Explore a conference talk on Ekya, a system for continuous learning of video analytics models on edge compute servers. Dive into the challenges of balancing inference and retraining tasks, addressing data drift, and optimizing resource allocation. Learn how Ekya outperforms baseline schedulers, achieving 29% higher accuracy gain and requiring 4x fewer GPU resources. Discover the intricacies of edge video analytics, the cost of continuous learning, and innovative scheduling decisions. Gain insights into Ekya's design, including its Thief Scheduler, and understand its performance in scaling with increasing video streams.
Ekya - Continuous Learning of Video Analytics Models on Edge Compute Servers
Overview
Syllabus
Intro
Video data is everywhere
Why video analytics at the edge?
Edge Video Analytics Setup
The cost of continuous learning
Resource demands of continuous learn
Summary thus far
Scheduling decisions to make
Working Example
Example - Fair Scheduler
Example - a smarter schedule
Key Takeaways
Ekya Design
Ekya Thief Scheduler Goal: Maximize mean inference accuracy across all jobs
Evaluation
Scaling with increasing video streams
Taught by
USENIX
