. The results are then combined (e.g., via a merge operation),
leading to a scan speed that is nearly the sum of the speeds of all individual
disks.

Benefits of Parallel Query Processing

• Faster Response Time: Queries finish sooner because the work is done at the
same time on multiple machines.
• Higher Throughput: The system can handle more requests at once, making it
effective for environments with many users.
• Efficient Resource Use: All available processors and storage devices are used
effectively, preventing bottlenecks.
• Scalability: As data or workload grows, more machines can be added to
continue processing efficiently without slowing down.

Challenges
While parallel query processing greatly improves performance, it also requires careful
management of how tasks are divided, how data is shared between machines, and how
results are combined. Improper task distribution can cause some machines to be
overworked while others are idle. Also, communication between machines can add
some overhead, which must be minimized for best results.

In summary, parallel query processing breaks down database requests to run many
5
parts at the same time over multiple nodes. This approach allows parallel databases
to quickly process huge volumes of data and hundreds of queries efficiently.

3.3 Parallel Database Architectures

Parallel databases use multiple processors and storage devices to handle large and
complex data workloads more efficiently. How these processors and resources are
organized defines the database architecture. The most common designs are Shared-
Memory, Shared-Nothing, Shared-Disk, and sometimes Hierarchical architectures.
Among these, Shared-Memory and Shared-Nothing are widely discussed due to their
distinct approaches and impacts on performance and scalability.

3.3.1 Shared-Memory Architecture

In a Shared-Memory architecture, multiple processors share the same global memory
and disks. This means each CPU can directly read and write to common memory
locations. Processors communicate by accessing shared variables, which simplifies
coordination and programming.
• Advantages:
o Easy to program because all processors access the same memory.
o Fast communication between processors due to shared memory.
o Suitable for small to medium systems with fewer processors.
• Disadvantages:
o Scalability is limited; as more processors are added, they compete to
access the shared memory, which slows performance.
o The single shared memory bus can become a bottleneck.
o Managing cache coherence among processors adds overhead.
Shared-Memory architectures are often used in Symmetric Multi-Processing (SMP)
systems and serve well where moderate parallelism is sufficient.

3.3.2 Shared-Nothing Architecture

In contrast, Shared-Nothing architecture assigns each processor its own private
memory and disk. No memory or disk space is shared between processors directly.
Instead, the processors communicate over a high-speed network.
• Advantages:
o Highly scalable; new nodes (processors plus memory plus disk) can be
added without affecting existing nodes.
o Eliminates contention for shared memory and disks, allowing parallel
tasks to proceed without conflict.
o Fault-tolerant, as failure in one node does not directly impact others.
o Commonly used in massively parallel processing (MPP) systems like
Teradata, Hadoop, and Spark.
• Disadvantages:
6
 o Inter-node communication costs can be higher, as data must be
transferred over the network.
o Programming and coordination are more complex compared to shared-
memory systems.

Shared-Disk Architecture
Shared-Disk systems are a middle ground where all processors share common disk
storage but have their own private memory. They allow independent CPU operations
with shared disk access, but face challenges in maintaining consistency and managing
disk contention.

Hierarchical Architecture
Hierarchical architectures combine the features of shared-memory, shared-disk, and
shared-nothing models to leverage the advantages of each. These systems offer
scalability and large memory but at higher costs and complexity.

Advantages and Disadvantages Table

Architecture Advantages Disadvantages Typical Use

Case
Shared- Simple programming, Scalability limits, Small to medium
Memory fast processor memory bus SMP systems
communication bottleneck
Shared- Excellent scalability, Higher communication Large-scale MPP
Nothing fault tolerance overhead, complex and big data
coordination
Shared-Disk Balanced access to Disk contention, Medium-scale
shared disks, complex consistency distributed
independent CPUs management DBMS
Hierarchical Combines strengths of Complex and Specialized high-
other architectures expensive performance
DBMS

Architecture choice depends on application size, workload type, hardware cost, and
scalability needs. Shared-Nothing architecture is preferred for today’s large,
distributed data systems due to its ability to scale easily and tolerate faults, while
Shared-Memory architectures serve smaller, tightly-coupled systems well.

7
 4. Conclusion

Parallel databases offer a powerful solution for handling the massive amounts of data
generated in today’s digital world. By using multiple processors and storage devices
working together, these systems can process queries and transactions much faster than
traditional single-server databases. The key techniques that enable this efficiency are
data partitioning, which breaks data into manageable pieces spread across different
machines, and parallel query processing, which executes many parts of a query at the
same time.
Choosing the right architecture plays a crucial role in the performance and scalability
of a parallel database system. Shared-memory architectures offer simpler
communication between processors but face limitations in scaling to large numbers
of machines due to resource contention. Shared-nothing architectures, where each
node operates independently with its own memory and disk, provide excellent
scalability and fault tolerance, making them the preferred choice for modern large-
scale data systems like cloud platforms and big data frameworks.
Overall, parallel databases combine these techniques and architectures to achieve high
performance, scalability, and reliability. They form the backbone of high-
performance data management systems used in business analytics, scientific research,
and real-time applications. As data volumes continue to grow, the importance of
parallel database systems will only increase, driving further innovation in distributed
computing and data processing technologies.

8
References
1. Ramakrishnan, R., & Gehrke, J. (2003). Database Management Systems.
McGraw-Hill, 3rd ed.
2. Stonebraker, M. (1986). "The Case for Shared-Nothing." IEEE Database
Eng. Bull., 9(1), 4-9.
3. DeWitt, D., & Gray, J. (1992). "Parallel Database Systems: The Future of
High Performance Database Processing." Communications of the ACM,
36(6), 85-98.
4. Oracle Corporation. (2024). "Oracle Database Data Warehousing Guide:
Parallel Execution." Oracle Documentation.
5. Apache Software Foundation. (2024). "Apache Hadoop
Documentation." hadoop.apache.org.
6. Apache Software Foundation. (2024). "Apache Spark: Cluster Mode
Overview." spark.apache.org.
7. Garcia-Molina, H., Ullman, J. D., & Widom, J. (2008). Database Systems:
The Complete Book. Prentice Hall, 2nd ed.