Accurately Modeling Workload Interactions for Deploying Prefetching in
Web Servers

Xin Chen and Xiaodong Zhang

Proceedings of 2003 International Conference on Parallel Processing
(ICPP'03), Kaohsung, Taiwan, China, October 6-9, 2003.


Abstract

Although Web prefetching is regarded as an effective method to improve
client access performance, the associated overhead prevents it from
being widely deployed. Specifically, a major weakness in existing Web
servers is that prefetching activities are scheduled independently
of dynamically changing server workloads. Without proper control and
coordination between the two kinds of activities, prefetching can
negatively affect the Web services and degrade Web access performance.
In this paper, we first develop an open queuing model to characterize
detailed transactions in Web servers. Using this model, we analyze server
resource utilization and average response time with different request
arrival rates when prefetching is involved under different kinds of
Web services.  Guided by this model, we then design a responsive and
adaptive prefetching scheme that dynamically adjusts the prefetching
aggressiveness in Web servers.  Our scheme not only prevents the Web
servers from being overloaded, but it can also minimize the average
server response time.  We have effectively implemented this scheme on an
Apache Web server.  Our measurement-based performance evaluation shows
our model can accurately predict the utilization of Web server resources
and the correspondent average response time.