Search and Placement in Tiered Cache Networks

TL;DR

This paper introduces an analytical framework for optimizing content search and placement in multi-tiered cache networks, addressing scalability and performance challenges in content dissemination systems.

Contribution

It develops a model combining random walk search and reinforced counters for placement, unifying existing strategies and deriving optimal policies.

Findings

Closed-form expressions for delay and load metrics.

Joint optimization of search and placement strategies.

Validation of known placement policies as special cases.

Abstract

Content distribution networks have been extremely successful in today's Internet. Despite their success, there are still a number of scalability and performance challenges that motivate clean slate solutions for content dissemination, such as content centric networking. In this paper, we address two of the fundamental problems faced by any content dissemination system: content search and content placement. We consider a multi-tiered, multi-domain hierarchical system wherein random walks are used to cope with the tradeoff between exploitation of known paths towards custodians versus opportunistic exploration of replicas in a given neighborhood. TTL-like mechanisms, referred to as reinforced counters, are used for content placement. We propose an analytical model to study the interplay between search and placement. The model yields closed form expressions for metrics of interest such as the average delay experienced by users and the load placed on custodians. Then, leveraging the model solution we pose a joint placement-search optimization problem. We show that previously proposed strategies for optimal placement, such as the square-root allocation, follow as special cases of ours, and that a bang-bang search policy is optimal if content allocation is given.