Math 627 - Introduction to Parallel Computing
Fall 2006 - Matthias K. Gobbert
Presentations of the Class Projects

Tuesday, December 19, 2006, 01:00 p.m., MP 401

  1. 01:00-01:20
    An Introduction to Parallel Computing with the Message Passing Interface
    Justin Newcomer, CIRC and Department of Mathematics and Statistics, UMBC
    The Message Passing Interface (MPI) is a library specification for message passing which provides a powerful, efficient, and portable way of expressing parallel programs. This talk will introduce the audience to parallel computing using MPI. It will discuss parallel systems and architecture as well as give several examples and applications to demonstrate the message passing philosophy.

  2. 01:25-01:45
    The Matrix-free Solution of the Three-dimensional Poisson Problem on a Beowulf Cluster: Effect of the Domain Cutting Scheme
    Maher D. Salloum, Department of Mechanical Engineering, UMBC
    The Poisson problem is encountered in many engineering and physics applications. This problem is often extremely expensive from the computational point of view to be simulated on a single processor. In this work, the Poisson equation is solved on a three dimensional domain using the matrix-free conjugate gradient method. The simulations are ran on a Beowulf parallel cluster. Using the C programming language, the solution is stored on a one dimensional vector where the nodes are ordered first in the x, y and z directions consecutively. Many domain cutting schemes are investigated. The best performance is obtained either by cutting the domain in the y and z directions or in all x, y and z directions. In both cases, the number of subdomains in the y directions should be bigger to obtain the best performance.

  3. 01:50-02:10
    Implementation of Fast Fourier Transform for Three Dimensional Cubic Matrices
    Alex Szatmary, Department of Mechanical Engineering, UMBC
    Three dimensional fast Fourier transforms are used to calculate Fourier components on discrete representations of functions of three variables; these algorithms are most often used in solving partial differential equations numerically. In this study, a three dimensional fast Fourier transform is implemented to run in parallel on a Beowulf cluster. For cubic matrices with 128 elements in each direction, the maximum speedup obtained was 3.9, while with matrices with 256 elements in each direction, the maximum speedup obtained was 12, using 32 processors in each case. Various methods of splitting the matrix were considered, and timing results demonstrated that splitting the matrix in the direction of the index that changes first is far superior to alternative splits.

  4. 02:15-02:35
    Atomistic Simulation of a Simple Nanostructure
    Kristi Harris, Department of Physics, UMBC
    As the sizes of electronic components and interconnects shrink further into the nanometer domain, the behavior of surface atoms plays an increasingly important role in interconnect failure. I will present a parallel implementation of a simulation of the random thermal motion of the surface atoms on a two-dimensional nanometer-scale wire. Achieving this simulation in parallel involves significant synchronization and communication as compared to the simple computation of the random walk algorithm. The current simulation exhibits decreased performance as compared to the serial implementation, making it useful only when memory limitations otherwise prohibit the modeling of the wire.

  5. 02:40-03:00
    The PETSc Libary on Kali
    Robert Newton, Department of Mathematics and Statistics, UMBC
    In my talk I will describe the PETSc library on Kali. I will give some details of its installation, and demonstrate some of its capabilities using the familiar Poisson problem as an example. We will see that PETSc is a powerful tool that can greatly simplify the parallel programming process.

Copyright © 2001-2006 by Matthias K. Gobbert. All Rights Reserved.
This page version 2.0, December 2006.