[Colloquium] Maksim Levental MS Presentation/Feb 9, 2022

[Megan Woodward]

This is an announcement of Maksim Levental's MS Presentation
===============================================
Candidate: Maksim Levental

Date: Wednesday, February 09, 2022

Time:  3 pm CST

Remote Location: https://uchicago.zoom.us/j/2681314906?pwd=MFBwTURuYXJKa1QzcHp5UVdJbS9LQT09  Meeting ID: 268 131 4906 Passcode: 123456

M.S. Paper Title: Memory Planning for Deep Neural Networks

Abstract: Deep neural networks (DNNs) are becoming increasingly memory intensive.
We study memory allocation patterns in DNNs and propose a “memorization" based technique, \texttt{MemoMalloc}, for optimizing both memory usage and inference latency.
Specifically, we use static memory planning techniques to reduce both peak memory consumption and heap mutex contention.
We present an implementation of MemoMalloc and evaluate memory consumption and execution performance on a wide range of DNN architectures. MemoMalloc substantially outperforms state-of-the-art caching allocators in terms of execution performance, by as much as 40%.

Advisors: Ian Foster

Committee Members: Ian Foster, Kyle Chard, and Raul Castro Fernandez



-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20220127/5f1044dc/attachment-0001.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: Memory_Planning_for_Deep_Neural_Networks__MS_thesis_.pdf
Type: application/pdf
Size: 1137085 bytes
Desc: Memory_Planning_for_Deep_Neural_Networks__MS_thesis_.pdf
URL: <http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20220127/5f1044dc/attachment-0001.pdf>