[Back]

J. Cito:
"Statistical Methods in Managing Web Performance";
Supervisor: S. Dustdar, Ph. Leitner; Institut für Informationssysteme, Distributed Systems Group, 2014.

Over the past few decades, the internet has grown to an enormous infrastructure connecting millions of people world-wide. The large scale of the internet has offered unique economic opportunities
by enabling the ability to reach a tremendous, global user base for businesses and individuals alike. The great success and opportunities also open up overwhelming challenges due to the drastic growth and increasing complexity of the internet in the last decade. The main
challenge for development and operations is to provide reliable and fast service, despite of fast growth in both traffic and frequency of requests. When it comes to speed, internet users have high demands on the performance of online services. Recent research has shown that 47% of online consumers expect load times of two seconds or less. This puts high pressure on web applications and websites in terms of performance and reliability, when even minor changes in performance
can have significant effects on how users perceive an application and decide whether or not to continue with their interaction. If the revenue of a business is generated through their online presence, performance directly correlates with loss of sales, and damage of brand perception.
With the growth of the internet and its user base, the underlying infrastructure has drastically transformed from single server systems to heterogeneous, distributed systems. Thus, the end performance depends on diverse factors in different levels of server systems, networks and infrastructure, which makes providing a satisfying end-user experience and QoS a challenge for large scale internet applications.
In this thesis, statistical methods are applied to deal with the management of web performance issues. In order to properly manage performance, the timely observation and analysis of performance
degradation is of key importance. Statistical changepoint analysis and exploratory visual analysis are applied to results obtained through synthetic monitoring of a simulation system to identify and discuss how performance issues in web applications can be detected.
Web Performance Benchmarking of competitors is discussed as a way to establish a baseline for an overall performance goal. A general methodology for web performance benchmarking is developed based on statistical analysis of finite collections of observed data from production systems. The presented approach is then applied in a case study for benchmarking search engine providers.