A scalable web server is one of popular information replication methods to accomodate the rapidly growing World Wide Web. It is basically a server clustering technique which supplies information from several servers with the illusion of one serving entity. Scheduling incoming requests is the crucial part for the load balancing.
DNS based approach utilizes name resolution process for request distribution. It is simple and low overhead scheduling method but it has limited control of load balancing because of the name resolution caching in the intermediate name servers. One popular approach is using centralized scheduler such as dispatcher. It provides full control of scheduling and fine-grained load balancing but it suffers from potential bottleneck. Another approach such as server based scheme has released the bottleneck problem by distributing the scheduling ability to all participating servers. However, the dispatching efficientcy degrades because the redirection process shares resources with data processing jobs.
We have developed an adaptive scheduling method that changes the number of scheduling decision entities according to different workload. It behaves exactly like centralized scheduling method with low or intermediate workload, taking advantage of full control of the scheduling. When the central decision entity itself has bacome a bottleneck point by bulk requests, some selected servers handle requests coming directly from clients to mitigate the service slowdown. In the simulation, with realistic workload and system parameters, the adaptive scheduling method shows flexible performance that is not hampered by a specific workload situation.