On the Asymptotic Content Routing Stretch in Network of Caches: Impact of Popularity Learning

Abstract

In this paper, we study the asymptotic average routing stretch for random content requests in a general network of caches. The key factor considered in our study is the need of learning content popularity in an online manner to consider time-varying changes of content popularity, where there exists a complex inter-play among (a) how long we should learn popularity, (b) how often we should change cached contents, and (c) how we use learnt popularity in caching contents over the network. We model this inter-play in a broad class of caching policies, called Repeated Learning and Placement (RLP), and aim at quantifying the asymptotic routing stretch of content requests under various external conditions. Our derivation of this scaling law in the routing stretch is made under different dependence of the speed of popularity change, average routing stretch in the network of caches, the shape of the popularity distribution, and heterogeneous cache budget allocation based on nodes’ geometric importance. We believe that our analytical results, even if they are asymptotic, provide additional ways and implications on understanding the delay performance of large-scale content distribution network (CDN) and information-centric network (ICN).