Energy efficiency of cache eviction algorithms for Zipf distributed objects

TL;DR

To improve energy efficiency in CDN caching, the paper compares LFU and PLFU eviction schemes using real and Zipf distributed data and introduces PLFUA, a PLFU variant with an admission policy. It measures cache performance via cache hit ratio and the novel total CPU time metric, analyzing both CHR and CPU time across varying cache sizes and object counts, with Zipf$(1.1)$ as the distribution. The results show PLFU increases CHR but at the cost of additional metadata and higher CPU time, while PLFUA achieves higher CHR with lower CPU time and smaller metadata. The findings suggest that smaller cache sizes can enhance energy efficiency and that PLFUA offers a practical, performant alternative for CDN caching.

Abstract

This paper presents a summary analysis of the Least Frequently Used (LFU) and Perfect Least Frequently Used (PLFU) cache eviction algorithms on real data, transferred on Content Delivery Nettworks (CDNs), as well as on Zipf distributed samples. In light of the growing emphasis on energy efficiency in CDNs in recent years due to rising energy costs, this paper considers and discusses the total CPU time required to run a cache algorithm. The total CPU time represents a novel metric for evaluating cache performance, and it is contrasted with the conventional Cache Hit Ratio (CHR) metric. Furthermore, a new algorithm with an admission policy and the eviction strategy that of PLFU is presented. The results demonstrate that it is a simple and straightforward algorithm to implement and offers high CHR and low CPU time.

Figures (7)

• Figure 1: Rank order probability distribution of ISP data
• Figure 2: Rank order of hits and misses on a loglog scale
• Figure 3: CHR values obtained from 12 random sample
• Figure 4: The CPU total time required
• Figure 5: PLFUA performance metrics