<html>
<head>
<meta http-equiv="Content-Type" content="text/html; charset=us-ascii">
</head>
<body>
<div class="" style="word-wrap:break-word">This is an announcement of Yuliana Zamora's Dissertation Defense.<br class="">
===============================================<br class="">
Candidate: Yuliana Zamora<br class="">
<br class="">
Date: Friday, June 03, 2022<br class="">
<br class="">
Time: 10:30 am CST<br class="">
<br class="">
Remote Location: <a href="https://urldefense.com/v3/__https://zoom.us/j/91528203161?pwd=K0U5RXBsUktORGtlVDVoSkQwMU9QZz09__;!!BpyFHLRN4TMTrA!4uIQk9mtzYahkKYoaMfyM4MsDJUqH-NtUDAFIMhWPSGuPXQ5W9soRQ0Ie8IREBj7ZHf_BWZ58tMsjGxc5osAySxDnv4DdUFoNODA$" class="">https://zoom.us/j/91528203161?pwd=K0U5RXBsUktORGtlVDVoSkQwMU9QZz09</a> Meeting
ID: 915 2820 3161 Passcode: T35nAc<br class="">
<br class="">
<br class="">
Title: Machine Learning for Performance Acceleration and Prediction in Scientific Computing<br class="">
<br class="">
Abstract: Scientific applications often require massive amounts of compute time and power. With the<br class="">
constantly expanding architecture landscape and growing complexity of application runs,<br class="">
understanding how to improve performance is vital. In this thesis, we will use machine<br class="">
learning in two distinct ways to improve science applications.<br class="">
First, we use a data-driven approach and leverage machine learning to understand and<br class="">
improve performance in high-performance computing applications. The goal of this work<br class="">
is to create a streamlined workflow of integrating machine learning surrogates into such<br class="">
applications. Using control theory to automatically and dynamically configure parameters,<br class="">
we can meet accuracy constraints while maximizing performance. This workflow, which we<br class="">
examine in the context of molecular dynamics simulations, will allow for faster and larger<br class="">
simulations by improving overall performance. By replacing typically high-cost functions,<br class="">
such as density functional theory calls or Hartree-Fock calls, with a low-cost machine learning<br class="">
inference call, our proposed workflow can reduce run-time while producing scientifically<br class="">
usable results. We create a decision engine that will automate finding the accuracy and<br class="">
performance trade-off relationship between using a high-fidelity, high-cost function call, and<br class="">
a lower fidelity machine learning inference.<br class="">
Second, in an effort to understand future performance, we focus on predicting performance<br class="">
across multiple architectures. Cross-architecture metric mapping is heavily studied<br class="">
with hopes of understanding application performance on future or untested machine architectures.<br class="">
Often, there are months or even years spent on porting applications to new<br class="">
architectures, that may or may not provide an increase in performance. Here, we will use a<br class="">
data-driven approach to predict total throughput across different GPU architectures in order<br class="">
to understand future success of these applications. Our goal is to create a general framework<br class="">
that can predict application performance on a target hardware, given performance metrics<br class="">
from a different hardware architecture, without expert input. In this thesis, we propose such <br class="">
a framework and use it to predict total throughput and IPC between two GPU architectures.<br class="">
<br class="">
Advisors: Ian Foster and Hank Hoffman <br class="">
<br class="">
Committee Members: Hank Hoffmann, Logan Ward, and Ian Foster
<div class=""><br class="">
<div class=""></div>
</div>
</div>
<div class="" style="word-wrap:break-word">
<div class="">
<div class=""></div>
<div class=""><br class="">
</div>
</div>
</div>
</body>
</html>