New tool predicts cloud capacity requirements for web services
Wed 30 Sep 2015
A Swedish scientist has developed a number of algorithms to auto-scale cloud capacity based on demand.
Ahmed Hassan, researcher at Umeå University, wanted to create a way to make fragile internet infrastructure more reliable, after studying the collapse of sites such as CNN, Google News, TMZ, Twitter and Wikipedia following the death of Michael Jackson in 2009, and the overloading of the Obamacare website after it went live in the U.S. in 2010. Even on a more local scale, Hassan notes the pressure experienced by Swedish university site, studera.nu, during the annual online application process.
Hassan’s study proposes two algorithmic methods to cope with fluctuating demand, which automatically add and remove resources to a web service as required. He explains that renting more cloud capacity than a website typically requires is an expensive option for businesses, while renting too little space can cause server overload and costly service disruption. The researcher suggests prediction algorithms could prove an effective way to respond to future demand.
During his research project, Hassan analysed server workloads from a group of major web services, noting spikes, ‘burstiness’ and invariants in traffic. He investigated the world’s seventh most popular website, Wikipedia, over five and a half years, monitoring capacity requirements during major events. Hassan also tracked demand at Swedish premium video-streaming site TV4. He found that users were extremely frustrated with video buffering, with around half of the users abandoning streaming content within less than 12 minutes.
The prototype workload analysis and auto-scale tool is able to provide predictions on the cloud capacity needed by a web service combining reactive and pro-active controls. Hassan argues that the system will allow web services to improve overall efficiency and performance, and therefore provide better services to their customers.
Hassan’s research was carried out in collaboration with students from Umeå University’s Department of Mathematics and Mathematical Statistics, as well as scientists from Lund University’s Department of Control and Department of Electrical and Information Technology.