A Flexible Alternative to Web Caching

The exponential growth of the WorldWide Web 1] is leading to a serious performance problem. Many studies have investigated the use of caching as a possible solution to reduce the access latency to Web documents, to decrease server loads, and to reduce bandwidth requirements. However, caching suuers from major drawbacks. Web caching is a general-purpose solution that ignores the diversity of Web objects: the same policy is applied to all objects even though they may have very diierent characteristics. For example, access patterns or update characteristics can be signiicantly diierent from one object to the other. Update characteristics are related to the way the object is modiied and the way its coherence is implemented. Such diierences lead to diierent requirements for replication and support for coherence. For example , it could be useful to massively replicate popular documents, i.e., those that are frequently accessed, whereas less popular ones hardly require any replication. Likewise, it could be eecient to implement an immediate update coherence policy for a seldom modiied Web object whereas a updating periodically is better for a frequently modiied Web object. Web caching already suuers from a lack of performance and this become even worse with more and diierent types of Web objects and users in the future. Taking such diierences into account by applying replication on a per-object basis is expected to increase performance. Two aspects of interest are the access patterns and the update characteristics, respectively. Access patterns can be derived automatically from access statistics thus allowing a replication policy to automatically adapt when necessary. However, update characteristics are hard to derive automatically and user involvement is generally necessary. This paper proposes an approach towards replication for the Web that is both exible and scalable. Our proposal is based on the use of the Globe architecture 3] designed to build wide-area applications. Web caching requires coherence to ensure that cached copies reeect changes of the original Web object 2]. In current caching policies, the decision to replicate a Web object or update a cache 1