Evangelos P. Markatos and Catherine E. Chronaki
Institute of Computer Science (ICS)
Foundation for Research & Technology - Hellas (FORTH)
Heraklio, Crete, GR-711-10 GREECE
tel: +30 81 391 655, fax: +30 81 391 661
Technical Report 173, ICS-FORTH. Available from http://www.ics.forth.gr/proj/arch-vlsi/www.html
In the World Wide Web bottlenecks close to popular servers are very common. These bottlenecks can be attributed to the servers' lack of computing power and the network traffic induced by the increased number of access requests. One way to eliminate these bottlenecks is through the use of caching. However, several recent studies suggest that the maximum hit rate achievable by any caching algorithm is just 40% to 50%. Prefetching techniques may be employed to further increase the cache hit rate, by anticipating and prefetching future client requests.
This paper proposes a Top-10 approach to prefetching, which combines the servers' active knowledge of their most popular documents (their Top-10) with client access profiles. Based on these profiles, clients request and servers forward to them, regularly, their most popular documents. The scalability of the approach lays in that a web server's clients may be proxy servers, which in turn forward their Top-10 to their frequent clients which may be proxies as well, resulting in a dynamic hierarchical scheme, responsive to users access patterns as they evolve over time. We use trace driven simulation based on access logs from various servers to evaluate Top-10 prefetching. Performance results suggest that the proposed policy can anticipate more than 40% of a client's requests while increasing network traffic by no more than 10% in most cases.