Derek L . Eager Michael C . Ferris Mary K . Vernon Optimized Caching in Systems with Heterogeneous Client Populations

Supporting on-demand access to large widely shared data, such as popular video objects, requires effective use of regional (proxy) servers that store some of the data close to the clients. The proxy caching problem is more complex in the context of continuous media files because of the need to consider bandwidth as well as storage constraints at the proxy servers, and because of the bandwidth sharing possibilities provided by recently proposed multicast delivery techniques. This paper develops new models fo determining optimal proxy cache content in such environments. Specifically, the models developed here provide insights for heterogeneous systems in which the proxy servers have differing client populations and server capabilities. The new results show that (1) in comparison to previous results for systems with homogeneous proxy workloads, it is even more often cost-effective to cache just the initial segments of many files, rather than the complete data for fewer of the most popular files, (2) to minimize total delivery cost, even in systems with quite strong heterogeneous features, it is often best for all proxy servers to store similar data, rather than to closely tailor each proxy cache content according to local client preferrences and server characteristics, (3) when minimizing total delivery cost, a (group of) regional server(s) with a distinct client workload can sometimes influence the data stored by the rest of the servers in unexpected ways, and (4) when minimizing the delivery cost for clients of an individual proxy server, the data to be stored at the proxy server may be quite different than the "socially optimal" set that would minimize total delivery cost to all clients. A. INTRODUCTION In systems that support on-demand access to widely-shared data, delivery cost can be greatly reduced through the use of regional (or proxy) servers that store some of the data close to the requesting clients. For popular objects, such as popular news clips, product advertisements, distance education content, and the like, significant savings in both server and network bandwidth can also be achieved through the use of multicast (or broadcast) delivery to multiple clients simultaneously. For large objects, such as video data, it is also advantageous to divide the file into segments, so that clients can receive and buffer some segments ahead of need, at a bandwidth (i.e., cost) savings because the multicast of those This paper addresses the question of which large popular widely shared data should be stored …