Understanding Cache Patterns and Caching Algorithm Performance for Content Delivery Networks

Abstract

This project aims to develop methods of evaluating the performance of caching algorithms for content delivery networks in order to understand what characteristics and patterns of cache traces affect caching algorithm performance. Using clustering algorithms on caching algorithm predictions and analyzing the types of objects evicted by different algorithms were two techniques that were able to determine what kinds of objects and access patterns different caching algorithms perform better or worse on. Heuristic caching algorithms, and even more sophisticated learned algorithms like Learned Relaxed Belady, each have their own strengths and weaknesses, performing better or worse on different types of object access patterns found in a cache trace.