Performance Evaluation of the Random Replacement Policy for Networks of Caches

TL;DR

This paper analyzes the hit ratio performance of the Random cache replacement policy in networks, providing exact and asymptotic formulas, and compares it with other policies like LRU through simulations.

Contribution

It offers the first comprehensive analysis of Random cache policies in network topologies, including asymptotic expressions and comparisons with LRU.

Findings

Asymptotic miss probability formulas for large caches and Zipf distributions.

Approximate miss probabilities for line and tree network topologies.

Random policy performance is comparable to LRU in large cache regimes.

Abstract

The overall performance of content distribution networks as well as recently proposed information-centric networks rely on both memory and bandwidth capacities. In this framework, the hit ratio is the key performance indicator which captures the bandwidth / memory tradeoff for a given global performance.This paper focuses on the estimation of the hit ratio in a network of caches that employ the Random replacement policy. Assuming that requests are independent and identically distributed, general expressions of miss probabilities for a single Random cache are provided as well as exact results for specific popularity distributions. Moreover, for any Zipf popularity distribution with exponent $\alpha$ > 1, we obtain asymptotic equivalents for the miss probability in the case of large cache size. We extend the analysis to networks of Random caches, when the topology is either a line or a homogeneous tree. In that case, approximations for miss probabilities across the network are derived by assuming that miss events at any node occur independently in time; the obtained results are compared to the same network using the Least-Recently-Used discipline, already addressed in the literature. We further analyze the case of a mixed tandem cache network where the two nodes employ either Random or Least-Recently-Used policies. In all scenarios, asymptotic formulas and approximations are extensively compared to simulations and shown to perform very well. Finally, our results enable us to propose recommendations for cache replacement disciplines in a network dedicated to content distribution. These results also hold for a cache using the First-In-First-Out policy.