Download as pptx, pdf, or txt
Download as pptx, pdf, or txt
You are on page 1of 10

WOLAITA SODO UNIVERSITY SCHOOL OF INFORMATICS

AND COMPUTING
WEEKEND PROGRAM (MSC IN COMPUTER SCIENCE)
PARALLEL PROCESSING ARTICLE REVIEW ON
(The Structure of Parallel Algorithms)
 
Reviewed by
Lewi Thomas
Afework Milkias
Gebre Worku

Submitted to: - Mr. Temesgen (Ass.Professor)


 
June/14/2023
Wolaita Sodo, Ethiopia
June 2023
School of Informatics and Computing Wolaita Sodo University
Wolaita Sodo, Ethiopia

Article Review on the Structure of Parallel Algorithms


H. T. KUNG Lewi Thomas, Afework Milkias, Gebre Worku
Department of Computer Science Wolaita Sodo University School of Informatics and
Carnegie-Mellon University Computing Department of Computer Science
Pittsburgh. Pennsylvania

ABSTRACT
The article provides an It aims to provide a
overview of parallel algorithms comprehensive understanding of
and their classifications based the basic issues and techniques
on three important attributes: in designing parallel algorithms
for different architectures.
 concurrency control
The abstract concludes by
 module granularity and stating that the article aims to be
 communication geometry a useful resource for readers
The article presents various interested in understanding the
examples of parallel algorithms basics of parallel algorithm
and studies them under a design and choosing the
appropriate parallel architecture
uniform framework.
Overview Of
.
The Article
Introduction
Parallel algorithms have been studied 1.Concurrency Control:
since the 1960s, but the emergence of concurrency control ensures the
large-scale parallel computers has correctness of concurrent execution in
increased interest in designing parallel parallel algorithms.
algorithms.
2. Module Granularity:
The article presents examples of parallel
Module granularity refers to the amount of
algorithms and aims to be a useful
resource for understanding the basic computation a task module can perform
issues and techniques in designing before needing to communicate with other
parallel algorithms. A parallel algorithm modules.
is a collection of independent task 3. Communication Geometry
modules that can be executed in parallel
Suppose that task modules of a parallel
and communicate with each other. The
algorithm are connected to represent
three orthogonal dimensions of the space
of parallel algorithms are concurrency intermodule communication. Then a
control, module granularity, and geometric layout of the resulting network
communication geometry is referred to as the communication
geometry of the algorithm.
Matching Parallel Algorithms with Parallel Architectures
The article discusses the matching With a MIMD machine such as the
along the dimensions of concurrency processors have independent instruction
control and module granularity. There counters, and operate asynchronously on
are three architectures and their shared memories.SIMD machines correspond
matching algorithms. to synchronous lock-step algorithms that
1.SIMD and MIMD Machines and require central controls, whereas MIMD
Algorithms machines correspond to asynchronous
The notion of single-instruction algorithms with relatively large granularities.
stream multiple-data stream (SIMD) 2 Systolic Machines and Algorithms
and multiple-instruction stream systolic machine is a network of simple and
multiple data stream (MIMD) is often primitive processors that circulate data in a
used for classifying parallel regular fashion. Each processor regularly
computers in MIMD One stream of pumps data in and out, each time performing
instructions issued by the central some short computation, so that a regular
control unit controls all the flow of data is kept up in the network
processors, each operating upon its
own memory synchronously
A Taxonomy for Parallel Algorithms
The article proposes taxonomy Algorithms Using One Dimensional
for parallel algorithms based on Linear Arrays
the three dimensions mentioned One-dimensional linear arrays are simple
earlier: concurrency control, and fundamental for connecting processors.
module granularity, and Four Examples of algorithms are discussed
communication geometry. in this article.
 Odd-Even Transposition Sort.
Algorithms for Synchronous
 Real-Time Finite Impulse Response
Parallel Computers (FIR) Filtering
This section focuses on parallel  Priority Queue
algorithms for synchronous  Recurrence Evaluation (Recursive
parallel computers, including Filtering)
systolic and SIMD machines. Following the discussions of the four
algorithms,we mention the use of linear
Algorithms will be classified
pipelines in the implementation of
based on their communication arithmetic operations.
geometries.
Algorithms Using Two-Dimensional Arrays
Two-dimensional
communication geometries that Algorithms Using Tree
are simple and regular are Structures
considered. In the article Tree structures are discussed
Algorithms using two- for chip implementation, and
dimensional hexagonal arrays
their advantages for
are presented, some of this are:
broadcasting, searching, and
 matrix multiplication
fan-in operations are
 LU decomposition
explained. Search algorithms
 transitive closure of a and systolic search trees are
Boolean matrix. highlighted as applications of
tree structures
Algorithms for Asynchronous
Multiprocessors
Concurrent Data Base Systems The major issues are the
In a concurrent data base system, correctness and efficiency of an
algorithm in then presence of the
a number of on-line transactions
asynchronous behavior of its
are allowed to run concurrently on processes. For the general data
a shared data base. base environment where only
Algorithms for Specific syntactic information can be
Problems used, the
When designing an asynchronous serialization approach is the
algorithm for solving a specific method for ensuring correctness.
problem, we have control over the Serialization can be achieved by
either locking or transaction
tasks that will be included in the
backup.
algorithm.
Analysis of the Article
Here is a breakdown of the main points  Classification of parallel algorithms
covered in each section: based on these dimensions.
 Introduction to parallel algorithms and  Correspondence between the
their historical development. Increasing classification of algorithms and parallel
interest in parallel algorithms due to the architectures
emergence of large-scale parallel  Proposal of a taxonomy for parallel
computers. algorithms in the systolic, SIMD,
MIMD communication geometries
 Mention of surveys and references on space.
parallel architectures and algorithms.  Algorithms for Synchronous Parallel
 The impact of large-scale integration Computers:
technology on parallel algorithms.  Introduction to synchronous parallel
 The Space of Parallel Algorithms: computers, including systolic and
Taxonomy and Relation to Parallel SIMD machines.
Architectures:  Classification of algorithms based on
communication geometries.
 Three important dimensions of parallel
 Examples of algorithms using one-
algorithms:concurrency control,module
dimensional linear arrays and two-
granularity, and communication dimensional arrays (hexagonal arrays
geometry. and trees).
Main Gaps of the Article
 The author not well reported Scalability  Portability: Parallel algorithms
issues: Many parallel algorithms suffer developed for one architecture may
from scalability issues when the size of not be portable to other
data or number of processors increases architectures due to differences in
beyond certain limits. There is a need for hardware and software.
new algorithmic techniques and CONCLUSUION
strategies to improve scalability. In general the paper presents
 The author not described Load taxonomy for parallel algorithms
balancing in which the work load is based on their concurrency control,
distributed evenly across all processors. module granularity, and
This is particularly challenging in communication geometry. It provides
dynamic systems where the workload examples of algorithms for different
may change over time. communication geometries, including
 The author not clarified Faults and error linear arrays, hexagonal arrays, and
that occur in parallel systems, and it is tree structures. The paper serves as a
important to develop algorithms that are valuable resource for understanding
resilient to such errors and can continue the design principles and techniques
to function correctly even in the of parallel algorithms for various
presence of faults. architectures.
THANK YOU!!

You might also like