<html>
<head>
<meta http-equiv="Content-Type" content="text/html; charset=us-ascii">
</head>
<body>
<div class="BodyFragment"><font size="2"><span style="font-size:11pt;">
<div class="PlainText">This is an announcement of Yuliana Zamora's Dissertation Defense.<br>
===============================================<br>
Candidate: Yuliana Zamora<br>
<br>
Date: Friday, June 03, 2022<br>
<br>
Time: 10 am CST<br>
<br>
Remote Location: <a href="https://zoom.us/j/91528203161?pwd=K0U5RXBsUktORGtlVDVoSkQwMU9QZz09">
https://zoom.us/j/91528203161?pwd=K0U5RXBsUktORGtlVDVoSkQwMU9QZz09</a> Meeting ID: 915 2820 3161 Passcode: T35nAc<br>
<br>
<br>
Title: Machine Learning for Performance Acceleration and Prediction in Scientific Computing<br>
<br>
Abstract: Scientific applications often require massive amounts of compute time and power. With the<br>
constantly expanding architecture landscape and growing complexity of application runs,<br>
understanding how to improve performance is vital. In this thesis, we will use machine<br>
learning in two distinct ways to improve science applications.<br>
First, we use a data-driven approach and leverage machine learning to understand and<br>
improve performance in high-performance computing applications. The goal of this work<br>
is to create a streamlined workflow of integrating machine learning surrogates into such<br>
applications. Using control theory to automatically and dynamically configure parameters,<br>
we can meet accuracy constraints while maximizing performance. This workflow, which we<br>
examine in the context of molecular dynamics simulations, will allow for faster and larger<br>
simulations by improving overall performance. By replacing typically high-cost functions,<br>
such as density functional theory calls or Hartree-Fock calls, with a low-cost machine learning<br>
inference call, our proposed workflow can reduce run-time while producing scientifically<br>
usable results. We create a decision engine that will automate finding the accuracy and<br>
performance trade-off relationship between using a high-fidelity, high-cost function call, and<br>
a lower fidelity machine learning inference.<br>
Second, in an effort to understand future performance, we focus on predicting performance<br>
across multiple architectures. Cross-architecture metric mapping is heavily studied<br>
with hopes of understanding application performance on future or untested machine architectures.<br>
Often, there are months or even years spent on porting applications to new<br>
architectures, that may or may not provide an increase in performance. Here, we will use a<br>
data-driven approach to predict total throughput across different GPU architectures in order<br>
to understand future success of these applications. Our goal is to create a general framework<br>
that can predict application performance on a target hardware, given performance metrics<br>
from a different hardware architecture, without expert input. In this thesis, we propose such
<br>
a framework and use it to predict total throughput and IPC between two GPU architectures.<br>
<br>
Advisors: Ian Foster and Hank Hoffman <br>
<br>
Committee Members: Hank Hoffmann, Logan Ward, and Ian Foster<br>
<br>
<br>
<br>
</div>
</span></font></div>
<div class="BodyFragment"><font size="2"><span style="font-size:11pt;">
<div class="PlainText"><br>
<br>
</div>
</span></font></div>
</body>
</html>