Math 627 - Introduction to Parallel Computing
Fall 2013 - Matthias K. Gobbert
Presentations of the Class Projects
Friday, December 13, 2013, 01:00 p.m., Math/Psyc 401
-
01:05-01:20
Low-Order Model of Biological Neural Networks as a Machine Learning
Algorithm with Parallel Architecture
Bryce Carey, Department of Mathematics and Statistics, UMBC
The Low-Order Model of biological neural networks is a
biologically plausible model of dendritic nodes, synapses, neurons, and
learning and retrieving mechanisms. The model is adapted into a machine
learning algorithm implemented as a C program with parallel architecture. A
sample problem involving digit recognition is used as a proof of concept.
Performance studies are conducted to demonstrate the benefits of parallel
computing applied to the algorithm. These efforts will enable the timely
acquisition of results in future studies.
This work is in collaboration with advisor Dr. James T. Lo.
-
01:20-01:35
Supremum Norm of Projections onto Third Order Difference Cones
with Applications to Constrained Minimax Theory
Teresa Lebair, Department of Mathematics and Statistics, UMBC
Shape constrained estimation receives increasing consideration
in applied mathematics and statistics. In this report, a B-spline
estimator subject to the constraint that the third derivative is non-negative
almost everywhere is proposed as an optimal estimator.
This estimator's optimality is equivalent to
projections onto third order difference cones being bounded in the
sup-norm independent of cone dimension. The sup-norm of these
projections is computed for these fixed values. Despite a large amount of
computing power, it was not feasible to determine enough of these
computationally expensive sup-norms to suggest that the boundedness result
holds. This work is being done in collaboration with Dr. Jinglai Shen and
Dr. Xiao Wang.
-
01:35-01:50
Parallelization of Mergesort Algorithm and Analysis of
Performance on Large Data Sets
Georgiy Frolov,
Department of Computer Science and Electrical Engineering, UMBC
A combination of parallel computing with classical sorting algorithms
can lead to significant improvements of average case performance. For
the past few years commercial systems faced an enormous increase in
amount of data and customers, which prompted the need for scalable
algorithms. Parallel mergesort truly utilizes divide-and-conquer
paradigm that it is based on by distributing the workload among a
number of processes and applying classical mergesort to each
individual chunk of data, regardless of data type. Performance
improvements may appear insignificant for small amount of processes,
however as the number of processes increases, results demonstrate
instant responses which can provide an edge in today's competitive market.
-
01:50-02:05
Parallelization of an UPFD-Method
for Advection-Diffusion-Reaction Equations
Sven Wallbaum, Institute of Mathematics, University of Kassel, Germany
Advection-diffusion-reaction equations occur in modeling physical and
biological systems. Often the unknowns represent properties that cannot be
negative. Benito M. Chen-Charpentier and Hristo V. Kojouharov proposed an
unconditionally positivity preserving finite difference scheme (UPFD) for
that purpose in 2011. The key is to discretize the terms outside of the
central term of the finite difference stencil at the old time step, which
combines the properties of ensuring positivity of the solution and
explicitness of the discretization.
We briefly review the analysis for the main properties of the UPFD method,
such as consistency and positivity. The consistency of the method is an
issue, because it is not per se consistent, but requires minor adjustments.
Afterwards, we analyze the parallel performance and convergence of the method
on the basis of different test cases with respect to real life applications.
This work is in collaboration with Dr. Andreas Meister and
Dr. Matthias K. Gobbert.
-
02:05-02:20
Parallel Performance Study of a Parabolic Test Problem Solved
Using the Conjugate Gradient Method
Timothy Munuhe, Department of Mechanical Engineering, UMBC
The conjugate gradient method is a widely-used tool in solving large,
sparse, linear systems of equations. This study uses the conjugate
gradient method to solve a forced heat equation, with a known true
solution, numerically in two spatial dimensions. A performance study
is performed utilizing the Unix cluster tara in parallel. Results
indicate that the conjugate gradient method scales well when solving
the parabolic test problem. It is recommended that someone utilizing
the conjugate gradient method to solve a parabolic test problem use as
many processes as possible. This class project was conducted in
collaboration with Dr. Matthias Gobbert.
Copyright © 2001-2013 by Matthias K. Gobbert. All Rights Reserved.
This page version 1.0, December 2013.