Bulk content delivery using co-operating end-nodes with upload/download limits

TL;DR

This paper proposes an optimization framework for cooperative content caching with upload/download limits, achieving near-optimal cost reduction and significant network load decrease through LP rounding and clustering techniques.

Contribution

It introduces a multi-criteria approximation method for data placement and request routing in cache networks with capacity constraints, improving efficiency over non-cooperative caching.

Findings

Cost within 5-15% of optimal in simulations

Reduces central server load by up to 55%

Outperforms non-cooperative caching by about 20%

Abstract

We study the problem of optimizing the cost of content delivery in a cooperative network of caches at end-nodes. The caches could be, for example, within the computers of users downloading videos from websites (such as Netflix, Blockbuster etc.), DVRs (such as TiVo, or cable boxes) used as part of video on demand services or public hot-spots (e.g. Wi-Fi access points with a cache) deployed over a city to serve content to mobile users. Each cache serves user requests locally over a medium that incurs no additional costs (i.e. WiFi, home LAN); if a request is not cached, it must be fetched from another cache or a central server. In our model, each cache has a tiered back-haul internet connection, with a usage cap (and fixed per-byte costs thereafter). Redirecting requests intended for the central server to other caches with unused back-haul capacity can bring down the network costs. Our goal is to develop a mechanism to optimally 1) place data into the caches and 2) route requests to caches to reduce the overall cost of content delivery. We develop a multi-criteria approximation based on a LP rounding procedure that with a small (constant factor) blow-up in storage and upload limits of each cache, gives a data placement that is within constant factor of the optimum. Further, to speed up the solution, we propose a technique to cluster caches into groups, solve the data placement problem within a group, and combine the results in the rounding phase to get the global solution.Based on extensive simulations, we show that our schemes perform very well in practice, giving costs within $5--15$% to the optimal, and reducing the network load at a central server by as much as 55% with only a marginal blow up in the limits. Also we demonstrate that our approach out-performs a non-cooperative caching mechanism by about 20%.