CS4395 INDEPENDENT STUDY PRESENTATION
Title: Toward an Intelligent Runtime System for Deep Learning Workflows on Heterogeneous Clouds
Presenter: Ethan Greene
Advisor: Dr. Tanzima Islam
Date/Time: Thursday, December 2 @11:00 a.m.
Location: ZOOM: https://txstate.zoom.us/j/92156883505
Deep learning workflows (DLWs) are applications that are comprised of assorted, unique models with different performance characteristics that can be scheduled on various heterogenous platforms. Not considering each models performance characteristics while scheduling models can cause a significant decline in performance. In this research, we have built a monitoring and analysis pipeline to collect performance metrics about a DLW in order to ascertain what specific characteristics indicate which platform a model should be run on to achieve its best performance. Utilizing these intrinsic characteristics a DLW application is able to improve its performance by up to 3.2x.
Deadline: Dec. 3, 2021, midnight