Ronald P. Doyle

Energy management for server clusters

The central point of this paper is that energy should be viewed as an important element of resource management for Web sites, hosting centers, and other Internet server clusters. In particular, we are developing a system to manage server resources so that cluster power demand scales with request throughput. This can yield significant energy savings because server clusters are sized for peak load, while traces show that traffic varies by factors of 3-6 or more through any day or week, with average load often less than 50% of peak. We propose energy-conscious service provisioning, in which the system continuously monitors load and adaptively provisions server capacity. This promises both economic and environmental benefits.

The Trickle-Down Effect: Web Caching and Server Request Distribution

Web proxies and Content Delivery Networks (CDNs) are widely used to accelerate Web content delivery and to conserve Internet bandwidth. These caching agents are highly effective for static content, which is an important component of all Web-based services. This paper explores the effect of ubiquitous Web caching on the request patterns seen by other components of an end-to-end content delivery architecture, including Web server clusters and interior caches. In particular, object popularity distributions in the Web tend to be Zipf-like, but caches disproportionately absorb requests for the most popular objects, changing the reference properties of the filtered request stream in fundamental ways. We call this the trickle-down effect. This paper uses trace-driven simulation and synthetic traffic patterns to illustrate the trickle-down effect and to investigate its impact on other components of a content delivery architecture, focusing on the implications for request distribution strategies in server clusters.