Date of Award
Doctor of Philosophy in Computer Information Systems (DCIS)
Graduate School of Computer and Information Sciences
Researchers have spent years understanding resource utilization to improve scheduling, load balancing, and system management through short-term prediction of resource utilization. Early research focused primarily on single operating systems; later, interest shifted to distributed systems and, finally, into web services. In each case researchers sought to more effectively use available resources. Since schedulers are required to manage the execution of multiple programs every second, short-term prediction has focused on time frames ranging from fractions of a second to several minutes.
The recent increase in the number of research studies about web services has occurred because of the explosive growth and reliance on these services by most businesses. As demand has moved from static to dynamic content, the load on machine resources has grown exponentially, periodically resulting in temporary loss of service. To address these short-term denial-of-service issues, researchers have tried short-term prediction to manage scheduling of service requests.
What researchers have not considered is that the same methods used for single step short-term prediction can also be used for long-term prediction if a coarse granularity of samples is used. Instead of using one or more samples per second, a coarser aggregate of minutes or hours more accurately emulates the long-term patterns. This research has shown that simple moving averages and exponential moving averages as a prediction technique can be used to more accurately predict hourly, daily, and weekly seasonal patterns of resource utilization for web servers.
Additionally, this research provides a foundation where using a resource prediction within a confidence interval range could be more useful to an administrator or system software than a single prediction point. When the focus shifts to a range, a set of probabilities can establish normal function within that system. For distributed systems, it will provide the ability to notify other systems when resource utilization is no longer normal before that system is unable to issue a notice of overloading. For web systems it can be used to provide a warning, permitting the instantiation of a second system to begin load balancing during unscheduled heavy loads. In both cases, the availability of the system can be improved by predicting a resource utilization level and the confidence interval within which that resource use has historically fallen.
Daniel Wayne Yoas. 2013. Using Forecasting to Predict Long-term Resource Utilization for Web Services. Doctoral dissertation. Nova Southeastern University. Retrieved from NSUWorks, Graduate School of Computer and Information Sciences. (343)