Explore a groundbreaking approach to content distribution network caching in this 20-minute USENIX NSDI '20 conference talk. Dive into the Learning Relaxed Belady (LRB) algorithm, which uses machine learning to approximate the Belady MIN algorithm for improved caching performance. Discover the innovative concepts of the Relaxed Belady algorithm, Belady boundary, and good decision ratio. Learn how LRB addresses system challenges in implementing machine learning for caching, including data collection, memory management, and efficient training and inference. Examine the results of LRB simulations using production CDN traces, showcasing significant reductions in WAN traffic compared to typical CDN cache designs. Gain insights into the practical implementation of LRB within Apache Traffic Server and its potential for deployment on current CDN servers.
Learning Relaxed Belady for Content Distribution Network Caching
Overview
Syllabus
Intro
Caching Remains Challenging
General Overview of our Approach
Generate Online Training Data
Solutions: Relaxed Belady Algorithm & Good Decision Ratio
Challenge: Hard to Mimic Belady (Oracle) Algorithm
Introducing the Relaxed Belady Algorithm
Good Decision Ratio: Directly Measures Eviction Decisions
Evaluate Design Decisions wlo Simulation
Past Information
Track Objects within a Sliding Memory Window
Sample Training Data & Label on Access or Boundary
ML Architecture
Solution 3: Feature & Model Selection
Eviction Candidates
Solution 4: Random Sampling for Eviction
Implementation
Evaluation Setup
Conclusion
Taught by
USENIX
