<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 Adrian Lehmann's MS Presentation<br>
===============================================<br>
Candidate: Adrian Lehmann<br>
<br>
Date: Friday, March 31, 2023<br>
<br>
Time: 2 pm CST<br>
<br>
Location: JCL 390<br>
<br>
M.S. Paper Title: Automatically parallelizing Diderot programs on CUDA targets<br>
<br>
Abstract: Diderot is a domain-specific language to perform scientific visualizations.
<br>
Its programs are structured largely like bulk-synchronous parallelism. In this pattern, multiple strands (often also called treads) run one update step in isolation, followed by a single global reduction step (similar to MapReduce).<br>
Currently, a compiler exists that transforms Diderot programs, along with the domain-specific operations, into C++.<br>
The compiler supports targeting both sequential and parallel CPU execution models.<br>
However, given the programming model's parallel nature, adding GPU support to Diderot's compiler is a natural step.<br>
<br>
Our work fills this gap. <br>
We add support for automatically parallelizing Diderot applications by modifying the compiler to be able to generate CUDA code.
<br>
We propose three strategies to schedule CUDA threads.<br>
One that closely follows the BSP model, one that runs strands to completion (assuming no reduction steps are needed), and one that builds on a work queue.<br>
We also propose a permutation mechanism for stochastic load distribution to mitigate strand divergence.<br>
We also create variants that utilize CUDA unified memory, an API to move memory pages between system and GPU memory.<br>
<br>
In benchmarks, we see speedups of 60-500x, where the queue-based approach outperforms other approaches.<br>
Further, we see differences in the performance of our approaches between benchmarks.<br>
We observe that permutation performance is highly dependent on the benchmark structure and the homogeneity of strand execution.<br>
Furthermore, we conclude that in our test, CUDA unified memory leads to a significant performance penalty for benchmarks with fewer strands while greatly simplifying the produced code.<br>
<br>
Advisors: John Reppy<br>
<br>
Committee Members: John Reppy, Hank Hoffmann, and Ravi Chugh<br>
<br>
<br>
</div>
</span></font></div>
<div class="BodyFragment"><font size="2"><span style="font-size:11pt;">
<div class="PlainText"><br>
<br>
<br>
<br>
</div>
</span></font></div>
</body>
</html>