Math 447/627 - Introduction to Parallel Computing

Fall 2023 - Matthias K. Gobbert
Presentations of the Class Projects

Tuesday, December 19, 2023, 01:00 p.m.

  1. 01:05-01:20
    Enhancing MOGONET: Parallel Computing for Multi-Omics Data Analysis
    Rodrigo Yepez Lopez, Department of Computer Science and Electrical Engineering
    The cutting edge machine learning framework MOGONET brings together different kinds of biological data to make it easier to classify illnesses and find biomarkers. MOGONET was made to deal with the problems that come with personalized medicine. MOGONET makes it easier to fully understand what causes sickness by combining different types of biological data, like mRNA, DNA methylation, and miRNA. Still, the huge amount of multi-omics data makes it hard to do computations, especially when it comes to processing speed and model accuracy. This project will use the Ada cluster for parallel computing to solve these issues. The goal is to make MOGONET better at handling data while keeping its diagnostic accuracy. We use CUDA's parallel processing to make our method work better, focusing on getting the most out of both model training and data flow. CUDA's efficient asynchronous processes and advanced memory management are important to make it possible to handle the growing amount of multi-omic data. A part of this study focuses on hyperparameter tuning, which involves finding settings that can preserve or potentially enhance the accuracy of MOGONET while handling larger datasets. The end result should be a better version of MOGONET that can handle bigger datasets more quickly while keeping or even improving its accuracy. These changes are very important for speeding up the processing of data and making the system better at handling complex multi-omics data. The successful completion of this project has the potential to result in more accurate and in-depth scientific studies.

  2. 01:20-01:35
    Spectral Energy Density Algorithm for a Steady State One Zone Emitting Model with Parallelization
    Rafael Diaz Brenes, Department of Physics
    This investigation follows the application of a numerical algorithm to calculate the spectral energy density for a One Zone Emission Model in order to fit and study observational data from active galactic nuclei (AGN). The numerical algorithm involved obtained the solution for a diffusion-loss equation that describes the electron energy distribution of relativistic electrons in a one zone emitting region. The numerical method involved using a difference equation with an implicit numerical scheme for obtaining the electron energy distribution. With the electron energy distribution, the luminosity and spectral energy density were calculated using parallelization with C MPI library functions. The program was able to solve the diffusion loss equation for a steady state with an isotropic power law injection of relativistic electrons. This consisted in being able to recreate the electron energy distribution and spectral energy density for a steady state subjected only to synchrotron losses. This simulates well the AGN case of radio and Seyfert galaxies. Parallelization allowed us to reduce calculation times approximately by half. This work is collaborative with Dr. Markos Georganopoulos and the UMBC Astro Group.

  3. 01:35-01:50
    Parameter Estimation of ODE Systems with Data Assimilation: A Case Study for Lorenz System
    Muhammad Jalil Ahmad, Department of Mathematics and Statistics
    This study introduces a novel approach to fortify the robustness and precision of Ordinary Differential Equation (ODE) models. A nudging term is introduced in the ODE model which makes it remarkably independent of initial conditions, which offers flexibility in choosing arbitrary starting points without compromising reliability. Moving to the crucial phase of parameter estimation, the study utilizes both nudged and original ODE systems to define a sophisticated loss function. This loss function is minimized using the Particle Swarm Optimization (PSO) algorithm. This methodology holds promise for applications where uncertainties in initial conditions and parameter estimates have traditionally posed limitations, providing a robust framework for advancing the utility and reliability of ODE models. This work is collaborative with Dr. Kathleen Hoffman and Dr. Animikh Biswas.

  4. 01:50-02:05
    Parallel Computation of the Eigenpairs of Perturbed Matrices
    Pradyoth Shandilya, Department of Computer Science and Electrical Engineering
    This report demonstrates an enhanced method for calculating eigenpairs of matrices, leveraging initial guesses to improve computational efficiency. Traditional linear algebra libraries typically do not have the option to input initial approximations for eigenpairs in their algorithms. However, in numerous scientific applications, rough estimates of these eigenpairs are often pre-existing. We demonstrate that by employing the eigenpairs of a previously established matrix as a starting point, we can more effectively compute a selection of the eigenpairs of a slightly altered (perturbed) matrix using the inverse iteration technique. Our findings indicate a notable improvement in efficiency compared to MATLAB's built-in functions, especially for matrices of smaller dimensions.

  5. 02:05-02:20
    Parallelization of Merge-Sort Algorithm
    Parisa Eslami, Department of Information Systems
    The merge-sort algorithm is a cornerstone of efficient sorting in computer science, renowned for its divide-and-conquer approach. By leveraging Message Passing Interface (MPI) processes, we evolve the serial Merge-Sort algorithm into its parallelized counterpart, aiming to enhance performance and efficiency. The crux of this exploration lies in comparing the execution times and efficiency between the serial and parallel versions, thereby encapsulating the transformative potential of parallel computing in algorithmic optimization.

  6. 02:20-02:35
    Adaptive Lift Regulation of a Pitchable Airfoil in Persistent Gusts
    Christopher Kasprzak, Department of Mechanical Engineering
    This report outlines a comprehensive study conducted to enhance the understanding of fluid dynamics using Computational Fluid Dynamics (CFD) simulations, facilitated by the high-performance computing capabilities of the University of Maryland, Baltimore County's (UMBC) Taki Cluster. The primary focus of the project was to delve into the complexities of CFD and to explore the efficacy of the Retrospective Cost Adaptive Control (RCAC) system in managing aerodynamic variables, particularly in scenarios involving persistent wind gusts. Key activities involved modifying and executing sophisticated CFD code, primarily written in C++, and the subsequent processing and analysis of simulation data using tools like MATLAB and TECPLOT. The study concentrated on monitoring the effects of different control strategies on the regulation of lift coefficients under varying aerodynamic conditions and controller performance. Simulations conducted at different Reynolds numbers aided in understanding the behavior of lift coefficients in the presence of wind gusts and the effectiveness of RCAC in stabilizing these coefficients. The research extended to comparing adaptive control with fixed gain controls and examining the impact of varying the frequency of controller adjustments. The results of this project contribute insights into the dynamics of fluid flow and control systems, particularly in complex airflow interactions, offering practical implications in various aerodynamic applications. This study underscores the importance of high-performance computing in conducting detailed and high-fidelity aerodynamic simulations and demonstrates the potential of advanced control mechanisms in optimizing aerodynamic performance under challenging conditions.

Copyright © 2001-2023 by Matthias K. Gobbert. All Rights Reserved.
This page version 1.0, December 2023.