Explore a conference talk from USENIX ATC '23 that delves into decentralized application-level adaptive scheduling for multi-instance Deep Neural Networks (DNNs) on open mobile devices. Learn about the challenges of running multiple DNN-powered apps simultaneously on common smartphones and tablets, and discover a novel approach to address scheduling issues in these scenarios. Understand how the proposed solution leverages Deep Reinforcement Learning to achieve a Nash equilibrium point, balancing gains among co-running apps while adapting to various running environments, operating systems, and hardware configurations. Gain insights into the experimental results demonstrating significant speedups and energy savings across different DNN workloads and hardware setups.
Overview
Syllabus
USENIX ATC '23 - Decentralized Application-Level Adaptive Scheduling for Multi-Instance DNNs on...
Taught by
USENIX