Web Caching and Replication

16.5. Why Choose a Server When You Can Have Them All?

Consider a client that has a choice of N servers from which it can download a page. Several studies indicate that active probing of servers just prior to download provides the best prediction of server performance [Dykes et al. 2000;Sayal et al. 1998;Carter and Crovella 1997]. Unfortunately, active measurements create extra network traffic and server load and add latency to the download itself. However, HTTP range requests (see Section 5.4) allow probes to perform useful work. Indeed, assuming that a client knows the size of the requested object, the client can send separate range requests for portions of the object to multiple servers as once. The only thing a client needs to know in advance ...