Math 627 - Introduction to Parallel Computing

Spring 2015 - Matthias K. Gobbert
Presentations of the Class Projects

Friday, May 11, 2015, 01:30 p.m., SOND 202

  1. 01:30-01:45
    Wearable Computing: A Parallel Implementation of the Discrete Wigner-Ville Distribution for Real-Time Facial Gesture Recognition
    Elishiah Miller, Department of Computer Science and Electrical Engineering
    Fast and efficient digital signal processing (DSP) techniques are needed for wearable computing that can recognize facial gestures to assist patients with paralysis. The discrete time Wigner-Ville distribution (DWVD) is implemented in parallel to develop a high performance facial movement recognition (FMR) system that can achieve fine resolutions in both time and frequency domains simultaneously. The system will make use of Doppler radars working in parallel to capture facial movement. The correctness, performance, and scalability of such a system will be determined using the high performance computing facility (HPCF) at the University of Maryland, Baltimore County (UMBC).

  2. 01:45-02:00
    Parallel Performance Studies for a Parabolic Test Problem on the Cluster maya 2013
    Ryan Day, Department of Mathematics and Statistics
    The heat equation with homogeneous Dirichlet boundary conditions is solved using a full finite difference discretization that uses a parallel implementation of the conjugate gradient method. The method uses a constant time step. The method has shown good performance results for the Poisson equation, so it is now implemented for the heat equation. Parallel performance studies are presented. The results show good convergence and good speedup. It is recommended that several different choices of time step be tested for this method.

  3. 02:00-02:15
    Improving the Speed of the LLL Lattice Basis Reduction Algorithm on the Maya Cluster
    Edward LaFemina, Department of Computer Science and Electrical Engineering and Department of Mathematics and Statistics
    This report discusses the performance of the LLL algorithm for lattice basis reduction used for cryptanalysis and in wireless technologies such as Wi-Fi. We aim to improve the speed of the original algorithm by following modifications made by Yixian Luo and Sanzheng Qiao. We modify the original to be a faster serial version which Luo and Qiao have parallelized on a shared memory system using pthreads. We attempt to implement this parallelized version on the maya cluster-a distributed memory system-using the Message Passing Interface (MPI) library, however we find that it does not easily adapt to the maya cluster's architecture and forcing it to conform introduces significant amounts of communication that we believe outweighs cost.

  4. 02:15-02:30
    Application of the Parallel Computing Algorithm to a Simulation of Optical Frequency comb Generation
    Zhen Qi, Department of Computer Science and Electrical Engineering
    The optical frequency Kerr comb generation in whispering-gallery-mode resonators is modeled by the so-called Lugiato-Lefever equation which is a variant of nonlinear Schrodinger equation. In time domain, pulse formation is shown to play an important role in comb generation, and the pulse evolution is critical for us to understand the process of comb generation. We apply the split-step Fourier transform method to simulate the pulse evolution in this spatiotemporal model. Using the high efficient parallel computing algorithm, we demonstate that it is applicable to our model and we are expecting a better computational efficiency if fftw3_mpi library is available.

  5. 02:30-02:45
    Solving Poisson's Equation with Parallel Computing using a GPU
    Mayur Darji, Department of Mathematics and Statistics and Department of Computer Science and Electrical Engineering
    Poisson's equation is a partial differential equation (PDE) found in many applications in physics and engineering. Thus, it is of interest to implement methods to efficiently solve this problem. We tested parallel C code with MPI (Message Passing Interface) commands to solve Poisson's equation numerically with set boundary conditions and a known solution to test our results against. We used a finite difference approximation to numerically obtain the solution to our PDE. This approximation transformed our problem into a system of linear equations that could be solved by the conjugate gradient (CG) method. The CG method was parallelized in code. Previously, the same parallel method was applied to our Poisson problem using CPU nodes. This yielded results that showed a speedup in the computation time as the number of processes increased. Our objective was to test our parallel computations on GPU (Graphics Processing Unit) nodes. A GPU itself is a parallel architecture that contains many cores that are designed to be efficient for parallel computation. The GPU used is the NVIDIA K20 with with 2496 computational cores and 5 GB of onboard memory. After modifying the parallel code for the CPUs to work with CUDA, a parallel computing platform for that works with GPUs, our Poisson problem was solved and the performance results were obtained and compared with the CPU results. The results show that there is a significant performance improvement when using the GPU in place of the CPU cores.

Copyright © 2001-2015 by Matthias K. Gobbert. All Rights Reserved.
This page version 1.0, May 2015.